🌟 Internal Project Selection#
Projects with main supervisors affiliated with Imperial College London are classified as internal.
All students must select internal projects by filling out the following form:
https://forms.cloud.microsoft/e/pNqs9YrVD8
The deadline to submit your internal projects selection is Monday, 13 April 2026, 10:00 BST. If you miss this deadline, you will be automatically allocated to an internal project based on availability.
Selection rules:
Enter project codes (e.g.
abcd-001orxyzw-122) in order of preference, with choice 1 being your most preferred. Do not enter project titles or supervisor names - only project codes.You can only select projects advertised for your MSc course.
You cannot select more than 3 projects with the same main or second supervisor.
Submitting earlier does not increase your chances of getting your top choice - take your time to make an informed decision.
Resubmissions: If you have already submitted the form and would like to make changes, submit it again. Only your most recent submission before the deadline will be considered.
To learn more about a project, contact the main supervisor directly. It is your responsibility to gather all necessary information from the main supervisor to make an informed decision. Please clarify with the main supervisor any doubts, such as project scope, expectations, frequency and mode (in-person/remotely) of supervision meetings, and any other relevant details.
Important
We do our best to allocate students to their top choices, but this is often not possible. Therefore, we strongly encourage you to find the main supervisor yourself and propose your own project (see 🌟 Student-Proposed Projects). You can approach any academic or professional at any university or company worldwide.
Allocation procedure#
Supervisors submit project proposals (deadline early February).
Projects advertised to students (February 2026).
Students select 8 projects in order of preference.
IRP Team allocates students to projects using the allocation algorithm described below.
Warning
The IRP Team reserves the right to adapt or revise the outlined procedures as necessary.
Allocation algorithm
When multiple students select the same project, the IRP Team uses an algorithm designed to maximise the number of students allocated to their top choices. Students with higher ECTS-weighted mean marks at the time of allocation are given priority for oversubscribed projects.
All students must select internal projects, even if already allocated to an external or student-proposed project. This ensures you can be re-allocated to an internal project if issues arise with your external or student-proposed project. Each student will be allocated to at most one project. Students already allocated to their proposed or external project will not receive an internal project allocation unless their original project falls through.
Available projects#
Warning
This provisional list may contain metadata inaccuracies but has been released early to facilitate your selection process. New projects will be added as processed, though the majority are already included.
3D low-frequency extrapolation with ML for breast ultrasound data#
Project code: osag-061
Main supervisor: Oscar Calderon Agudo,
o.calderon-agudo14@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Torsten Hopp,
torsten.hopp@kit.edu, Karlsruhe Institute of Technology, GermanyAvailable to: ACSE
This project may accept multiple students.
Three-dimensional ultrasound computed tomography (3D USCT) offers transformative potential for breast imaging and breast cancer detection, enabling artefact-free reconstructions, arbitrary slice orientations, improved lesion characterisation, and faster data acquisition compared to conventional 2D approaches. However, current 3D reconstructions lack sufficient quality for diagnostic use and the application of high-quality 3D full-waveform inversion (FWI) at clinically relevant high-frequencies (>0.5 MHz) remains computationally prohibitive and limited by the limitations in the acquisition systems.
This project, a collaboration between the Imperial College London and the Karlsruhe Institute of Technology (KIT), aims to lay the groundwork for overcoming this challenge by utilising machine-learning (ML) models to predict the missing low-frequency components of the ultrasound data and improve the convergence and accuracy of full-waveform inversion for breast 3D USCT data. In the first stage, the project will investigate how existing 2D ML approaches for low frequency extrapolation—developed in geophysics and for 2D USCT—can be extended to 3D USCT. One of these methods will then be implemented, applied, and validated on a realistic 3D simulated breast USCT dataset, which the student will generate using the Stride library and the acquisition characteristics of the KIT USCT III device.
The results of this work will help assess whether combining 3D USCT with advanced FWI algorithms can offer a realistic, non ionising, and less painful alternative to the current gold standard mammography.
Reasoning Beyond Language: Exploring Physically-Grounded Knowledge via Video Generation#
Project code: roar-003
Main supervisor: Rossella Arcucci,
r.arcucci@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Che Liu,
che.liu21@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project does not accept multiple students.
Current multimodal evaluations often overlook “Reasoning Beyond Language,” where accurate visual generation acts as a rigorous test of internal world modeling , particularly for “Type III: Physical & Temporal Reasoning” involving dynamics like gravity and collision. To address the limitations of “dream-logic” in current video generation, we aim to explore a Physically-Grounded Video Generation framework. This project will investigate a hybrid architecture that leverages Autoregressive (AR) models for logical causal planning and Diffusion models for high-fidelity rendering, focusing on “Visual Conditioning” tasks to strictly test the model’s adherence to physical laws rather than open-ended imagination. Consequently, we expect the student to have a strong background in Diffusion, Autoregressive (AR) models, and video/image generation to successfully implement these complex deep learning architectures and evaluate them against the Generative Reasoning Hierarchy.
Learning patient-specific breathing dynamics from paired static CT through coupled transformer and physics-informed modelling#
Project code: roar-049
Main supervisor: Rossella Arcucci,
r.arcucci@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Ollie Pitts,
o.pitts23@imperial.ac.uk, Department of Computing, Imperial College LondonAvailable to: ACSE
This project may accept multiple students.
Ventilation issues in patients with chronic airway diseases arise from structural abnormalities and present with regional variation. Current pulmonary function testing methods such as spirometry and oscillometry are neither sensitive nor specific enough to characterise this heterogeneity. While visual analysis of CT scans can identify localised structural issues, this qualitative approach provides limited insight into breathing dynamics. At present, there is no clinically practical method to infer patient-specific breathing dynamics from routinely acquired imaging, highlighting a need for models that link structure to physiology and underlying fluid mechanics. A patient-specific digital twin of breathing would enable precise localisation of ventilation defects, improved characterisation of lung function, and prediction of bronchodilator responsiveness. Previous approaches have relied on 4D computed tomography (CT), which is rarely collected in airway disease and offers noisy motion fields with hazy structural detail. Physics-informed motion models derived from static CT have also been explored but typically assume global homogeneity in lung motion and are insufficiently bound by patient-specific physiology. In contrast, quantitative CT (qCT), provides voxel-level numerical insights to lung parenchyma and the airway tree that lends itself to accurate personalised modelling. Paired static inspiratory and expiratory CTs, combined with qCT-derived structural constraints, provide a widely available imaging format and a direct link to patient-specific abnormalities. Transformer-based models are uniquely suited to this setting as they can learn global relationships between airway geometry, parenchymal structure, and motion that cannot be captured by convolutional or rule-based approaches. Supported by recent advances in medical imaging foundation models, and available data from the ATLANTIS and UBIOPRED cohorts, we propose a project developing a computer-vision transformer that learns a three-dimensional deformation vector field (DVF) between expiratory and inspiratory states. Time-query tokens interpolate this deformation to generate a continuous breathing cycle, producing physiologically grounded transformation fields suitable for downstream physics-based airflow modelling. This project would suit two students working together towards:
A novel transformer framework to infer continuous breathing dynamics from paired static CT scans and quantitative lung metrics.
Integrating this output with physics-informed airflow simulation to infer lung function or ventilation patterns.
The students would develop strong python coding skills, a sound understanding of deep learning architectures, as well as familiarity with CT imaging, quantitative CT analysis software and statistical analysis.
Assessment of Robust Deep Learning–Based Change Detection Techniques Using Satellite Remote Sensing Datasets#
Project code: shaw-027
Main supervisor: Shubham Awasthi,
sawasthi@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Philippa Mason,
p.j.mason@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML GEMS READY
This project may accept multiple students.
Satellite remote sensing is a fundamental tool in geoscience for observing and monitoring changes on the Earth’s surface caused by natural processes and human activities. Accurate detection of such changes is particularly important for disaster monitoring and risk assessment related to landslides, floods, earthquakes, urban expansion, and environmental degradation. However, reliable change detection from multi-temporal satellite imagery remains challenging due to atmospheric effects, sensor noise, seasonal variability, illumination differences, and complex surface conditions. This project aims to assess robust deep learning–based change detection techniques using multi-temporal satellite remote sensing data. Optical datasets (e.g., Sentinel-2 or Landsat) and/or Synthetic Aperture Radar (SAR) data (e.g., Sentinel-1) will be utilised to capture complementary surface information relevant to geoscience and disaster monitoring applications. The study will focus on the implementation and evaluation of established deep learning models for change detection, such as convolutional neural networks (CNNs), Siamese networks, and encoder–decoder architectures (e.g., U-Net), alongside selected conventional change detection methods for benchmarking. The workflow will include data pre-processing steps such as radiometric and geometric correction, image co-registration, normalization, and noise reduction to ensure consistency across multi-temporal images. Deep learning models will be trained using labelled or semi-labelled datasets where available, and applied to generate change maps highlighting significant surface changes. The results will be analysed within a GIS framework to examine spatial patterns of change and their association with geophysical processes, land-use dynamics, or disaster events. The expected outcomes of the project include accurate change detection maps, a comparative performance assessment of deep learning–based techniques, and practical insights into their robustness across different sensors and scenarios. The project will provide students with hands-on experience in satellite data processing, deep learning implementation, and interpretation of change detection results for geoscience and disaster monitoring applications.
Land Deformation Analysis in Himalayan Towns Using Time-Series SAR Interferometry#
Project code: shaw-026
Main supervisor: Shubham Awasthi,
sawasthi@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Philippa Mason,
p.j.mason@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML GEMS READY
This project may accept multiple students.
Rapid urbanization, fragile geology, steep slopes, and intense hydro-meteorological processes make Himalayan towns highly vulnerable to land deformation phenomena such as subsidence, uplift, and slow slope movements. These processes pose serious risks to buildings, roads, and other critical infrastructure, yet continuous ground-based monitoring in mountainous terrain is often limited due to accessibility and cost constraints. In this context, spaceborne Synthetic Aperture Radar (SAR) interferometry provides a powerful, all-weather and day-night solution for large-area deformation monitoring. This project aims to analyse land deformation patterns in selected Himalayan towns using time-series SAR interferometry techniques based on multi-temporal Sentinel-1 data. Advanced time-series approaches such as Persistent Scatterer Interferometry (PS-InSAR) or Small Baseline Subset (SBAS) methods will be employed to retrieve millimetre-level ground deformation rates and deformation time series. The processing workflow will include SAR data pre-processing, interferogram generation, phase unwrapping, atmospheric phase screen mitigation, and deformation velocity estimation. The derived deformation products will be integrated with geographic information system (GIS) datasets to identify deformation hotspots and assess their spatial association with urban infrastructure, topography, and geomorphological settings. Particular emphasis will be placed on detecting zones of persistent subsidence or instability that may indicate underlying geotechnical or hydro-geomorphological processes. The expected outcomes of the project include deformation velocity maps, cumulative displacement time series, and deformation hotspot maps for the selected Himalayan town. The study will demonstrate the practical applicability of time-series InSAR techniques for urban hazard assessment in mountainous regions and provide students with hands-on experience in SAR data processing, geospatial analysis, and interpretation of land deformation processes relevant to disaster risk reduction and sustainable urban planning in the Himalaya.
Python-based domain-specific language for atomistic Hamiltonians#
Project code: mabe-175
Main supervisor: Marijan Beg,
m.beg@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor is not yet assigned.
Available to: ACSE EDSML
This project may accept multiple students.
In specific scenarios, computational studies are the only practical approach to tackle complex research challenges and efficiently design a range of products and systems. In nanomagnetism, simulations have become a prominent tool, often the only method for investigating diverse magnetic phenomena. Researchers employ several models in computational magnetism, each with advantages and disadvantages. Compared to continuous models, atomistic simulations enable us to simulate the magnetic moments of individual atoms and their interactions with other magnetic moments and the environment. However, the capabilities of simulation tools often limit scientists and engineers. For instance, if they wanted to include a new energy term in the energy equation, the developers did not implement it. Extending the capabilities of simulation tools is non-trivial - it requires dedicated resources and expert programming knowledge. Therefore, in this project, we will design and implement a domain-specific Python language that enables scientists and engineers to define custom atomistic energy terms. Our work would enable scientists and engineers to simulate any energy term without having to dive into the computational backend. No prior physics knowledge is necessary - all required physics concepts will be covered in the introductory learning sessions with the supervisor during the first few weeks. Supervision consists of introductory group learning sessions as well as group and individual supervision meetings. Please contact the main supervisor, Marijan Beg, if you would like to discuss this project in more detail.
Landau-Lifshitz-Gilbert-based micromagnetic computational backend#
Project code: mabe-176
Main supervisor: Marijan Beg,
m.beg@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor is not yet assigned.
Available to: ACSE EDSML
This project may accept multiple students.
Computational micromagnetics has become an essential tool in academia and industry, supporting fundamental physics research and the design and development of devices for data storage, information processing, sensing, and medicine. We have designed and developed a human-centred research environment called Ubermag. With Ubermag, scientists can control different micromagnetic computational backends. The complete simulation workflow, including definition, execution, and data analysis of simulation runs, can be performed within a single Python session. Furthermore, numerical libraries co-developed by the computational and data science community can be used immediately for micromagnetic data analysis within Ubermag. In this project, we will develop a Python-based micromagnetic computational backend in Python. Although this model simulates magnetic nanosystems at zero temperature, it is necessary to simulate the system’s time evolution. Using finite differences, we will extend Ubermag to compute different energy terms and simulate the system’s time evolution by integrating the Landau-Lifshitz-Gilbert equation. The main deliverable will be a well-tested, reliable, and documented extension that allows scientists and engineers to explore magnetic phenomena, both static and dynamic. No prior physics knowledge is necessary - all required physics concepts will be covered in the introductory learning sessions with the supervisor during the first few weeks. Supervision consists of introductory group learning sessions as well as group and individual supervision meetings. Please get in touch with the main supervisor, Marijan Beg, if you would like to discuss this project in more detail.
Enhancing Feedback in Higher Education with LLMs#
Project code: mabe-172
Main supervisor: Marijan Beg,
m.beg@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor is not yet assigned.
Available to: ACSE EDSML
This project may accept multiple students.
Large Language Models (LLMs) are transforming higher education, offering innovative ways to support both students and educators. However, while student numbers continue to rise, the number of teaching staff remains limited. This creates a major challenge: high-quality, personalised feedback is essential for learning, yet the time available for staff to provide it is shrinking. Delegating feedback entirely to LLMs might seem like a solution, but it could make students feel they are not receiving “value for money” and that they have no contact with teaching staff. Instead, this project aims to enhance human feedback with AI support rather than replace it. We will develop an AI-assisted feedback expansion tool in which markers provide concise feedback points via a Graphical User Interface (GUI). The system will then use an LLM to expand this feedback by incorporating explanations, concrete examples, and tailored exercises. Importantly, the LLM will be based on lecture notes and the provided reading list, ensuring all generated feedback aligns with the course content. This approach ensures that feedback is detailed, pedagogically relevant, and contextualised. Supervision consists of group and individual supervision meetings.
Field manipulation of emergent magnetic monopoles#
Project code: mabe-174
Main supervisor: Marijan Beg,
m.beg@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor is not yet assigned.
Available to: ACSE EDSML
This project may accept multiple students.
The prediction of topologically stable magnetic skyrmions being used to change how we store and process data has led to materials with the Dzyaloshinskii-Moriya interaction becoming the focus of intensive research. It was recently demonstrated that nanodisks consisting of two layers with opposite chirality could host a stable emergent magnetic monopole - Bloch point (a three-dimensional singularity of the magnetisation field). For future data storage and information processing devices, it is necessary to explore the manipulation of Bloch points using various driving methods. In this project, we will implement a Python-based micromagnetic simulation tool to investigate whether it is possible to manipulate the Bloch point in helimagnetic nanostrips and under what conditions this can be achieved. More precisely, we will simulate the translational motion of Bloch points in planar nanostructures. The results of this work will allow us to determine whether a Bloch point can be an information carrier in racetrack-like data storage devices. No prior physics knowledge is required - all necessary physics concepts will be covered in the introductory learning sessions during the first few weeks. Supervision consists of introductory group learning sessions as well as group and individual supervision meetings. Please contact the main supervisor, Marijan Beg, if you would like to discuss this project in more detail.
Eigenmode analysis of topologically stable quasi-particles#
Project code: mabe-173
Main supervisor: Marijan Beg,
m.beg@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor is not yet assigned.
Available to: ACSE EDSML
This project may accept multiple students.
Topologically stable quasi-particles could fundamentally change how we store and process data. However, for their applications in future data storage and information processing devices, it is necessary to develop a detection (reading) method to determine whether they encode bit 0 or 1. One possible method is ferromagnetic resonance, in which the magnetic sample is excited by an external magnetic field, and the presence of a topologically stable quasiparticle is determined from the ferromagnetic resonance response. In this project, we will develop a computational and data analysis tool in Python to analyse magnetisation dynamics data. This tool will allow us to analyse resonance frequencies and spatially resolved eigenmodes, and to propose a method for reading data from storage devices. No prior physics knowledge is needed - all required physics concepts will be covered in the introductory learning sessions with the supervisor during the first few weeks. Supervision consists of introductory group learning sessions and group and individual supervision meetings. Please contact the main supervisor, Marijan Beg, if you would like to discuss this project.
Bloch point exploration using the mean-field model of interacting non-collinear spins#
Project code: mabe-169
Main supervisor: Marijan Beg,
m.beg@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor is not yet assigned.
Available to: ACSE EDSML
This project may accept multiple students.
Complex magnetic materials hosting topologically nontrivial particle-like objects, such as skyrmions, hopfions, and Bloch points, are under intensive research. They could fundamentally change how we store and process data. One crucial material class are helimagnetic materials with Dzyaloshinskii-Moriya interaction, which emerges in systems lacking symmetry. It was recently demonstrated that nanodisks consisting of two layers with opposite chirality can host a stable Bloch point. So far, simulations have been performed using micromagnetic models at zero temperature. However, since the magnetisation norm vanishes at the Bloch point, it is necessary to explore the stability and Bloch-point structure using different models. In this project, we will implement a mean-field model of interacting non-collinear spins and explore whether Bloch points are stable in nanodisks or are merely an artefact of micromagnetic simulations. In addition, the mean-field model will allow us to examine the room-temperature stability of the Bloch point, which is necessary for applications in future data storage and information-processing devices. No prior physics knowledge is needed - all required physics concepts will be covered in the introductory learning sessions with the supervisor during the first few weeks. Supervision consists of introductory group learning sessions and group and individual supervision meetings. Please contact the main supervisor, Marijan Beg, if you would like to discuss this project in more detail.
Real-Time LLM-Powered Feedback for Asynchronous Python Learning#
Project code: mabe-170
Main supervisor: Marijan Beg,
m.beg@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor is not yet assigned.
Available to: ACSE EDSML
This project may accept multiple students.
Large Language Models (LLMs) are transforming higher education, offering new ways to support student learning. Today, much of learning happens asynchronously, meaning that students engage with course materials, complete exercises, and seek help at different times rather than in a live classroom setting. While this flexibility is beneficial, one key challenge remains: students need immediate, high-quality feedback to learn and improve effectively. Traditional feedback often comes with delays, leaving students uncertain about their mistakes and hesitant to experiment. Automated feedback can remove this barrier, allowing students to make mistakes freely and request feedback as many times as needed - without fear of judgment. This fosters a more interactive and iterative learning experience. In this project, we will develop an AI-powered feedback tool for learning basic Python programming, embedded directly into a Jupyter Notebook. The system will compare student solutions against hidden reference solutions provided by the instructor. The LLM-based system will generate formative feedback that identifies errors, helps students understand their mistakes, and improves their approach over time. This ensures that students receive real-time guidance tailored to their specific needs. Supervision consists of group and individual supervision meetings.
Unlocking satellite ground motion data to evaluate tectonic processes, environmental change, and climate susceptibility#
Project code: rebe-044
Main supervisor: Rebecca Bell,
rebecca.bell@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: James Wood,
james.wood18@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: EDSML GEMS READY
This project may accept multiple students.
(Supervised by Rebecca Bell, Valerie Locher and James Wood)
Understanding how the surface of the Earth responds to dynamic processes is fundamental to analysing active tectonics, environmental change, and climate sensitivity. The quantification of ground motion has been revolutionised in recent years through the development of satellite-derived, ground motion time series datasets that exploit repeat-pass InSAR imagery to measure change at high spatial and temporal resolution on annual to decadal timescales. Since 2021, availability to such datasets has expanded significantly thanks to the release of the European Ground Motion Service (EGMS). The EGMS provides 5-year InSAR time series data (in Line of Sight, Vertical and East-West components of ground motion) across most of Europe, including the tectonically active and water-scarce country of Greece where this dataset is currently being explored in detail.
Projects using EGMS data could aim to use data science and/or machine learning techniques to: • Isolate and evaluate potential earthquake precursors (i.e. indicators happening before an earthquake) from ground motion data • Extract the seasonal component of ground motion time series to explore how hydrological cycling and environmental / climate sensitivity are recorded in the dataset, and provide a de-seasonalised time series dataset to better evaluate non-hydrological processes (e.g., tectonic ground motion) • Detrend the the East-West component dataset to remove overprint from continental-scale plate motion and unlock tectonic analysis at fault-scale (i.e. 10s km) • Automate coastal susceptibility assessment to rising sea levels and monitoring of at-risk areas
The EGMS dataset can be combined with other data to support analysis (i.e., rainfall, soil moisture, GPS and seismological data). This project will provide experience applying computational techniques to an extremely large, remotely sensed satellite dataset with applications across tectonic geology, environmental science, hydrology and engineering.
Characterizing the erosion and basal fill of deepwater channels; Implications for sediment transport to the deep oceans and heterogeneity of channel-fills.#
Project code: rebe-051
Main supervisor: Rebecca Bell,
rebecca.bell@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Mike Mayall,
m.mayall@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: GEMS
This project does not accept multiple students.
Channels, which are the main arteries of submarine fans, are commonly kilometres wide and hundreds of meters deep. The initial deposits, “basal lag” in a channel are poorly understood. They reflect a complex period when sedimentary processes were both eroding and depositing sediment, with potential for significant sediment bypass. and may include extra-formational conglomerates, mudclast conglomerates, mud drapes, slumps and debrites, as well as different styles of sandstone. There is no consensus on why these facies are so variable and the cause(s) of this variability. In modern systems this part of the fill can be the focus for the accumulation of detrital carbon and anthropomorphic pollutants. In ancient systems this part of the fill may control the gross permeability architecture of a sandbody, which are used as sites for carbon capture and storage. This project will be particularly suited to a geoscientist, with an interest in sedimentary facies and reservoir characterization.
The University of Leeds Deep Marine Architectural Knowledge Store (DMAKS) database includes information on the dimensions and fill of thousands of different channel-fills from across the globe and through geological time. The data base contains over 6,000 data entries. It provides a unique opportunity to statistically analyse the facies, sequences of facies and their relationship to numerous variable controls such as channel dimensions, grain size and tectonic setting using data science techniques. The key deliverables will be -
Produce a comprehensive statistical analysis of the nature of the facies in channel basal and their geological setting
Documentation of the range of facies and facies associations observed in the basal lags for a number of contrasting channel systems
Interpretation of the range of basal lag facies and their inferred sediment transport processes under different bounding conditions, using for example, multivariate statistics to investigate the relationships
Analytical analysis of imbibition in mixed-wet media#
Project code: mabl-011
Main supervisor: Martin Blunt,
m.blunt@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Branko Bijeljic,
b.bijeljic@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: GEMS
This project may accept multiple students.
You will extend the analysis of spontaneous imbibition analytically and numerically that you will be taught in the GEMS2 course. The emphasis of the project will be to study mixed-wet media and compare predictions of imbibition rate and saturation profiles with experimental data. You will be asked to constrain multiphase flow properties - specifically the water relative permeability - to match the measured imbibition rate. The work will integrate analytical and numerical methods with a critical analysis of the literature with the use of experimental data.
AUV-based seismic acquisition and full-waveform inversion for offshore wind farm development#
Project code: edca-031
Main supervisor: Edward Caunt,
edward.caunt15@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Rebecca Bell,
rebecca.bell@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML GEMS READY
This project does not accept multiple students.
The energy transition has introduced a massively increased requirement for alternative energy sources, resulting in the proliferation offshore wind farms worldwide. This has been accompanied by increased demand for ultra-high-resolution (UHR) seismic data to characterise the shallow subsurface to de-risk offshore wind turbine construction (Caselitz 2025). Whilst traditional seismic processing workflows such as Kirchoff migration are widely used in this domain, modern, wave- equation-based workflows have been found to yield superior results (Espin 2022). In particular, full-waveform-inversion (FWI) imaging (Zhang 2020) has been demonstrated to yield extremely sharp imaging (Espin 2022), with a nominal meter-scale lateral resolution (Jiang 2024). From a high-resolution subsurface model, obtained via high-frequency FWI, impedance contrasts are calculated, directly yielding a reflectivity image with balanced illumination (Zhang 2020), utilising the full wavefield (including diving waves, multiples, and ghosts) for image building (Espin 2022), and negating the need for preprocessing steps such as regularisation and demultiple.
However, such workflows incur higher costs than conventional approaches, being both substantially more computationally expensive, and requiring specialised acquisition geometries featuring both long offsets for deep updates (Warner 2013) and source-over-spread (Lie 2020) acquisition to reduce minimum offsets (Espin 2022). In recent years, autonomous marine vehicles have been explored as an alternative to traditional seismic survey ships (Moldovenanu 2017, Cheong 2021), offering low costs, reduced environmental impact, and potential for increased endurance. The use of autonomous underwater vehicles (AUVs) enables 3D positioning of sensor arrays in the water column (Moldovenanu 2017), whilst providing greater flexibility of acquisition geometry and enabling acquisition at very long offsets (Moldovenanu 2017). Furthermore, FWI workflows based on compute distributed across the robotic platforms used in the acquisition has been demonstrated (Shin 2024), highlighting the potential for more intelligent seismic acquisition systems applying iterative decision-based adaption of the acquisition geometry to maximise coverage and image quality.
The goal of this project is to conduct a modelling and synthetic study exploring the feasibility of AUV-based seismic acquisition for the low-cost acquisition of datasets suitable for FWI imaging of UHR seismic data. Such an approach could have substantial implications for the availability of high-quality seismic images for subsurface characterisation to the renewables sector. The core of this project will involve the evaluation of geometries and acquisition strategies made possible by the AUV acquisition paradigm, alongside the effect of uncertainties such as positioning error, exploring their impact on imaging quality to inform potential future hardware and surveys. This work will utilise Devito, a domain-specific language and compiler for finite-difference computations, to build a specialised modelling setup and FWI workflow tailored to the problem at hand.
Mixed-precision tiling strategies for high-performance seismic wave simulation#
Project code: edca-054
Main supervisor: Edward Caunt,
edward.caunt15@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Gerard Gorman,
g.gorman@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE
This project does not accept multiple students.
Modern seismic imaging algorithms are underpinned by numerical solvers, using finite-difference methods to solve an array of wave equations. Finite-difference methods are typically memory-bound, that is, the performance bottleneck which ultimately determines the speed of the kernel is the rate at which values can be read from and written to memory. By utilising half-precision floating-point values (FP16) to store the wavefields, memory pressure is accordingly halved, yielding an approximate doubling of throughput. However, the restricted dynamic range and precision of FP16 versus the more typical FP32 is liable to result in overflows or rounding errors if the problem is not suitably scaled - in practice this requires either introduction of scaling factors to the equations themselves or careful selection of units (or even non-dimensionalisation). Mixed-precision strategies have proliferated in high-performance computing applications, retaining precision where it is needed and discarding where it is not. Indeed, this approach is often partially utilised in seismic applications, whereby time-invariant material parameters are stored in reduced-precision formats. Mixed-precision representation the unscaled wavefields requires a more sophisticated approach; wavefronts feature large amplitudes and dynamic range, whilst surrounding regions may be substantially more subdued. By dividing the computational domain into tiles of varying precision, high-dynamic-range regions can be represented in single or even double precision, whilst those with low dynamic range can be represented in half. Furthermore, the precision of each tile can be selected dynamically as the simulation progresses, minimising memory pressure whilst preserving precision where it is needed.
This project would suit a student interested in performance engineering, numerical methods, and designing high-performance computing strategies for modern hardware including the latest GPU models. This project will make use of Devito, a domain-specific language and compiler for finite-difference computations, embedded in Python, to develop and evaluate an implementation and evaluation of this tiling approach.
Assessing the impact of clouds on climate change#
Project code: pace-167
Main supervisor: Paulo Ceppi,
p.ceppi@imperial.ac.uk, Department of Physics, Imperial College LondonSecond supervisor is not yet assigned.
Available to: EDSML
This project may accept multiple students.
Clouds have a profound impact on Earth’s radiation budget, by regulating how much solar radiation is absorbed and how much thermal infrared is emitted to space. As the climate warms, clouds change in subtle ways that can amplify or dampen global warming – a phenomenon known as cloud feedback. This is at the heart of uncertainties in climate change projections, particularly because global climate models struggle to accurately simulate clouds and their future changes.
In this project, we will assess an aspect of cloud feedback that has been understudied – the seasonality of the cloud response to warming; previous analyses have largely focused on annually-averaged changes. We know that certain key cloud types exhibit a strong seasonal cycle, so it is likely that their response is also seasonally dependent. Combined with the seasonally-varying insolation, it is plausible that seasonal variations in the cloud response are important for the feedback, but whether this is the case is currently unknown.
In previous work, I have developed a statistical learning framework to assess cloud feedback from global satellite observations, and applied this framework to constrain the climate model representation of this feedback. The method involves assessing relationships between clouds and meteorological factors, using regularised regression. As part of this project, you will extend and refine this framework; we can also discuss alternative methods, for example based on neural networks.
The analysis will use existing model simulations of global climate under present-day and elevated CO2 conditions. The analysis will be performed in Python (or similar) and will require basic statistics skills. The project will also involve learning the basics of atmospheric and cloud physics in order to interpret the modelling results.
Multiphase flow modelling using AI4PDEs#
Project code: boch-166
Main supervisor: Boyang Chen,
boyang.chen16@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Nathalie Carvalho Pinheiro,
n.pinheiro23@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
AI4PDEs is an in-house computational fluid dynamics (CFD) solver embedded into Neural Physics framework [1] that solves discretized systems of partial differential equations using convolutional neural network (CNN) architectures. Unlike conventional machine-learning approaches, AI4PDEs does not rely on data-driven training: the network weights are directly derived from the governing physical equations. This approach enables benefiting from the high computational efficiency and parallelization implemented in modern AI libraries, while retaining full interpretability and physical consistency [2]. In this project, students will apply AI4PDEs to simulate multiphase flows—systems in which two or more phases (e.g., liquid and gas) interact dynamically within a shared domain. Multiphase flow plays a key role in many Earth-science and engineering applications, including collapsing dam and 3D flooding scenarios, geothermal systems, carbon dioxide transport, and industrial pipeline flows such as those found in nuclear power plants. Variations in boundary and initial conditions can give rise to distinct flow regimes in a pipeline, and understanding these patterns is essential for improving system efficiency, safety, and cost effectiveness. The project is hands-on: students will implement targeted modifications to the AI4PDEs codebase in Python, design and run numerical test cases, and develop pre- and post-processing tools for visualization and analysis (with the option to build a graphical user interface if time permits). The project is well suited for students interested in scientific computing and the application of modern AI techniques to physical problems in Earth sciences and engineering. [1] Chen, B., Heaney, C. E., & Pain, C. C. (2026). Neural Physics: Using AI Libraries to Develop Physics-Based Solvers for Incompressible Computational Fluid Dynamics. Computers & Fluids, 106981. [2] Chen, B., Heaney, C. E., Gomes, J. L., Matar, O. K., & Pain, C. C. (2024). Solving the discretised multiphase flow equations with interface capturing on structured grids using machine learning libraries. Computer Methods in Applied Mechanics and Engineering, 426, 116974.
Controlling ventilation in buildings using Generative Networks (GANs/VAEs/AAEs) to produce comfortable healthy environments that are energy efficient#
Project code: boch-151
Main supervisor: Boyang Chen,
boyang.chen16@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Christopher Pain,
c.pain@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
Energy-efficient buildings are central to implementing successful carbon reduction strategies. In 2019, the UK’s total energy consumption was 142 Mtoe (energy equivalent to burning 142 Mt of oil), and buildings account for 40% of this (IEA 2021). [1] and [2] have shown that full-automation can save 50-60% of HVAC (Heating, ventilation, and air conditioning) energy consumption. An AI control system might expect to meet that saving with associated reduction in CO2 emissions and thus greatly contribute to net zero. Smart controls and connected devices could save 230 EJ in cumulative energy savings by 2040 (IEA), lowering energy consumption of buildings and the associated carbon footprint by as much as 10% globally, while improving comfort for occupants. With this in mind, in this project, we will apply generative neural networks (GANs/VAEs/AAEs and the lattest latent diffusion models) to (1) Predict the transient spatial distribution of temperature, relative humidity, CO2 concentration and pollution concentration within a building given the outside weather conditions and ventilation settings. (2) Assimilate sensor data (e.g. temperature, CO2) and room occupancy data into the AI model in order to predict the future room conditions. (3) Perform control of the ventilation settings in order to save energy and produce healthy living conditions. (4) Perform uncertainty quantification in order to determine uncertainties on controls and predictions. (5) Apply generative AI methods in order to produce priors of initial room conditions (much like weather prediction models) and integrate this with an AI4PDE model [3,4,5] of the detailed air flow and temperature, relative humidity and CO2 distribution within a room.
Project 1 will complete tasks 1 and 2, project 2 will complete tasks 3 and 4, and project 3 will address task 5. These projects would suit students with an interest in computational fluid dynamics, energy including net-zero objectives, neural networks (especially generative networks) and optimisation (such as data assimilation and control).
[1] Schiavon, Melikov, Sekhar (2010) Energy analysis of the personalized ventilation system in hot and humid climates. Energy and Buildings, 42(5):699-707.
[2] Khalil (2020) Computer Simulation of Air Distribution and Thermal Comfort in Energy Efficient Buildings.
[3] Chen, Heaney, Pain (2024) Using AI libraries for Incompressible Computational Fluid Dynamics, in preparation.
[4] Chen, Heaney, Gomes, Matar, Pain (2024) Solving the Discretised Multiphase Flow Equations with Interface Capturing on Structured Grids Using Machine Learning Libraries, arXiv preprint. https://doi.org/10.48550/arXiv.2401.06755
[5] Phillips, Heaney, Chen, Buchan, Pain (2023) Solving the Discretised Neutron Diffusion Equations Using Neural Networks, International Journal for Numerical Methods in Engineering 124(21):4659-4686. https://doi.org/10.1002/nme.7321
Using machine learning to distinguish between clouds and smoke from wildfires in raw satellite imagery#
Project code: ancl-077
Main supervisor: Andrew Clelland,
a.clelland@imperial.ac.uk, Department of Computing, Imperial College LondonSecond supervisor: Rossella Arcucci,
r.arcucci@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project does not accept multiple students.
Smoke from wildfires affects air quality, contributes to global warming, damages aircraft engines and can force large emergency responses. Distinguishing smoke from clouds in raw, near real-time satellite imagery remains challenging due to their visual similarity, semi-transparent structure and regional variability in atmospheric conditions and fire behaviour. Although machine learning approaches for cloud masking and smoke detection exist, most models are trained and evaluated within limited geographical contexts, and their ability to generalise across regions with contrasting fire regimes is poorly understood. This project investigates whether state-of-the-art deep learning architectures can robustly discriminate wildfire smoke from clouds across several geographical areas with distinct fire and atmospheric properties, such as boreal forests, the Western United States, and the Amazon rainforest.
Raw multispectral satellite imagery from Sentinel-2 and/or Landsat-8 and -9 can be accessed via Google Earth Engine. Additional training datasets for clouds (e.g. CloudSEN12) and smoke/fire (e.g. FASDD) detection are publicly available. The choice of the machine learning model is subject to the student’s discretion, however the code should be written in Python. Examples of suitable model choices include vision transformers or a hybrid CNN-transformer model, which can be compared against a baseline UNet model. Model performance shall be quantified using appropriate metrics.
By the end of the project, the student will have created a state-of-the-art machine learning model which will, to a reasonable extent, be able to distinguish between clouds and wildfire smoke from raw, near real-time satellite imagery. The student will have evaluated and compared the performance of the model against a baseline UNet model in at least two contrasting geographical locations, providing insights into the model’s strengths, limitations and uncertainty.
Full-waveform inversion for brain vascular super-resolution#
Project code: cacu-206
Main supervisor: Carlos Cueto,
c.cueto@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Qi Gao,
qi.gao24@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE
This project may accept multiple students.
Motivation:
Blood flow, and its disruption, is the key physiological process that sustains the healthy and pathological function of the brain. Blood is delivered to the brain tissue through the microvasculature (the smallest vessels in the brain, with a diameter below 100 microns). It is by imaging the microvasculature that it is possible to understand, diagnose and prevent most neurological disorders; currently, no imaging modality can map the adult human microvasculature.
Ultrasound, applying Nobel prize-winning advances in super-resolution together with safe contrast agents, has been used to successfully image the microvasculature throughout the human body. However, the skull, a sound distorter, has prevented its application to brain imaging. Recently, it has been shown that full-waveform inversion (FWI), a physics-driven imaging methodology, can enable safe ultrasound imaging through the skull; this project will exploit these advances to demonstrate, for the first time, that ultrasound FWI can be used to super-resolve the brain microvasculature.
Method:
The project will combine well-established tools for acoustic modelling and for modelling microbubble contrast agents —including scattering and dynamic response simulation— with advanced FWI imaging algorithms to generate a computational proof-of-concept for FWI super-resolution imaging.
Building on forward models that explicitly account for microbubble scattering, motion, and related acoustic effects, the project investigates how microstructural information can be effectively recovered during the FWI inversion process, and how high-resolution representations relevant to vascular structure and flow can be extracted from the reconstructed fields. To ensure verifiability, the project will rely on simulation-based phantom toolchains with known ground truth, enabling systematic assessment of reconstruction quality and stability under varying noise levels, sparsity, and acquisition conditions.
Differentiable imaging of fluid flows#
Project code: cacu-207
Main supervisor: Carlos Cueto,
c.cueto@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Qi Gao,
qi.gao24@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE
This project may accept multiple students.
Motivation:
Particle image velocimetry (PIV) and image-sequence-based vector flow imaging (VFI) methods are the key computational techniques used to image the velocity of fluid flows (their magnitude and direction), with applications ranging from physics research to medical imaging. Classical PIV and VFI pipelines, based on correlation- or matching-based approaches, are robust and interpretable in practice, but they rely heavily on expert knowledge and numerous heuristic hyperparameters, often requiring repeated trial-and-error under noisy conditions, varying particle densities, or complex flow patterns.
By contrast, modern learning-based optical flow and deep PIV approaches offer higher automation and resolution, but are typically harder to interpret and more sensitive to data distribution shifts. The project will use modern differentiable-programming paradigms, like JAX or PyTorch, to redesign the key steps of classical PIV/VFI pipelines into a set of composable, differentiable modules, preserving the structure and interpretability of traditional methods while enabling more systematic optimisation and integration with learning-based techniques.
Method:
Following the theme of a “differentiable classical pipeline,” the project explores which standard processing steps can be directly implemented as differentiable operators, and which require smooth or robust differentiable approximations. This design allows end-to-end gradients to pass through matching and displacement estimation stages, making it possible to analyse parameter sensitivity and error propagation within a unified framework. Physical consistency, such as smoothness, incompressibility, or boundary-related constraints, can then be introduced as differentiable terms that influence the estimation process itself rather than acting solely as post-hoc filters.
The project aims to transform many hand-tuned design choices in traditional workflows into learnable, testable, and ablatable components. The resulting framework will provide a controlled research baseline that links observational consistency and physical consistency within a single optimisation view, while leaving room for future extensions that combine these differentiable pipelines with lightweight learning modules or data-driven priors.
Meta-learning full-waveform inversion#
Project code: cacu-204
Main supervisor: Carlos Cueto,
c.cueto@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Ben Moseley,
b.moseley@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE
This project does not accept multiple students.
Motivation:
Ultrasound full-waveform inversion (FWI) is a powerful imaging technique with the potential to provide high-resolution, safe, and portable alternatives to conventional brain imaging methods like MRI and CT scans. However, FWI is computationally expensive, requiring iterative numerical optimisation to reconstruct high-quality images. Additionally, its accuracy is often limited by the ill-posedness of the inverse problem, and FWI can struggle to converge to high-fidelity solutions.
Recent advances in machine learning (ML) and differentiable physics offer exciting opportunities to address these challenges. By leveraging data-driven methods, we can learn more efficient and accurate inversion algorithms, potentially transforming FWI into a faster and more reliable tool for medical imaging.
Method:
In this project, the student will investigate the use of meta-learning and differentiable physics to improve FWI algorithms. The key idea is to learn better optimization strategies from a large dataset of synthetic brain velocity models, which serve as ground truth. Specifically, the project will explore two approaches:
Learning gradient descent steps: using meta-learning techniques, we will optimize the steps of the gradient descent process to accelerate convergence and improve the reconstruction quality of FWI.
Exploring alternative search algorithms: beyond gradient descent, the student will investigate more advanced search methods, such as tree search, which may provide better solutions for the highly non-convex FWI problem.
The project will include implementing differentiable physics models to incorporate physical constraints and ensure that learned algorithms are grounded in the underlying wave physics.
Advanced Machine Learning for Urban Atmospheric Flow Prediction and Environmental Management#
Project code: fafa-021
Main supervisor: Fangxin Fang,
f.fang@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Linfeng Li,
l.li20@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
This project focuses on the prediction of urban velocity fields with explicit representation of three-dimensional building geometry and the influence of the atmospheric boundary layer, with direct relevance to applications such as safe and efficient drone operations in complex urban environments. Accurate modelling of wind flow in cities is critical for understanding ventilation, pollutant dispersion, pedestrian comfort, and the performance of low-altitude aerial vehicles operating near buildings. Numerical models play a vital role in environmental forecasting by providing a quantitative framework for simulating complex atmospheric processes. In urban settings, these models enable detailed analysis of flow–structure interactions, long-term environmental trends, and the impacts of human activities on local climate and air quality. Such capabilities support informed urban planning and environmental management. However, high-fidelity numerical simulations that resolve detailed urban geometry are computationally expensive, which limits their suitability for real-time prediction and operational applications. Machine learning (ML) models offer a powerful complementary approach to traditional physics-based methods. By leveraging large datasets from observations and simulations, ML algorithms can efficiently capture complex, non-linear relationships in urban atmospheric flows arising from heterogeneous building layouts, surface roughness, and thermal effects. These data-driven models can provide rapid predictions, making them particularly attractive for real-time applications such as drone navigation and urban monitoring. Despite these advantages, ML models face challenges related to interpretability, uncertainty quantification, and their ability to consistently respect physical laws governing atmospheric dynamics. This MSc project aims to address these challenges by integrating physical understanding with data-driven methods for real-time urban atmospheric prediction. Advanced machine learning techniques—such as Transformers, graph neural networks, generative models, and data assimilation frameworks—will be explored to improve velocity field prediction in complex urban environments. The outcomes of this research will support applications in drone navigation, urban environmental monitoring, and sustainable city management.
Geomechanics and Fractured Porous Media Flow Modelling#
Project code: adfa-108
Main supervisor: Ado Farsi,
ado.farsi@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Robert Zimmerman,
r.w.zimmerman@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: GEMS READY
This project may accept multiple students.
This MSc project, supervised by Dr Ado Farsi and Prof Robert Zimmerman FREng, in collaboration with the startup Tanuki Technology Ltd., focuses on advancing computational geomechanics and fractured porous media flow modelling for next-generation subsurface simulation tools. These models underpin critical geo-energy and subsurface applications such as geothermal energy production, underground hydrogen storage, geological CO₂ storage, and broader reservoir and waste-isolation problems. Students will be assigned one of two key objectives:
Robust Numerical Methods for Fractured Porous Media Flow
Develop and implement advanced finite element / Discontinuous Galerkin (DG) methods for flow in fractured porous media using reduced-dimensional fracture representations.
Formulate stable matrix–fracture coupling strategies (interface fluxes, penalty stabilisation, conservation across fractures) suited to high permeability contrasts and complex fracture networks.
Investigate stability and accuracy (consistent fluxes, mass conservation, convergence behaviour) and validate using benchmark and manufactured-solution tests.
Build a verification suite and report accuracy and computational performance through systematic mesh refinement and parameter studies.
Coupled Hydro-Mechanical (Geomechanics) Modelling of Fractured Media
Extend fractured flow models to include geomechanics effects (e.g., poroelasticity, effective stress, pressure–deformation coupling) relevant to subsurface engineering and geo-energy applications.
Incorporate fracture constitutive behaviour such as compliance/closure and pressure-dependent fracture transmissivity or permeability.
Implement and compare coupling strategies, focusing on robustness and scalability.
Validate coupled simulations using canonical poroelastic benchmarks and fractured-media test problems, demonstrating physically meaningful behaviour under realistic parameter regimes.
This project offers hands-on experience in geomechanics, fractured porous media modelling, and high-performance scientific computing for subsurface simulation, providing valuable skills for careers in research and industry.
Parallel compilation strategies for high-performance finite-difference codes#
Project code: gego-056
Main supervisor: Gerard Gorman,
g.gorman@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Edward Caunt,
edward.caunt15@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE
This project does not accept multiple students.
Devito is a domain-specific language (DSL) and compiler for finite-difference methods, generating optimised C code from a high-level symbolic Python specification. Initially developed in the department of Earth Science and Engineering at Imperial College London, Devito has gone on to be widely used across industry and academia in a range of applications. However, as these applications have grown in variety and scope, the potential for particularly complex equations to stress the compiler performance has become apparent, in many cases as a consequence of the sequential nature of the Python language in which the compiler is written.
As of Python 3.13, it is possible to disable the Global Interpreter Lock (GIL), enabling multiple threads to execute concurrently; this capability could be used to accelerate elements of the Devito compiler which are currently bottlenecked by sequential execution. Several obstacles to leveraging this capability remain however, including thread safety, memory management, and automated selection of threaded/sequential backends for backward compatibility.
This project will develop and evaluate multithreaded parallelism strategies to accelerate compilation of finite difference codes, with particular focus on routines such as factorisation and common-subexpression elimination. It will require knowledge of parallel programming (experience working with OpenMP and MPI for example). This project represents and opportunity to contribute to a widely-used open-source codebase, develop good research software engineering practices, and gain experience of working with established, real-world codebases.
AI-Assisted Reengineering of Legacy Scientific Codes Using the Horseshoe Model and LLM Agents#
Project code: gego-184
Main supervisor: Gerard Gorman,
g.gorman@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Rhodri Nelson,
rhodri.nelson@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE
This project may accept multiple students.
Problem: Critical scientific and engineering software — in geophysics, climate modelling, computational physics, and beyond — remains locked in legacy codebases (Fortran, C, C++) that are difficult to maintain, extend, and port to modern hardware. These codes often contain architecture-specific optimisations, hardware intrinsics, and platform-dependent idioms that tightly couple algorithmic intent to obsolete targets, making portability and maintainability increasingly costly. Manual rewriting typically requires years of expert effort, while naive automated translation produces unidiomatic, unverifiable output. The Horseshoe Model (Kazman et al., 1998) provides a well-established framework for code modernisation — abstracting legacy code upward to a semantic representation, then regenerating downward into modern targets — but its upper levels have historically required manual expert intervention, limiting practical adoption.
Methodology: This project will develop and evaluate an agentic AI pipeline that operationalises the Horseshoe Model for legacy scientific code. The student will implement a multi-stage workflow: (1) LLM-driven semantic lifting of legacy routines, using computational graph analysis to separate algorithmic intent from platform-specific optimisations and identify core primitives (stencil operations, linear algebra, spectral transforms); (2) progressive construction of test suites during the abstraction phase, capturing input-output behaviour at each computational node; and (3) iterative LLM-driven regeneration into a modern, portable target (e.g., Python/NumPy or a domain-specific language such as Devito), with CI-driven correction cycles validating functional equivalence against captured legacy behaviour.
Kazman, R., Woods, S.G. & Carrière, S.J. (1998). “Requirements for Integrating Software Architecture and Reengineering Models: CORUM II.” Proc. Fifth IEEE Working Conference on Reverse Engineering, pp. 154–163. ResearchGate Wagner, C. (2014). Model-Driven Software Migration: A Methodology. Springer. Link Zhang, N., Rao, S., Franusich, M. & Franchetti, F. (2025). “Towards Semantics Lifting for Scientific Computing: A Case Study on FFT.” arXiv:2501.09201 Guo, D. et al. (2021). “GraphCodeBERT: Pre-Training Code Representations with Data Flow.” ICLR 2021. arXiv:2009.08366 Ben-Nun, T., Jakobovits, A.S. & Hoefler, T. (2019). “Stateful Dataflow Multigraphs: A Data-Centric Model for Performance Portability on Heterogeneous Architectures.” Proc. SC ‘19. arXiv:1902.10345 Wang, Y. et al. (2023). “SANN: Algorithm Detection via Subtree Extraction and Two-Way Embedding.” ACM CIKM 2023. Alon, U., Zilberstein, M., Levy, O. & Yahav, E. (2019). “code2vec: Learning Distributed Representations of Code.” POPL 2019. arXiv:1803.09473 Zhang, J. et al. (2019). “A Novel Neural Source Code Representation Based on Abstract Syntax Tree (ASTNN).” ICSE 2019. Chen, L., Lei, B., Zhou, D., Lin, P.-H., Liao, C., Ding, C. & Jannesari, A. (2024). “Fortran2CPP: Automating Fortran-to-C++ Translation using LLMs via Multi-Turn Dialogue and Dual-Agent Integration.” arXiv:2412.19770 Gupta, S., Kamalakkannan, K., Moraru, M., Shipman, G. & Diehl, P. (2025). “From Legacy Fortran to Portable Kokkos: An Autonomous Agentic AI Workflow.” arXiv:2509.12443 Dhruv, A. & Dubey, A. (2024). “Leveraging Large Language Models for Code Translation and Software Development in Scientific Computing.” arXiv:2410.24119 Rozière, B. et al. (2022). “Leveraging Automated Unit Tests for Unsupervised Code Translation” (TransCoder-ST). ICLR 2022. arXiv:2110.06773 Bhatia, S., Qiu, J., Hasabnis, N., Seshia, S.A. & Cheung, A. (2024). “Verified Code Transpilation with LLMs” (LLMLift). NeurIPS 2024. arXiv:2406.03003 Eniser, H.F. et al. (2024). “Towards Translating Real-World Code with LLMs: A Study of Translating to Rust” (FLOURINE). arXiv:2405.11514 Diggs, A. et al. (2024). “LLM-Generated Documentation for Legacy Code.” arXiv:2411.14971
AI-driven compiler fuzzing for robust high-performance code generation#
Project code: gego-055
Main supervisor: Gerard Gorman,
g.gorman@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Edward Caunt,
edward.caunt15@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE
This project does not accept multiple students.
Fuzzing is the process of exposing bugs in software by generating a large number of random inputs and executing with them to elicit crashes, exceptions, or incorrect execution, enabling pre-emptive rectification of issues which may not be apparent from conventional testing. Whilst the earliest fuzzers simply generated random data or parameters, fuzzers with awareness of program structure and API can more achieve more effective coverage, for example, by providing input which is syntactically correct but nonsensical or unexpected. This is particularly useful when considering compilers, which must take some given piece of code and generate a lower-level representation which correctly implements the top-level specification for an essentially immeasurable range of inputs. The emergence of LLMs enables rapid generation of potential codes to test, using semantic knowledge from existing examples to generate mutated or permuted inputs which span a wider range than that which could be covered by programmatic mutation of an input. By drawing on strategies such as retrieval-augmented-generation, an LLM-driven fuzzer can be imbued with an understanding of the syntax of the programming language, generating large number of codes which would not necessarily be created in the course of normal use, but which are nonetheless valid.
This project will explore the development of such a fuzzer for Devito, a domain-specific language and compiler framework for finite-difference computations. Devito generates highly-optimised C-level code from a high-level symbolic specification, streamlining development and enabling performance portability across a wide range of hardware. Development of a suitable fuzzer will help in identifying oversights and bugs in the codebase, ultimately improving the robustness of a broad suite of high-performance computing codes based on Devito.
Wireline-log characteristics of clinoform surfaces in shallow-marine sandstone reservoirs, Rannoch and Etive formations, Brent Group, North Sea#
Project code: gaha-117
Main supervisor: Gary Hampson,
g.j.hampson@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor is not yet assigned.
Available to: GEMS
This project may accept multiple students.
The Brent Group has been the most prolific hydrocarbon-bearing reservoir interval in the North Sea basin, offshore UK and Norway. Many Brent Group reservoirs have been abandoned or are reaching the end of their hydrocarbon-producing life, but have the potential to be repurposed for storage of CO2 and other fluids. Such repurposing requires robust reservoir characterisation using legacy oil and gas data.
The lower part of the Brent Group contains shallow-marine sandstones of the Rannoch and Etive Formations, which record the advance of a wave-dominated shoreface system. Oil production in this interval has been heterogeneous, as a result of permeability baffles along clinoform surfaces that dip from top to base of the Rannoch-Etive reservoir interval. Although clinoform surfaces have been documented in outcrop analogues, they have yet to be characterised fully in subsurface core and wireline log data. Locally the clinoforms are characterised by heavy mineral concentrations, marked by high gamma-ray values in wireline logs, in foreshore and upper shoreface deposits. These heavy mineral concentrations are not laterally extensive, but may be correlative to: (1) depositionally equivalent mica concentrations and thin mudstones in lower shoreface deposits; and (2) thin kaolinite-rich zones produced by diagenetic alteration of heavy minerals and mica.
This project will use publically available wireline log data and core photographs from the Rannoch-Etive reservoir interval in the Brent Field, UK North Sea to develop a model for identification and characterisation of potential clinoform surfaces in wireline logs. The candidate will develop code that uses data science techniques to classify sandstone lithofacies types in the Rannoch-Etive reservoir interval using wireline-log combinations, with the goal of identifying subtle variations in wireline-log character that express the depositional and diagenetic mineralogy of clinoform surfaces. If time allows, the resulting classification will be used to test and augment existing geological interpretations of clinoforms in the Rannoch-Etive reservoir interval of the Brent Field (e.g. to generate maps of diagnostic wireline-log characteristics along individual clinoform surfaces) and/or the resulting classification will be tested against wireline-log data from the Rannoch-Etive reservoir interval in the Murchison Field.
Mapping pollution in the urban environment#
Project code: clhe-119
Main supervisor: Claire Heaney,
c.heaney@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Christopher Pain,
c.pain@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
Air pollution is damaging to our health and can cause a variety of adverse health outcomes such as an increased risk of respiratory infections, heart disease, lung cancer and an increased response to allergens. In urban environments, vehicles contribute significantly to air pollution through noxious gases such as nitrogen oxides, carbon dioxide as well as through particulate matter.
This project will consider a number of air pollution datasets and investigate correlations between measured pollution levels and traffic-related data, such as vehicle density, road networks and temporal traffic patterns. By analysing these relationships, the project aims to better understand how traffic contributes to spatial and temporal variations in air quality within urban areas.
The project will also explore methods for generating higher-resolution pollution maps from sparsely distributed air quality monitoring stations. This will involve the use of data-driven modelling approaches, including machine learning and generative techniques, to infer fine-scale pollution patterns. Where appropriate, computational fluid dynamics concepts will be incorporated to capture the influence of urban form and airflow on pollutant dispersion.
For this project, an interest in the following would be beneficial: air pollution, neural networks, generative AI and computational fluid dynamics. The coding will be done in PyTorch and Python.
Neural Physics and AI Surrogates on Large-Scale GPU Systems for Urban Airflow, Heat and Pollution Modelling in Delhi, India#
Project code: clhe-230
Main supervisor: Claire Heaney,
c.heaney@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Christopher Pain,
c.pain@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
This MSc project applies Neural Physics (AI4PDEs) and Foundational AI Surrogate models on large-scale GPU systems to simulate atmospheric airflow, heat transport and air pollution across Delhi, India. Delhi experiences severe particulate pollution, extreme heat waves, and complex atmospheric stagnation, driven by dense urbanisation, seasonal meteorology and regional emissions.
Traditional Computational Fluid Dynamics (CFD) methods struggle to deliver timely, city-scale predictions under such complexity. This project replaces classical solvers with Neural Physics, whose neural network weights are defined analytically from the governing physical equations, enabling neural architectures to directly solve discretised Partial Differential Equations.
Students will develop scalable multi-GPU pipelines to capture urban heat island effects, low-wind pollution accumulation and seasonal variability. The focus will be on temperature, relative humidity and air flow modelling in Dehli’s complex city.
Project 1: Neural Physics for Urban PDEs Focus on implementing Neural Physics solvers on multi-GPU clusters, modelling airflow, heat and pollution transport across Delhi.
Project 2: Foundational AI Surrogates for Real-Time Environmental Prediction Train and deploy surrogate models for rapid city-scale inference, including model compression and real-time pollution forecasting.
These projects would suit students with interests in urban flows, scientific AI and GPU computing.
Modelling the Extraction of NO2 from the Urban Environment Using Neural Physics#
Project code: clhe-120
Main supervisor: Claire Heaney,
c.heaney@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Andreas Kafizas,
a.kafizas@imperial.ac.uk, Department of Chemistry, Imperial College LondonAvailable to: ACSE EDSML
This project does not accept multiple students.
Air pollution is damaging to our health and can cause a variety of adverse health outcomes such as increased risk of respiratory infections and heart disease. Nitrogen oxides (NOx) contribute to air pollution and are released when fossil fuels undergo combustion in vehicles and power stations. NOx also contributes to the formation of particulate matter and ground level ozone.
Photocatalytic materials offer an exciting potential for the removal of NO2 from the air by turning it into N2 and O2 in the presence of light [1]. A trial was carried out by a section of the M1 motorway, which showed a reduction in NO2 due to a barrier coated with photocatalytic material [2]. Building materials including brick and glass can also be sprayed with photocatalytic coatings, and this project will attempt to analyse the effectiveness of photocatalytic coatings on buildings in removing NO2 from the air.
In this project, we will investigate a number of urban layouts and model the potential reduction of NO2 from the air due to the use of photocatalytic coatings. To model the air flows, we will use an in-house computational fluid dynamics code, called NN4PDEs [3,4,5], which solves discretised systems of equations using neural networks based on the Neural Physics approach. The weights of the networks are determined in advance through the choice of discretisation method, so no training is needed. The deposition rate of NO2 onto the photocatalyst is known from experiment, and this will need to be included in the AI4PDEs code to model how much NO2 can be removed.
For this project, an interest in several of the following would be beneficial: air flows and pollution transport in urban neighbourhoods, computational fluid dynamics, neural networks and air pollution. The coding will be done in PyTorch and Python. No specialist knowledge of photocatalytic materials is required.
[1] Towards Purer Air https://eic-uk.co.uk/media/baecbnd4/towards-purer-air.pdf
[2] Kafizas, Rhys-Tyler (2023) https://nationalhighways.co.uk/media/qp3fr5mg/smogstop-report-final-002.pdf
[3] Chen, Heaney, Pain (2024) https://doi.org/10.48550/arXiv.2402.17913
[4] Chen et al. (2024) https://doi.org/10.48550/arXiv.2401.06755
[5] Phillips et al. (2023) https://doi.org/10.1002/nme.7321
Solving inverse problems in bio-physics using Neural Physics#
Project code: clhe-121
Main supervisor: Claire Heaney,
c.heaney@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Pranav Mamidanna,
p.mamidanna22@imperial.ac.uk, Department of Bioengineering, Imperial College LondonAvailable to: ACSE EDSML
This project does not accept multiple students.
When assessing and promoting muscle health during rehabilitation, it is preferable to avoid invasive procedures which could increase risk, increase recovery time and increase the potential for complications. Surface electromyography measures the electric potential generated when muscles contract, which can be used to assess muscle health and guide therapy without the need for invasive techniques [1].
During this project, the student will solve Maxwells equations using the Neural Physics approach [2] and use backpropagation algorithms to solve the inverse problem.
For this project, an interest in several of the following would be beneficial: numerical methods, optimisation and inverse modelling, neural networks and bio-physics (no specialist knowledge of the latter is required). The coding will be done in PyTorch and Python. This project will show how knowledge and code applied to fluid dynamics can be re-applied to bio-physics.
[1] Maksymenko, Clarke, Mendez Guerra, et al (2023) https://doi.org/10.1038/s41467-023-37238-w
[2] Chen, Heaney, Pain (2026) Neural Physics: Using AI Libraries to Develop Physics-Based Solvers for Incompressible Computational Fluid Dynamics, Computers & Fluids, 106981.
Neural Physics and AI Surrogates on Large-Scale GPU Systems for Urban Airflow Modelling in Salvador, Brazil#
Project code: clhe-227
Main supervisor: Claire Heaney,
c.heaney@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Christopher Pain,
c.pain@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
This MSc project applies Neural Physics (AI4PDEs) and Foundational AI Surrogate models on large-scale GPU systems to simulate atmospheric airflow across Salvador, Brazil, a unique coastal city surrounded by the Atlantic Ocean. Salvador’s geography creates complex interactions between sea breezes, urban heat islands, humidity and coastal turbulence, making it an ideal testbed for next-generation AI-driven environmental modelling.
Traditional Computational Fluid Dynamics (CFD) struggles to resolve these coupled land–sea atmospheric processes at city scale in real time. This project replaces classical solvers with analytically weighted Neural Physics networks that solve discretised Partial Differential Equations, combined with GPU-accelerated AI surrogates trained on high-fidelity simulations. Together, these methods enable rapid prediction of wind fields, thermal transport, and pollutant dispersion influenced by Salvador’s maritime climate.
Students will build scalable GPU pipelines to model how ocean-driven airflow penetrates dense urban regions, affects air quality, and moderates temperature. The resulting system forms a digital twin of Salvador, supporting studies in coastal climate resilience, thermal comfort, pollution transport, and emergency response under extreme weather conditions.
Project 1: Neural Physics for Coastal Urban PDEs Focus on implementing AI4PDE Neural Physics solvers on multi-GPU clusters, capturing sea–land interaction wind flows within the complex city building and terrain geometries.
Project 2: Foundational AI Surrogates for Real-Time Coastal City Inference Training and deploy surrogate models for fast prediction of Salvador’s airflow and pollution fields, emphasising model compression and real-time inference of wind flows within the complex city building and terrain geometries.
These projects would suit students with interests in urban flows, scientific AI and GPU computing.
Developing Urban Land-Surface Models and Green-Blue-Grey Models for the Urban Environment#
Project code: clhe-135
Main supervisor: Claire Heaney,
c.heaney@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Christopher Pain,
c.pain@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
Urban land-surface processes and green-blue-grey infrastructure (such as vegetation, soil, buildings, drainage networks and surface water) play a critical role in regulating urban heat, flooding, air quality and water resources. Accurate modelling of these coupled systems is essential for climate adaptation, sustainable city design and resilient infrastructure planning. In this project, we aim to develop: (i) AI-enabled modelling frameworks for urban land-surface and green infrastructure processes, and (ii) AI solvers for predicting coupled hydrological, thermal and pollutant transport processes across urban environments.
We will develop aspects of an AI4PDEs urban land-surface model that takes high-resolution imaging or structural data (e.g. X-ray CT, LiDAR-derived representations or synthetic urban microstructures) and assigns solid, fluid or vegetative phases to each node or cell. The AI4PDEs framework [1] will then be used to solve coupled flow, heat transfer and scalar transport processes relevant to urban environments at low Reynolds numbers. In addition, the project will develop efficient surrogate models using scale-independent convolutional filters, enabling models trained on fine-scale urban structures to be applied at neighbourhood to city scales while retaining key physical behaviour.
Possible projects include:
Physics-based model Development of a large-scale AI4PDEs multiphase urban land-surface and GGBI model capable of running in parallel on GPUs, resolving coupled hydrological, thermal and pollutant transport processes across heterogeneous urban domains.
Surrogate AI solver Construction of surrogate models using scale-independent convolutional filters and state-of-the-art latent diffusion generative neural network methods to rapidly emulate high-resolution urban land-surface and infrastructure simulations.
Optimisation of green infrastructure for heat island and pollution reduction Development of AI-driven optimisation frameworks that couple the AI4PDEs and surrogate models with design variables representing green infrastructure characteristics (e.g. vegetation type, canopy density, soil moisture, green roof and wall configurations). These tools will be used to identify optimal green infrastructure layouts and management strategies that minimise urban heat island intensity and reduce the dispersion and accumulation of air and surface pollutants.
These projects are suited to students with interests in physics, heat transfer in the urban environment, machine learning. The projects will use PyTorch and python.
[1] Chen, Heaney, Pain (2024) Neural Physics: Using AI libraries for Incompressible Computational Fluid Dynamics, https://arxiv.org/abs/2402.17913.
Parameter estimation of the subsurface using Neural Physics#
Project code: clhe-122
Main supervisor: Claire Heaney,
c.heaney@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Donghu Guo,
donghu.guo21@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML GEMS
This project does not accept multiple students.
A novel approach for solving discretised governing equations has emerged called Neural Physics. This method sets the weights of convolutional layers according to finite element or finite difference theory instead of using training data. The weights are chosen from theory in such a way that the network can exactly replicate physics-based solvers. Neural Physics models have several advantages over standard physics-based solvers, including (1) an ability to run easily on GPUs and also the latest AI processors (2) differentiability of the model through autodiff and the computational graph which makes optimisation tasks such as data assimilation and parameter estimation more manageable (3) ease of combining with trained surrogate models.
In this project we aim to exploit the differentiability of Neural Physics problems and build a framework for parameter estimation, seeking to make improvements to the efficiency of the model. In physics-based solvers, many calculations are performed over many iterations resulting in a huge computational graph. We will seek to reduce the memory of the graph to make the method applicable to real-world problems. The application area suggested is full waveform inversion which tries to discover physical properties of the subsurface from seismic waves.
For this project, an interest in several of the following would be beneficial: numerical methods, optimisation and inverse modelling, neural networks and backpropagation. The coding will be done in PyTorch and Python.
Combining Human Expertise and Model-Based Optimization for Experimental Design in Bioprocessing#
Project code: lahe-074
Main supervisor: Laura Helleckes,
l.helleckes@imperial.ac.uk, Department of Chemical Engineering, Imperial College LondonSecond supervisor is not yet assigned.
Available to: ACSE EDSML
This project does not accept multiple students.
Problem description: Bioprocess development is expensive and involves high-dimensional op- timization problems; therefore, efficient experimental design (ED) is critical [1]. Classical space- filling designs can allocate resources inefficiently, while data-driven optimizers such as Bayesian Optimization (BO) [2] can be difficult to interpret. Both approaches also lack the incorporation of domain expertise; this motivates interpretable, human-steered sampling for bioprocesses. Pareto Front Guided Sampling (PFGS) [3] is a model-guided ED strategy that makes the exploration– exploitation trade-off explicit by producing a Pareto set of candidate experiments that balance predicted performance (posterior mean) and information gain (posterior uncertainty). In practice, selecting among Pareto candidates requires domain expertise (e.g., feasibility, risk, cost, and safety constraints) that is hard to encode in a single acquisition function. This project will rigorously evaluate PFGS in a human-in-the-loop (HITL) setting and determine when expert selection among Pareto alternatives improves outcomes relative to automated selection.
Computational methodology: The student will build a web-based HITL framework in which PFGS runs on a server and a browser-based interface allows experts to review the Pareto candidate set, apply constraints/filters, make selections, and optionally annotate their rationale. The system will log candidates, surrogate model states, expert decisions, and interaction timing for reproducible analysis. Coding tasks include developing a modular web interface to test HITL algorithms and implementing comparable HITL methods for benchmarking. Evaluation will use in-silico bioprocess simulators and other relevant case studies to compare (i) expert-guided PFGS, (ii) automated PFGS selection rules, (iii) classical optimization algorithms, and (iv) other selected HITL baselines.
Expected outcomes: Deliverables are (i) an open, modular web framework to conduct HITL studies, (ii) an empirical assessment of PFGS performance in a HITL setting, and (iii) quantitative comparisons of HITL vs. classical optimization that characterize how expert interaction affects optimization trajectories.
[1] A. Kasemiire, H. T. Avohou, C. De Bleye, P.-Y. Sacre, E. Dumont, P. Hubert, and E. Ziemons, “Design of experiments and design space approaches in the pharmaceutical bioprocess optimiza- tion,” European Journal of Pharmaceutics and Biopharmaceutics, vol. 166, pp. 144–154, Sept. 2021. [2] M. Siska, E. Pajak, K. Rosenthal, A. del Rio Chanona, E. von Lieres, and L. M. Helleckes, “A Guide to Bayesian Optimization in Bioprocess Engineering,” Biotechnology and Bioengineering. [3] S. Stricker, “Optimal Design of Experiments for Bio-Processes using Hybrid Models,” Master’s thesis, ETH Zurich, Zurich, Sept. 2024.
Seismology on the Moon: Moonquake Simulations#
Project code: doki-194
Main supervisor: Doyeon Kim,
doyeon.kim@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Rhodri Nelson,
rhodri.nelson@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
Artemis III marks humanity’s return to the Moon and opens a new era for exploring its interior structure, thermal evolution, and resource potential. Seismology provides the most direct way to probe planetary interiors. On Earth, seismograms record the timing and amplitudes of seismic phases that sample the deep interior, and comparisons between observed data and synthetic seismograms from wavefield simulations are used to refine models of Earth’s structure. The same approach can be applied to the Moon, but existing tools are largely optimised for Earth and are poorly suited for simulating the high frequency wavefields required to characterise the Moon’s highly scattering near-surface, produced by its fragmented and powdered regolith. With a new generation of missions, including NASA’s Artemis III Lunar Environmental Monitoring Stations, the Artemis IV South Pole Seismic Station, the Farside Seismic Suite, and China’s Chang’e-7, deploying state-of-the-art broadband seismometers, there is a timely need for new wavefield analysis tools that can be applied directly to forthcoming lunar data. This MSc project will help to develop a next-generation, global-scale wavefield simulation strategy based on finite-difference methods, ultimately designed to be flexible across rocky bodies with different boundary conditions and velocity structures. Starting from a planar isotropic elastic code, the candidate will adapt and validate the approach for lunar structure and generate synthetic recordings of different classes of lunar seismic events. These will be benchmarked against spectral-element tools such as SPECFEM3D and AXISEM3D, the current standard in terrestrial seismology. The resulting framework will be directly applicable to upcoming lunar seismic datasets, helping prepare lunar seismology for the Artemis era.
Seismology on the Moon: Developing a Next-Generation Moonquake Location Method#
Project code: doki-177
Main supervisor: Doyeon Kim,
doyeon.kim@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor is not yet assigned.
Available to: ACSE EDSML
This project does not accept multiple students.
Artemis III marks humanity’s historic return to the Moon and opens a new era for exploring its interior structure, thermal evolution, and resource potential. Seismology provides the most direct way to probe planetary interiors: on Earth, precise earthquake locations from dense seismic networks have enabled high-resolution imaging of internal structure and tectonic processes.
Lunar seismology, however, faces a unique challenge. Unlike Earth, future lunar seismic coverage will be sparse and assembled opportunistically through a small number of single-station missions (e.g. Artemis III-LEMS, Artemis IV-SPSS, Farside Seismic Suite, Chang’e-7), rather than a global network. As a result, reliable single-station event location methods are essential.
While NASA’s InSight mission on Mars demonstrated that single-station seismology can revolutionise our understanding of planetary interiors, applying these techniques to the Moon is far more difficult. Strong seismic scattering in the highly fractured lunar regolith obscures phase arrivals, leading to large uncertainties in the locations of shallow and deep moonquakes as well as meteoroid impacts. Because event location is the foundation for all subsequent interpretations, these uncertainties currently limit our ability to constrain the Moon’s interior and evolution.
This MSc project is dedicated to developing a next-generation moonquake location algorithm designed for future lunar missions and seismic data analysis. Using Apollo seismic recordings of artificial impacts, such as Saturn booster and spacecraft stages deliberately crashed at precisely known locations on the lunar surface, the project will improve the robustness and accuracy of single-station event locations and rigorously validate the proposed methodology. The resulting approach will be directly applicable to upcoming lunar seismic datasets, helping prepare lunar seismology for the Artemis era.
Acceleration of Mesh Adaptation through Machine Learning#
Project code: stkr-203
Main supervisor: Stephan Kramer,
s.kramer@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Matthew Piggott,
m.d.piggott@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE
This project may accept multiple students.
This project combines advanced methods in machine learning and (PDE-based) computational modelling. Solving complex Partial Differential Equations (PDEs) accurately and efficiently is an essential and challenging problem in all scientific and engineering disciplines. Mesh movement methods, which move the mesh nodes to increase resolution to better focus on areas of importance in the domain, provide the capability to improve accuracy without increasing the overall degree of freedom count. The movement process itself can be very expensive however and recent research in our group has demonstrated how machine learning can be used to dramatically accelerate this process, see https://erizmr.github.io/UM2N/. This project will provide the opportunity to work on several aspects of the UM2N (Universal Mesh Movement Networks) framework, improving its performance and the quality of the produced meshes through choices in the architecture, retraining on better datasets, and a better evaluation of the current capabilities. There are opportunities to focus purely on the (pytorch based) ML aspects, or on the application to solving PDEs, specific science or engineering applications, or a combination of these.
Data-Driven Modelling of Anthropogenic CO2 Fluxes Across Urban and Land Use Landscapes#
Project code: lili-022
Main supervisor: Linfeng Li,
l.li20@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Fangxin Fang,
f.fang@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
Understanding the impact of land use change on carbon dioxide (CO₂) emissions is increasingly important in the context of global climate change and rapid urbanisation. Different land use types, including urban areas, agricultural land, and forests, strongly influence anthropogenic CO₂ fluxes through variations in energy consumption, transportation, industrial activity, and ecosystem carbon storage. Urban expansion, in particular, leads to concentrated emission sources associated with buildings, infrastructure, and human activities, making accurate assessment of land use-driven CO₂ emissions essential for climate mitigation and sustainable planning. Recent advances in remote sensing technologies have significantly enhanced the ability to monitor land use dynamics and associated carbon emissions across spatial and temporal scales. Multispectral and multi-temporal satellite imagery enables detailed mapping of land cover changes, while emerging technologies such as Unmanned Aerial Vehicle (UAV) remote sensing and Light Detection and Ranging (LiDAR) provide high resolution information on surface structure, building density, and urban morphology. These data sources allow improved characterisation of anthropogenic emission patterns linked to land use. Deep learning methods, including Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and self-attention-based models, have demonstrated strong capability in extracting spatial features and temporal dependencies from large remote sensing datasets. These techniques improve land cover classification accuracy and support reliable estimation of CO₂ fluxes associated with different land use types. A key challenge, however, is ensuring the transferability of trained models across different regions, spatial resolutions, and time periods. Despite recent progress, challenges remain, including data redundancy, high computational cost, and reduced performance in high-dimensional feature spaces, commonly known as the Hughes Phenomenon. This MSc project aims to address these challenges by developing robust deep learning frameworks that integrate multispectral remote sensing data with land use and energy-use indicators, enabling transferable and scalable assessment of anthropogenic CO₂ fluxes to support climate mitigation and sustainable land management.
Reasoning Beyond Language: Exploring Physically-Grounded Knowledge via Video Generation#
Project code: chli-002
Main supervisor: Che Liu,
che.liu21@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor is not yet assigned.
Available to: ACSE EDSML
This project may accept multiple students.
Current multimodal evaluations often overlook “Reasoning Beyond Language,” where accurate visual generation acts as a rigorous test of internal world modeling , particularly for “Type III: Physical & Temporal Reasoning” involving dynamics like gravity and collision. To address the limitations of “dream-logic” in current video generation, we aim to explore a Physically-Grounded Video Generation framework. This project will investigate a hybrid architecture that leverages Autoregressive (AR) models for logical causal planning and Diffusion models for high-fidelity rendering, focusing on “Visual Conditioning” tasks to strictly test the model’s adherence to physical laws rather than open-ended imagination. Consequently, we expect the student to have a strong background in Diffusion, Autoregressive (AR) models, and video/image generation to successfully implement these complex deep learning architectures and evaluate them against the Generative Reasoning Hierarchy.
Privacy-Preserving Machine Learning Using Autoencoders#
Project code: pelo-080
Main supervisor: Petra Loncar,
p.loncar@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Rossella Arcucci,
r.arcucci@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
The goal of this Independent Research Project is to explore the innovative field of privacy-preserving machine learning using latent-space representations from autoencoders. The project encompasses diverse domains, engaging with various types of data. The data type will be agreed upon with the student. A significant emphasis is placed on relevant honest computing practices, as the widespread use of digital systems has increased concerns regarding data management, privacy and security.
During this computational project, the student will explore the application and impact of autoencoders in suppressing sensitive information while maintaining adequate utility of application data. In the computational practical part, the student will independently build a complete autoencoder model and evaluate its accuracy, efficiency, and overall performance. Each participating student will focus on a particular application area, contributing to a collective methodological and conceptual framework that underpins the project.
The final requirement is to provide a comprehensive description of the complete model solution and the technologies used. The results generated from this project must be documented, presented, compared, analysed, and described in detail.
Neural Networks for Submarine Groundwater Discharge (2 projects)#
Project code: trmc-053
Main supervisor: Tristan Mckenzie,
t.mckenzie@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor is not yet assigned.
Available to: ACSE EDSML GEMS READY
This project may accept multiple students.
The coastal ocean is highly vulnerable to climate change due to superimposed terrestrial, climatic, and oceanic forcings. Coastal groundwater systems, which include the coastal aquifer and flow of groundwater to the coastal ocean (submarine groundwater discharge or SGD), are particularly vulnerable to contamination yet relatively underexplored. Field observations of geochemical groundwater tracers such as 222Rn (radon) are one way to estimate groundwater flow rates, which is most often achieved through a mass balance approach. Integration of SGD into more traditional hydrological budgets and policy frameworks will require increasing (both spatial and temporal) monitoring as well as predictive power through models. Towards this goal, two Python-based projects are available. Both projects contribute towards building a real-time predictive SGD network.
Develop a physics-informed neural network (PINN) to predict temporal variations in SGD. Both synthetic and real data will be used along with groundwater flow equations to encode physical relationships. This will build upon past work using non-physics informed FCNNs and CNNs to predict coastal radon concentrations, and if successful, has the potential to be a groundbreaking contribution to the subfield.
Create an automated pipeline for processing field-collected (temporal and spatial) radon data and calculating SGD. This will include incorporating and QC’ing data from multiple sensors and data sources, designing criteria for selecting the correct mass balance, and conducting an uncertainty analysis on the result. There are multiple ways to do this, the student will be expected to evaluate model performance and comparison between approaches.
Separating Semantic and Predictive Components in the Late-Layer Representation Space of Large Language Models#
Project code: sime-069
Main supervisor: Sina Mehrdad,
s.mehrdad@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Rossella Arcucci,
r.arcucci@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project does not accept multiple students.
Large language models (LLMs) generate text token-by-token using high-dimensional hidden representations. Intuitively, these representations must simultaneously encode (a) relatively stable information about the topic/intent of the ongoing response and (b) rapidly changing information needed for local syntactic planning and next-token prediction. This project investigates whether these two roles correspond to distinguishable structure in the LLM’s late-layer representation space.
The central hypothesis is that, during coherent generation, a “semantic state” component varies slowly (remaining consistent across many tokens and updating more strongly at clause/sentence boundaries), while a “predictive/syntactic” component varies quickly as the model commits to local wording. The goal is to test whether late-layer activations can be decomposed into subspaces or directions that differentially capture stable semantic content versus short-horizon predictive structure, and whether these components exhibit distinct temporal dynamics.
The work will use representation-space analysis grounded in linear algebra: examining subspace structure, alignment, and time-evolution of hidden states across tokens, prompts, and model conditions. Readout-based analyses will connect hidden states to predicted tokens and intermediate “vocabulary views” (building on lens methods), while probing-style analyses will test what linguistic signals are linearly recoverable at different positions and layers. Where appropriate, causal intervention techniques (e.g., activation patching) will be used to validate whether identified components are functionally involved in semantic continuity versus local next-token selection.
This project aims to provide empirical evidence for (or against) a structured separation between semantic persistence and token-level prediction in LLM internal representations, contributing to interpretability and reliability of generative models.
Diagnosing Stepwise Regime Shifts in ERA5 Radiative Flux Fields Using Autoencoder Drift and Explainable AI#
Project code: sime-071
Main supervisor: Sina Mehrdad,
s.mehrdad@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Rossella Arcucci,
r.arcucci@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: EDSML
This project does not accept multiple students.
When a machine-learning model is trained on one climate dataset and evaluated on another, systematic changes in reconstruction error can reveal non-stationarities in the target data that are not captured by the learned representation. In earlier work, separate autoencoders were trained on Northern Hemisphere top-of-atmosphere (TOA) net radiation from multiple CMIP6 models and from ERA5, all remapped to a common grid. When these models were used to reconstruct ERA5, the reconstruction loss increased over time in a small number of pronounced step-like “leaps” that appeared consistently across models and occurred at similar time periods.
This project investigates the origin and physical meaning of these stepwise increases in reconstruction loss. Two competing explanations will be examined: (i) artifacts arising from changes in the observing system and reanalysis production, and/or (ii) real climate regime changes that modify the statistical structure of TOA net radiation, such as the reported stepwise behavior of Arctic amplification.
The workflow will proceed as follows. First, an ensemble of autoencoder models will be retrained using TOA net radiation from multiple CMIP6 models and ERA5 to establish a robust reference set of learned representations. Second, changepoints in reconstruction error will be rigorously detected and dated, and related to known observing-system transitions (e.g., satellite-era milestones). Third, interpretable and explainable machine-learning techniques will be used to identify the spatial and physical patterns most responsible for the error increases. Finally, the inferred patterns will be compared against established physical narratives (e.g., Arctic amplification and circulation regime changes) and, where feasible, against alternative datasets or reanalyses.
The expected outcome is a process-aware diagnosis of whether the observed reconstruction-error “leaps” reflect true climate shifts, reanalysis inhomogeneities, or a combination of both, transforming an ML anomaly into a physically interpretable signal.
Self-Supervised Transformer Representations of Sea-Level Pressure for Climate Index Readout and Attention-Based Teleconnection Diagnostics#
Project code: sime-070
Main supervisor: Sina Mehrdad,
s.mehrdad@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Rossella Arcucci,
r.arcucci@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: EDSML
This project does not accept multiple students.
Large-scale climate variability is often summarized using climate indices derived from spatial patterns in gridded fields. The North Atlantic Oscillation (NAO) is one of the most prominent examples, commonly defined from mean sea-level pressure (MSLP) contrasts between fixed regions. While such definitions are useful for standardization, they impose a predefined spatial structure and may obscure alternative spatial dependencies or links to broader teleconnection patterns. Understanding how large-scale pressure anomalies interact across regions, and how such interactions relate to climate indices, remains an active area of research.
This project investigates the usability and interpretability of transformer self-attention mechanisms in a climate-science context by learning representations from daily MSLP fields using self-supervised learning (SSL). A transformer backbone will be pre-trained on a large daily CMIP6 MSLP dataset using an SSL objective, enabling the model to learn general-purpose spatial representations without relying on predefined climate indices. The trained backbone will then be frozen and used as a feature extractor.
Validation will proceed in two complementary ways. First, climate indices, starting with the NAO, will be tested for linear decodability from the frozen representations using lightweight linear readout heads, assessing whether the learned feature space encodes physically meaningful index information. Second, the model’s self-attention patterns will be analyzed during index readout to evaluate whether spatial attribution highlights known NAO centers of action and remains robust across seasons and extreme phases.
If these validations are successful, the project will extend attention-based analysis beyond NAO to explore whether persistent long-range attention dependencies provide interpretable signals of teleconnections in daily pressure variability. The expected outcome is a proof-of-concept framework demonstrating when and how transformer attention can support transparent, physically grounded climate diagnostics from gridded atmospheric fields.
Cross-Scale Consistency of Dense Patch Representations in Self-Supervised Vision Transformers#
Project code: sime-068
Main supervisor: Sina Mehrdad,
s.mehrdad@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Rossella Arcucci,
r.arcucci@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project does not accept multiple students.
Recent self-supervised vision foundation models, such as DINOv3, produce high-quality dense patch representations that remain stable across varying input resolutions. This property enables high-resolution inference and multi-scale training without architectural changes, relying instead on the internal structure of the learned representation space. However, the relationship between patch features extracted at different spatial resolutions remains poorly understood.
This project aims to investigate how dense patch representations change with scale and whether lower-resolution patch features can be systematically related to, or expressed through, higher-resolution patch features extracted from the same image. Intuitively, when an image is processed at higher resolution, multiple fine-scale patch features correspond to a single coarse-scale patch at lower resolution. This raises a fundamental question: does the lower-resolution representation lie within the subspace spanned by its corresponding higher-resolution features, and if so, to what extent?
The study will analyze dense features extracted from a frozen self-supervised vision transformer under multiple input resolutions. For corresponding spatial regions, the relationship between groups of high-resolution patch features and their lower-resolution counterparts will be examined from a representation-space perspective. Rather than assuming a predefined aggregation rule, the project will explore whether consistent structural relationships emerge across scales, such as shared subspaces, linear dependencies, or stable similarity structures.
The analysis will focus on identifying scale-consistent properties of the learned feature space and understanding how spatial resolution influences semantic and geometric information encoded in patch tokens. The findings are expected to provide insight into why large self-supervised models can support high-resolution training and inference without explicit multi-scale supervision.
By clarifying the cross-scale behavior of dense representations, this project contributes to a deeper theoretical understanding of vision foundation models and may inform future model design, evaluation strategies, and applications requiring reliable multi-scale representations.
Evaluating the Robustness of Self-Supervised Vision Transformers to Adversarial Perturbations: A Study of Gram Anchoring in DINOv3#
Project code: sime-067
Main supervisor: Sina Mehrdad,
s.mehrdad@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Rossella Arcucci,
r.arcucci@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project does not accept multiple students.
Self-supervised vision foundation models such as DINOv2 and DINOv3 have demonstrated remarkable performance across a wide range of visual tasks without task-specific supervision. A key contribution of DINOv3 is the introduction of Gram anchoring, a regularization strategy designed to preserve patch-level consistency and prevent degradation of dense feature maps during long training schedules. While Gram anchoring has been shown to substantially improve dense representations and stability, its impact on robustness to adversarial perturbations remains unexplored.
This project aims to systematically evaluate the resistance of DINOv3 to adversarial attacks and assess whether Gram anchoring contributes to improved robustness compared to earlier self-supervised models. The focus will be on feature-space and input-space perturbations that challenge patch-level consistency, such as gradient-based adversarial noise (e.g., FGSM or PGD) and localized or structured perturbations affecting spatial coherence. Robustness will be evaluated using frozen backbones, in line with the foundation-model paradigm, and measured through changes in downstream performance (e.g., linear probing for classification or segmentation) as well as feature-space stability metrics.
The study will primarily analyze DINOv3 models and, where feasible, compare them to DINOv2 to partially isolate the effect of Gram anchoring. Special attention will be paid to how adversarial perturbations affect patch similarity maps and Gram matrices, providing insight into whether enforcing similarity structure improves resilience against adversarial feature corruption.
The expected outcome is a clearer understanding of the relationship between self-supervised regularization strategies and adversarial robustness. This work will contribute to the growing discussion on the reliability and safety of large vision foundation models, particularly in scientific and safety-critical applications where robustness is essential.
Education data mining: explaining errors in student self-study#
Project code: mame-086
Main supervisor: Marcus Messer,
m.messer@imperial.ac.uk, Department of Mechanical Engineering, Imperial College LondonSecond supervisor is not yet assigned.
Available to: ACSE
This project may accept multiple students.
Lambda Feedback is a self-study platform providing automated formative feedback. Using our unique dataset of 500,000 algebraic expressions from students, this project will explore how students learn with Lambda Feedback, by analysing and classifying series of student submissions, investigating the learning curve of topics across different domains aThis analysis will build on a prior MSc project and a recent summer project, which used parsing of expressions into abstract syntax trees, finding the edit distance to the correct answer, and clustering against a small number of cases (~1,000). The output will be recommendations for improving algorithms for automated feedback to enhance the educational experience for future students at Imperial and beyond.
Deterministic Feedback vs Large Language Model-Generated Feedback on Mathematical Expressions#
Project code: mame-087
Main supervisor: Marcus Messer,
m.messer@imperial.ac.uk, Department of Mechanical Engineering, Imperial College LondonSecond supervisor is not yet assigned.
Available to: ACSE
This project may accept multiple students.
Lambda Feedback is a self-study platform with over 500,000 algebraic expressions and associated automated formative feedback. Currently, this feedback is provided by rule-based evaluation functions, which compare a student’s submission to a teacher’s reference answer. This project will investigate in which cases a well-defined prompt and selected large language models (LLMs) perform better than our existing rules-based approach. The output of this work will be a detailed comparison of rules-based and LLM feedback on the same questions, to determine where a rules-based approach outperforms an LLM and when an LLM outperforms it.
Zonal Electricity Price Forecasting in the Nordic Market Using Advanced Machine Learning#
Project code: samo-076
Main supervisor: Samaneh Abolpour Mofrad,
s.abolpour-mofrad@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Simon Warder,
s.warder15@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: EDSML READY
This project may accept multiple students.
Problem description: Electricity prices and traded volumes in the Nordic power market vary across interconnected bidding zones due to transmission constraints, demand patterns, generation mix, and cross-border power flows. Accurate zonal forecasting is essential for market participants, system operators, and policymakers, particularly as power systems become more complex with higher renewable penetration and increasing congestion. Traditional forecasting methods often rely on zone-specific models or manually engineered spatial features, which limits their ability to capture complex interdependencies across the power system. This project addresses the need for data-driven, scalable forecasting approaches that can model both temporal dynamics and spatial interactions between bidding zones.
Computational methodology: The project will develop an end-to-end forecasting pipeline using time series data from the ENTSO-E Transparency Platform. The student will implement automated data ingestion and preprocessing in Python, including data validation, temporal alignment, and feature engineering to produce clean and analysis-ready datasets. Multiple advanced machine learning approaches will be explored and compared, including tree-based models, deep learning time-series architectures (e.g. recurrent or convolutional models), and graph-based methods. In particular, Graph Neural Networks (GNNs) will be investigated as one option to explicitly model spatial dependencies between bidding zones using transmission topology, observed power flows, or statistical relationships. The project involves substantial programming in Python, covering data engineering, model development, training, and evaluation.
Expected outcomes: Expected outcomes include a reproducible data and modelling pipeline for Nordic electricity market forecasting, a systematic comparison of advanced machine learning methods for zonal price and volume prediction, and insights into the role of spatial coupling and congestion in forecasting performance. The results will be relevant to both academic research and practical energy market applications.
Graph Neural Network for Weather Forecasting in Ocean Areas and Sensor Placement Optimization#
Project code: samo-247
Main supervisor: Samaneh Abolpour Mofrad,
s.abolpour-mofrad@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Matthew Piggott,
m.d.piggott@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: EDSML READY
This project does not accept multiple students.
Despite significant advancements in weather and climate forecasting over the past half-century, accurately predicting weather in maritime areas for vessels remains challenging. State-of-the-art data-driven approaches, such as Graph Neural Networks (GNNs), have recently shown success in similar forecasting tasks.
This project aims to develop a deep understanding of how effectively GNNs improve local weather forecasting in maritime areas compared to existing methods, both in terms of results and computational efficiency. It also seeks to optimize the number and placement of weather sensors in the ocean and understand the relationship between the distance of weather stations from a ship and the forecast time horizon.
Students will become familiar with different GNNs and utilize them with real data. They will work with various ML methods for time-series analysis and gain knowledge of different weather forecasting methods.
Zonal Electricity Price Forecasting in the Nordic Market Using Advanced Machine Learning#
Project code: samo-081
Main supervisor: Samaneh Abolpour Mofrad,
s.abolpour-mofrad@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Simon Warder,
s.warder15@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE READY
This project may accept multiple students.
Problem description:
Electricity prices and traded volumes in the Nordic power market vary across interconnected bidding zones due to transmission constraints, demand patterns, generation mix, and cross-border power flows. Accurate zonal forecasting is essential for market participants, system operators, and policymakers, particularly as power systems become more complex with higher renewable penetration and increasing congestion. Traditional forecasting methods often rely on zone-specific models or manually engineered spatial features, which limits their ability to capture complex interdependencies across the power system. This project addresses the need for data-driven, scalable forecasting approaches that can model both temporal dynamics and spatial interactions between bidding zones.
Computational methodology:
The project will develop an end-to-end forecasting pipeline using time series data from the ENTSO-E Transparency Platform. The student will implement automated data ingestion and storage in a PostgreSQL database, including data validation, feature engineering, and temporal alignment. Multiple advanced machine learning approaches will be explored and compared, including tree-based models, deep learning time-series architectures (e.g. recurrent or convolutional models), and graph-based methods. In particular, Graph Neural Networks (GNNs) will be investigated as one option to explicitly model spatial dependencies between bidding zones using transmission topology, observed power flows, or statistical relationships. The project involves substantial programming in Python, covering data engineering, model development, training, and evaluation.
Expected outcomes:
Expected outcomes include a reproducible data and modelling pipeline for Nordic electricity market forecasting, a systematic comparison of advanced machine learning methods for zonal price and volume prediction, and insights into the role of spatial coupling and congestion in forecasting performance. The results will be relevant to both academic research and practical energy market applications.
Neurophysiological Synchronisation Within and Between Individuals During Music Exposure: A Multi-Modal EEG–ECG Investigation#
Project code: jamo-052
Main supervisor: Jazmin Morrone,
j.morrone@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Oscar Calderon Agudo,
o.calderon-agudo14@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE
This project may accept multiple students.
This project explores how brain and body signals become synchronised within a person and between different people while listening to music. The main focus is on using computational modelling, signal processing, and data-driven analysis to study these effects. Music is a complex, multisensory stimulus, making it a useful way to examine how brain activity and physiological responses interact in real-world scenarios. Although there is increasing interest in how physiological signals align between individuals, there are still limited computational methods for reliably measuring connections within and between people using brain and heart data.
The project will use multimodal electroencephalography (EEG) and electrocardiography (ECG) datasets recorded during exposure to multiple music genres (such as rock, classical, jazz). Students will develop from scratch, or substantially extend, a computational analysis pipeline to quantify synchronisation across three levels: (1) inter-personal physiological synchrony (ECG–ECG), (2) inter-personal neural synchrony (EEG–EEG), and (3) intra-personal neurophysiological coupling (electrode–electrode). This will involve implementing advanced signal preprocessing, time–frequency analysis, phase- and amplitude-based coupling metrics, and graph-based or statistical models of synchrony.
A central computational challenge is to design and validate a quantitative evaluation of neural synchronisation that is robust to noise, inter-individual variability, and genre-dependent dynamics. Students may explore machine learning or dimensionality-reduction techniques to identify genre-specific neurophysiological signatures and determine which features best predict synchronisation strength. Emphasis will be placed on algorithmic accuracy, reproducibility, and scalability to larger datasets.
This work aims to advance understanding of how acoustic waves influence human neurophysiology at both local and global scales, and the broader implications of these interactions. The project aligns closely with computational science, data analysis, and applied machine learning. Anticipated outcomes include applications for evaluating sound-based interventions across clinical, educational, and performance contexts, while providing students with a rigorous, end-to-end computational research experience.
Differentiable Physics for Inertial Microbubble Dynamics#
Project code: bemo-092
Main supervisor: Ben Moseley,
b.moseley@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Callum Rhys Tilbury,
c.tilbury25@imperial.ac.uk, Department of Bioengineering, Imperial College LondonAvailable to: ACSE EDSML
This project does not accept multiple students.
Treating neurological disorders is difficult due to the blood-brain barrier (BBB). Though designed to protect the brain, ironically it also prevents the entry of most therapeutic drugs. Transcranial focused ultrasound (tFUS) in conjunction with intravascular microbubbles is being explored as a mechanism for drug delivery into the brain. The bubbles, in the presence of ultrasound, oscillate: shrinking in the compressional phase, growing in the rarefactional phase. This behaviour has been shown to transiently and locally change the permeability of the BBB in a non-invasive way. Modelling these bubble dynamics is vital to the safety and efficacy of this therapeutic direction.
We have developed jbubble as a differentiable simulator in JAX, using the Rayleigh-Plesset equation, to relate the ultrasound pressure wave to the bubble’s radius. Under this model, there are two notable regimes of oscillation: non-inertial cavitation, with smooth gradients in the radius; and inertial cavitation, where the bubble collapses rapidly. It is hypothesised that inertial cavitation is key to therapeutic effects; however, this regime is computationally difficult, due to stiffness (numerically requiring multiple time scales) and physical discontinuities (due to non-linear shell behaviour), both of which cause issues with differentiability. Nonetheless, we would like to leverage gradient-based optimisation and design techniques in this regime.
This project will build upon—and significantly extend—jbubble, focusing on differentiability during inertial cavitation. The student will: create synthetic data of collapse trajectories; clearly demonstrate and explain associated gradient failure modes; use gradient-relaxation techniques to smooth the optimisation landscape, softening the stiff and discontinuous bubble dynamics; and finally, use these extensions to perform gradient-based parameter estimation in inertial cavitation contexts.
The student’s work will enable a powerful optimisation pipeline for microbubbles in a therapeutically relevant context. This would be an important step towards precise and effective treatments for a wide range of neurological disorders.
3D surrogate models for ultrasound brain full-waveform inversion#
Project code: bemo-205
Main supervisor: Ben Moseley,
b.moseley@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Carlos Cueto,
c.cueto@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project does not accept multiple students.
Motivation:
Full-waveform inversion (FWI) has emerged as a promising technique for ultrasound-based imaging of the brain, with the potential to overcome the limitations of conventional methods based on MRI and CT. However, FWI is very sensitive to noise, modelling errors and poor initialisation due to its non-linear and ill-posed nature.
Recently, machine learning methods have shown significant potential in stabilising and accelerating FWI, improving convergence and robustness. A major barrier to progress in this area is the lack of large, realistic and well-annotated training datasets. Unlike other medical imaging domains, there are currently no available datasets that provide physically meaningful acoustic property models of the human brain suitable for training these models.
This project will create a large-scale, three-dimensional dataset of realistic brain models derived from MRI data, enabling the development and evaluation of AI-assisted FWI techniques for transcranial ultrasound imaging.
Methods:
The project will begin with the identification and curation of open-access, high-resolution three-dimensional MRI brain datasets from established neuroimaging repositories. These datasets will be pre-processed to ensure consistent spatial resolution, orientation, and quality.
Advanced image processing and segmentation techniques, including classical algorithms and modern deep learning approaches, will be explored to delineate key anatomical structures such as skull, cerebrospinal fluid, grey matter, white matter, and major vascular regions. The segmented tissue classes will then be translated into physically meaningful acoustic property maps (e.g. sound speed, density, and attenuation) using literature-based models, empirical relationships and machine learning approaches.
The final output will be a large library of realistic 3D acoustic brain models, which will represent a foundational resource for future research into AI-enhanced FWI and transcranial ultrasound imaging.
Improved numerical simulation of the Rayleigh-Taylor instability in geological CO2 storage#
Project code: anmu-008
Main supervisor: Ann Muggeridge,
a.muggeridge@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor is not yet assigned.
Available to: GEMS
This project does not accept multiple students.
The project will investigate the mixing of CO2 in brine by the unstable gravitational fingering (the Rayleigh-Taylor instability). It will involve the development, testing and application of an existing C/C++ simulator. It would suit someone with a first degree in mathematics, physics of reservoir engineering
Clustering/Grouping of upscaled two phase flow functions for field scale performance prediction of geological CO2 storage#
Project code: anmu-045
Main supervisor: Ann Muggeridge,
a.muggeridge@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Alistair Jones,
alistair.jones@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: GEMS
This project does not accept multiple students.
Upscaling of two phase flow functions is required to represent the influence of small scale geological heterogeneities on field scale flows. Current upscaling approaches based on steady scale upscaling produce a different set of flow functions for every field scale grid block, resulting in the requirement to input several thousand sets of curves into the field scale simulation. This project will investigate whether it is possible to reduce this number of different functions by grouping either based on the characteristics of the functions (curve shape or fractional flow and total mobility) and/or the underlying fine scale heterogeneity pattern. This could be achieved using simple statistical metrics or machine learning
Using Machine Learning to quantify experimental mineral magnetic data#
Project code: admu-007
Main supervisor: Adrian Muxworthy,
adrian.muxworthy@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Lesleis Nagy,
lesleis.nagy@liverpool.ac.uk, University of Liverpool, UKAvailable to: ACSE EDSML GEMS READY
This project does not accept multiple students.
The mineral magnetic signature of a rocks and sediments is relatively quick to measure. These measurements can tell us about the magnetic minerals in a sample, plus the grain size distribution of these magnetic minerals. Such information can be used as a rapid proxy for a range of environmental signals or as quality check in palaeomagnetic studies. For example, the magnetic signatures of rocks can be used to track monsoon migration paths over the glacial/inter-glacial cycle. We have recently developed a method (FORCINN) for inverting a type of magnetic hysteresis data called first-order reversal curve (FORC), using a machine-learning approach (Pei et al., 2025). Unfortunately, there is no single magnetic measurement that uniquely quantifies the magnetic mineralogy and the corresponding grain size distributions; FORCINN relies on knowing the magnetic mineralogy. Ideally, we want a system that combines all types of data. Currently a suite of quantitative and qualitative measurements are made, e.g., magnetisation versus temperature or magnetic field; leaving the end-user/scientist to interpret the data. This is currently done manually on a sample-by-sample basis. The aim of the project is to develop a machine learning approach to the interpretation of the data. This will make data interpretation less subjective to user subjectivity, and more consistent. Whilst some training sets can be numerically generated, e.g., magnetisation vs. field, other datasets are derived experimentally, given rise to limited datasets. The challenge of the project is develop a robust algorithm that combines both of these types of data.
Pei, Z., W. Williams, L. Nagy, G. A. Paterson, R. Moreno, A. R. Muxworthy and L. Chang (2025), FORCINN: First-Order Reversal Curve inversion of magnetite using Neural Networks, Geophysical Research Letters, 52(3), e2024GL112769, doi:10.1029/2024GL112769.
Developing machine learning models for using crystal textures to investigate magmatic evolution leading to explosive volcanic eruptions#
Project code: chna-048
Main supervisor: Chetan Nathwani,
chetan.nathwani@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor is not yet assigned.
Available to: ACSE EDSML GEMS READY
This project does not accept multiple students.
The textures of crystals in volcanic rocks may vary between those erupted by hazardous explosive eruptions versus non-hazardous effusive eruptions. Quantifying variations in the nature and abundance of crystal textures may therefore provide critical information for volcano monitoring to indicate changes in magma dynamics and anticipate shifts towards hazardous eruption styles. However, until now, most studies of crystal textures in volcanic rocks depend on manual data acquisition and are thus time-consuming and subject to sampling bias.
This project will aim to further develop a machine learning pipeline for a) segmentation of crystals from thin section images of volcanic rocks for quantification of crystal size distributions and b) develop unsupervised and self-supervised approaches to recognise different populations of crystals in volcanic rocks based on their texture. A library of scans of thin sections has been collected for volcanic rocks, largely from active Aegean volcanoes. An initial model is currently in place which requires optimisation, and new state-of-the-art models will be trained and tested to compare performance. The developed models will be used to track changes in the makeup and textures of crystal cargo over the history of a volcano’s lifetime. This will test whether crystals erupted by explosive eruptions (compared to effusive eruptions) have specific textures that may be diagnostic of perturbations triggering explosive volcanism.
The project will be well suited to a student interested in applying deep learning and computer vision to a topical geoscience issue. It will also involve collaborations with partners at ETH Zurich, the Natural History Museum and the University of Southampton.
Ultrasound Full‑Waveform Inversion for Cardiac‑Style Imaging with Devito#
Project code: rhne-165
Main supervisor: Rhodri Nelson,
rhodri.nelson@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor is not yet assigned.
Available to: ACSE EDSML
This project may accept multiple students.
Full‑waveform inversion (FWI) is a powerful way to recover spatially varying material properties from wave measurements, but it is best known in geophysics. This project explores whether the same ideas can be prototyped for medical ultrasound imaging, with a focus on a simplified “heart” scenario, using Devito for fast wave simulation and adjoint gradients.
You will:
Set up a 2D acoustic (or visco‑acoustic) wave model appropriate for ultrasound frequencies and tissue‑like wave speeds.
Create synthetic “phantoms” that mimic key cardiac structures (or use an open phantom dataset) and generate simulated transducer data.
Implement an FWI loop: define an objective (misfit), compute gradients via adjoint propagation, and update a parameter field (e.g., sound speed) with regularisation.
Study practical issues: limited‑view acquisition, noise, cycle skipping, and the effect of frequency continuation / multiscale strategies.
Deliverable: a reproducible prototype (notebooks + code) that can reconstruct simple phantoms and documents what acquisition setups and regularisation choices work best. This suits students who like inverse problems, PDEs, and fast computational kernels; it avoids clinical data/ethics by using synthetic or open data.
Learning Data‑Driven Metrics for Schemas and Taxonomies#
Project code: rhne-164
Main supervisor: Rhodri Nelson,
rhodri.nelson@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor is not yet assigned.
Available to: ACSE EDSML
This project may accept multiple students.
Debates about “the right” schema, taxonomy, or ontology often get stuck because the criteria are implicit and subjective. This project flips the problem: instead of hand‑crafting a taxonomy and then arguing about it, learn an explicit metric (objective function) from data that captures what “good” means in a given context.
You will:
Choose a concrete setting (e.g., organising a document collection, building a label hierarchy for classification, or aligning two schemas).
Define a family of candidate metric features that are easy to compute on a proposed structure (coverage, consistency, depth/balance, predictiveness for a downstream task, annotation cost, etc.).
Collect supervision signals (pairwise preferences, “gold” examples, or downstream performance scores) and learn a metric that best explains them. This could be a simple weighted model, a small neural network, or a transformer that outputs metric parameters from a textual specification.
Use the learned metric to generate or optimise a schema/taxonomy (via search or gradient‑based optimisation if differentiable) and compare it to hand‑designed baselines.
Deliverable: an end‑to‑end pipeline (data → learned metric → improved structures), with a clear discussion of interpretability and failure modes. Ideal for students interested in ML, representation learning, and “meta” evaluation problems.
Metric‑Driven Typing of Student Questions to LLMs in Higher Education#
Project code: rhne-179
Main supervisor: Rhodri Nelson,
rhodri.nelson@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Alexandra Neagu,
alexandra.neagu20@imperial.ac.uk, Department of Mechanical Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
Students ask LLMs for many different things: explanations, worked examples, debugging help, feedback on writing, brainstorming, revision planning, and sometimes assistance that crosses academic‑integrity boundaries. Existing “intent/taxonomy” classifiers often reuse generic chat schemas, but these are not designed around what matters in HE (learning value, assessment risk, and what intervention the system should take). It’s also rarely clear what those schemas are optimising beyond label accuracy on an arbitrary label set.
This project develops a question‑type classifier where the training signal is explicitly tied to a metric that reflects HE goals. You will:
Propose a compact, HE‑relevant label set (e.g., conceptual explanation, procedural/how‑to, debugging, critique/feedback, summarisation, planning, admin, assessment‑answer seeking, integrity‑risk).
Define a metric that captures “decision usefulness” rather than just accuracy, e.g. a cost‑sensitive objective where high‑risk misclassifications (assessment‑answer seeking → normal help) are penalised more than benign ones, or a metric based on how well the predicted type routes the user to the right support strategy (hinting vs full solution, source‑grounded responses, refusal/escalation).
Build a small dataset (consented/anonymised logs if available, or curated prompts + expert annotation), plus an annotation guide and inter‑rater analysis.
Train baselines (standard supervised classifier) and a metric‑aligned model (cost‑sensitive learning / direct metric optimisation), then compare on both standard scores and your decision metric.
Deliverable: a reproducible dataset + taxonomy, a clearly stated metric, a trained classifier, and practical recommendations for how HE LLM tools should detect and handle different student question types.
Feedback Control of Vortex Shedding with Reduced‑Order PDE Models in Devito#
Project code: rhne-163
Main supervisor: Rhodri Nelson,
rhodri.nelson@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Zoe Leibowitz,
zoe.leibowitz21@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE
This project does not accept multiple students.
Laminar flow past a cylinder exhibits vortex shedding beyond a critical Reynolds number, producing an unsteady wake. This project investigates linear feedback strategies to suppress (or delay) shedding using a tractable PDE model implemented in Devito.
Rather than aiming for a full CFD solver, you will work with a reduced but physically meaningful model, for example:
a linearised perturbation model about a steady base flow, or
a wake surrogate such as the (complex) Ginzburg–Landau equation.
You will:
Implement the chosen PDE model in Devito and validate it against known behaviours (growth/decay rates, onset of oscillations).
Design a feedback controller (e.g., LQR, pole placement, or simple proportional feedback) using measurements at a small number of “sensor” locations.
Test closed‑loop performance under disturbances and parameter changes; quantify stability margins and control effort.
Produce clear visualisations of controlled vs uncontrolled wake dynamics.
Deliverable: a reproducible set of notebooks/scripts that go from model → controller → closed‑loop simulation, plus a short report on what’s captured by the model and what would be needed for higher‑fidelity extensions. Good for students who like PDEs, control, and scientific computing.
Prototyping a Devito Back‑End for iSALE‑Style Impact Modelling#
Project code: rhne-161
Main supervisor: Rhodri Nelson,
rhodri.nelson@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Gareth Collins,
g.collins@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
iSALE is a widely used shock‑physics impact modeller, but much of its runtime is spent in repeated, grid‑based update kernels. Devito offers an alternative route: express finite‑difference stencils at a high level in Python and generate optimised low‑level code. This project explores what it would take to migrate (part of) an iSALE‑style solver to a Devito back‑end.
You will:
Identify one or two self‑contained numerical kernels that are good candidates for Devito (e.g., a pressure/velocity update for a simplified compressible model, artificial viscosity, or a stress update).
Implement the chosen kernel(s) in Devito with clear interfaces so they could be called from a larger code base.
Validate against standard test problems (e.g., 1D shock tube / 2D blast wave) and, where feasible, compare outputs to reference iSALE runs.
Benchmark performance (CPU, and optionally GPU) and document what helps/hurts (memory layout, halo regions, operator fusion).
Deliverable: a working prototype library + test suite, a short validation report, and a “migration note” describing architectural options for a larger port. Best for students interested in scientific software engineering and HPC.
Spatial SEIRD Epidemic Modelling in Devito#
Project code: rhne-159
Main supervisor: Rhodri Nelson,
rhodri.nelson@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor is not yet assigned.
Available to: ACSE EDSML
This project may accept multiple students.
Classic SEIRD models treat a population as “well mixed”, but real outbreaks spread through space via mobility and local contact structure. In this project you will build and explore a spatially dependent SEIRD model (reaction–diffusion / advection–diffusion) and implement it efficiently using Devito (finite‑difference code generation in Python).
You will:
Formulate a spatial SEIRD PDE model, choose boundary conditions, and implement stable time stepping on a 1D/2D grid.
Verify the implementation against limiting cases (no diffusion → ODE SEIRD; simple manufactured solutions).
Run scenario studies inspired by COVID‑19 (e.g., localised seeding, mobility changes, regional interventions) and visualise how parameters affect peak timing and spatial patterns.
Optional extension: fit a small number of parameters to publicly available case data for one region, and discuss identifiability and uncertainty.
Deliverable: a documented, reproducible simulator (with tests and notebooks), plus a short write‑up of model behaviour and performance. Suitable for students who like PDEs, scientific Python, and computational modelling.
Personalised Learning Paths with AI Study Assistants: Custom Tooling vs Off‑the‑Shelf#
Project code: rhne-158
Main supervisor: Rhodri Nelson,
rhodri.nelson@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor is not yet assigned.
Available to: ACSE EDSML
This project may accept multiple students.
Students increasingly use AI tools to learn: summarising, self‑testing, and planning what to study next. But general tools don’t know your module structure, prerequisites, or what you’ve already mastered. This project explores how far we can get with off‑the‑shelf assistants (e.g., NotebookLM‑style tools) versus a lightweight custom system tailored to UG and PGT curricula.
You will:
Specify what a “learning path” means in practice (prerequisite graph, target outcomes, time budget, assessment dates).
Prototype two approaches: (i) configure an existing study assistant on a curated set of course materials, and (ii) build a small custom planner that recommends next topics + exercises based on a learner profile.
Evaluate with realistic scenarios (different backgrounds, missed lectures, exam revision) using rubrics such as coverage, sequencing quality, faithfulness to sources, and usability.
Deliverable: a small working prototype (web app or notebooks), a set of example learning paths for selected modules, and a short report discussing what should be built “in house” vs what can be delegated to commercial tools. Good fit for students interested in education tech, LLMs, and UX.
Iterative (Agentic) RAG vs Knowledge Graphs for Multi‑Hop QA#
Project code: rhne-157
Main supervisor: Rhodri Nelson,
rhodri.nelson@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor is not yet assigned.
Available to: ACSE EDSML
This project may accept multiple students.
Knowledge graphs are attractive because they can connect concepts that are logically related even when the underlying text is not semantically similar. Retrieval‑augmented generation (RAG), on the other hand, is often “one shot”: retrieve, read, answer. This project asks whether an agentic / iterative RAG loop—where the model repeatedly retrieves, writes down intermediate findings, and expands its own context—can match or beat a knowledge‑graph pipeline on multi‑hop questions.
You will:
Build two baselines over the same corpus: (i) KG‑assisted QA (entity linking + relation extraction + graph traversal) and (ii) iterative RAG (query rewriting + tool calls + “memory” summarisation).
Design an evaluation set emphasising “bridge” queries (A relates to B via C), plus standard factoid questions.
Compare accuracy, evidence traceability, latency and cost; then run ablations to identify what actually helps (chunking, retrieval depth, memory format, stopping criteria).
Deliverable: a reproducible benchmark (data + scripts), a concise set of design recommendations, and a working demo notebook/CLI. Ideal for students interested in LLM systems, information retrieval, and evaluation.
Neural Physics and AI Surrogates for Urban Wind and Flood Modelling with AKTII#
Project code: chpa-234
Main supervisor: Christopher Pain,
c.pain@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Claire Heaney,
c.heaney@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
This MSc project, delivered in collaboration with AKT-II, explores the use of Neural Physics (AI4PDEs) and Foundational AI Surrogates to accelerate high-resolution modelling of urban wind and flooding over realistic city geometries. Traditional CFD approaches are computationally expensive and slow to deploy in design workflows. This project replaces classical solvers with Neural Physics models (neural networks whose weights are analytically derived from governing physical equations) combined with AI surrogates trained on high-fidelity simulations.
Students will work with real architectural datasets, including 3D building and terrain geometries (typically within 400–500 m radius) and atmospheric boundary layer profiles. Benchmark studies will include AIJ Case E (a standardised urban wind test case) and a real AKT-II project. Primary outputs are pedestrian-level wind speed time histories, with secondary outputs including surface pressure time histories on key buildings. Students will use detailed geometries alongside rainfall statistics and boundary conditions to model transient surface water flow.
The project aims to deliver fast, AI-driven urban digital twins capable of predicting wind comfort, structural pressures, and flood risk in near real time (Projects 1–3).
Project 1: Neural Physics for Urban Wind Implement analytically weighted Neural Physics solvers to predict transient wind fields and pedestrian-level wind speeds over complex cityscapes.
Project 2: AI Surrogates for Rapid Design Feedback Train surrogate models for fast inference of wind and pressure histories, enabling rapid evaluation of multiple design scenarios.
Project 3: Neural Physics for Urban Flooding Develop Neural Physics models for rainfall-driven surface flooding, integrating geometry, terrain, and rainfall statistics.
Project 4: Neural Physics for Urban Sound Propagation and Acoustic Comfort This option focuses on modelling urban sound propagation and acoustic comfort using Neural Physics implementations of the acoustic wave equation over realistic city geometries provided by AKT-II. Students will develop analytically weighted neural networks that directly solve the discretised wave equation, enabling prediction of transient sound fields in complex urban environments. Using detailed 3D building and terrain models (typically within a 400–500 m radius), the project will simulate how traffic, construction, or environmental noise sources propagate, reflect, diffract, and attenuate through streets and around buildings. Neural Physics solvers will enable rapid evaluation of multiple design scenarios, such as changes in building layout, façade treatments, or sound barriers. Primary outputs include time histories of sound pressure levels at pedestrian locations, spatial acoustic maps, and frequency-dependent metrics relevant to urban comfort. Results will be benchmarked against conventional acoustic solvers where available. This option supports applications in urban noise mitigation, building design and healthy city planning, providing students with hands-on experience in Neural Physics, wave-based modelling and AI-driven digital twins for smart cities.
Students will use Python and PyTorch on GPU systems and should have an interest in scientific AI, and urban modelling.
Neural Physics and AI Surrogates on Large-Scale GPU Systems for Urban Heat and Humidity Mapping in Westminster, London#
Project code: chpa-235
Main supervisor: Christopher Pain,
c.pain@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Claire Heaney,
c.heaney@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
This MSc project applies Neural Physics (AI4PDEs) and Foundational AI Surrogate models on large-scale GPU systems to generate high-resolution, transient maps of temperature and relative humidity across Westminster, London—one of the most densely built and well-instrumented districts in the UK. Westminster combines complex urban geometry, heavy pedestrian and transport activity, and extensive environmental monitoring, making it an ideal testbed for AI-driven urban microclimate modelling.
Urban heat islands, street canyons, and heterogeneous land use create strong spatial variability in thermal comfort and humidity. Traditional CFD methods struggle to provide real-time predictions at neighbourhood scale. This project replaces classical solvers with Neural Physics networks whose weights are analytically derived from governing heat and moisture transport equations, combined with GPU-accelerated AI surrogates trained on high-fidelity simulations and OpenWeather datasets.
Students will develop scalable GPU pipelines integrating OpenWeather atmospheric variables (temperature, relative humidity, wind speed, solar radiation) with detailed Westminster building and terrain geometries to model heat accumulation and moisture transport through streets, squares, and transport corridors. The resulting system forms a high-resolution digital twin of Westminster’s urban microclimate, supporting applications in thermal comfort assessment, heatwave resilience planning, and exposure analysis for walking routes, bus corridors, and Underground stations.
Performance will be benchmarked against local sensor observations in Westminster, enabling rigorous validation of temperature and humidity predictions. Students will also explore uncertainty quantification and scenario analysis, such as tree planting, shading interventions, or altered traffic patterns.
Project 1: GPU-Based Neural Physics for Urban Heat and Moisture PDEs Implement analytically weighted Neural Physics solvers on multi-GPU clusters to model transient temperature and humidity fields across Westminster’s complex cityscape.
Project 2: Foundational AI Surrogates for Real-Time Microclimate Inference Train GPU-based surrogate models for rapid prediction of neighbourhood-scale heat and humidity, enabling near–real-time urban climate mapping.
Project 3: Tree planting, shading interventions.
Students should have an interest in scientific AI, GPU computing and applied urban climate modelling.
Solving PDEs using machine-learning techniques#
Project code: chpa-155
Main supervisor: Christopher Pain,
c.pain@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Claire Heaney,
c.heaney@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
Neural networks struggle with long-term time series prediction, as the error in the solution accumulates over time. Physics-informed approaches attempt to enforce the neural network solutions to satisfy physical laws (such as conservation of mass and/or momentum) with the hope that this will improve the long-term forecasting ability of the network [1,2,3].
In this project we will integrate AI4PDEs, a solver for computational fluid dynamics (CFD) that is written as a neural network. This will be used to calculate the residual that will be included in the loss function of the physics-informed neural network (PINN). This offers an elegant approach of combining CFD (an untrained neural network) with a surrogate model (trained neural network). We will also look at alternative methods to PINNs, such as minimising the discrete residual of the governing equations evaluated with the output of the neural network [4,5]; developing geometry- and grid-invariant foundational AI models that can be treated like a CFD model; particle and molecular dynamics surrogates.
The student(s) will be provided with CFD data and build AI models in Python and PyTorch.
[1] Raissi, Perdikaris, Karniadakis (2019) Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378:686-707.
[2] Arthurs, King (2021) Active training of physics-informed neural networks to aggregate and interpolate parametric solutions to the Navier-Stokes equations. Journal of Computational Physics, 438.
[3] Chen, Wang, Hesthaven, Zhang (2021) Physics-informed machine learning for reduced-order modeling of nonlinear problems, Journal of Computational Physics, 446:110666.
[4] Xiao, Fang, Buchan, Pain, Navon, Du, Hu (2010) Non-linear model reduction for the Navier-Stokes equations using residual DEIM method, Journal of Computational Physics, 263:1-18.
[5] Sipp, de Pando, Schmid (2020) Nonlinear model reduction: A comparison between POD-Galerkin and POD-DEIM methods. Computers & Fluids, 208
Using generative AI to develop a predictive tool for the impact of pollution on health#
Project code: chpa-156
Main supervisor: Christopher Pain,
c.pain@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Boyang Chen,
boyang.chen16@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
The challenge: Up to 90% of the world’s population breathe air with high levels of both indoor and outdoor pollution that kills ~7 million people each year worldwide. In the UK, it is rated as one of the most serious threats to public health with only cancer, obesity and heart disease eclipsing it. The health risks associated with fine and ultrafine particulate matter (PM2.5 and PM0.1) include development and exacerbation of respiratory diseases such as chronic obstructive lung diseases including asthma, respiratory infections and lung cancer.
Maximising health of individuals: Using cohort data, we will develop an AI model of exposure and health response to pollution to make generic predictions. At the individual level, we utilise the increasing range of sensor information from smart watches and mobile phones to personalise the predicted health response to levels of personal pollution exposure. Furthermore, as individual sensors collect more information about an individual and how they respond to their environment, an AI model can be refined making it specific to that individual, accounting for their medical condition and history where available, to make predictions and mitigation suggestions for that individual. This information can then be uploaded to the more generic generative neural network [2] to greatly improve its predictive ability over time. In this way the system becomes increasingly smart as time goes by and is able to better: (1) diagnose health issues e.g. viral infections, respiratory problems with the nuances in the patterns the AI system learns from sensitivity to individuals age and season, (2) provide advice and develop new health insights in terms of exposure to pollution, potentially provide health advice, to exercise or not.
Data from AI-Respire app: The data from wearable devices currently includes: temperature, blood oxygen saturation, respiratory rate, heart rate, movement, position (using GPS) plus user-added information on gender, weight and age. These are embodied in, for example, Fitbit, Garmin and Apple watch software with inclusion of air quality data from OpenWeather
Projects: Using this data the project will develop further the generative AI models [2] (e.g. GANs, VAEs, Latent Diffusion) in order to address challenges 1 and 2 above. Specific projects: Project 1: Form a generative autoencoder that is able predict health responses given pollution conditions for specific classes of individuals (e.g. asthmatics, healthy people). Then tailor this model (using transfer learning) so that it is able to predict the responses of a specific individual e.g. an asthmatic of healthy person. Project 2: Develop generative AI methods for diagnosing potential benefits (of actions that reduce pollution exposure) and provide uncertainties associated with those benefits. Further develop this for an individual. Project 3: Perform optimisation of a persons environment in order to maximise their health e.g. their commute through a city. Project 4: Integrate personal health data with UK BIO bank data in order to predict long terms consequences of pollution exposure for individuals. Project 5: Integrate personal ECG data with pollution exposure for individuals.
PyTorch and Python skills needed.
[1] https://www.imperial.ac.uk/news/246893/government-funding-revolutionise-ai-healthcare-research/
[2] Goodfellow, Pouget-Abadie, Mirza, Xu, Warde-Farley, Ozair, Courville, Bengio (2014) Generative adversarial networks, arXiv preprint: arxiv:1406.2661.
High‑Resolution (Millimetre Scale) Neural Physics and AI‑Driven Geometry for Turbulent Airflow Around a Vehicle with Dynamically Optimised Distorted Grids#
Project code: chpa-233
Main supervisor: Christopher Pain,
c.pain@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Boyang Chen,
boyang.chen16@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
This MSc project focuses on developing millimetre‑scale turbulent airflow models around a car using Neural Physics (AI4PDEs) and AI‑based dynamic grid optimisation with realistic 3D geometry representation. The automotive aerodynamics community routinely uses detailed meshes such as those from the AutoCFD4 benchmark case for the DrivAer notchback model, which includes volume meshes refined around key features like wheels, seals, and near‑wall boundary layers. See :https://autocfd4.s3.eu-west-1.amazonaws.com/test-cases/case2/meshes/AutoCFD4_UpdatedMesh.pdf Students will replace traditional CFD solvers with Neural Physics networks whose weights are analytically derived from the governing Navier–Stokes equations and turbulence closures. These models will be coupled with AI world models and dynamic grid representations that can adaptively refine and distort computational grids around critical regions (e.g., wake, separation zones) to resolve flow structures at millimetre resolution. The project combines neural geometry learning with a Neural Physics PDE solvers to generate transient turbulent flow predictions that resolve vortices, separation, and pressure fluctuations in the near‑body and wake regions.
Outcomes could include: • A neural geometry module capable of representing complex car surfaces and surrounding distorted grids suitable for flow prediction. • A Neural Physics solver integrated with dynamic grid optimisation to produce transient turbulent airflow fields at ~1 mm resolution around the vehicle. • Comparisons against established CFD results (e.g., RANS/LES benchmarks) to validate accuracy, flow separation prediction, and computational efficiency. • A documented framework demonstrating how AI can dynamically adapt grid structures based on flow features.
These projects suit people with an interest in scientific machine learning, high‑fidelity flow modelling and AI‑driven geometry representation for advanced aerodynamic simulation.
Traffic modelling using Foundational Grid Invariant Neural Networks#
Project code: chpa-154
Main supervisor: Christopher Pain,
c.pain@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Boyang Chen,
boyang.chen16@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
This project will model traffic using some of the transformative foundational surrogate modelling methods that have been applied recently to model fluid flows. These models include methods based on AI diffusion models (https://en.wikipedia.org/wiki/Diffusion_model) that have an architecture that ensures that the model can learn and be applied on any grid resolution. Thus the model can be applied to model flows around a few buildings but then re-applied to model flows on the city scale. The same approach will be attempted with the modelling of traffic. In this way it is hoped that the model will be able to learn the behaviour of vehicles and their interactions. The modelling will learn from (i) data collected from real traffic systems (ii) the VSM traffic simulator.
After completion of the project we hope to have a model of vehicle movement that can be used to determine the pollution levels emitted by vehicles as well as to be able to assess drivers’ behaviour by comparing with the neural network model in order to provide advice to improve driving economy or help determine if there is a fault with the vehicles.
The coding will be in PyTorch and Python.
There is a possibility of working with colleagues in the Civil Engineering Department at Imperial College as well as working with colleagues at University College London.
Solving PDEs on unstructured and distorted meshes with AI libraries#
Project code: chpa-152
Main supervisor: Christopher Pain,
c.pain@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Boyang Chen,
boyang.chen16@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
Due to new hardware technologies, increases in computing power and developments in AI software, the benefits of combining AI techniques with traditional numerical methods for solving governing equations of dynamical systems are becoming apparent. For example, Cerebras have just released a new ‘AI computer’ which has about 1 million cores on a single chip with vastly increased computational speed, yet which requires much less energy than GPUs or CPUs, making it a tantalising prospect for researchers wishing to run demanding computations in an energy efficient manner. If the potential of combining the new AI computers and AI software with traditional numerical methods can be harnessed, one can expect a revolution in computational physics across disciplines and a ‘must have’ next generation technology. Based on the latest techniques in AI software that are particularly suited to exascale computing, this project develops a potentially revolutionary approach to the discretisation and solution of differential equations, and the formation of surrogate models. Our approach implements models, such as Computational Fluid Dynamics (CFD), using AI software with the aim of simplifying the software development and building on the very substantial developments already made in AI software. This simplification would greatly increase the number of developers capable of developing current CFD/nuclear codes, speeding up the implementation of developments and – crucially – their parallel scalability. This new approach also enables relatively simple development of digital twins, using the optimisation engine and sensitivities embedded in AI software. The digital twins can be used to optimise systems, form error measures, assimilate data and quantify uncertainty in a relatively straightforward manner with this approach. This enables the major deficiency (formation of an error estimate) in current modelling approaches for safety critical systems in nuclear engineering and the environment to be addressed.
Progress has been made with this approach on structured meshes [1,2,3], and in these projects we look to extend this approach to unstructured meshes. For these projects, an interest in several of the following would be beneficial: computational fluid dynamics, neural networks and numerical methods. The coding will be in PyTorch and Python.
The individual projects on offer are: (i) Multi-grid for solving large systems of equations (ii) Automatic formation of adjoint sensitivities using backpropagation (iii) Parallelisation (iv) Forming a Computational Fluid Dynamics solver (v) Cube-sphere representations for capturing the dynamics of global atmosphere using distorted structured grids (viii) Mesh movement for capturing fluid dynamics using distorted structured grids
[1] Chen, Heaney, Pain (2024) Using AI libraries for Incompressible Computational Fluid Dynamics, https://arxiv.org/abs/2402.17913.
[2] Chen, Heaney, Gomes, Matar, Pain (2024) Solving the Discretised Multiphase Flow Equations with Interface Capturing on Structured Grids Using Machine Learning Libraries, Computer Methods in Applied Mechanics and Engineering 426: 116974. https://doi.org/10.1016/j.cma.2024.116974
[3] Chen, Nadimy, Heaney et al. (2024) Solving the Discretised Shallow Water Equations Using Neural Networks, Advances in Water Resources, accepted. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4956116
Neural Physics and AI Surrogates on FPGAs for Greater London Airflow Modelling#
Project code: chpa-226
Main supervisor: Christopher Pain,
c.pain@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Boyang Chen,
boyang.chen16@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
This project explores the use of Field-Programmable Gate Arrays (FPGAs) to accelerate Neural Physics (AI4PDEs) and Foundational AI Surrogate models for city-scale simulation of atmospheric airflow across the Greater London area. Traditional CFD methods are computationally expensive and typically rely on GPUs. FPGAs offer a compelling alternative, providing significantly improved energy efficiency, low latency, and compact deployment, making them ideal for real-time environmental modelling and edge computing. The project will be run with the leading FPGA manufacturer Macnica DHW. To solve discretised Partial Differential Equations, students will work with neural networks whose weights are determined analytically instead of by training, thereby replacing classical solvers with Neural Physics architectures. These models will be combined with AI surrogates trained on high-fidelity simulations to enable rapid prediction of wind fields, heat transport and pollutant dispersion. The project contributes toward creating a digital twin of Greater London, supporting applications in air quality assessment, climate resilience and emergency response.
Students will collaborate with FPGA developers to implement and optimise these models on FPGA hardware, focusing on parallelisation, memory efficiency, and numerical precision. Performance will be evaluated against GPU-based implementations in terms of speed, power consumption and accuracy. Successful outcomes may contribute to an open FPGA developer platform for Neural Physics.
Project 1: FPGA Implementation of Neural Physics (AI4PDEs) Focus on mapping analytically weighted Neural Physics networks onto FPGAs. Tasks include hardware-aware model design, kernel optimisation and benchmarking for airflow PDEs over urban domains.
Project 2: FPGA Deployment of Foundational AI Surrogates Concentrate on deploying pretrained AI surrogate models on FPGAs for rapid urban airflow inference. Emphasis is placed on model compression, precision tuning and real-time prediction of Greater London wind and pollution fields. Both options offer strong foundations in AI, hardware acceleration and environmental modelling, preparing students for careers in scientific AI and high-performance computing.
AI4PDEs for Viscous Incompressible Flow#
Project code: chpa-101
Main supervisor: Christopher Pain,
c.pain@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Donghu Guo,
donghu.guo21@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML GEMS READY
This project may accept multiple students.
The student(s) are expected to explore the use of AI4PDEs to generate simulations for viscous incompressible flow problems governed by the Navier-Stokes equations, such as flow past a cylinder, multiple cylinders, or buildings, starting in 2D and potentially expanding to 3D. Also, the students are expected to compare the results with those of traditional numerical solvers such as ICFERST. Furthermore, if time allows, the student could use these generated simulations to train some of the AI surrogate models.
AI4PDEs is an in-house computational fluid dynamics (CFD) solver, which solves discretised systems of equations using neural networks. The weights of the networks are determined in advance through the choice of discretisation method (e.g. first-order finite elements, second-order finite differences etc), so no training is needed. The solutions are the same as those obtained from Fortran or C++ codes (to within solver tolerances).
Generative Models, e.g., Scored-based Models, Flow Matching, for Viscous Incompressible Flow#
Project code: chpa-102
Main supervisor: Christopher Pain,
c.pain@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Donghu Guo,
donghu.guo21@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML GEMS READY
This project may accept multiple students.
In recent years, generative models such as score-based models and flow matching have had a significant impact across various fields. Their application to fluid flow problems has also gained increasing attention. This project aims to leverage these state-of-the-art models for viscous incompressible flow. There are two main streams in the project.
Superresolution – The student(s) could explore the use of generative models to perform superresolution tasks, enhancing simulations by mapping between different grid sizes.
Prediction – The student(s) could investigate the application of generative models for time-step forecasting. There are multiple possible approaches, including: Autoregressive prediction, where the model takes data from the previous time step, predicts the next, and iteratively repeats this process to forecast further. Multi-step prediction, where the model directly predicts multiple future time steps in a single forward pass.
Additionally, different conditioning strategies and inference methods for the generative models could be explored and compared.
Agentic AI#
Project code: chpa-225
Main supervisor: Christopher Pain,
c.pain@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Jucai Zhai,
jucai.zhai25@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
Project 1: CO AI-ARCHITECT for Digital Twins and Neural Physics This project develops a collaborative AI framework (CO AI-ARCHITECT) capable of creating and managing digital twins for complex systems spanning environment, health, and engineering. Using advanced neural networks and uncertainty-aware AI, the system will overcome the curse of dimensionality to support modelling, risk analysis, and design optimisation. A key aim is to explore “neural physics”: AI models that can automatically generate simplified physics representations and integrate them into agent-based workflows. The student will build prototype pipelines where AI forms models, analyses scenarios (e.g., climate or energy systems), and produces actionable recommendations, demonstrating scalable, accessible scientific modelling.
Project 2: CO AI-MEDIC for One Health Decision Support This project focuses on applying AI to the One Health paradigm, linking human health, environmental quality and planetary wellbeing. The student will develop a CO AI-MEDIC digital twin that integrates environmental data and health indicators to support diagnosis, treatment optimisation, and population-level risk assessment. The system will simulate scenarios such as pandemics, flooding or pollution events, enabling rapid model formation and response planning. Emphasis will be placed on uncertainty quantification, personalised recommendations and ethical deployment of AI in healthcare and public health decision-making.
Project 3: AI-Author Platform for Science–Medicine–Art Communication This project creates an interactive AI-Author platform that translates complex AI outputs into accessible reports, visualisations, and artistic narratives. The system will generate tailored content for audiences ranging from the public to researchers, integrating real-time environmental mapping and decision-support tools. The student will explore AI-driven knowledge dissemination and creative expression, demonstrating how science, medicine and art can converge to empower individuals and communities while supporting sustainable planetary development.
Modelling Birds: Behaviour, Appearance, and Interaction through Art–Science Collaboration#
Project code: chpa-224
Main supervisor: Christopher Pain,
c.pain@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: James Coupe,
james.coupe@rca.ac.uk, Royal College of Art, UKAvailable to: ACSE EDSML
This project may accept multiple students.
Birds occupy a unique position at the intersection of ecology, motion, perception and cultural meaning. Their flight dynamics, collective behaviours, vocalisations and visual appearance have inspired scientific inquiry, artistic practice and design innovation for centuries. Yet contemporary modelling approaches often fragment these aspects, treating movement, form, behaviour and representation as separate problems.
This project proposes a collaborative research programme between ESE and the Royal College of Art to develop integrated models of birds that combine physical motion, behavioural rules, visual appearance and expressive representation including aspects such as action and behaviour, form and appearance, and interaction and expression.
Possible project directions • Animated and spatial representations of bird behaviour Creating dynamic visual works that explore flight, flocking and interaction in physical or virtual space. • Urban birds and the built environment Modelling how birds inhabit, navigate and respond to architectural and urban landscapes, with outcomes informing design speculation and public installations. • Speculative avian forms Exploring imagined or future bird morphologies as a response to environmental change, climate pressure or human influence. • Birds as interfaces and agents Investigating birds as interactive agents within responsive environments, performances or digital systems. • Artistic tools for modelling living systems
Developing modelling frameworks that are accessible to artists and designers, enabling experimentation with behaviour, motion and form without requiring conventional scientific workflows.
Solving PDEs on unstructured and distorted meshes using Quantum Convolutional Neural Networks and AI libraries#
Project code: chpa-223
Main supervisor: Christopher Pain,
c.pain@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Boyang Chen,
boyang.chen16@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
Rapid advances in quantum computing, AI software, and heterogeneous hardware architectures are opening new pathways for solving partial differential equations (PDEs) at scales that are currently far beyond the reach of classical computing. Alongside the growth of exascale and AI-centric platforms, quantum computing offers the potential for fundamentally new computational paradigms, particularly for high-dimensional, nonlinear and multiscale dynamical systems.
This project explores a quantum-enabled approach to PDE solution, based on Quantum Convolutional Neural Networks (QCNNs) implemented using the PennyLane quantum machine-learning framework and integrated with PyTorch. The focus is on developing scalable hybrid quantum–classical solvers for complex flow systems.
The project develops a novel computational framework in which QCNNs are used as numerical operators for PDE solution on unstructured and distorted meshes. Quantum convolutional layers encode local spatial interactions, fluxes, and nonlinear couplings associated with the governing equations, while classical AI components manage mesh connectivity, domain decomposition, and time integration. PennyLane provides the interface between quantum circuits and classical execution environments, enabling QCNNs to be embedded within large-scale simulation workflows.
Key features of the approach include quantum-enhanced representation of nonlinear flow dynamics; compatibility with unstructured and distorted meshes; and extreme-resolution simulation. The governing equations of atmospheric dynamics, such as the Navier-Stokes equations, thermodynamics and moisture transport, are incorporated through physics-based numerical formulations coupled with QCNN-based operators, rather than relying solely on classical discretisation techniques.
Projects include 1. Quantum convolutional architectures for PDE solution Design and evaluation of QCNN architectures for representing nonlinear PDE operators on unstructured and distorted computational meshes. 2. Hybrid quantum–classical fluid solvers Development of solvers in which QCNNs compute local flow updates on mesh elements while classical components manage mesh connectivity, boundary conditions, and time stepping. Here we will extend multi-grid method implemented as a U-Net to make them more suitable for quantum computing. 3. Execution strategies for quantum hardware Investigation of how QCNN-based solvers can be deployed on near-term quantum devices and scaled toward future fault-tolerant quantum computers.
[1] Chen, Heaney, Pain (2024) Using AI libraries for Incompressible Computational Fluid Dynamics, https://arxiv.org/abs/2402.17913.
[2] Chen, Heaney, Gomes, Matar, Pain (2024) Solving the Discretised Multiphase Flow Equations with Interface Capturing on Structured Grids Using Machine Learning Libraries, Computer Methods in Applied Mechanics and Engineering 426: 116974. https://doi.org/10.1016/j.cma.2024.116974
[3] Chen, Nadimy, Heaney et al. (2024) Solving the Discretised Shallow Water Equations Using Neural Networks, Advances in Water Resources, accepted. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4956116
Surrogate models for carbon storage in porous media#
Project code: chpa-112
Main supervisor: Christopher Pain,
c.pain@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Nathalie Carvalho Pinheiro,
n.pinheiro23@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML GEMS READY
This project may accept multiple students.
Modelling solid–fluid interactions relies on solving coupled systems of partial differential equations (PDEs) to capture the evolution of flow and deformation across a domain. High-fidelity numerical simulations of these processes—such as fluid flow in porous media—typically require fine spatial discretisation, resulting in high computational cost and memory demands. This limits their applicability in large-scale or time-sensitive Earth-science problems. Recent advances in artificial intelligence offer a promising alternative. In particular, machine-learning–based surrogate models can approximate complex physical simulations at a fraction of the computational cost, while retaining key dynamical features of the system. In this project, students will develop, train, and evaluate machine-learning surrogate models to predict fluid flow in porous media for a carbon storage scenario, using data from existing numerical simulations that modelled the flow behaviour and its reaction with the surrounding rock [1]. The project has a strong programming focus, using Python/PyTorch to create AI-based models. Students will compare different forecasting strategies, including recursive single-timestep prediction and multi-timestep prediction, and assess their accuracy, stability, and computational efficiency, comparing the models with a previously developed model. The methods developed are broadly applicable to a wide range of Earth-science and engineering problems involving subsurface flow, making this project an excellent opportunity for students who aim to combine machine learning and geoscience applications.
References [1] J. Maes, C. Soulaine, H. P. Menke, Improved volume-of-solid formulations for micro-continuum simulation of mineral dissolution at the pore-scale, arXiv preprint 2204.07019 (2022).
Using Generative AI to Develop Predictive Tools for the Impact of Urban Heat on Health and Transport in London#
Project code: chpa-231
Main supervisor: Christopher Pain,
c.pain@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Claire Heaney,
c.heaney@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
The Challenge
Extreme heat is an increasing public health threat, intensified by climate change and urban heat island effects. London experiences elevated temperatures in dense built environments and underground transport systems, particularly the Tube, where heat accumulation can exceed surface temperatures. Heat exposure is associated with dehydration, cardiovascular stress, respiratory problems, reduced cognitive performance, and increased mortality, especially among vulnerable populations such as the elderly and those with chronic conditions. The project will be supervised with Transport For London. This project focuses on modelling how urban heat affects individual health and transport usage across London, combining environmental data with wearable sensor information and generative AI.
Maximising Individual Health Using cohort-level data, students will develop AI models linking heat exposure to physiological response. At the personal level, data from smart watches and mobile devices (temperature, heart rate, blood oxygen, respiratory rate, movement, GPS, age, weight, gender) will be used to personalise predictions. As more data is collected, the model adapts to each individual via continual learning, accounting for medical history where available. These personalised models feed into a larger generative neural network, enabling improved diagnosis of heat stress, fatigue, and cardiovascular strain, while providing mitigation advice such as route changes, travel timing, hydration strategies, or reduced exertion during peak heat. Transport environments (Tube stations, trains, buses, walking routes) are explicitly included to quantify cumulative heat exposure during commutes.
Projects Using PyTorch and Python, students will develop generative AI models (GANs, VAEs, Latent Diffusion) addressing: Project 1: Build a generative autoencoder predicting health responses under heat for population groups, then personalise via transfer learning. Project 2: Develop generative models to evaluate benefits of heat-mitigation actions (e.g., altered travel routes or schedules), including uncertainty quantification. Project 3: Optimise personal urban environments and commutes to minimise heat exposure and maximise health. Project 4: Integrate personal data with UK Biobank to predict long-term health impacts of repeated heat exposure. Project 5: Combine ECG and wearable data with thermal exposure to assess cardiovascular risk. This project supports climate-resilient urban design and healthier transport systems in London.
Flooding Modelling With Neural Networks#
Project code: chpa-124
Main supervisor: Christopher Pain,
c.pain@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Boyang Chen,
boyang.chen16@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
Recently, AI4PDEs (also known as NN4PDEs) has shown great success in being able to solve very large scale floods across UK cities [1]. Based on Neural Physics [2,3], is an in-house computational fluid dynamics code that solves discretised systems of equations using neural networks. The weights of the networks are determined in advance through the choice of discretisation method (e.g. first-order finite elements, second-order finite differences etc), so no training is needed. The solutions are the same as those obtained from Fortran or C++ codes (to within solver tolerances). The code runs on GPUs as well as CPUs and AI processors. The benefits of reliable flood prediction are an ability to save lives by providing early flood warnings and also providing suitable flood resilience through flood defence measures.
Project 1: Coupling drainage systems and flood models A next step in the development of AI4PDEs for flooding is its application to model the drainage systems under our cities. This project has two stages: (a) establishing a simplified drainage model from a multiphase AI4PDE model; (b) coupling the drainage models with the 2D shallow water equation models both formed using AI4PDEs.
Project 2: Country-scale models Another important development of AI4PDEs for flooding is to be able to model flooding events at the country scale. The project aims to take a step towards country-scale modelling of large scale floods. The model is formed from a patchwork of shallow water equation models in different subdomains and covering the river catchment areas across a large area, for example, over Britain.
Project 3: 3D Flooding models 3D Flooding modelling enabling interaction with the atmosphere, underground drainage and rapid flows (e.g. rivers) to be modelled accurately.
For these projects, an interest in several of the following would be beneficial: computational fluid dynamics, neural networks, flooding and numerical methods. The coding will be done in PyTorch and Python.
[1] Chen, Nadimy, Heaney et al. (2024) Solving the Discretised Shallow Water Equations Using Neural Networks, Advances in Water Resources, accepted. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4956116
[2] Chen, Heaney, Pain (2024) Using AI libraries for Incompressible Computational Fluid Dynamics, https://arxiv.org/abs/2402.17913.
[3] Chen, Heaney, Gomes, Matar, Pain (2024) Solving the Discretised Multiphase Flow Equations with Interface Capturing on Structured Grids Using Machine Learning Libraries, Computer Methods in Applied Mechanics and Engineering 426: 116974. https://doi.org/10.1016/j.cma.2024.116974
Solving PDEs on unstructured and distorted meshes using Rational ConvFEM and AI libraries#
Project code: chpa-222
Main supervisor: Christopher Pain,
c.pain@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Boyang Chen,
boyang.chen16@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
Recent advances in hardware architectures, computing power, and AI software systems are fundamentally changing how partial differential equations (PDEs) can be discretised and solved. Novel AI-centric computing platforms, such as the recently released Cerebras AI computer, offer unprecedented opportunities for large-scale scientific computing. When combined with modern AI software frameworks, these platforms enable new numerical methods that go beyond the limitations of traditional discretisation techniques.
Conventional finite element methods (FEM), while widely used, suffer from several fundamental limitations: they rely on low-order polynomial basis functions, struggle with geometric fidelity on distorted meshes, and can only approximate curved geometries such as spheres and other conic sections. These limitations introduce geometric and numerical errors that propagate into the solution, particularly for nonlinear PDEs and multi-physics problems.
This project explores a next-generation discretisation framework based on Rational Convolutional Finite Element Methods (Rational ConvFEM). These methods employ rational basis representations that are inherently nonlinear, substantially more accurate than classical FEM, and capable of exactly representing conic sections and curved geometries, including spheres, cylinders, and ellipsoids. By embedding Rational ConvFEM formulations directly within AI software libraries, we aim to create a powerful, scalable, and flexible approach for solving PDEs on unstructured and highly distorted meshes.
This approach delivers several key advances including an exact geometric representation; a nonlinear, high-fidelity discretisation; an AI-native implementation; and simplified software development. The complexity of traditional FEM codebases is dramatically reduced, lowering the barrier to developing advanced CFD and multi-physics solvers. Building on recent success with AI-based solvers on structured grids [1,2], this project focuses on extending these ideas to unstructured and distorted meshes, where Rational ConvFEM offers decisive advantages over polynomial FEM.
Possible projects include 1. Rational ConvFEM multi-grid solvers Development of multigrid methods tailored to Rational ConvFEM discretisations for the efficient solution of large nonlinear systems of equations on unstructured meshes. 2. Automatic adjoint sensitivities via AI backpropagation Exploiting the differentiable nature of AI frameworks to automatically construct adjoint models for Rational ConvFEM solvers, enabling optimisation, data assimilation, and rigorous error estimation. 3. Parallelisation on AI-centric hardware Designing and optimising Rational ConvFEM solvers for execution on massively parallel AI architectures, including GPU-systems. 4. Rational ConvFEM-based CFD solver Construction of a high-fidelity CFD solver using Rational ConvFEM discretisations, capable of accurately resolving complex geometries and nonlinear flow physics on distorted meshes. 5. Exact geometry representations for spherical and curved domains Application of Rational ConvFEM to problems requiring exact spherical or conic geometries, such as global-scale atmosphere and ocean modelling, without the geometric errors inherent in traditional FEM. 6. Mesh movement and deformation Development of Rational ConvFEM formulations for moving and deforming meshes, enabling accurate simulation of fluid–structure interaction and evolving geometries while preserving exact geometric fidelity.
[1] Chen, Heaney, Pain (2024) Using AI libraries for Incompressible Computational Fluid Dynamics, https://arxiv.org/abs/2402.17913.
[2] Chen, Heaney, Gomes, Matar, Pain (2024) Solving the Discretised Multiphase Flow Equations with Interface Capturing on Structured Grids Using Machine Learning Libraries, Computer Methods in Applied Mechanics and Engineering 426: 116974. https://doi.org/10.1016/j.cma.2024.116974
Foundational AI Surrogate model for Geothermal modeling (ATES)#
Project code: chpa-129
Main supervisor: Christopher Pain,
c.pain@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Yueyan Li,
yl222@ic.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
SCALED is a foundational surrogate model for computational physics that is scalable, grid-invariant, and geometry-invariant. It has already been successfully applied in domains such as urban flow modeling, particle simulations, and wildfire dynamics. In this project, we focus on geothermal systems, with a particular emphasis on ATES (Aquifer Thermal Energy Storage). The goal is to apply the SCALED surrogate framework to geothermal modeling in order to enable fast inference and seamless integration with digital twin systems. This project aims to explore the application of a foundational AI surrogate model to geothermal simulations, enabling efficient prediction of complex subsurface thermal and flow processes while maintaining robustness across varying geometries and spatial discretizations. Students are strongly encouraged to contact us in advance to discuss project scope and expectations before selecting this project. 📧 Contact email: yl222@ic.ac.uk
Neural Physics and AI World Models for Coupled Solid–Fluid Mechanics in 3D Environments#
Project code: chpa-232
Main supervisor: Christopher Pain,
c.pain@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor is not yet assigned.
Available to: ACSE EDSML
This project may accept multiple students.
Five MSc projects explores the integration of solid mechanics and fluid mechanics using Neural Physics models, combined with AI-based 3D geometry representations and AI world models, to create next-generation simulation and digital twin technologies. Traditional numerical methods for coupled solid–fluid systems (e.g. aeroelasticity, structural deformation under flow, urban wind–structure interaction) are computationally expensive and difficult to scale. Neural Physics offers an alternative by representing governing Partial Differential Equations (PDEs) using neural networks whose weights are analytically defined from physical laws, enabling fast, stable, and interpretable solvers.
The project will investigate how these Neural Physics models can be embedded within AI world models (i.e. learned representations of dynamic physical environments that evolve in time). Advanced 3D geometry representations (e.g. implicit neural fields, signed distance functions, neural meshes) will be used to represent complex, deforming geometries and enable automatic coupling between solids, fluids, and boundaries.
Students will develop modular AI systems capable of simulating physical interactions, predicting system evolution, and supporting design and decision-making in complex environments. The focus will be on predicting fluids flows, moving solid-geometries subject to physical laws (PDEs) within AI geometry generated worlds. Applications include digital twins, robotics, climate-resilient infrastructure, and engineering design.
Project Options Option 1: Neural Physics for Solid Mechanics Develop analytically weighted neural networks for elastic or viscoelastic solid deformation in 3D. Option 2: Neural Physics for Fluid Mechanics Implement neural solvers for incompressible or compressible flow, focusing on stability and scalability. Option 3: Coupled Fluid–Structure Interaction Integrate solid and fluid Neural Physics models to simulate two-way coupling using shared neural geometry. Option 4: AI-Based 3D Geometry Representation Develop neural implicit representations for complex and evolving geometries interacting with physics models. Option 5: AI World Models for Physical Systems Construct AI world models that learn system dynamics from Neural Physics outputs for fast prediction and control.
The project requires strong Python and PyTorch skills and an interest in scientific AI and computational mechanics.
Producing optimal adaptive meshes for finite element methods using convolutional generative Neural Networks#
Project code: chpa-150
Main supervisor: Christopher Pain,
c.pain@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Boyang Chen,
boyang.chen16@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project does not accept multiple students.
To model fluid flow problems that are turbulent and/or have complex geometries or interfaces, meshes that change their resolution can be an efficient way to obtain accurate results (known as mesh adaptivity or mesh optimisation) [1,2,3]. In order to determine where the high and low areas of resolution should be, an error measure is formed, sometimes based on the gradient of solution variables. We use mesh optimisation in order to optimise the size and shape quality of every tetrahedral element in the finite element mesh in order to meet the demands of the error measure e.g. achieve a 1% accuracy in a solution variable. Here a convolutional autoencoder is proposed to form a new approach to adapt a mesh optimally in response to the error measure.
Convolutional methods have been used to represent differing material properties [4]. If these materials have convex shapes (e.g. grains in a sand stone), then the centres of the different materials can represent nodes of a finite element tetrahedral mesh. If different materials are next to one another then we assume there is an edge (in the tetrahedral mesh) between the associated nodes. In this way a tetrahedral mesh can be formed in 3D. Moreover, if the convolutional methods are able to gauge the size and shape quality of the tetrahedral elements then these can have their sizes and shapes optimised thus producing an optimal mesh.
For this project an interest in some of the following would be beneficial: computational fluid dynamics, numerical methods and neural networks. The coding will be done mostly in PyTorch and Python.
[1] Pain, Umpleby, de Oliveira, Goddard (2001) Tetrahedral mesh optimisation and adaptivity for steady-state and transient finite element calculations, Computer Methods in Applied Mechanics and Engineering, 190(29):3771-3796.
[2] Kampitsis, Adam, Salinas, Pain, Muggeridge, Jackson (2020) Dynamic adaptive mesh optimisation for immiscible viscous fingering, Computational Geosciences, 24:1221-1237.
[3] Salinas, Regnier, Jacquemyn, Pain, Jackson (2021) Dynamic mesh optimisation for geothermal reservoir modelling, Geothermics, 94:102089
[4] Gayon-Lombardo, Mosser, Brandon, Cooper (2020) Pores for thought: generative adversarial networks for stochastic reconstruction of 3D multiphase electrode microstructures with periodic boundaries. Computational Materials, 6(1):82.
Applying convolutional autoencoders to data on unstructured meshes#
Project code: chpa-149
Main supervisor: Christopher Pain,
c.pain@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Boyang Chen,
boyang.chen16@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
Autoencoders have been in use for over 20 years [1,2], and are neural networks which aim to learn the identity map through an architecture that has a central `bottleneck’ layer (with a dimension lower than that of the input and output layers). Currently, they are used for their ability (i) to generate realistic-looking images (variational AEs); (ii) to extract noise from signals (denoising AEs); (iii) and in reduced-order modelling as a tool for dimensionality reduction. In the latter, dimensionality reduction (DR) methods aim to extract features from data (often solutions to PDEs, for example velocity or vorticity fields). One such DR method is Proper Orthogonal Decomposition (also known as Principal Component Analysis), which is based on Singular Value Decomposition (SVD) and a linear combination of basis functions. Autoencoders are a natural extension to the SVD, with the ability to better capture features in the flow due to the nonlinear activation functions.
Data from fluid flow simulations is often stored on unstructured, adapted meshes (as this can be a more efficient way of modelling these problems). However, many numerical techniques have been developed for structured grids, including convolutional neural networks. In order to apply convolutional autoencoders (CAE) to data on unstructured meshes, space-filling curves (SFC) can be used to to reorder the data before application of the autoencoder (SFC-CAE) [3]. A number of projects are now proposed to extend the SFC-CAE, including (i) applying SFC-CAE to unstructured, adapted meshes; (ii) finding optimal mappings between unstructured and structured meshes based on the earth mover’s distance (the Wasserstein metric).
For these projects, an interest in several of the following would be beneficial: numerical modelling, computational fluid dynamics and neural networks.
[1] Bourlard, Kamp (1988) Auto-association by multilayer perceptrons and singular value decomposition, Biological Cybernetics 59:291-294.
[2] Brunton, Noack, Koumoutsakos (2020) Machine Learning for Fluid Mechanics Annual Review of Fluid Mechanics, 52(1):477-508.
[3] Heaney, Li, Matar, Pain (2021) Applying Convolutional Neural Networks to Data on Unstructured Meshes with Space-Filling Curves, arxiv preprint.
Urban World Geometry Generation with 3D Generative Models#
Project code: chpa-127
Main supervisor: Christopher Pain,
c.pain@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Yueyan Li,
yl222@ic.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
This project aims to leverage remote sensing imagery together with urban foundational information, and apply 3D generative models and 3D scene reconstruction techniques to generate high-fidelity, scalable 3D assets of urban environments. The generated assets include urban buildings, vegetation, and critical infrastructure. The project would suit students interested in machine learning and computer graphics. It will involve foundational theories and methodologies related to graphics and 3D generation, and external supervisors with expertise in computer graphics will be invited to participate in supervision. The project will also include GIS-related data processing and spatial information handling. Relevant works include:
Skyfall-GS Skyfall-GS: Synthesizing Immersive 3D Urban Scenes from Satellite Imagery
UrbanGen UrbanGen
SAM3D SAM 3D: 3Dfy Anything in Images Upon completion, the generated urban geometry assets will be integrated with our Neural Physics frameworks and Foundational AI Surrogate Models to support downstream physics-based simulations. Students are strongly encouraged to contact us in advance to discuss project scope and expectations before selecting this project. 📧 Contact email: yl222@ic.ac.uk
Compressing particles and converting particles to a continuum using AI4PDEs, AI4Particles and Convolutional Autoencoders#
Project code: chpa-147
Main supervisor: Christopher Pain,
c.pain@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Boyang Chen,
boyang.chen16@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
Recently AI4PDEs [1,2,3,4] has shown great success in being able to solve very large fluid systems in great detail. Also, the AI4Particle modelling method has been able to describe particle motion in turbulent flows. AI4Particles can also be used as a Discrete Element Method (DEM) [5] and as a molecular dynamics model. AI4PDEs and AI4Particles offer potentially revolutionary advantages over conventional modelling methods. However, the compression of cloud of particles is an unexplored area as is the conversion of a cloud of particles to a continuum description, which could be more accurate. Here we will use the outputs of these two modelling approaches in order to attempt to compress the velocity and position of particles and also relate particle descriptions of a concentration field (e.g. air pollution) to a continuum description.
AI4PDEs is an in-house computational fluid dynamics (CFD) solver, which solves discretised systems of equations using neural networks. The weights of the networks are determined in advance through the choice of discretisation method, so no training is needed. The solutions are the same as those obtained from Fortran or C++ codes (to within solver tolerances).
A new type of scale independent convolutional autoencoder will be used within this project. This autoencoder, after training, can be applied to very large systems even when the original training was conducted on much smaller systems.
Two student projects are offered here: (i) particle compression; (ii) converting particles to a continuum.
For these projects, an interest in several of the following would be beneficial: computational fluid dynamics, particle dynamics, neural networks and numerical models. The coding will be done in PyTorch and Python.
[1] Chen, Heaney, Pain (2024) Using AI libraries for Incompressible Computational Fluid Dynamics, https://arxiv.org/abs/2402.17913.
[2] Chen, Heaney, Gomes, Matar, Pain (2024) Solving the Discretised Multiphase Flow Equations with Interface Capturing on Structured Grids Using Machine Learning Libraries, Computer Methods in Applied Mechanics and Engineering 426: 116974. https://doi.org/10.1016/j.cma.2024.116974
[3] Chen, Nadimy, Heaney et al. (2024) Solving the Discretised Shallow Water Equations Using Neural Networks, Advances in Water Resources, accepted. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4956116
[4] Phillips, Heaney, Chen, Buchan, Pain (2023) Solving the Discretised Neutron Diffusion Equations Using Neural Networks, International Journal for Numerical Methods in Engineering 124(21):4659-4686. https://doi.org/10.1002/nme.7321
[5] Naderi, Chen, Yang, Xiang, Heaney, Latham, Wang, Pain (2024) A discrete element solution method embedded within a Neural Network, Powder Technology 448: 120258. https://doi.org/10.1016/j.powtec.2024.120258
Developing AI based Sub-Grid-Scale Models for Particles moving in a Fluid using AI4PDEs, AI4Particles and Scale Independent Convolutional Autoencoders#
Project code: chpa-148
Main supervisor: Christopher Pain,
c.pain@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Boyang Chen,
boyang.chen16@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project does not accept multiple students.
Recently AI4PDEs [1,2,3] has shown great success in being able to solve very large fluid systems in great detail. Also, the AI4Particle modelling method has been able to describe particle motion in turbulent flows. AI4Particles can also be used as a Discrete Element Method (DEM) and as a molecular dynamics model. AI4PDEs and AI4Particles offer potentially revolutionary advantages over conventional modelling methods. However, there is still a need to incorporate, error prone, correlations for the drag forces between the particles and the fluid flow. This project attempts to overcome this drawback by using a Sub-Grid-Scale Model (SGS) of the fluid flow around the particles and thus resolve the drag forces. A new breed of scale independent convolutional autoencoders will be used to resolve the difference between the course grid AI4PDE model and a resolved model of the flow around the particles. This autoencoder, after training, can be applied to very large systems even when the original training was conducted on much small systems.
AI4PDEs is an in-house computational fluid dynamics (CFD) solver, which solves discretised systems of equations using neural networks. The weights of the networks are determined in advance through the choice of discretisation method (e.g. first-order finite elements, second-order finite differences etc), so no training is needed. The solutions are the same as those obtained from Fortran or C++ codes (to within solver tolerances). The AI4Particles model is similar in its architecture to AI4PDEs but resolves the interaction and motion of particles using convolutional neural networks. The AI4PDEs and AI4Particles codes run on GPUs as well as CPUs and AI processors.
The student(s) will use CFD data and build AI models in Python and PyTorch. For this project, an interest in several of the following would be beneficial: computational fluid dynamics, particle dynamics, neural networks and numerical models. The coding will be done in PyTorch and Python.
[1] Chen, Nadimy, Heaney et al. (2024) Solving the Discretised Shallow Water Equations Using Neural Networks, Advances in Water Resources, accepted. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4956116
[2] Chen, Heaney, Pain (2024) Using AI libraries for Incompressible Computational Fluid Dynamics, https://arxiv.org/abs/2402.17913.
[3] Chen, Heaney, Gomes, Matar, Pain (2024) Solving the Discretised Multiphase Flow Equations with Interface Capturing on Structured Grids Using Machine Learning Libraries, Computer Methods in Applied Mechanics and Engineering 426: 116974. https://doi.org/10.1016/j.cma.2024.116974
Incorporating automatic code generation to solve PDEs using neural networks#
Project code: chpa-146
Main supervisor: Christopher Pain,
c.pain@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Boyang Chen,
boyang.chen16@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project does not accept multiple students.
AI4PDEs [1,2,3] is an in-house computational fluid dynamics (CFD) solver, which solves discretised systems of equations using neural networks. The weights of the networks are determined in advance through the choice of discretisation method (e.g. first-order finite elements, second-order finite differences etc), so no training is needed. The solutions are the same as those obtained from Fortran or C++ codes (to within solver tolerances). The code runs on GPUs as well as CPUs and AI processors.
In this project the aim is to solve a general system of discretised differential equations. We will do this using SymPy (used within Devito) for the interface into which the user types their desired differential equations. The resulting sets of set of coupled equations are solved using a multi-grid method that is applied using a U-Net neural network architecture. The focus of the work will be on how to solve efficiently the resulting linear system of equations and maintaining the advantages of neural networks (e.g. the ability to run on CPUs or GPUs).
In this project the aim is to solve a general system of discretised differential equations. We will do this using SymPy (used within Modulus and Devito) for the interface that generates the differential equations and the resulting sets of set of coupled equations are solved using a multi-grid method that is applied using a U-Net architecture. The focus of the work will be on how to solve efficiently the resulting linear system of equations and maintaining the advantages of neural networks (e.g. the ability to run on CPUs or GPUs).
For this project, an interest in the following would be beneficial: computational fluid dynamics, neural networks and symbolic maths. During this project the student will use the SymPy, Modulus and PyTorch packages. On successful completion of the project, the solution of even complex PDEs will be made relatively easy impacting across computational physics. There might be the possibility to link with NVIDIA engineers on PhysicsNeMo development.
[1] Chen, Heaney, Pain (2024) Using AI libraries for Incompressible Computational Fluid Dynamics, https://arxiv.org/abs/2402.17913.
[2] Chen, Heaney, Gomes, Matar, Pain (2024) Solving the Discretised Multiphase Flow Equations with Interface Capturing on Structured Grids Using Machine Learning Libraries, Computer Methods in Applied Mechanics and Engineering 426: 116974. https://doi.org/10.1016/j.cma.2024.116974
[3] Chen, Nadimy, Heaney et al. (2024) Solving the Discretised Shallow Water Equations Using Neural Networks, Advances in Water Resources, accepted. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4956116
Foundational AI Surrogate Model for Weather and Climate Modeling#
Project code: chpa-130
Main supervisor: Christopher Pain,
c.pain@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Yueyan Li,
yl222@ic.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
SCALED is a foundational surrogate model for computational physics that is scalable, grid-invariant, and geometry-invariant, and has demonstrated strong performance in domains such as urban flow modeling, particle simulations, and wildfire dynamics. This project aims to extend the SCALED framework to weather and atmospheric modeling, with a focus on learning surrogate representations of complex, high-dimensional, and multiscale atmospheric processes. The goal is to enable fast, robust, and resolution-independent inference for weather prediction tasks, complementing or augmenting traditional numerical weather prediction (NWP) models. Key research directions include:
AI surrogate modeling of atmospheric dynamics (e.g., wind fields, temperature, pressure, humidity)
Learning models that generalize across spatial resolutions, grids, and geometries
Integration with global or regional weather datasets, including reanalysis and simulation data
Exploring connections with neural operators, graph-based models, and foundation weather models The project emphasizes physics-aware learning, scalability, and generalization, and is particularly relevant to students interested in AI for climate and weather, geophysical fluid dynamics, and large-scale scientific foundation models. Upon completion, the developed surrogate models may be integrated into broader digital twin systems or used to support downstream environmental and climate impact assessments. Students are strongly encouraged to contact us in advance to discuss project scope and expectations before selecting this project. 📧 Contact email: yl222@ic.ac.uk
Short term weather forecasting in the presence of forward model error#
Project code: jape-229
Main supervisor: James Percival,
j.percival@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor is not yet assigned.
Available to: ACSE EDSML
This project may accept multiple students.
In this project you will investigate techniques to improve predictive skill in the case of forecasts and analyses which exhibit significant model biases. This could be either in highly simplified models of the atmosphere such as quasi-geostrophic models, or (data and time permitting) more realistic and complete models such as WRF (https://www.mmm.ucar.edu/models/wrf), and through machine learning or traditional methods.
If you are interested in a project on this theme, you are very strongly encouraged to discuss the topic with me before making your decisions on which projects to request, please feel free to email me to find a time to discuss in more detail.
Machine learning for subgridscale parameterization using simple models#
Project code: jape-100
Main supervisor: James Percival,
j.percival@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor is not yet assigned.
Available to: ACSE EDSML
This project may accept multiple students.
Parameterization is a process in numerical modelling in which the effects of inputs present in the real system which cannot be appropriately represented in the modelled scales (for example sub-gridscale turbulence) are modelled in terms of the prognostic variables the model does have. (for example the large scale winds and temperature patterns. Whether and how Machine Learning models can be used to generate new and better parameterizations is a topic of ongoing research. In this project you will investigate the problem in simplified systems (such as the the two tier Lorenz 96 model, which can be run quickly to generate both true and model states, as well as observation data.
See Parthipan et al. (https://gmd.copernicus.org/articles/16/4501/2023/) for an example study.
Computational Fluid Dynamics projects using OpenFOAM#
Project code: jape-099
Main supervisor: James Percival,
j.percival@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor is not yet assigned.
Available to: ACSE EDSML
This project may accept multiple students.
OpenFOAM (https://openfoam.org/) is an open source, user-extendable C++ based solver for Computational Fluid Dynamics (CFD), solving variations of the Navier-Stokes equations governing viscous fluid flow. It comes with solvers for a number of classes of problem:
Steady state/time averaged
Rotating fluids
Moving mesh problems plus a number of others.
Depending on the interest of the student, you’ll run the code on and analyse the results you obtain. The precise terms of the project are fairly flexible, based on the student’s interests within the overall topic and other data sets which become available. If you are interested in the project, you are strongly encouraged to discuss the topic with me before making your decisions on which projects to request, please feel free to email me to find a time to discuss.
Some References
Numerical investigation of particle lateral migration in straight channel flows using a direct-forcing immersed boundary method (https://doi.org/10.1016/j.jfluidstructs.2020.103110)
Turbulence modeling in OpenFOAM:Theory and applications (http://www.wolfdynamics.com/training/turbulence/OF2021/turbulence_2021_OF8.pdf)
Comprehensive segmentation of Magellan Synthetic Aperture Radar images on Venus#
Project code: gepe-060
Main supervisor: Gerard Gallardo I Peres,
g.g.peres@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Philippa Mason,
p.j.mason@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project does not accept multiple students.
In the context of the forthcoming EnVision and VERITAS SAR missions to Venus, we are in need of tools to automatically detect and classify surface units that might be characteristic of geological units on the surface of the planet. Supervised, geological mapping of surface units has been extensively pursued using the Magellan dataset SAR data, but these units might not align with alternative classifications constructed purely from information contained in radar images such as mean backscatter or texture. The objective of this project is therefore to explore the provided set of Magellan backscatter images and devise a strategy to segment them exploiting only the information present in the images (i.e. without geological context). Additionally, geological masks adapted from co-registered with the images will be provided in order to compare segmentation outputs with the supervised geological mapping results and provide a means for cross-validation. In this project, you will:
• Understand SAR backscatter image characteristics, including the concept of speckle noise and radar texture, and how can they be mathematically defined in the context of SAR image segmentation.
• Explore segmentation strategies that can be applied to SAR data of geological terrain on Venus. Potential strategies include threshold-based, region-based, edge-based or cluster-based segmentation algorithms, but also Machine Learning or Deep Learning approaches. A set of potential algorithms will be proposed, but you can choose your own preferred method for the task.
• Implement a code pipeline (Python) for the chosen segmentation strategy. The only requirement is that the algorithm is able to classify terrain units into a finite set of classes, for the given Magellan SAR test images.
• Assess the results of the algorithm by means of a performance analysis tailored to the intrinsic functioning of your algorithm of choice. Furthermore, compare the results with the provided geological mapping masks.
Developing a Data Processing Pipeline for FTIR data#
Project code: elph-118
Main supervisor: Elizabeth Phillips,
elizabeth.phillips@imperial.ac.uk, The Grantham Institute for Climate Change, Imperial College LondonSecond supervisor is not yet assigned.
Available to: ACSE EDSML GEMS READY
This project may accept multiple students.
Researchers who use Fourier-transform infrared (FTIR) spectroscopy spend a significant amount of time processing data – including water vapour subtraction and baseline correction. Furthermore, this process can be highly subjective, impacting the accuracy of interpretations using FTIR data. Baseline subtraction can significantly impact the interpretations made using FTIR data.
An open-source Python package (ProSpecPy) has been initiated to replicate, as much as possible, manual processing of FTIR data that significantly decreases processing time. In this MSc project, the accuracy of ProSpecPy will be tested against other existing baseline subtraction and manual methods. The results of this testing will be used to substantially improve the baseline subtraction method, both in terms of time required and accuracy of interpretations. Existing FTIR data to study reaction mechanisms of hydrogenases, enzymes that reversibly oxidise hydrogen and serve as inspiration for catalyst development for hydrogen fuel cells and electrolysers, will be used to test and calibrate the data processing pipeline. Additional FTIR spectra will be used to ensure the method is generalisable to other systems.
The output will be an open-source pipeline for data processing from raw data to baseline-subtracted spectra with processing parameters that can easily be compared between users. A further aim of this study is to initiate an effort for a community-led contributions to develop best practices in processing and reporting FTIR data, including but not limited to; best practices in collecting FTIR data (e.g., water vapour measurements, blanks), methods for water vapour subtraction, and methods to determine and report uncertainty in FTIR data.
ICL x Google Building a Super-Labelled Validation Dataset for Geospatial Foundation Models: Fusing Drone and Satellite Modalities#
Project code: mapi-218
Main supervisor: Matthew Piggott,
m.d.piggott@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Raul Adriaensen,
raul.adriaensen17@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project does not accept multiple students.
Part of the ICL x Google Benchmark Island Dataset collaboration. This project will contribute to creating a multi-modal ground-truth dataset purpose-built for validating and fine-tuning Geospatial Foundation Models (GeoFMs). Current FM benchmarks rely on satellite-only labels that lack the spatial precision for rigorous evaluation. By fusing sub-decimeter drone imagery, 3D photogrammetric reconstructions, LiDAR point clouds, multispectral data, and field-verified annotations, this project will produce a “super-labelled” dataset that sets a new standard for FM evaluation. A successful project will contribute to publications and deliver a dataset of lasting scientific value.
The Challenge & Innovation: • Multi-Modal Fusion Pipeline: Workflows to co-register and fuse drone orthomosaics (RGB + multispectral), digital surface models, 3D point clouds, and satellite imagery into a unified geospatial framework. • Annotation & Label Generation: Systematic protocols to produce fine-grained semantic labels (coral rubble vs. reclaimed sand, seagrass density gradients, building typologies) that cannot be distinguished from satellite alone. • 3D-to-2D Label Transfer: Methods to project rich 3D labels (elevation, surface roughness, vegetation height) onto 2D satellite views for FM training and validation. • FM Benchmarking: Evaluating GeoFMs against super-labels vs. conventional labels to expose systematic failure modes that standard benchmarks miss.
Data & Resources: • Field survey data from multiple Maldives island sites: drone imagery (RGB + multispectral), LiDAR, 3D models, and field annotations. • Satellite imagery stack (PlanetScope, Sentinel-2). OpenDroneMap/QGIS processing infrastructure. • Multiple related projects under the same collaboration provide a built-in research community. • Possibly google FMs
Impact Fills a critical gap in the geospatial AI community: high-fidelity, multi-modal ground truth for FM validation. Directly contributes to the Google-ICL Benchmark Island Dataset and provides a lasting open-science resource for AI-driven environmental monitoring in Small Island Developing States. Skill development: Deep Learning, Photogrammetry (OpenDroneMap/WebODM), LiDAR processing, 3D point cloud analysis, GIS (QGIS), multi-modal data fusion, foundation model evaluation, dataset curation.
ICL x Google Scaling Geospatial Foundation Models: Cross-Resolution Validation from Satellite to centimetre Drone Imagery#
Project code: mapi-219
Main supervisor: Matthew Piggott,
m.d.piggott@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Raul Adriaensen,
raul.adriaensen17@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project does not accept multiple students.
Part of the ICL x Google Benchmark Island Dataset collaboration. This project investigates how Geospatial Foundation Models (GeoFMs), typically pre-trained on 10–30m satellite imagery, perform when applied across a resolution spectrum spanning three orders of magnitude down to 2–5cm drone imagery. Using a unique multi-modal dataset from the Maldives, the project will quantify cross-resolution generalisation and develop strategies to bridge the domain gap. A successful project will contribute to publications and directly inform the Google-ICL research programme.
The Challenge & Innovation: • Cross-Resolution Benchmarking: Systematic performance profiling of publicly available GeoFMs across a resolution ladder from Sentinel-2 (10m) through PlanetScope (3m) to drone orthomosaics (2–5cm), with potential access to Google in-house FM models. • Context-Aware Approaches: Developing metadata embedding strategies and hierarchical multi-scale encoding architectures to anchor high-resolution patches within their geographic context. • Adaptation Strategies: Exploring efficient fine-tuning approaches for cross-resolution information transfer. Data & Resources: • Benchmark Island Dataset: drone orthomosaics, 3D point clouds, and concurrent satellite imagery from the Maldives. • Multi-resolution satellite stack (Sentinel-2, PlanetScope) covering identical locations. • Possibly google FMs
Impact Addresses a fundamental open question in geospatial AI with direct relevance to disaster response, urban planning, and environmental monitoring in data-scarce regions. Findings will inform the wider Google-ICL research programme and the future development of foundation models for Earth observation.
Skill development: Foundation models, Vision Transformers, fine-tuning (LoRA, adapters), multi-scale representation learning, satellite and drone remote sensing, Deep Learning.
ICL x UNEP x GFW PlumeWatch: Deep Learning for Automated Detection and Characterisation of Dredging-Induced Sediment Plumes#
Project code: mapi-217
Main supervisor: Matthew Piggott,
m.d.piggott@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Raul Adriaensen,
raul.adriaensen17@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE
This project does not accept multiple students.
Description A collaboration between Imperial College London and UNEP Grid-Geneva targeting a critical gap in the Marine Sand-watch initiative. The platform currently tracks where dredging occurs via vessel AIS but has no capability to observe the resulting environmental impact. This project will build the first automated plume detection pipeline and global database of dredging-induced sediment plumes — no such database or labelled training dataset currently exists in the literature. A successful project will contribute to a high-impact publication.
The Challenge & Innovation: • Plume Classification: Deep learning models (CNNs, U-Nets, Vision Transformers) to distinguish dredging plumes from algal blooms, cloud shadows, shallow bathymetry, and natural turbidity in satellite imagery. Reducing high false-positive rates in existing heuristic approaches. • Plume Characterisation: Quantitative extraction of plume attributes including spatial extent, dispersion direction, optical density gradients, and temporal persistence. • Global Database Generation: Systematic, queryable database of classified plume events with associated metadata, linked to dredging vessel activity via AIS. • Temporal Back-testing: Retrospective classification across multi-year satellite archives for trend analysis and long-term impact assessment. Data & Resources: • Exclusive access to multi-year global AIS datasets and the world’s first extensively labelled dredging behaviour library. • Planet 3m satellite imagery; in-house labelled plume dataset (the only such collection known to exist) • Deployment pathway via Global Fishing Watch and Marine Sand-watch platforms.
Impact Creates a complete cause-and-effect monitoring chain linking plume observations to vessel tracking — something no existing platform provides. Extensive downstream applications in sediment transport modelling, coral reef impact assessment, and international environmental policy.
Key skills: Deep learning, computer vision, satellite remote sensing, geospatial database design, environmental data science, large-scale deployment.
ICL x UN-EP x GFW Deep Learning for Global Dredging Classification: Multi-Modal Fusion of AIS and Satellite Imagery#
Project code: mapi-216
Main supervisor: Matthew Piggott,
m.d.piggott@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Raul Adriaensen,
raul.adriaensen17@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE
This project does not accept multiple students.
A collaboration between Imperial College London, UNEP Grid-Geneva and Global fishing Watch to advance the Marine Sand-watch initiative. This project will transition the current heuristic dredging classification pipeline to a high-performing deep learning framework for real-time, planetary-scale surveillance of maritime sand extraction. A successful project will contribute to a high-impact publication.
The Challenge & Innovation: • Multi-Modal Data Fusion: Harmonising global AIS time-series with Planet 3m satellite imagery for dual-verification classification. • Advanced Architectures: Exploration of Transformers, LSTMs, and CNNs optimised for spatiotemporal maritime data, with temporal back-testing for validation. • Deployment (Agentic AI ): Development of an autonomous deployment wrapper targeting continuous, real-time global updates. Can be trialled with Agentic workflows Data & Resources: • Exclusive access to multi-year global AIS datasets and the world’s first extensively labelled dredging behaviour library. • Planet 3m satellite imagery for ground-truthing and feature extraction. • Direct deployment pathway via Global Fishing Watch and Marine Sand-watch platforms.
Impact Opportunity to co-author a high-impact publication with UNEP and Imperial. A successful model will be deployed as an operational global environmental monitoring tool under international environmental policy frameworks.
Skill development: Deep learning, time-series classification, satellite remote sensing, multi-modal data fusion, AIS maritime data, model deployment, agentic AI.
ICL x Google Pushing the Limits: Resolution Sensitivity Analysis of Drone Photogrammetry for Small Island Mapping#
Project code: mapi-220
Main supervisor: Matthew Piggott,
m.d.piggott@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Raul Adriaensen,
raul.adriaensen17@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project does not accept multiple students.
Part of the ICL x Google Benchmark Island Dataset collaboration. This project investigates the practical limits of drone-based photogrammetric reconstruction for environmental mapping of small islands. Modern SfM/MVS pipelines can produce centimetre-level outputs, but the relationship between processing parameters, output resolution, and downstream analytical value is poorly understood. This project will systematically quantify what is gained by pushing resolution limits, and critically compare orthomosaics against raw imagery as inputs for AI classification. A successful project will produce practical guidelines for all future field campaigns and contribute to scientific publications.
The Challenge & Innovation: • Resolution Ladder Experiments: Systematic variation of photogrammetric parameters (GSD, overlap, point cloud density, mesh resolution) and quantification of impact on geometric accuracy and thematic content. • Mosaics vs. Raw Imagery: Comparative evaluation of AI classification performance using stitched orthomosaics (with blending/distortion artefacts) versus raw drone frames (with preserved radiometric fidelity). • Cost-Benefit Analysis: Profiling computational requirements across the resolution spectrum against measurable analytical gains to establish operational “sweet spots”. • Downstream Task Sensitivity: Evaluating resolution impact on coastline delineation, building extraction, vegetation classification, and change detection.
Data & Resources: • Raw drone imagery and processed outputs from Maldives island sites (DJI Mavic 3M, RGB + multispectral). • Established WebODM/OpenDroneMap processing infrastructure with guidance. • High-resolution Planet satellite imagery for cross-scale comparison.
Impact Delivers evidence-based processing guidelines directly informing all future Google-ICL field campaigns. Findings are broadly valuable to the drone remote sensing community, particularly in resource-constrained environments where optimising data quality against operational cost is critical. Skill development: Drone photogrammetry (SfM/MVS), OpenDroneMap/WebODM, 3D reconstruction, remote sensing, geospatial AI, experimental design, computational performance profiling.
Physics-Guided Deep Optical Flow for Large-Scale Particle Image Velocimetry via Test-Time Optimization#
Project code: mapi-250
Main supervisor: Matthew Piggott,
m.d.piggott@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Zhi Wang,
zhi_wang@zju.edu.cn, Zhejiang University, ChinaAvailable to: ACSE EDSML
This project may accept multiple students.
Deep learning (DL) methods in standard particle image velocimetry (PIV) often suffer from domain shift when applied to large-scale PIV (LS-PIV) cases due to complex lighting and sparse tracers. To address this challenge, this project proposes a physics-guided DL framework employing test-time optimization to bridge the gap between laboratory data and real-world (e.g. river) environments. Building on pre-trained optical flow networks (e.g. RAFT), the model aims to adapt to specific scenes by minimizing physics-informed losses, such as mass conservation and smoothness constraints. Furthermore, uncertainty quantification will be incorporated to evaluate prediction reliability, ensuring robust flow reconstruction in complex conditions.
This project is particularly suitable for students with an interest in (environmental) fluid dynamics.
Tailored, Trustworthy LLM-Driven Synthesis and Generation of Mathematical and Algorithmic Training Material#
Project code: mapi-202
Main supervisor: Matthew Piggott,
m.d.piggott@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Rhodri Nelson,
rhodri.nelson@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE
This project may accept multiple students.
This project will focus on the development of a scalable pedagogical framework for teaching material at the intersection of mathematical theory and algorithmic implementation. The objective is to bridge the gap between abstract theoretical frameworks and their practical computational execution for students entering from diverse academic backgrounds. By employing a “NotebookLM” inspired architecture, the system should enforce trustworthy content generation by grounding the LLM’s outputs in a curated corpus of legacy Jupyter notebooks, treating institutional material as the definitive “source of truth” to prevent mathematical drift, inconsistencies or algorithmic errors. Recognizing that students arrive with different foundational degrees, the system should adapt the narrative of the material, mapping core mathematical concepts onto application domains familiar to the student (e.g., physical sciences, engineering, or finance). By synthesizing rigorous theory with adaptive, implementation-focused instruction, this project aims to research the potential for AI based personalised teaching assistants and content creators that ensure students can best master the complex interplay between mathematical abstraction and computational reality.
Agentic AI for Computational Fluid Dynamics/Computational Engineering/Design Engineering#
Project code: mapi-201
Main supervisor: Matthew Piggott,
m.d.piggott@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Rhodri Nelson,
rhodri.nelson@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE
This project may accept multiple students.
This project will consider the development of agentic systems to conduct robust prompt-based computational modelling. This will involve problem decomposition, design of key output variables, model and mesh generation, solution verification (and mesh refinement), and output visualisation and diagnostics to inform human-in-the-loop iteration for science or design engineering tasks. This project will build on a number of recent publications in the areas of Agentic CFD, Autonomous Computational Engineers etc. There is an opportunity for students to work on the overall framework or to concentrate on a specific task/sub-agent.
Farm2Market — A Digital Infrastructure for Data-Driven Smallholder Agriculture#
Project code: yvpl-110
Main supervisor: Yves Plancherel,
y.plancherel@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Jazzie Jao,
jazzie.jao@dlsu.edu.ph, De La Salle University, PhilippinesAvailable to: ACSE EDSML
This project may accept multiple students.
Small-scale farmers produce a substantial share of the world’s food, yet often lack access to digital tools that could help them manage inputs, understand environmental conditions, and connect efficiently to markets. While remote sensing, environmental data, and data analytics are increasingly powerful, these capabilities rarely reach farmers in a usable, low-cost, and context-appropriate form. This project addresses that gap by developing Farm2Market, a prototype digital infrastructure—implemented as a web platform or mobile app—that enables small-scale farmers to digitise their own farm data and benefit directly from data-driven insights.
The core objective is to design and build a working prototype that can be tested with realistic use cases. Students will develop tools that allow farmers to record basic information on farm location, crops, planting dates, yields, fertiliser or pesticide use, and management practices. These user-provided data will be combined with open-access environmental information, including satellite-derived vegetation indices, soil moisture, rainfall, temperature, and drought indicators, to provide simple, actionable analyses and visualisations tailored to farm-scale decision-making.
Additional project components may include computer-vision pipelines for analysing user-uploaded images of crops, enabling applications such as crop health assessment, pest or disease detection, and weed identification. Other tracks may focus on logistics and market access, including route optimisation for produce delivery, co-transport or aggregation options shared among farmers, and exploratory models for short-term price trends in local markets. These components highlight the integration of environmental data, machine learning, and optimisation within a single digital ecosystem.
Multiple MSc students can work on complementary aspects of Farm2Market, such as frontend and backend development, data integration, machine-learning modules, or user experience design. Emphasis is placed on low-cost, scalable, and transparent solutions. The project aims to demonstrate how digital infrastructure can empower farmers, improve resource efficiency, reduce waste, and strengthen links between farms and markets, while providing students with hands-on experience at the interface of data science, remote sensing, and real-world impact.
Machine Learning with High-Resolution Drone Data for Next-Generation Geospatial Models#
Project code: yvpl-107
Main supervisor: Yves Plancherel,
y.plancherel@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Matthew Piggott,
m.d.piggott@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
The rapid development of geospatial foundation models depends not only on algorithmic advances, but critically on the availability of high-quality, high-resolution benchmark datasets that capture real-world structure and heterogeneity. While satellite imagery underpins most current large-scale geospatial models, its spatial resolution often limits the representation of fine-scale ecological and geomorphological processes. Drone-based observations provide a powerful complementary data source: centimetre-scale multispectral, thermal, and LiDAR measurements resolve surface structure, vegetation complexity, and micro-topography that are otherwise invisible, making them uniquely valuable for developing, stress-testing, and benchmarking machine-learning approaches that can later be scaled up or transferred to satellite contexts.
This set of MSc independent projects builds on an existing and diverse collection of drone datasets that have already been acquired across a range of environments. These include, for example, low-lying coral islands in the Maldives, a legacy mining site studied within the Bio+Mine project (https://bioplusmine.earth/), a British winery, and small-scale agricultural and mountainous systems in the Philippines. Together, these sites span contrasting ecological, geomorphological, and socio-environmental settings, providing an unusually rich testbed for generalisation and transfer learning. Students will work with high-resolution multispectral imagery, thermal data, and/or LiDAR products, engaging with the full machine-learning pipeline from preprocessing to evaluation. A key emphasis is that different data modalities require different methodological approaches: 2D image-based methods for multispectral or thermal data; point-based, graph-based, or voxel-based methods for 3D LiDAR point clouds; and surface or mesh-based approaches for photogrammetric reconstructions. Students will explore how these representations affect model design, performance, and transferability, and how multimodal fusion can improve robustness.
Multiple MSc students can work in parallel on complementary tracks, sharing common datasets, preprocessing workflows, and evaluation protocols. Individual projects may focus on semantic segmentation or object detection of ecological or geomorphological features, multimodal learning across spectral, thermal, and structural data, or self-supervised and weakly supervised strategies to reduce reliance on labelled data. Collectively, the work aims to produce reusable datasets, benchmarks, and modular pipelines that contribute both to applied environmental science, change detection analysis, and to the longer-term development of geospatial foundation models.
Remote Sensing and Material Flow Analysis of Illegal Gold Mining and Associated Infrastructures#
Project code: yvpl-109
Main supervisor: Yves Plancherel,
y.plancherel@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Sesinam Dagadu,
s.dagadu@snoocode.com, SnooCODE LimitedAvailable to: ACSE EDSML
This project may accept multiple students.
Illegal and artisanal gold mining is a major global environmental and socio-economic challenge, driving deforestation, land degradation, river pollution, and severe human health impacts. It is also closely linked to organised crime, illicit financial flows, and human exploitation, as highlighted by organisations such as INTERPOL and UNEP. Because activities are often remote, informal, and deliberately hidden, illegal gold mining remains difficult to monitor, regulate, and address using conventional approaches. This project explores how Earth observation and modelling can provide actionable evidence to support environmental protection, crime prevention, and improved livelihoods.
The project aims uses open-access satellite data (optical, multispectral, radar, night-time lights) to analyse the development and impacts of illegal gold mining, with a focus on regions such as Ghana, while allowing for regional or global extensions. Students will work with multi-year to multi-decadal satellite time series to detect, map, and characterise the spatio-temporal evolution of artisanal and small-scale gold mining (ASGM), including site emergence, expansion, relocation, and abandonment.
One track focuses on remote sensing and machine learning: mapping mining activity through change detection and time-series classification; scaling algorithms across large areas and long time periods using cloud platforms such as Google Earth Engine; analysing post-mining recovery and regreening trajectories; and assessing downstream impacts such as sediment plumes, river turbidity, and floodplain disturbance. These analyses can help identify hotspots, track enforcement outcomes, and evaluate environmental recovery.
A complementary track focuses on adapting Bayesian material flow analysis (MFA) methods, a fundamentally different approach that targets the economic and industrial systems behind illegal mining and associated infrastructures. By combining satellite-derived estimates of mining activity with data on gold trade, mercury and cyanide production, transport, and use, Bayesian MFA can quantify hidden flows and uncertainties. This perspective helps reveal supply chains that underpin illegal mining, supporting efforts to disrupt organised crime, improve transparency, and design interventions that protect communities.
By linking environmental signals to economic and criminal networks, the project highlights how geospatial science can contribute not only to environmental monitoring, but also to crime prevention and to policies that reduce exploitation and support more sustainable livelihoods in mining-affected regions.
This project is well-suited to multiple students interested in working on complementary tracks towards a common theme.
Farm2Market: A Digital Infrastructure for Data-Driven Smallholder Agriculture#
Project code: yvpl-111
Main supervisor: Yves Plancherel,
y.plancherel@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Jazzie Jao,
jazzie.jao@dlsu.edu.ph, De La Salle University, PhilippinesAvailable to: ACSE EDSML
This project may accept multiple students.
Small-scale farmers produce a substantial share of the world’s food, yet often lack access to digital tools that could help them manage inputs, understand environmental conditions, and connect efficiently to markets. While remote sensing, environmental data, and data analytics are increasingly powerful, these capabilities rarely reach farmers in a usable, low-cost, and context-appropriate form. This project addresses that gap by developing Farm2Market, a prototype digital infrastructure—implemented as a web platform or mobile app—that enables small-scale farmers to digitise their own farm data and benefit directly from data-driven insights.
The core objective is to design and build a working prototype that can be tested with realistic use cases. Students will develop tools that allow farmers to record basic information on farm location, crops, planting dates, yields, fertiliser or pesticide use, and management practices. These user-provided data will be combined with open-access environmental information, including satellite-derived vegetation indices, soil moisture, rainfall, temperature, and drought indicators, to provide simple, actionable analyses and visualisations tailored to farm-scale decision-making.
Additional project components may include computer-vision pipelines for analysing user-uploaded images of crops, enabling applications such as crop health assessment, pest or disease detection, and weed identification. Other tracks may focus on logistics and market access, including route optimisation for produce delivery, co-transport or aggregation options shared among farmers, and exploratory models for short-term price trends in local markets. These components highlight the integration of environmental data, machine learning, and optimisation within a single digital ecosystem.
Multiple MSc students can work on complementary aspects of Farm2Market, such as frontend and backend development, data integration, machine-learning modules, or user experience design. Emphasis is placed on low-cost, scalable, and transparent solutions. The project aims to demonstrate how digital infrastructure can empower farmers, improve resource efficiency, reduce waste, and strengthen links between farms and markets, while providing students with hands-on experience at the interface of data science, remote sensing, and real-world impact.
Developing ML Emulators to Reconstruct Earth’s Cryosphere through Space and Time#
Project code: frri-116
Main supervisor: Fred Richards,
f.richards19@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor is not yet assigned.
Available to: ACSE EDSML
This project may accept multiple students.
Problem description:
Understanding how polar ice sheets have responded to past periods of extreme warmth is critical for accurately predicting future sea-level change. Ancient shoreline elevations provide an important window into past sea levels but connecting them to ice volumes is complicated by two processes that vertically deflect shorelines away from global mean sea level. The first is “dynamic topography” (DT), vertical motions driven by mantle convection; the second is “glacial isostatic adjustment” (GIA), sea-level and topography changes caused by Earth’s viscoelastic response to shifting ice and ocean loads. We can now simulate them in sufficient detail to correct ancient shoreline datasets for their impacts so that the impact of past global warming on ice volumes can be reconstructed. However, it is currently impossible to quantify uncertainties, since each simulation requires >1000 CPU hours, prohibiting application of probabilistic approaches. This project will overcome this obstacle by building computationally efficient “emulators” (ML-based surrogate models) that accurately approximate these simulations in seconds.
Computational methodology:
The computational expense and global coverage of the simulations necessitate emulators that can handle a small training ensemble (~400 simulations), while accurately capturing spatiotemporal covariance on a sphere. These requirements will be satisfied by marrying two novel ML techniques: neural architecture search (NAS), a deep learning approach that automatically optimises neural network architecture to maximise prediction accuracy for small training sets, and spherical convolutional neural networks (SCNNs), which correctly account for spatial autocorrelation and the spherical geometry of Earth’s surface.
Expected outcomes:
By incorporating these ML emulators into a Bayesian inverse framework, all sources of model and data uncertainty will be propagated into global-scale reconstructions of past ice volume. The resulting estimates will provide the IPCC with new benchmarks for improving ice-sheet models and their sea-level projections.
Unveiling Earth’s internal dynamics with the adjoint method#
Project code: frri-133
Main supervisor: Fred Richards,
f.richards19@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor is not yet assigned.
Available to: ACSE EDSML
This project does not accept multiple students.
Problem description:
Since its formation 4.5 billion years ago, the Earth has been cooling, solidifying and segregating into a dense iron core and silicate mantle. Over these aeons, thermochemical convection of the mantle is the primary mechanism by which material and heat have been transferred from our planet’s deep interior to its surface, and it continues to govern the locations of critical mineral deposits, the stability of our climate, and the distribution of natural hazards. Advances in seismic imaging now allow us to directly observe the mantle’s present-day structure, while computational innovations mean we can simulate its dynamics. Although these developments have improved our understanding of Earth’s inner workings, the physical properties of imaged structures remain very uncertain and considerable (lively!) debate continues over whether they mostly represent thermal or chemical anomalies. This project aims to shrink this uncertainty by combining cutting-edge optimisation techniques with new geophysical datasets to place unprecedented constraint on the physical state of Earth’s mantle.
Computational methodology:
We will use the state-of-the-art G-ADOPT Python package and a range of geophysical datasets (topographic, geodetic, and seismic) to develop inversions for present-day mantle temperature, density, and viscosity. This revolutionary software exploits the adjoint method, an efficient technique for computing the gradient of a misfit function with respect to variations in N model parameters that only requires solution of a single set of forward and adjoint equations. Crucially, this approach makes gradient-based optimisation of mantle flow simulations—where N~106–109—computationally feasible.
Expected outcomes:
New models of Earth’s internal dynamics emerging from this project have the potential to settle major debates around the long-term evolution of our planet and will underpin ongoing efforts to better predict: a) future stability of polar ice sheets; and b) locations of critical mineral deposits key to the green transition.
Leveraging AI Libraries and GPUs for Improved Accuracy-Cost Balance in Ocean Modeling#
Project code: pasa-193
Main supervisor: Parastoo Salah,
p.salah@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Yezhang Li,
yezhang.li22@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
This project aims to investigate the impact of on the accuracy-cost balance in ocean modeling. As numerical software based on the finite element method (FEM) such as Thetis/Firedrake have been widely used for ocean simulations, their computational efficiency and scalability remain challenging. This project seeks new insights into how advanced FM approaches implemented by AI-library and GPU-optimized techniques can enhance accuracy while reducing computational costs, contributing to a deeper understanding of high-performance simulation in coastal ocean modeling. In this study, benchmarking of shallow water scenarios (including wet/dry interface and tsunami case) is performed and implemented by both traditional finite element (FE) based solver (Thetis/Firedrake) and partial differential equation solver with convolutional layers, leveraging AI/GPU Techniques (NN4PDEs). Performance metrics such as accuracy, computational cost, and scalability (GPU vs. CPU speedup) are analyzed using high-performance computing platforms. Uncertainty quantification will be further performed based on NN4PDEs. (Prof. Matthew Piggott/Dr Stephan Kramer will be additional collaborator.)
AI-Assisted Acceleration of Numerical Hazard Modeling#
Project code: pasa-236
Main supervisor: Parastoo Salah,
p.salah@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Matthew Piggott,
m.d.piggott@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML GEMS READY
This project may accept multiple students.
This project aims to develop artificial intelligence (AI)–based methods to accelerate numerical hazard modeling while reducing computational cost and uncertainty propagation time. Traditional hazard simulations often rely on large-scale Monte Carlo methods, requiring thousands of realizations to capture uncertainty. Although accurate, these approaches are computationally expensive and limit rapid decision-making.
The proposed research explores AI-assisted model reduction and surrogate modeling techniques to compress large uncertainty ensembles into a smaller, representative set that can be efficiently processed by numerical solvers. First, clustering and dimensionality reduction methods will be investigated to identify dominant patterns in high-dimensional stochastic inputs. These reduced samples will serve as optimized substitutes for large Monte Carlo ensembles.
Second, machine learning–based surrogate models will be developed to approximate complex physical simulations. Hybrid architectures combining Graph Neural Networks (GNNs), Long Short-Term Memory (LSTM), and Convolutional Neural Networks (CNNs) will be evaluated for learning spatial–temporal hazard dynamics. In parallel, Physics-Informed Neural Networks (PINNs) will be explored to embed governing equations directly into the learning process, improving physical consistency and generalization.
The project will compare traditional Monte Carlo simulations with AI-assisted frameworks in terms of accuracy, speed, and robustness. Performance will be assessed using benchmark hazard modeling problems, focusing on uncertainty quantification and predictive reliability.
ClimateTech Startup-Investor Matching Tool#
Project code: masc-075
Main supervisor: Maiko Schaffrath,
m.schaffrath@imperial.ac.uk, The Grantham Institute for Climate Change, Imperial College LondonSecond supervisor is not yet assigned.
Available to: ACSE EDSML
This project may accept multiple students.
Matching ClimateTech startups with relevant investors is a complex ranking problem involving multiple competing criteria and diverse data sources. Our existing approach relies on simple rule-based filters (sector, geography, ticket size) that cannot capture nuanced compatibility factors or learn from outcomes. The objective of this project is to design, implement, and evaluate a matching system that substantially improves upon rule-based baselines by applying computational ranking algorithms, natural language processing, and similarity learning techniques. A basic rules-based prototype exists using Airtable and JavaScript; this project will develop ranking algorithms, implement robust data processing pipelines, and systematically evaluate performance improvements.
Methodology The student(s) will design and implement: Feature extraction and representation: Development of a feature engineering pipeline to extract and encode startup and investor characteristics from structured data (sector, stage, geography, ticket size, investment history) and unstructured text (pitch decks, company descriptions). Investigation of appropriate representations for textual content using NLP techniques such as embeddings, topic models, or keyword extraction. Similarity metrics and ranking algorithms: Design and implementation of similarity functions that combine multiple dimensions of compatibility. Development of ranking algorithms that integrate rule-based constraints with similarity scoring. Exploration of different weighting schemes and combination strategies. If sufficient labeled data is available, investigation of supervised learning approaches (e.g., learning-to-rank methods). Algorithm evaluation: Implementation of an experimental framework to compare algorithm variants against baselines using standard ranking metrics (precision@k, NDCG, MRR). Analysis of performance across different conditions including cold-start scenarios, sparse data, and varying investor preferences. Collection and incorporation of feedback data to assess real-world matching quality. System implementation: Integration of the matching algorithms with existing infrastructure (Airtable forms for data input, email for delivery). Implementation of data validation, processing pipelines, and automated matching workflows. The technical stack will be chosen by the student based on project requirements.
Expected Outcomes -A working matching system that demonstrably improves upon rule-based baselines, with quantitative performance evaluation. -Analysis of which features and similarity approaches are most effective for investor-startup matching under different data conditions. -Well-architected, documented codebase suitable for deployment and future extension. -Critical evaluation of the system’s performance, limitations, and potential improvements.
Developing, Testing and Implementing an Automated Defacing Workflow for Magnetic Resonance Imaging Research Data.#
Project code: jase-012
Main supervisor: Jan Sedlacik,
j.sedlacik@imperial.ac.uk, Institute of Clinical Sciences, Imperial College LondonSecond supervisor is not yet assigned.
Available to: ACSE
This project does not accept multiple students.
Increased MRI scan quality and the capability of facial recognition software to search the internet make it necessary to remove facial features from neurological research scans to better protect subject privacy when sharing MRI data.
The aim of this project is to develop an automated workflow for defacing MRI scans using Python. The workflow needs to work with single and multi-frame DICOM files and without writing temporary interim data, e.g. Neuroimaging Informatics Technology Initiative (NIfTI) formatted data, to the file system. The code should be implemented as a locally running command line script on a image series folder on the file system. Furthermore, the code should be implemented as an pipeline on the eXtensible Neuroimaging Archive Toolkit (XNAT) imaging data platform. The defaced DICOM data should then be archived as a separate image series on XNAT. The workflow should be optimised with respect to low computational and memory demands, since the XNAT archive server, as well as the local desktop PCs, have limited hardware specifications. Furthermore, it needs to be investigated which type of MRI scans (2D, 3D, structural, functional) have sufficient facial information for possible deidentification and, therefore, will need to be defaced. The defaced data needs to be checked and tested for the successful removal of the facial features and if it affects the subsequent structural or functional analysis of the image data.
The student is expected to work independently but in close communication with the supervisors. The student should conduct a thorough literature research and test all available and potentially useful defacing tools using anonymised image data scanned at our facility. Literature research and writing of the final report can be done remotely. However, testing and implementing the workflow needs to be done on site, since the image data is not available remotely.
Developing a processing pipeline for the quantitative assessment of hemodynamic parameters from 4D-flow cardiovascular magnetic resonance data.#
Project code: jase-013
Main supervisor: Jan Sedlacik,
j.sedlacik@imperial.ac.uk, Institute of Clinical Sciences, Imperial College LondonSecond supervisor is not yet assigned.
Available to: ACSE
This project does not accept multiple students.
4D flow data contains complex hemodynamic information which is difficult to break down and extract in a reliable quantified manner for further analysis. Currently available tools quantify only a limited number of hemodynamic parameters at limited locations.
The aim of this project is to develop a workflow for the comprehensive assessment of the hemodynamic information from 4D-flow cardiovascular magnetic resonance data, especially for the pulmonary arteries, using Python and/or Matlab. The user-interactive data segmentation and visualisation can be done by already available software tools like 3D Slicer and ParaView. The processing pipeline needs to accept single and multi-frame Digital Imaging and Communications in Medicine (DICOM) files as well as Neuroimaging Informatics Technology Initiative (NIfTI) formatted data. The quantified hemodynamic parameters need to be exported in a well structured CSV (comma-separated values) file for further analysis and data visualisation. The code should run independently of the operation system. The workflow should be optimised with respect to low computational and memory demands since the local desktop PCs have limited hardware specifications.
The student is expected to work independently but in close communication with the supervisors. The student should conduct a thorough literature research and test all available and potentially useful software tools using anonymised image data scanned at our facility. Literature research and writing of the final report can be done remotely. However, testing and implementing the workflow needs to be done on site, since the image data is not available remotely.
Further Reading: https://doi.org/10.1186/s12968-018-0451-1
Software Examples: JulioSoteloParraguez/4D-Flow-Matlab-Toolbox
Foundational GeoAI: Leveraging Visual Language Models for Global Environmental Mapping And Resilience#
Project code: misi-040
Main supervisor: Minerva Singh,
minerva.singh07@imperial.ac.uk, Centre for Environmental Policy, Imperial College LondonSecond supervisor: Adriana Paluszny,
apaluszn@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML GEMS READY
This project may accept multiple students.
This project explores integrating Foundational AI and GeoAI to address critical global challenges across the Earth sciences. By leveraging the power of Visual Language Models (VLMs) and multimodal deep learning, students will develop frameworks capable of interpreting complex spatial-temporal data to solve problems specific to their chosen track: ACSE, EDSML, GEMS, or READY.
The core of the project involves utilising large-scale pre-trained models trained on massive satellite, geophysical, and textual datasets to perform “zero-shot” or “few-shot” analysis on unstructured geospatial information. In the context of Climate Hazards, students may build predictive systems for flood or wildfire risk by fusing historical meteorological data with real-time aerial imagery. For Habitat Monitoring, the focus shifts to automated biodiversity assessment, using VLMs to describe ecological changes and identify shifts in species distribution from remote sensing feeds. Alternatively, those interested in the energy transition can apply these models to Mineralisation Potential Mapping to identify subsurface signatures of critical minerals required for renewable technologies.
Technically, the project will require students to move beyond standard regression, implementing advanced architectures like Vision Transformers (ViTs) and Graph Neural Networks (GNNs) to model geological connectivity and environmental dependencies. By bridging the gap between foundational computer vision and traditional geosciences, this project prepares students to deploy scalable, AI-driven solutions for resource management and planetary resilience. Whether focused on carbon sequestration in GEMS or geothermal exploration in READY, the student will master transforming raw Earth observation data into actionable, high-level intelligence.
Engineering a Proof-of-Concept Digital Warning System for High-Risk Prescribing in Oncology#
Project code: bosu-237
Main supervisor: Bowen Su,
b.su@imperial.ac.uk, Cancer Research UK Convergence Science Centre, Imperial College LondonSecond supervisor: Parastoo Salah,
p.salah@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
This project contributes to the HEADSPACE research programme, which focuses on improving patient safety during cancer treatment through digital health innovation. During intensive treatments such as chemoradiotherapy, some patients receive multiple psychotropic medications (e.g. hypnotics, anxiolytics, antidepressants) within short timeframes. These prescribing patterns may increase the risk of treatment interruption and unplanned hospital admissions. A digital “track-and-trigger” warning system could help identify such high-risk scenarios early and support safer clinical decision-making. As part of the project, a proof-of-concept (POC) risk algorithm has been developed to identify potentially high-risk psychotropic prescribing patterns based on predefined clinical rules and risk flags informed by real-world data analysis and clinician input. Building on this algorithm, the project will focus on the engineering and implementation of a prototype digital warning system that demonstrates how such risk signals could be operationalised in practice.
LLM analysis of the barriers to offshore wind deployment as framed by industry, government and the media#
Project code: siwa-090
Main supervisor: Simon Warder,
s.warder15@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Matthew Piggott,
m.d.piggott@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML GEMS READY
This project may accept multiple students.
The ongoing transition to renewable energy sources encounters a wide range of barriers including technical, economic, regulatory and social factors. These issues are frequently raised by a variety of stakeholder groups including industry, government and the media. This project will apply LLMs to the analysis of industry reports, government statements, policy documents, and news articles, with the aim of identifying how the key barriers to offshore wind development are framed and evaluated by different stakeholder groups, and how these patterns evolve over time.
LLMs will be applied to:
Extract and cluster recurring issues and barriers discussed across documents
Analyse sentiment and stance towards these issues, organised by stakeholder groups, to analyse alignment or divergence between stakeholders
Identify patterns of responsibility attribution, including issues that are framed as systemic or left unresolved
With the above as a starting point, this project will be open-ended with a broad scope for the development and application of new LLM-based tooling.
Integration of renewable technologies into energy system models#
Project code: siwa-089
Main supervisor: Simon Warder,
s.warder15@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Matthew Piggott,
m.d.piggott@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML GEMS READY
This project does not accept multiple students.
Energy system models are essential tools underpinning the expansion of renewable energy technologies, and their integration into energy grids. They allow us to simulate power flow through energy networks and optimise future capacity and grid expansion for the renewable energy transition. However, the accurate representation of renewable generators such as wind and tidal within these systems is limited by their high spatial and temporal variability, as well as long-range physical interactions between different generators; the addition of a new wind or tidal installation in one location can affect the power generated by others elsewhere. This makes these renewable energy sources challenging to integrate into energy system models, compared with traditional sources such as fossil fuel-powered generators where no such interaction occurs.
Potential aspects of this project include:
Comparison of current energy system models, including investigation of their adequacy in representing wind and tidal energy sources
Quantifying the sensitivity of such models to renewables representation, alongside the uncertainty inherent in modelling energy systems
Developing new functionality to properly represent renewables in energy system models
Bias correction of reanalysis datasets for wind farm energy production#
Project code: siwa-088
Main supervisor: Simon Warder,
s.warder15@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Matthew Piggott,
m.d.piggott@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML GEMS READY
This project does not accept multiple students.
So-called “reanalysis” products (such as ERA-5) constitute our best estimate of the state of the atmosphere over the last several decades, and are commonly used in wind energy resource assessment. However, such products still contain a number of biases, due e.g. to limited spatial or temporal resolution, missing physics (e.g. wind turbine/farm “wake” effects), and uncertain parameterisations. This project will explore methods to correct biases in such reanalysis datasets, including statistical methods leveraging real-world data, physics-based modelling, and potentially ML methods to reproduce missing physics.
Wind farm layout optimisation under uncertainty#
Project code: siwa-085
Main supervisor: Simon Warder,
s.warder15@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Matthew Piggott,
m.d.piggott@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML GEMS READY
This project may accept multiple students.
Wind farm layouts are often carefully optimised to minimise the energy losses which arise when one wind turbine is situated in the downstream wake of another. However, the potential gains from such layout optimisation are limited by (i) the variability of wind speed and direction at the wind farm site, (ii) uncertain future developments of other wind farm sites nearby, (iii) costs and constraints which depend on the specific turbine positions. This creates a highly complex constrained optimisation problem which will be addressed in this project through a combination of:
Utilisation of existing wake modelling tools, and/or development of new ML approaches
Application of optimisation algorithms from the “robust optimisation” literature, which may also include ML approaches (e.g. reinforcement learning)
Comparison of wind distributions globally to address challenges in regions with developing wind industries
Wind energy assessment accounting for wake effects#
Project code: siwa-084
Main supervisor: Simon Warder,
s.warder15@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Matthew Piggott,
m.d.piggott@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML READY
This project may accept multiple students.
Wind energy development requires accurate estimates of potential future energy yield. However, this is limited by high levels of uncertainty in future wake effects. Wakes arise due to the extraction of energy by a wind turbine or farm, which reduces the energy available at another installation downstream. These effects reduce wind energy yields by 20% or more, but are highly uncertain due to unknown future wind projects. Wind energy resource assessment datasets often neglect these effects altogether. This project will develop new machine learning models and tooling to address these gaps.
Potential aspects of this project:
Development and application of new ML wake modelling approaches
Simulation of wind turbine efficiency using existing wake modelling methods
Development of a resource assessment tool/interface, accounting for uncertain wake effects
Use of open-source wind energy deployment datasets to track/model wind energy efficiency over decadal time scales
There is opportunity for timely and impactful research on this topic, with scope for multiple projects, and for interested students to identify particular research directions suited to their interests.
Predictive Modelling and Software Development for Contaminant Removal in Water Treatment#
Project code: dowe-208
Main supervisor: Dominik Weiss,
d.weiss@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Jay Bullen,
jbullen@anglianwater.co.uk, Anglian WaterAvailable to: ACSE EDSML
This project may accept multiple students.
Our research group develops predictive models and software tools to better understand and improve chemical processes used in water treatment, with a particular focus on adsorption and column transport. A key application of our work is the removal of PFAS (per- and polyfluoroalkyl substances), widely known as “forever chemicals, which pose major challenges for drinking water treatment worldwide. We offer four MSc research projects, all of which can be adapted to match a student’s background in chemistry, environmental science, engineering, or data science:
Modelling PFAS adsorption on water-treatment sorbents. Students will develop surface complexation models to describe how PFAS interact with existing and emerging sorbent materials. The project combines experimental data with chemical modelling to predict adsorption behaviour under different water chemistries.
Machine-learning prediction of PFAS breakthrough in adsorption columns. This project focuses on applying machine-learning methods to predict PFAS breakthrough in granular activated carbon (GAC) columns. Students will learn how to link influent chemistry, operational conditions, and material properties to column performance.
Kinetic modelling of adsorption processes. Students will develop kinetic models describing the adsorption of inorganic and organic contaminants onto a range of sorbents. This project involves solving ordinary differential equations (ODEs) and is well suited to students with a strong quantitative or mathematical background.
Software development for water-treatment data analysis. This project involves creating user-friendly software tools for the quantitative analysis of batch and column adsorption data, with direct relevance to laboratory experiments and real-world water treatment systems. Depending on projects, MSc students will gain practical experience in regression analysis, machine learning, numerical modelling, and scientific coding. Expected outcomes include a validated predictive model or software tool, a reproducible data-analysis workflow, and a high-quality MSc thesis. These projects provide excellent preparation for careers in water treatment, environmental consulting, industry, or PhD research
Micronutrient and Contaminant Cycling in Complex Environments: Modeling Metal Speciation and Transport#
Project code: dowe-211
Main supervisor: Dominik Weiss,
d.weiss@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor is not yet assigned.
Available to: ACSE EDSML
This project may accept multiple students.
Our group investigates the factors controlling micronutrient and contaminant cycling across Earth systems, combining molecular-level insights with environmental-scale processes. Understanding these mechanisms is critical for predicting the mobility, bioavailability, and retention of essential and toxic elements in soils, waters, and sediments. Project themes include:
Siderophore-mediated micronutrient uptake: We explore the role of siderophore exudation in the acquisition of essential metals, focusing on molybdenum (Mo) and vanadium (V). Key questions include whether siderophore production significantly enhances uptake and how this mechanism operates under varying environmental conditions.
Impact of salinization on metal speciation: We quantify how increasing salinity affects metal complexation, adsorption, and transport, helping to predict metal behavior in saline or impacted environments. Students will gain practical experience in extracting and integrating experimental and computational data, performing multi-variable regression analyses, and implementing geochemical models into coding frameworks. These projects combine precise numerical modeling, chemical speciation, and data-driven analysis to uncover the processes governing micronutrient and contaminant cycling. The expected outcome is a high-quality MSc thesis providing new insights into metal mobility and bioavailability, while developing advanced skills in geochemical modeling, coding, and quantitative data analysis.
Quantifying Metal Movement Using Isotopes: A Data-Driven Approach to Understanding Isotope Fractionation#
Project code: dowe-210
Main supervisor: Dominik Weiss,
d.weiss@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor is not yet assigned.
Available to: ACSE EDSML
This project may accept multiple students.
Our group investigates the factors controlling metal isotope fractionation in natural and engineered systems, from molecular structure to large-scale geochemical processes. Understanding how metals fractionate is essential for predicting contaminant transport, assessing water quality, and tracing the cycling of both essential and toxic trace metals in the environment. By integrating modeling, experimental data, and reactive transport frameworks, we study the molecular and chemical mechanisms that govern how metals move, react, and partition in soils, sediments, and waters.
Project themes include:
Metal–ligand complexes: Predicting how isotopes partition depending on bonding, coordination, and solvation, providing insights into metal reactivity and speciation.
Transport processes: Quantifying isotope fractionation during advective–dispersion transport in diverse aqueous solutions to understand the mobility and retention of trace metals.
Kinetic isotope effects: Modeling fractionation during dissolution, precipitation, and adsorption using ODEs and reactive transport models, linking molecular-scale reactions to environmental behavior.
Skills and experience gained: MSc students will develop expertise in extracting and integrating experimental, literature, and quantum chemical data, solving linear and nonlinear regressions, implementing scientific code, performing precise numerical modeling of subtle isotope effects, simulating chemical speciation under varying environmental conditions, and conducting uncertainty and sensitivity analyses.
These projects provide a quantitative understanding of the relationships between molecular structure, chemical processes, and isotope fractionation. Students gain hands-on experience in geochemical modeling, coding, and data analysis, preparing them for careers in water treatment, environmental consulting, industry, or PhD research.
Expected outcome: A high-quality MSc thesis demonstrating how metals fractionate during chemical reactions and transport, combining advanced computational skills with data-driven insights into trace metal cycling and environmental contaminant behavior.
Data-Driven Approaches to Tropospheric Aqueous Phase Chemistry#
Project code: dowe-209
Main supervisor: Dominik Weiss,
d.weiss@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Parastoo Salah,
p.salah@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
Our research group studies aqueous chemical processes in the atmosphere, with a strong focus on air pollution, atmospheric particles, and their role in the Earth system. We are particularly interested in how airborne particles chemically evolve as they are transported through the atmosphere and interact with water under a wide range of environmental conditions. This work sits at the interface of atmospheric chemistry, environmental geochemistry, and data science. We currently offer three distinct research projects, which can be adapted to suit students with different backgrounds and interests:
Controls on particle dissolution during atmospheric aging. This project aims to identify the key chemical and physical parameters—such as pH, ionic strength, and temperature—that control the dissolution of atmospheric particles as they age and are transported over long distances. Students will work with experimental and observational datasets to quantify how environmental conditions influence dissolution rates and metal release.
Kinetic modelling of atmospheric particle dissolution. In this project, students will develop rate laws and kinetic models to describe particle dissolution during atmospheric aging. The work involves translating laboratory and field data into predictive models that can be applied across a range of atmospheric conditions.
Coupling kinetics with geochemical speciation models. This project focuses on integrating experimentally derived rate laws into geochemical speciation models, with a particular emphasis on iron oxidation in low-pH aqueous solutions relevant to atmospheric aerosols and cloud water. Across all projects, students will gain hands-on experience in multivariate regression analysis, machine-learning approaches, and numerical modelling. A key component of the work is the integration of custom scripts and kinetic formulations into established geochemical modelling software, especially PHREEQC. These projects provide strong training in quantitative analysis, coding, and process-based thinking, and are well suited for students interested in atmospheric chemistry, environmental modelling, and Earth system science.
Synchrotron-powered insights into how early eukaryotes became fossils#
Project code: chwo-103
Main supervisor: Christina Woltz,
c.woltz@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor is not yet assigned.
Available to: ACSE EDSML
This project does not accept multiple students.
For more than 85% of Earth’s biological history, life was almost entirely microbial. It was not until Earth’s “middle ages”, around 1,700 million years ago (Ma), that the first eukaryotes appeared in the fossil record. Although modern eukaryotes include large, complex organisms—plants, animals, algae, and fungi—most ancient and living eukaryotes are single celled. Fossilized organic remains of these organisms provide direct insights into early eukaryotic evolution and ecology.
Early eukaryotes lived in dynamic ocean environments shaped by fluctuating oxygen levels and evolving biogeochemical cycles. These conditions can influence the likelihood that their remains become preserved as fossils. Understanding how local environmental conditions control fossilization is essential for reconstructing early eukaryotic evolution and discovering new microfossil assemblages.
Goal: This project aims to determine how local redox conditions affect the preservation of organically preserved microfossils in shale and mudstone.
Methodology: We will use high resolution synchrotron based x ray fluorescence and x ray absorption spectroscopy collected at the Stanford Synchrotron Radiation Lightsource. The dataset includes nine fossil bearing samples from the ~750 Ma Uinta Mountain Group (Utah) and Chuar Group (Arizona, USA).
Key components include:
• Elemental abundance maps— including Zn, Cu, Fe, Mn, Ca, K, S, P, Si, Al, and Mg— collected over ~1 cm² polished rock regions at 75 µm² resolution, with higher resolution mapping (35µm² and 5µm²) focused on fossil bearing areas.
• Redox sensitive Fe and S speciation using targeted energies that maximize contrast among species (Fe-oxides, -silicates, and -sulfides; pyrite, monosulfides, organic sulfur, sulfoxides, sulfonates, and sulfates)
In total, we have over 200,000 spatially resolved datapoints of elemental abundances and redox sensitive species across fossil bearing samples. This project will explore a new frontier: using artificial intelligence to recognize geochemical patterns linked to microfossil preservation. The discoveries made here could transform how—and where—we hunt for Earth’s earliest eukaryotic life.
Particle Modelling using Neural Networks - AI4Particles#
Project code: jixi-153
Main supervisor: Jiansheng Xiang,
j.xiang@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonSecond supervisor: Christopher Pain,
c.pain@imperial.ac.uk, Department of Earth Science and Engineering, Imperial College LondonAvailable to: ACSE EDSML
This project may accept multiple students.
Our vision for the AI4Particles is the formation of a general particle based method that can model any particle system from solids structures and fractured Discrete Element Modelling structures to Monte-Carlo for radiation transport. This project builds on a new approach that solves for particles on a convolutional grid and using a neural network with analytically defined weights [1]. Other applications are vast including molecular dynamics, Smooth Particle Hydrodynamics, Particle in a cell modelling to understanding turbulent mixing by seeding particles in flows to plankton in the oceans and people behaviour. Particles are an example of complex system behaviour with emerging dynamic behaviours and as such the approach AI4Particles may be re-applied to systems modelling in general. We will form a focus team in order to help develop these other application areas and take advantage of all the massive benefits of AI and associated workflows. In addition, it so often common place to represent a systems dynamics or to discover systems dynamics that is represented by a PDE, see [2], e.g. for forming a turbulent model closure. However, particle systems could equally be discovered with the dynamics of interest and thus this work is a major step forward in this direction. e.g. SPH particles or Lattice Boltzmann methods used to model fluids. The Monte Carlo and DEM are developed as extreme case in which the other examples are closely related to one of these and thus relatively universally applicable. Monte Carlo and DEM are also of particular interest in themselves but other particle models can be developed from these. We also address the link to the continuum. The benefits of ANN-based solvers without necessitating training data, enabling GPU-accelerated computations, enhanced programmability, platform interoperability, full differentiability, seamless multi-physics integration, and compatibility with surrogate models based on trained neural networks. Projects on offer include: (1) Monte-Carlo modelling of nuclear reactors (2) SPH modelling (3) Molecular dynamics The coding will be in PyTorch and Python.
[1] Naderi, Chen, Yang, Xiang, Heaney, Latham, Wang, Pain (2024) A discrete element solution method embedded within a Neural Network, Powder Technology 448: 120258. https://doi.org/10.1016/j.powtec.2024.120258.
[2] Chen, Heaney, Gomes, Matar, Pain (2024) Solving the Discretised Multiphase Flow Equations with Interface Capturing on Structured Grids Using Machine Learning Libraries, Computer Methods in Applied Mechanics and Engineering 426: 116974. https://doi.org/10.1016/j.cma.2024.116974