Overview and Background

Heatwaves (i.e. sustained periods of exceptionally hot weather) are a well-recognised public health hazard. Strong evidence shows that the co-occurrence of extreme air pollution events during heatwaves amplifies health risks [1,2]. The 2022 European heatwave, when the UK recorded its first ever temperature >40°C, was accompanied by a widespread deterioration in air quality, with air pollutants at ground level exceeding safe limits across much of the continent3. Causal relationships between extreme temperature and air pollution are complex, involving weather patterns that promote air stagnation, pollutant emissions (e.g. from wildfires) and atmospheric photochemistry [4, 5]. These factors are not yet well understood but are important to detangle as the frequency and intensity of summer heatwaves is expected to rise due to climate change [6].

This project’s overarching goals are to (1) characterize the response of air pollutants to heatwaves across Europe and to assess their combined health impacts, (2) provide new process-level insight into the causal relationship between extreme temperature and air pollution, and (3) evaluate and improve current systems for forecasting extreme air pollution events. This will be achieved through ultra-high-resolution air quality model simulations, analysis of air pollutant measurement data, and by exploring data-driven approaches to air quality forecasting and inference, including machine learning.

The successful candidate will join LEC’s vibrant atmospheric science research group (AtMOS) and benefit from a diverse supervisory team. This includes a placement with a UK industry partner (Reliable Insights Ltd www.reliable-insights.com / https://www.tangentworks.co.uk/) and partnership with the UK Health Security Agency. 

Methodology and Objectives

Teaser Project 1: What drives adverse air quality during heatwaves?
In Year 1, the goal will be to characterise the observed response of ground-level ozone (an important air pollutant) during European heatwaves. Using the severe summer 2022 heatwave as a case study, an analysis of surface temperature and ozone measurements from hundreds of sites across Europe will be performed (e.g. exploiting the extensive TOAR-II measurement database). Output from the UCI chemical transport model (CTM), a global model developed and maintained by the Lancaster atmospheric science group, will also be analysed. The ability of the model to simulate the behaviour of ozone during the heatwave, including the observed ozone-temperature relationship, will be determined. This work will provide a strong grounding in the observational datasets and atmospheric model used in subsequent years.

In Year 2 emphasis will be placed on understanding the various processes (meteorological, chemical, physical) responsible for elevating ozone during heatwaves. This will be achieved through carefully designed model sensitivity experiments that allow these factors to be detangled and quantified. One enquiry will be to assess the importance of temperature-dependent ozone precursor emissions, such as emissions from wildfires (CO, NOx) and volatile organic compound emissions from stressed vegetation. Another will be to assess long-range transport of ozone into continental Europe from other world regions. As the project progresses in years 2 and 3, the scope will expand to consider other notable heatwave years and air pollutants. The ability of regional scale models to forecast extreme air pollutant events will be explored, as will innovative data-driven approaches (e.g. machine learning).

Teaser Project 1 Objectives:

• Characterize the ozone-heatwave response across multiple European summers using surface and satellite measurements.
• Assess the ability of atmospheric models to capture extreme ozone events and the observed ozone-temperature relationship.
• Interpret the observed ozone-heatwave responses using high resolution model simulations and assess the roles of meteorological, chemical and physical factors.
• Explore data-driven approaches to air quality forecasting.

Teaser Project 2: Optimising air quality forecasts: exploring data-driven methods

Process-based air quality models are frequently used to simulate past (i.e. ‘hindcast’) air quality. This provides information needed to quantify how air pollutant levels have changed over time (e.g. due to policy interventions) and the implications for human exposure and health. Such models are also increasingly used to alert the public in advance of upcoming air pollution episodes (i.e. ‘forecast’). Like weather forecasts, air quality forecasts are provided up to several days ahead, though confidence generally decreases as the forecast range increases. As the skill of air quality models is often inadequate, particularly for the most ‘extreme’ episodes, various approaches to ‘bias correct’ model forecasts (before issued) have emerged [7, 8].

In Year 1, the goal will be to examine the ability of WRF-Chem to simulate UK air quality in hindcast mode, with an emphasis on quantifying the model’s skill during recent heatwaves. WRF-Chem is a well-evaluated and widely adopted process model suitable for national-scale high resolution simulations. The model’s skill in simulating extreme air pollutant levels will be assessed using the UK’s extensive network of air pollutant measurements and by considering a range of key performance metrics (e.g. hit rate, false alarm rate etc.). This will provide a solid grounding on the strengths and weaknesses of air quality models and approaches to evaluate them.

In Year 2, the effectiveness of a range of bias correction techniques applied to the WRF-chem simulations will be explored, including machine learning based approaches [9]. Methods that improve the tail of the ozone distribution (e.g. ‘quantile mapping’) will be examined. Bias-corrected hindcasts will be produced and the annual mortality burden attributable to long-term air pollutant exposure, along with the health impacts of elevated air pollution in conjunction with heatwaves, will be quantified [10].

In Year 3, the focus of the project will be to explore how data science approaches can improve the skill of air quality forecasts across a range of forecast lead times (24 to 96 hours). Important predictor variables for adverse UK air quality will be assessed and ranked, including the potential for their near real-time assimilation (e.g. surface and satellite data). These data will be used to train machine learning models and the resulting data-driven forecasts will be evaluated against process models. A key question will be whether a data-driven approach outperforms a traditional process model forecast and if so under what settings. Whether the two may be combined to produce optimal results will also be examined.

The development of data-driven solution to air quality forecasting is a rapidly evolving field with very strong opportunity for impactful science. During year 3, a placement with Reliable Insights (a leading Lancaster based time-series data specialist) will be undertaken. This placement will provide an excellent opportunity to develop industry links, to implement

Teaser Project 2 Objectives:

• Evaluate the skill of the WRF-Chem model in simulating air pollutants during recent UK heatwaves employing a range of metrics.
• Test and apply bias correction techniques to produce optimised WRF-Chem hindcasts and use these to quantify the human health effects of UK air pollution over time.
• Explore a data-driven air quality forecasting system for the UK, exploiting machine learning and other innovative data science approaches.

References & Further Reading

[1] Schnell, J.L., and Prather, M.J. (2017). Co-occurrence of extremes in surface ozone, particulate matter, and temperature over eastern North America. Proc. Natl. Acad. Sci., 114, 2854-2859, https://doi.org/10.1073/pnas.1614453114
[2] Gouldsbrough, L., Hossaini, R., Eastoe, E., & Young, P.J.Y. (2022). A temperature-dependent extreme value analysis of UK surface ozone, 1980-2019. Atmos. Env., 273, 118975.
[3] https://atmosphere.copernicus.eu/copernicus-scientists-warn-very-high-ozone-pollution-heatwave-continues-across-europe
[4] Pope, R. J., et al. (2023). Investigation of the summer 2018 European ozone air pollution episodes using novel satellite data and modelling, Atmos. Chem. Phys., 23, 13235-13253, https://doi.org/10.5194/acp-23-13235-2023.
[5] Otero, N., Jurado, O. E., Butler, T., and Rust, H. W. (2022). The impact of atmospheric blocking on the compounding effect of ozone pollution and temperature: a copula-based approach, Atmos. Chem. Phys., 22, 1905-1919, https://doi.org/10.5194/acp-22-1905-2022.
[6] Doherty, R.M., Heal, M.R., and O’Connor, F.M. Climate change impacts on human health over Europe through its effect on air quality, Environ. Health, 16, https://doi.org/10.1186/s12940-017-0325-2.
[7] Staehle, C., et al. (2024). Technical note: An assessment of the performance of statistical bias correction techniques for global chemistry–climate model surface ozone fields, Atmos. Chem. Phys., 24, 5953-5969, https://doi.org/10.5194/acp-24-5953-2024.
[8] Neal, L.S., Agnew, P., Moseley, S., Ordóñez, C., Savage, N.H. and Tilbee, M. (2014). Application of a statistical post-processing technique to a gridded, operational, air quality forecast, Atmos. Env., 98, 385-393, https://doi.org/10.1016/j.atmosenv.2014.09.004.
[9] Gouldsbrough, L., Hossaini, R., Eastoe, E., Young, P.J.Y. & Vieno, M. (2023). A machine learning approach to downscale EMEP4UK: analysis of UK ozone variability and trends. Atmos. Chem. Phys., 24, 3163-3196, https://doi.org/10.5194/acp-24-3163-2024.
[10] Macintyre, H.L., et al. (2023). Impacts of emissions policies on future UK mortality burdens associated with air pollution. Environ. Int., 174, 107862, https://doi.org/10.1016/j.envint.2023.107862.

Overview and Background

​​This project will develop machine learning and statistical methods for real-time forecasting via data fusion with uncertainty quantification for water catchments. It will use and develop advanced AI models (building on and expanding the work such as Allen, et al., 2025 and relevant foundation models) to fuse in situ sensors and satellite data (optical and/or radar) for hydrology and surface water quality in river catchments. The project will also assess the generalisability of developed methods across multiple river catchments. 

​Data demands can be very large with, for example, data collected approximately every 15 minutes from multiple in situ sensors and 10m resolution for satellite. To obtain fast (real-time or near real-time), reliable and computationally efficient (and therefore environmentally friendly) models, this project specifically targets the development of GPU programming for scalable analytics and will consider the advantages of cloud GPUs and other related platforms, for example EDITO, to support transferability and impact. 

​This project is a collaboration with domain expert colleagues at University of Stirling and Scottish Water and will link to NERC projects such as SenseH20 and MOT4Rivers and the Forth-ERA digital observatory of the Forth Catchment.​ 

Methodology and Objectives

The AI-based models will be developed using the latest generation of deep learning approaches (such as transformer models, physics–informed losses, etc). As detailed in the background, this project would leverage existing model frameworks (like Aardvark), foundation models that can be specialised, and construct domain–specific pipelines as appropriate. The fusion methodologies will be tested on a number of downstream tasks, most notably predicting values of water quantity/quality far from the sensor locations (validated by cross validation) and forecasting. While the frameworks generated will be tested on the dataset in question, the assumption would be that the models can be transferred to other localities, and testing the portability of these approaches will be part of the later stages of this project.

​Teaser Project 1: 

​The first teaser project will focus on developing an initial methodology for data fusion, based on taking an off-the-shelf deep learning approach, and applying it to the dataset in question to explore the performance. This will be compared with standard statistical approaches (e.g. Kriging and hierarchical Bayesian spatiotemporal models) to understand the relative advantages and disadvantages of this approach both in accuracy of prediction and computational time. Based on these initial findings and limitations of the approach, we will consider additional changes to the architecture including software optimisations (including GPU programming) or indeed the development of a different approach which will naturally extend into a full PhD project should the student decide to pursue this. 

​Teaser Project 2:  

​The second teaser project is strongly rooted in uncertainty quantification, i.e. understanding how certain we should be about the model’s predictions. AI-based approaches while regularly delivering high accuracy often lack the strong probabilistic frameworks to give good uncertainty quantification. To do this, we will employ a mixture of approaches from Monte Carlo based and variational inference approaches leveraging the fast inference time of AI models, to emulation-based approaches. Given the data fusion-based pipelines we will be developing this will require understanding both the uncertainty induced by the model and the uncertainty in the observations themselves. Computationally this is quite intensive, and therefore part of this teaser project will be understanding this complexity and optimising it, both through computational means (i.e. GPU coding, HPC etc etc) and through statistical techniques to most efficiently use computational resources (and limit their environmental impact).​ 

References & Further Reading

Allen, A., Markou, S., Tebbutt, W., Requeima, J., Bruinsma, W. P., Andersson, T. R., … & Turner, R. E. (2025). End-to-end data-driven weather prediction. Nature, 641(8065), 1172-1179. 10.1038/s41586-025-08897-0 

Andersson, T. R. et al. (2021) Seasonal Arctic sea ice forecasting with probabilistic deep learning. Nature Communications, 12, 5124. (doi: 10.1038/s41467-021-25257-4) (PMID:34446701) (PMCID:PMC8390499) 

Colombo, P., Miller, C., Yang, X., O’Donnell, R., & Maranzano, P. (2025). Warped multifidelity Gaussian processes for data fusion of skewed environmental data. Journal of the Royal Statistical Society Series C: Applied Statistics, 74(3), 844-865. 10.1093/jrsssc/qlaf003 

Wilkie, C. J., Miller, C. A., Scott, E. M., O’Donnell, R. A., Hunter, P. D., Spyrakos, E., & Tyler, A. N. (2019). Nonparametric statistical downscaling for the fusion of data of different spatiotemporal support. Environmetrics, 30(3), e2549. 10.1002/env.2549 

Forth Environmental Resilience Array (Forth-ERA) project and digital observatory, University of Stirling 

MOT4Rivers: Monitoring, Modelling and Mitigating Pollution Impacts in a Changing World: Science and Tools for Tomorrow’s Rivers 

SenseH2O: a scalable, integrated systems-based approach to monitoring water quality from headwaters to river outlets 

Overview and Background

Antarctic ice shelves regulate the discharge of grounded ice into the ocean, and their collapse can accelerate sea level rise (Berthier et al., 2012). The Larsen C Ice Shelf (LCIS) in particular is thought to be vulnerable to surface melt and hydrofracture, processes which may lead to eventual collapse and which are controlled by the firn layer’s ability to store and refreeze meltwater. Current firn models simplify these processes to reduce computational demand (e.g. Verjans et al., 2019), but this introduces uncertainty and risks misrepresenting thresholds for saturation and collapse. This project will exploit GPU acceleration to test and improve meltwater physics in the Community Firn Model (CFM, Stevens et al., 2020) against field and satellite observations and use the developed model to simulate the evolution of the firn layer on the LCIS under predicted climate warming. By resolving these processes at scale, the project will deliver new predictions of LCIS stability under future climate forcing. 

Methodology and Objectives

​​This PhD project integrates novel field observations (Hubbard et al., 2016), satellite data, and firn modelling to address a key uncertainty in Antarctic climate science: when will the firn layer of the Larsen C Ice Shelf (LCIS) saturate, and what are the consequences for its stability? The programme begins with two exploratory projects of approximately six months each, both centred on the Community Firn Model (CFM). These projects will provide complementary experience in model physics and high-performance computation. Following this training year, the student will select one pathway to pursue in depth from Year 2 onwards, ultimately delivering a substantive contribution to predicting ice-shelf vulnerability. 

​Teaser Project 1: Advancing meltwater physics in the Community Firn Model 
Firn modelling in high-melt environments is fundamentally limited by how liquid water processes are represented. Meltwater percolation, refreezing, and the possible development of firn aquifers are central to whether surface melt is buffered or contributes directly to destabilisation. Current CFM schemes necessarily simplify these processes, often assuming one-dimensional percolation and instantaneous refreezing, which can underestimate storage depth and persistence. 

​This project will benchmark the CFM’s existing meltwater parameterisations against borehole-derived density profiles and refrozen ice layers sampled during field campaigns on LCIS, together with satellite observations of meltwater ponding (e.g. Corr et al., 2022) and surface elevation change. The student will then develop an inventory of physically more complete options, such as multi-phase flow and coupled thermodynamics which could be implemented within the CFM framework using GPU acceleration to reduce the computational penalty of these enhancements, enabling higher spatial and temporal resolution 

​Teaser Project 2: GPU-accelerated simulations of firn evolution under extreme melt 
The second project develops expertise in scalable Earth system modelling by implementing the CFM on GPU-enabled high-performance clusters. Starting from borehole-constrained initial conditions, the model will be forced with outputs from regional climate models covering the past two decades. Particular emphasis will be placed on extreme melt events associated with atmospheric rivers, which have played a central role in past ice-shelf collapse (Wille et al., 2022). 

​The student will design sensitivity experiments to explore firn response to a range of forcing scenarios, including anomalously warm summers, rainfall events, and multi-year accumulation variability. By exploiting GPU acceleration, simulations will be scaled to high spatial resolution across LCIS, and large ensembles will be run to explore uncertainty. These experiments will provide new insight into the thresholds governing firn saturation and the role of rare but intense events in destabilising ice shelves. 

​Beyond Year 1: 
At the end of Year 1, the student will select a pathway for further development. If Project 1 is chosen, the PhD will concentrate on improving the physics of meltwater flow in the CFM and integrating field and satellite observations to constrain model predictions. If Project 2 is chosen, the emphasis will be on producing GPU-accelerated projections of firn evolution and collapse risk to 2100 under multiple climate scenarios and will include an assessment of uncertainty in predictions. 

​Long-term objectives 
Whichever pathway is selected, the overarching aims of the PhD are to: 

• ​Advance the representation of firn processes in high-melt Antarctic environments. 
• ​Develop innovative methods for combining borehole and satellite data with physically based modelling. 
• ​Exploit exascale computing to enable continent-scale, ensemble firn simulations with improved physics. 
• ​Provide new projections of Antarctic ice shelf vulnerability, ultimately contributing to sea-level rise assessments

References & Further Reading

Berthier, E., Scambos, T. A., & Shuman, C. A. (2012). Mass loss of Larsen B tributary glaciers (AP) unabated since 2002. Geophysical Research Letters, 39, 6. 

Corr, D., Leeson, A., McMillan, M., Zhang, C., and Barnes, T.: An inventory of supraglacial lakes and channels across the West Antarctic Ice Sheet, Earth Syst. Sci. Data, 14, 209–228, https://doi.org/10.5194/essd-14-209-2022, 2022. 

Wille, J.D., Favier, V., Jourdain, N.C. et al. (2022) Intense atmospheric rivers can weaken ice shelf stability at the Antarctic Peninsula. Commun Earth Environ 3, 90. https://doi.org/10.1038/s43247-022-00422-9 

Hubbard, B., Luckman, A., Ashmore, D. et al. Massive subsurface ice formed by refreezing of ice-shelf melt ponds. Nat Commun 7, 11897 (2016). https://doi.org/10.1038/ncomms11897 

Stevens, C. M., Verjans, V., Lundin, J. M. D., Kahle, E. C., Horlings, A. N., Horlings, B. I., and Waddington, E. D.: The Community Firn Model (CFM) v1.0, Geosci. Model Dev., 13, 4355–4377, https://doi.org/10.5194/gmd-13-4355-2020, 2020. 

Verjans, V., Leeson, A. A., Stevens, C. M., MacFerrin, M., Noël, B., and van den Broeke, M. R.: Development of physically based liquid water schemes for Greenland firn-densification models, The Cryosphere, 13, 1819–1842, https://doi.org/10.5194/tc-13-1819-2019, 2019. 

​

Overview and Background

​Rogue waves, exceptionally high ocean waves, whose height exceeds twice the significant wave height, are rare, short-lived events that pose serious risks to shipping, fishing, and maritime infrastructure, including offshore platforms and wind turbines. Understanding their formation and improving prediction are essential for safe marine operations.  

​Despite advances in theoretical and experimental studies, the physical mechanisms driving rogue wave formation in real seas remain poorly understood, making prediction challenging. This PhD project aims to address these gaps by analysing extensive field data, developing advanced non-linear dynamic techniques, and utilising high-performance computing. The aim is to improve understanding of rogue wave dynamics and enhance forecasting capabilities, thereby contributing to the safety and resilience of marine operations. 

​[115] 

Methodology and Objectives

​The PhD project will address the following questions: How can data processing and modelling be accelerated to enable non-linear analysis and modelling with higher spatial and temporal resolution? Which of the sea state parameters predicted by existing operational wave models are useful for detecting the formation of rogue waves? How do the formation and predictability of rogue waves depend on physical conditions?  

​Teaser 1: 

​This data-intensive project aims to accelerate novel time-localised analysis methods to investigate physical mechanisms underlying rogue waves and predict their occurrence.  

​O1: Exploit GPU Accelerated Computing to parallelise algorithms for time-localised phase coherence and couplings between waves recorded in many spatial points, enabling scaling to higher-resolution and near real time analysis.  

​O2: Isolate the mechanisms leading to the formation of rogue waves using algorithms developed in O1. 

​O3: Develop in-situ feature detection for automated analyses exploiting GPU and assess the relationship between the occurrence of rogue waves and their characteristics from time-series measured under different physical conditions. 

​O4: Develop a time-series-based prediction modelling approach, using the relationships identified in O2-3 and assess its ability to predict the occurrence of rogue waves. 

​Methods: 

​The numerical modelling and algorithms for time-series analysis will exploit GPU Accelerated Computing; exascale will then allow near real-time practical applications. The Multiscale Oscillatory Dynamics Analysis (MODA) toolbox for non-linear and time-localised phenomena in time-series (e.g. phase coherence, coupling and wave energy exchange [3&4]) will be parallelised and used to identify rogue wave mechanisms. Automated pattern analysis and feature engineering will be applied using Tangent to detect anomalous sea surface elevations and enable an easily deployable computationally light forecasting solution using processed data. The methods will be first applied to laboratory data (e.g. [1]) and then to publicly available field measurements (e.g. Free Ocean Wave Dataset with more than 1.4 billion wave measurements). The newly developed prediction modelling approach will be systematically validated with measured data.  

​Teaser 2: 

​This is a data-intensive project focused on the computational optimisation of time series analyses for dynamic systems and the relationship between rogue wave properties and environmental conditions. 

​O1: Assess the current performance of the numerical tools included in MODA and Tangent in terms of their relevance for detecting the mechanisms of rogue waves and their computational efficiency. 

​O2: Optimise the algorithms of the tools identified in O1 with multiple GPU to improve time to improve computation time and experimental throughput, enabling large-scale ensemble time-series analyses. 

​O3: Develop and apply a GPU version of MODA to field measured data to isolate mechanisms that lead to the formation of rogue waves.   

​O4: Assess the relationship between the occurrence of rogue waves and concurrent ocean and atmospheric data.  

​Methods: 

​The Multiscale Oscillatory Dynamics Analysis (MODA) toolbox offers several high-order methods for time-series analysis, some based on wavelets. The high computational demands of uncertainty evaluation methods limits their use for operational purposes.  Optimised algorithms, GPU-acceleration and Exascale facilities will enable higher resolution and practical applications. MODA will identify the mechanisms underlying rogue wave formation using field measured time-series of surface elevations (e.g. Free Ocean Wave Dataset). The concurrent environmental data (e.g. surface ocean currents, wind and atmospheric pressure) will be collated either from field measurements or from the operational forecast models provided by meteorological offices. The correlation between the occurrence of rogue waves and environmental parameters, as well as ‘causal’ relationships between the identified mechanisms and the environmental conditions, will be investigated using the Tangent modelling engine which can be incorporated into predictions in the future.  

​[577] 

References & Further Reading

1. Luxmoore, J.F., Ilic, S. and Mori, N., 2019. On kurtosis and extreme waves in crossing directional seas: a laboratory experiment. Journal of Fluid Mechanics, 876, pp.792-817. 
2. Mori N., Waseda, T., Chabchoub A.(eds.) (2023) Science and Engineering of Freak Waves, Elsevier (https://doi.org/10.1016/C2021-0-01205-0). 
3. Newman, J., Pidde, A. and Stefanovska, A., 2021. Defining the wavelet bispectrum. Applied and Computational Harmonic Analysis, 51, pp.171-224. 
4. Stankovski, T., Pereira, T., McClintock, P.V. and Stefanovska, A., 2017. Coupling functions: universal insights into dynamical interaction mechanisms. Reviews of Modern Physics, 89(4), p.045001. 
5. Yang X., Rahmani H., Black S., Williams B. M. Weakly supervised co-training with swapping assignments for semantic segmentation. In European Conference on Computer Vision 2025 (pp. 459-478). Springer, Cham. 
6. Jiang Z., Rahmani H., Black S., Williams B. M. A probabilistic attention model with occlusion-aware texture regression for 3D hand reconstruction from a single RGB image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition 2023 (pp. 758-767). 
7. Jiang Z., Rahmani H., Angelov P., Black S., Williams B. M. Graph-context attention networks for size-varied deep graph matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition 2022 (pp. 2343-2352). 

Overview and Background

​Understanding the stability of ecosystems and how they are impacted by climate and land use change can allow us to identify sites where biodiversity loss will occur and help to direct policymakers in mitigation efforts. Our current digital twin of plant biodiversity – https://github.com/EcoJulia/EcoSISTEM.jl – provides functionality for simulating species through processes of competition, reproduction, dispersal and death, as well as environmental changes in climate and habitat, but it would benefit from enhancement in several areas. This project would target improving the speed of model runs and the feasible scale of model simulations to allow stochastic modelling to quantify uncertainty, to develop techniques for scalable inference of missing parameters, and to allow more complex models. It would do this through two approaches – on the one hand porting the code to run on GPUs for higher computational efficiency, and on the other hand applying techniques such as mesh refinement and partitioning to allow models to run at high resolution only where required by the ecosystem complexity. 

Methodology & Objectives

​Teaser Project 1 Objectives: GPU: Port core EcoSISTEM code to GPU 

​This project will analyse the core CPU routines in EcoSISTEM, and port them to GPU. This will use packages from the JuliaGPU ecosystem, that provide a relatively easy user interface to NVIDIA and AMD GPUs, which are available on EPCC’s HPC system that the student will have access to. The main branch of the EcoSISTEM code is already efficiently parallelised for CPUs, and a preliminary assessment has suggested that the porting task should be feasible within a teaser project. This teaser project can be extended in a variety of ways to a full PhD: 

​On the one hand, once the GPU port speed-ups have been realised, the student can add major new components to EcoSISTEM. For instance, the student can investigate uncertainty quantification and parameter inference techniques within EcoSISTEM. 

​On the other hand, there is a more sophisticated (dev) development branch of EcoSISTEM that is not currently well optimised but allows greater flexibility of how interactions can occur between components of the model. Porting this to GPUs will be a significantly harder task, but will allow richer interactions to be more easily modelled between ecosystem components. 

​Teaser Project 2 Objectives: Scalability improvements: mesh refinement and partitioning 

​Different types of ecosystem require different scales of spatial resolution to model them adequately. For example, hedgerow and rivers are linear features that may be very species dense compared to the surrounding terrain. Currently EcoSISTEM uses a uniform mesh size, which means that the whole model has the same spatial resolution. To be able to increase the fidelity of the model in a scalable way, it must increase the spatial resolution only where needed. This teaser project would begin by capturing the requirements for non-uniform meshing and prototype an implementation, possibly in a simple proxy code.  

​With a non-uniform mesh, scaling to many thousands of processes require load balancing techniques, either static or dynamic, depending on the use case. The teaser project will investigate the suitability of existing mesh partitioners and adaptive meshing techniques, and how they can interface with the existing Julia code.  

References & Further Reading

1. ​​​Digital twins of the natural environment​​  
2. Dynamic virtual ecosystems as a tool for detecting large-scale responses of biodiversity to environmental and land-use change 
3. Effective extensible programming: Unleashing Julia on GPUs 
4. Strong phylogenetic signals in global plant bioclimatic envelopes 

Overview and Background

The drainage of supraglacial lakes plays an important role in modulating ice velocity, and thus the mass balance, of the Greenland Ice Sheet. To date, research has primarily focused on drainage events and their impacts on ice motion during the summer melt season, but recent research has shown that drainage events during winter can also affect ice dynamics. However, current approaches are limited to a single season and/or 1–2 satellite sensors, limiting observations when year-round and multi-sensor monitoring are required to fully understand lake processes and their impacts. This PhD will utilise petabytes of available Earth Observation data and exascale computing to perform ice-sheet–scale analysis of year-round supraglacial lake drainage events on the Greenland Ice Sheet and produce scalable workflows that can be used to assess the impact of lake drainage events in other glaciological environments. 

Methodology and Objectives

This PhD project will advance capabilities in the detection of supraglacial lake drainage events on the Greenland Ice Sheet (GrIS) and assess the impact of these drainage events on ice dynamics. The PhD will commence with two exploratory projects (~6 months each), providing complementary experience in deploying exascale compute and performing big-data analysis. 

Teaser Project 1: Near-real-time, automated, multi-sensor, year-round monitoring of supraglacial lake drainage 

Current assessments of supraglacial lake drainage on the GrIS are largely restricted to a single season and/or satellite sensor, particularly during winter, where existing approaches remain constrained to low-volume pipelines that analyse single orbits and provide limited temporal and spatial coverage. However, the plethora of remotely sensed imagery now available provides the opportunity to detect supraglacial lake drainage events at high temporal and spatial resolution across the entire ice sheet in near real-time through exascale big-data analysis. This project will scale a SAR-based methodology for supraglacial lake drainage detection to ice-sheet-wide monitoring, using a high-data-volume approach to achieve near-daily temporal resolution by leveraging all available orbits and both C- and L-band SAR. The project will exploit access to exascale compute to deploy GPU-accelerated machine learning methods (e.g., convnets or Unets), able to extract spatiotemporal patterns from large and complex volumes of multi-frequency inputs, to support robust, scalable detection of drainage events across diverse glaciological settings. Validation and training will draw on timestamped ArcticDEM strips and ICESat-2 altimetry, ensuring reliable accuracy assessments at ice-sheet scale, using methods for (cross-) validation across space and time (e.g., Otto et al. 2024). 

Teaser Project 2: Near-real-time evaluation of the impact of supraglacial lake drainage on the GrIS 

As supraglacial lakes form increasingly farther inland under increasing atmospheric temperatures, the year-round impact of their drainage events on ice dynamics is not yet well understood. Recent research has shown that winter drainage events are numerous, often occur as “cascade events”, and can result in short-term increases in ice velocity (Dean et al., under review). However, a systematic, year-round analysis of the impact of supraglacial lake drainage events on ice dynamics across the GrIS has not yet been undertaken. This project will set up a pipeline on a sub-region of the ice sheet to analyse the impact of drainage events on ice dynamics in real-time, allowing later scaling up to an ice-sheet-scale. Access to exascale compute will enable comparison of the database of drainage events created for Project 1 with climate and glaciological data (e.g., temperature, precipitation, surface energy balance, ice surface velocity, ice thickness, and bed elevation). By applying scalable statistical methods like changepoint analysis (offline detection), statistical process monitoring (online surveillance), and anomaly detection using deep learning methods, the impact of lake drainage events will be evaluated in real-time and assessed over a range of timescales. Additionally, the trained DNNs from Project 1 can be used for dimensionality reduction and process monitoring based on data depths, which can indicate further sources/reasons for detected changes (Malinovskaya et al. 2024). 

Long-term pathway and objectives 

At the end of Year 1, the student will select a pathway for further development. If Project 1 is chosen, the PhD will focus on leveraging additional compute resources and machine learning methods to scale the analysis to include additional, larger, and multi-modal data sources in our drainage detection methodology (such as optical to enhance summer detection) and deploying our method over other regions, such as Antarctic ice shelves, where real-time monitoring of supraglacial lake drainage events and cascades could be a useful precursor for forecasting ice-shelf disintegration (Banwell et al., 2013). If Project 2 is chosen, the PhD will focus on exploiting access to exascale compute and machine learning methods to scale the analysis to an ice-sheet-wide scale, enabling near-real-time assessment of the impact of supraglacial lake drainage events on the GrIS and potentially other glaciological environments.  

For either pathway, the aims of the PhD are to: 

• Advance understanding of year-round supraglacial lake drainage events on the Greenland Ice Sheet. 

• Exploit exascale compute resources and machine learning to enable real-time detection of supraglacial lake drainage events at an unprecedented scale and assess their impacts. 

• Produce scalable workflows that can be applied to supraglacial lake drainage events in other glaciological regions. 

References & Further Reading

Banwell, A. et al., (2013), Breakup of the Larsen B Ice Shelf triggered by chain reaction drainage of supraglacial lakes, Geophysical Research Letters, 40, 22, 5872-5876, doi.org/10.1002/2013GL057694 

Christoffersen, P. et al., (2018), Cascading lake drainage on the Greenland Ice Sheet triggered by tensile shock and fracture, Nature Communications, 9, 1064, doi.org/10.1038/s41467-018-03420-8 

Dean, C. et al. (under review), A decade of winter supraglacial lake drainage across Northeast Greenland using C-band SAR, The Cryosphere Discussions 

Dunmire, D. et al., (2025), Greenland Ice Sheet wide supraglacial lake evolution and dynamics: insights from the 2018 and 2019 melt seasons, Earth and Space Science, 12, 2, doi.org/10.1029/2024EA003793 

Leeson, A. et al., (2015), Supraglacial lakes on the Greenland ice sheet advance inland under warming climate, Nature Climate Change, 5, 51–55, doi.org/10.1038/nclimate2463 

Malinovskaya, A., Mozharovskyi, P., & Otto, P. (2024). Statistical process monitoring of artificial neural networks. Technometrics, 66(1), 104-117, doi.org/10.1080/00401706.2023.2239886 

Miles, K., et al., (2017), Toward monitoring surface and subsurface lakes on the Greenland Ice Sheet using Sentinel-1 SAR and Landsat-8 OLI imagery, Frontiers in Earth Science, 5, doi.org/10.3389/feart.2017.00058 

Otto, P., Fassò, A., & Maranzano, P. (2024). A review of regularised estimation methods and cross-validation in spatiotemporal statistics. Statistic Surveys, 18, 299-340, doi.org/10.1214/24-SS150 

Overview and Background

​Around 12,900 years ago, a rapid cooling of the climate triggered widespread glaciation across the Northern Hemisphere. This cold phase, known as the Younger Dryas, ended abruptly with a dramatic warming. This event provides a valuable natural experiment for understanding how Earth systems respond to sudden shifts in climate, a question highly relevant to today’s ongoing climate change. 

​Scotland preserves one of the most detailed geological and geomorphological records of Younger Dryas glaciation anywhere in the world. This project will leverage these records by developing a GPU-accelerated computational workflow that integrates numerical simulations of glaciation with glaciological, geological, and geomorphological datasets. The aim is to reconstruct and test interactions among glaciation, climate, and topography during the Younger Dryas, improving our understanding of Earth’s response to abrupt climate change. 

Methodology and Objectives

​​Although previous research has produced a wealth of observations on the Younger Dryas glaciation in Scotland, efforts to integrating them with numerical modelling remain limited. This is largely because most model-data integration methods require running a large group of simulations, while traditional glaciation simulations are computationally demanding. A recently developed GPU-based ice flow model offers an opportunity to address this challenge. By applying Physics-Informed Machine Learning techniques, the model leverages the power of modern GPU hardware to efficiently simulate glaciation. This project will build on this model and focus on exploring model-data integration methods on exascale computers. 

​The overarching goal of this PhD is to develop a robust and scalable model-data integration workflow that can integrate the GPU-based model with glaciological, geomorphological, and geological datasets in Scotland. Depending on their interests, the student may also pursue further model developments, such as optimizing the workflow for efficient multi-GPU simulations or coupling it with other Earth system models, including climate and landscape evolution models. 

​Teaser Project 1: Numerical inversion of past climate conditions in Scotland 

​Although the glaciation event in Scotland broadly coincides with the global Younger Dryas period, evidence is emerging that the exact timing and magnitude of climate shift in Scotland differed from those in other parts of northwestern Europe, and a better understanding of regional climate conditions in Scotland has important implications for understanding the interactions between different Earth systems during abrupt climate changes. The objective of this Teaser Project is to use numerical inversion methodology to infer past climate conditions from mapped ice extent during Younger Dryas glaciation in Scotland. This project will integrate GPU-based simulations of glaciation with ice extent constraints through Markov Chain Monte Carlo method, a computational method that explores many possible scenarios to determine which ones best fit the evidence. The goal is to constrain the climate conditions needed to produce glaciation that are consistent with field observations in Scotland. 

​The long-term pathway of this Teaser Project will focus on optimizing the workflow by 1) integrating additional geomorphological and geological observations (e.g., moraine positions, sedimentary records, and chronology data), and 2) combing glaciation simulations with climate simulations to capture the interaction between ice growth/decay and climate variations.  

​Teaser Project 2: Using data assimilation to reveal the dynamic evolution of glacier systems 

​The evolution of Younger Dryas glaciation in Scotland provides a valuable natural experiment for understanding how glacier systems respond to rapid climate fluctuations. Field evidence has helped reconstruct the general pattern and timing of ice advance and retreat in some areas, but these records remain spatially patchy and only partially time-resolved. This Teaser Project aims to combine such sparse chronological observations with physically based ice flow models by using data assimilation techniques (e.g., Ensemble Kalman Filter). Data assimilation is a method that updates a glacier simulation over time by combining the model’s predictions with observations, such as exposure dating, to produce the most likely evolution of the glacier. By applying this approach, the project aims to develop a framework capable of reconstructing the timing, extent, and dynamics of glaciation in unprecedented detail.  

​The long-term pathway of this Teaser Project is to create a transferable data–model integration workflow that can assimilate diverse glaciological, geological, and geomorphological datasets, ultimately improving our understanding of the dynamic response of ice flow in Scotland during the Younger Dryas. 

​References & Further Reading

https://www.antarcticglaciers.org/glacial-geology/british-irish-ice-sheet/younger-dryas-loch-lomond-stadial/ 

https://www.antarcticglaciers.org/glacial-geology/british-irish-ice-sheet/younger-dryas-loch-lomond-stadial/the-loch-lomond-stadial/ 

Leger, T. P. M., Jouvet, G., Kamleitner, S., Mey, J., Herman, F., Finley, B. D., Ivy-Ochs, S., Vieli, A., Henz, A., & Nussbaumer, S. U. (2025). A data-consistent model of the last glaciation in the Alps achieved with physics-driven AI. Nature Communications, 16(1), 848. https://doi.org/10.1038/s41467-025-56168-3 

Overview and Background

​Reclaimed lands (e.g., post-mining and post-industrial sites) are expanding globally and are particularly susceptible to colonisation by invasive plant species, which may cause negative effects on restoration, biodiversity, and ecosystem functions. We propose to evaluate and extend AlphaEarth Foundations, i.e., Google’s Satellite Embedding V1 (annual, 10m by 10m, 64-Dimension), for fine-resolution identification of invasive plant species in reclaimed areas worldwide and for tracking temporal change in species composition and spread. AlphaEarth encodes multi-sensor Earth Observation (EO) time series into consistent, analysis-ready embeddings during 2017-2024, supporting scalable classification and efficient change detection. We will combine embedding-based models with climate and reclamation histories to identify key drivers, quantify uncertainty, and align the results with ExaGEO’s “exascale model & big-data coupling” platform.  

Methodology and Objectives

​Data & Methods Used: 

​We will use Google Earth Engine’s Satellite Embedding V1 as the core feature space; embeddings are unit-length, consistent over years, and produced by AlphaEarth Foundations (Brown, Christopher F., et al., 2025), i.e., multi-modal assimilation across optical, radar, and LiDAR, facilitating both classification and dot-product/angle-based change metrics. We will (i) assemble global reclaimed-area masks, beginning with open global-scale mining polygons (v2) (Maus, Victor, et al., 2022) and complementary sources; (ii) compile invasive plant labels from the Global Invasive Species Database (GISD) linked to the Global Biodiversity Information Facility (GBIF) occurrences; (iii) develop label-efficient methods (e.g., positive-unlabelled, semi-supervised, and weak-supervision with quality controls); (iv) build temporal transformers over annual embeddings for trend/change analysis; and (v) perform GPU-accelerated distributed training/inference with rigorous uncertainty quantification (e.g., deep ensembles) and domain shift tests across regions and biomes.  

​Teaser Project 1  

​Suggested headings: Fine-resolution invasive species mapping on global reclaimed lands  

​Objectives:  

1. ​Benchmark AlphaEarth embeddings vs. standard EO features for species-level classification on reclaimed sites (start with mining areas, then generalise), measuring macro-F1, calibration, and cross-region transfer. 
2. ​Build a label-efficient pipeline that fuses GISD species lists with spatially filtered GBIF occurrences and negative sampling within reclaimed buffers; evaluate sensitivity to sampling bias. 
3. ​Develop multi-scale embedding aggregation (tile/patch pooling and linear composability) for species distinguishability; compare tree-based methods vs. shallow nets on the 64-Dimension space for compute efficiency.
4. Produce regional probability maps (e.g., Scotland, Ruhr, Appalachia, Hong Kong, Shantou, and Jiangsu) with uncertainty maps to guide validation and management.  

​How it becomes a full PhD: Global expansion across reclaimed typologies (e.g., mines, landfills, and brownfields), deeper label curation, and active learning with end-user feedback; deliver a reproducible mapping stack and evaluation framework.  

​Teaser Project 2  

​Suggested headings: Temporal dynamics and early detection of invasive spread post-reclamation 

​Objectives: 

1. ​Use annual embeddings (2017–2024) to compute intra-site temporal similarity (e.g., cosine/dot-product shifts) and detect emergence or range expansion of invasive species; and integrate climate covariates with Dr Zhang’s modelling to attribute underlying drivers.  
2. Train temporal sequence models (temporal transformers on annual 64-Dimension vectors) to predict next-year probabilities and time-to-detection, with quantified uncertainty. 
3. Quantify management impacts by comparing trajectories across reclamation strategies (where metadata exist) and by disentangling the influence of climate anomalies vs. anthropogenic disturbance signals.
4. ​Deliver global change products and a watchlist of high-risk sites for early intervention. 

​How it becomes a full PhD: Scale temporal modelling globally, generalise to additional taxa, and operationalise early-warning thresholds with end-users. 

References & Further Reading

​Brown, Christopher F., et al. “Alphaearth foundations: An embedding field model for accurate and efficient global mapping from sparse label data.” arXiv preprint arXiv:2507.22291 (2025). 

​Maus, Victor; da Silva, Dieison M; Gutschlhofer, Jakob; da Rosa, Robson; Giljum, Stefan; Gass, Sidnei L B; Luckeneder, Sebastian; Lieber, Mirko; McCallum, Ian (2022): Global-scale mining polygons (Version 2) [dataset]. PANGAEA, https://doi.org/10.1594/PANGAEA.942325 

Overview and Background

​Landslides and debris flows threaten human life, infrastructure, and economies worldwide, especially as climate change drives more extreme rainfall events. Timely prediction of such mass movements is therefore crucial for disaster risk reduction. Achieving this requires combining accurate physics-based modelling of granular flows with rich observational data: for example, remote-sensing change-detection of slope deformation. Recent advances in data-driven modelling and physics-informed machine learning offer new opportunities to integrate physical laws with observational and simulation data. This project aims to establish a unified hybrid modelling framework that bridges high-fidelity particle-scale simulations and large-scale field data for improved prediction of landslide initiation and runout, thereby laying the foundation for future digital twinning of landslides.  

Methodology and Objectives

​The overarching goal of this PhD project is to develop a hybrid modelling framework that combines data-driven learning with physics-based constitutive modelling to describe the transition between the solid-like and fluid-like behaviour, which is critical for predicting slope failure and runout, and to bridge the gap between microscale simulations and macroscale landslide observations.  

Teaser Project (TP) 1: Discrete Element Simulations of Slope Flows

This project focuses on performing high-fidelity discrete element method (DEM) simulations to capture the failure, flow, and deposition processes in granular slopes under different inclinations. Simulations will be conducted using the open-source LAMMPS software, with GPU acceleration implemented to enhance computational efficiency. The objective is to generate a comprehensive dataset describing particle-scale kinematics, contact forces, and evolving stress fields during slope instability. By varying slope angles and material parameters, the study will investigate the onset of failure, the transition from solid-like to fluid-like flow, and the subsequent deposition patterns. These DEM results will provide detailed micro-mechanical insights and serve as training data for subsequent data-driven model development.

A GPU-accelerated solver will be developed to optimize LAMMPS for large-scale granular simulations. The solver will utilize domain decomposition and parallel computation to handle millions of particles efficiently. The output data—velocity, stress, strain, and microstructure—will be analysed to characterize flow regime transitions. This will enable formulation of rheological indicators linking particle-scale dynamics to continuum measures of deformation.

Teaser Project 2: Physics-Informed Machine Learning for Granular Rheology

The second project will apply physics-informed neural networks (PINNs) or sparse regression techniques to learn rheological models for granular flow where constitutive relations are already known, such as the μ(I) rheology for steady-state flow. The objective is to test and validate the physics-informed learning methodology by comparing the discovered models against the analytical forms of these known relationships.

Synthetic datasets will be generated using controlled numerical experiments from TP1 to train the PINNs or sparse regression models. By incorporating physical constraints such as positive energy dissipation, the learned models are expected to exhibit improved generalization and interpretability. This project will demonstrate how hybrid modelling can faithfully recover known constitutive relationships while providing a robust foundation for future discovery of new rheological forms from experimental and DEM data.

Together, these teaser projects establish a methodological framework for subsequent years of research, which will extend toward modelling large-scale landslides by learning with data from both DEM simulations and field observations, coupling learned rheologies with continuum-scale solvers and validating against field data. The integration of physics-based and data-driven models will ultimately enable prediction of landslide initiation and runout with improved accuracy and computational efficiency. 

References & Further Reading

1. Iverson, R. M. The physics of debris flows. Reviews of Geophysics 35, 245 296 (1997). 
2. Forterre, Y. & Pouliquen, O. Flows of Dense Granular Media. Annual Review of Fluid Mechanics 40, (2008). 
3. Chialvo, S., Sun, J. & Sundaresan, S. Bridging the rheology of granular flows in three regimes. Phys Rev E 85, 021305 021305 (2012). 
4. Raissi, M., Perdikaris, P. & Karniadakis, G. E. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. J. Comput. Phys. 378, 686–707 (2019). 

Overview and Background

Climate change is showing tangible consequences on our environment. As glaciers retreat and rainfalls intensify, the frequency and scale of landslides and debris flows are rising worldwide. Their destructive power extends beyond the areas where these phenomena are initiated: unstable masses may travel several kilometres, depending on the evolving characteristics of the solids involved and their interaction with water; when masses plunge into large water bodies, they can unleash catastrophic tsunamis. Yet, our understanding and prediction of granular flows are limited by knowledge gaps in how fluids and solids interact. Key challenges include unravelling how grain shape and breakage affect flow mobility, and how mass, momentum, and energy are transferred during violent impacts. Using the GPU-accelerated multiphase solver MFIX-Exa, this project will pioneer next-generation simulations and physics-based laws to transform landslide hazard forecasts in a changing climate. 

Methodology and Objectives

Teaser Project 1: When Earth Hits Water — Geophysical Flows Triggering Tsunamis 

Objective: During the initial six-month project, the focus will be on developing and validating a simplified GPU-accelerated Volume of Fluid (VoF) module within the MFIX-Exa framework to represent two-phase air–water interactions during a solid-body impact. This will establish the numerical infrastructure and performance benchmarks needed to later include granular particles. Using canonical test cases (e.g., water entry of wedges, deformable intrusions), we will assess how impact geometry and velocity control wave generation and energy transfer. 

This short-term work will lay the foundation for the full PhD, which will extend the solver to fully three-phase (solid–gas–liquid) conditions, include granular rheology and pore-pressure coupling, and simulate natural examples such as pyroclastic flows entering the sea. The long-term goal is to derive physics-based coupling laws that can inform exascale tsunami forecasting and hazard models. 

Teaser Project 2: Evolving Grain Size and Shape in Geophysical Granular Flows 

Objective: The first six months will focus on implementing and testing a basic bonded-sphere representation of irregular grains in MFIX-Exa, without breakage. The aim is to quantify how initial particle shape (aspect ratio, angularity) modifies packing density and stress transmission under simple shear. Benchmark simulations and comparisons with existing experimental datasets will be used to verify the new contact model and establish computational efficiency on GPUs. 

This work forms the foundation for a PhD that would progressively incorporate fracture and breakage physics, enabling the grain size and shape to evolve dynamically. Later stages would explore how fragmentation alters permeability, pore-fluid pressure response, and bulk rheology in flows such as landslides, debris avalanches, and pyroclastic currents, ultimately yielding improved continuum closures for natural-hazard prediction. 

References & Further Reading

Svennevig, Kristian, et al. “A rockslide-generated tsunami in a Greenland fjord rang Earth for 9 days.” Science 385.6714 (2024): 1196-1205. 

https://www.exascaleproject.org/research-project/mfix-exa/ 

https://github.com/NREL/BDEM.git 

Musser, J., Almgren, A. S., Fullmer, W. D., Antepara, O., Bell, J. B., Blaschke, J., … & Syamlal, M. (2022). MFIX-Exa: A path toward exascale CFD-DEM simulations. The International Journal of High Performance Computing Applications, 36(1), 40-58. 

Lu, L., Gao, X., Shahnam, M., & Rogers, W. A. (2021). Simulations of biomass pyrolysis using glued-sphere CFD-DEM with 3-D intra-particle models. Chemical Engineering Journal, 419, 129564. 

Overview and Background

​Machine learning (ML) lets us generate large ensemble simulations of extreme weather events quickly and efficiently. This can help assess the risk of floods, which constitute almost half of the world’s weather-related disasters. Rather than physical laws, ML models are determined by training data, which are at coarser resolutions than flood impact models. Scale transformations, through the climate-to-impacts processing pipeline, add uncertainties and highlight the challenge of representing extreme events. Are ML models realistic enough for vital applications in disaster risk reduction in present and future climates, given these challenges? You will address this with state-of-art global ML models and key industry partners JBA and NVIDIA to test the physical and statistical fidelity of extreme weather simulations of flooding and climate risk. 

Methodology and Objectives

​This PhD will use statistical machine learning to interrogate the data produced by an AI-driven weather simulation that generates large ensembles of synthetic events for flood impact analysis under different climate conditions. The full pipeline, which has global capability, uses NVIDIA’s ‘Earth-2’ ML platform to transform (i) a combination of 75 single- and pressure-level atmospheric variables at 0.25 degrees resolution to (ii) precipitation fields, which are aggregated over river basins, to (iii) drive hydrological models of flood flows at points situated at between ~1 and ~10 km spacing on the river network, and finally (iv) flood impact analysis over variable-resolution spatial grids. Despite this cutting-edge risk analysis capability, there are important uncertainties and, hence, potential to improve the processing chain through increased understanding of the character of the various data layers and their inter-relations, specifically measurement processes, scaling processes, the statistical validity of distributions for extremes, and semantic interpretations and their fit to real-world phenomena. This is crucial since misrepresented processes and states can lead, for example, to prediction biases or omissions of consequential key events, which, in the global flood impact context, can mean loss of life or assets.  

​The PhD will develop and use a range of statistical and machine learning methods with which to interrogate the data in the climate-to-flood risk processing chain and suggest improvements in data representation. The student will work with JBA with input from NVIDIA to fully understand the Earth-2-platform. The PhD will then employ a combination of methods, including latent Gaussian processes (GPs) within a Bayesian inference framework to diagnose data support (that is, the spatial and temporal scales at which quantities are measured or modelled) and scaling relationships; machine learning approaches to characterise and represent dependencies between datasets; and natural language processing (NLP) coupled with ontological hierarchies to explore the meaning of, and relationships between, data layers. The representation of key atmospheric processes and large-scale phenomena – such as storm development, blocking, and energy balance – will be considered to ensure that the statistical and machine-learning analyses remain grounded in physical realism. 

​You will work with data and tools from the Earth-2 suite – the SFNO large ensembles model (Mahesh et al.), downscaling (Mardani et al.), and AFNO ‘diagnostic’ precipitation model (Kurth et al.) – alongside finer scale data, potentially including CEH-GEAR 1km UK rainfall, rain gauges, river flow gauges, and high-resolution flood maps produced by JBA. 

​Teaser Project 1: Data sampling, measurement processes and scale transformation to enhance realism of weather and flood risk simulations 

​Flood risk analysis depends crucially on earlier parts of the processing chain that transforms climate ensembles at coarse resolution into reliable flood simulations, and their interpretation relative to asset and population distributions. You will use the power of latent GP models, coupled with GPU-accelerated, scalable to exascale, compute to disentangle measurement processes (such as spatial discretisation, data model, spatial support, uncertainty) from the signal of interest as a key requirement for improved inter-operability between datasets. You will focus on climate ensembles and their transformation through to precipitation fields and flood extents and impacts using GPs to diagnose the measurement processes at each stage augmented with machine learning to represent the transforms and enable their improvement. This will improve interoperability between data layers and reduce artefactual loss of information, including on extremes, during scaling and transformation.  

​Specific goals include: 

• ​Specification of the practical measurement support for each data layer in the processing chain, allowing its exploitation in subsequent processing.
• Increased understanding of why extreme events are under-represented and how to inject lost structure to recover them to an improved extent.  

​Teaser Project 2: Extracting representations from weather simulations to test against physical phenomena 

​Flood risk analysis depends on the feeding of ensemble datasets through a processing chain to decision-makers who have responsibility for interpretation of the risk information provided by the Earth-2 platform. You will take the climate ensemble datasets ‘as read’ and specify and fit models to characterise the information, and minimise the losses of information that occur, in the processing chain. You will use the tools of ML and large language models (LLMs) to learn latent processes and represent the optimal data inter-operability pathways that are both internally ‘coherent’ and maximise the utility of the information sets provided. This may require interaction with decision-makers, facilitated through JBA, and inversion of the forward look through a learning framework (e.g., with reinforcement learning to optimise data transformations for maximal information retention).    

​Specific goals include: 

• ​Greater understanding of the semantic meaning within datasets that are of most interest to decision-makers.  
• More useful characterisations of the data (e.g., extremes) that can be inverted to optimise the parameters in the processing chain.  

References & Further Reading

Context and links to technical details of the modelling stack: 

Ashcroft, J. et al., 2025, AI weather forecasting: Can it reveal unseen flood risks? JBA Risk Management blog. 

Descriptions of the underlying ML models and example methods: 

Chácon-Montalván, E.A., Atkinson, P.M, Nemeth, C., Taylor, B.M, Moraga, P., 2024, Spatial latent Gaussian modelling with change of support, arXiv, https://arxiv.org/abs/2403.08514. 

Kurth, T. et al., 2023, FourCastNet: accelerating global high-resolution weather forecasting using adaptive Fourier neural operators. In Platform for Advanced Scientific Computing Conference (PASC ’23), 2023, Davos, Switzerland. ACM, https://doi.org/10.1145/3592979.3593412. 

Mahesh, A. et al., 2024, Huge ensembles Part I: design of ensemble weather forecasts using spherical Fourier neural operators, arXiv, https://doi.org/10.48550/arXiv.2408.03100.

Mardani, M. et al., 2025, Residual corrective diffusion modeling for km-scale atmospheric downscaling. Commun. Earth Environ. 6, 124. https://doi.org/10.1038/s43247-025-02042-5. 

Overview and Background

​​The Philippines are one of the most densely populated and economically fastest growing regions in Southeast Asia while they constantly face the threat of natural hazards such as strong climatic (e.g., monsoon) and geological forces (e.g., earthquakes and volcanic activity) in parallel to experiencing anthropogenic alteration (i.e. mining, river management) of the natural landscape. Any signals emerging from these will be communicated across a landscape primarily through their fluvial systems.  

​This project aims to develop novel live-monitoring techniques that bridge process-based modelling of the fluvial and anthropogenic systems with AI-driven data analysis to model fluvial morphology and change across Luzon. These efforts will provide state-of-the-art computational tools to inform policy makers in the Philippines to better adapt to climate change.

Methodology and Objectives

​​Digital shadow of landscape evolution of Luzon (Teaser Project 1); AI-driven prediction of geomorphic change in the Philippines (Teaser Project 2) 

​Methods Used: landscape evolution modelling, AI-enhanced remote sensing, The surface processes models employed in this effort will reach unprecedented metre-scale resolution, hence having high data volume throughput at the TB level. This requires GPU capacities that take advantage of latest flow routing routines adopted on parallelised computational architecture reducing model completion time by up to two orders of magnitude relative to the current generation of landscape evolution models. This enables the use of more efficient inversion schemes as well as lays the groundwork for potential exascale application in the future. 

​Teaser Project 1 Objectives: Landscape evolution models have been essential in developing our insights into interconnectivity of earth systems and surface processes. However, humans now have an unprecedented impact on our landscapes, now moving vast quantities of material across earth surface (e.g., mining). Teaser project 1 will establish the necessary boundary conditions to produce a mid-resolution (500 m / cell) representation of the Quaternary landscape evolution of the island of Luzon. A particular focus will be placed on accurately predicting to a first-order major present-day fluvial erosional and depositional systems including offshore basin stratigraphic records in tropical environments. Novel ensemble Kalman inversion schemes will be employed that leverage GPU capacities to be built in the project to computationally efficiently achieve fast parameter convergence using novel parallelised flow routing routines that lead to present-day geomorphological conditions.     

​Teaser Project 2 Objectives: A key challenge when working with remote sensing (satellite) based datasets is the availability of ‘cloud free’ images to perform high temporal resolution analysis on geomorphic change in river catchments. As a result, there is a bias in analysis of geomorphic change to datasets that have climates that are more conducive to ‘cloud free’ weather. Teaser project 2 will employ an AI approach to fill spatial and temporal data gaps in satellite remote sensing data due to dense cloud coverage. The Philippines has been selected as a location to develop the method due to the dense cloud coverage associated with its tropical monsoonal climate, common in regions near the equator.  This AI curated data set will then be used to train an artificial neural networks (ANNs) to make near-future (up to 500 yr) geomorphological predictions for the river catchments in the Philippine archipelago. GPU capacities will be essential in training the models. 

References & Further Reading

Lachowycz, S. (2024). Utility of artificial intelligence in geoscience. nature geoscience, 17(10), 953-955. 

Braun, J., & Deal, E. (2023). Implicit algorithm for threshold stream power incision model. Journal of Geophysical Research: Earth Surface, 128(10), e2023JF007140. 

Evensen, G. (2009). The ensemble Kalman filter for combined state and parameter estimation. IEEE Control Systems Magazine, 29(3), 83-104. 

​

Overview and Background

​​The UK has set ambitious targets to expand woodland cover to nearly 20% by 2050, recognising the multiple benefits of forests for carbon storage, biodiversity, and climate resilience. Delivering these targets requires more than simply planting trees: new and existing woodlands must be resilient to climate stress and safeguarded against invasive pests and pathogens. At the same time, globalisation and climate change are accelerating the spread of invasive species, threatening the long-term success of woodland expansion strategies. This project will develop next-generation, GPU-accelerated environmental models that couple tree growth dynamics, pest spread, and climate forcing within an Earth system framework. By integrating large-scale observation data, the project will generate fine-grained risk forecasts to inform national woodland policy and sustainable land-use planning. 

Methodology and Objectives

​​The project will develop a new spatial modelling (open-source) framework that couples a biological growth model for tree stands with a spatially explicit SIR-type epidemiological model for invasive pests and pathogens using a fully differentiable approach. Specifically, we will adapt existing 3PG models (to to represent stand-level woodland growth) into a differentiable GPU based code. This model accounts for temperature, drought, CO₂, and management. However, at present, a major limitation is explicitly linking canopy openness and stand structure to pest and disease vulnerability. Addressing this shortcoming is one of the main aims of the project. 

​Climate projections will provide environmental forcing, while observation data from project partners at Forest Research will constrain key model parameters in both the 3PG and SIR models and their coupling. Data integration will account for changes in resolution and modality, enabling high-resolution simulations of coupled pest–climate–tree interactions at scales relevant for both national policy assessments. Differentiability allows the direct use of gradient-based optimisation and Hamiltonian Monte Carlo (HMC), making it feasible to perform efficient parameter calibration and full Bayesian uncertainty quantification in a high-dimensional, spatially explicit model. It also facilitates the development of risk forecasts and scenario analyses for woodland expansion strategies.  

​Teaser Project 1 Objectives: Woodland Expansion and Policy Support 

​This project strand focuses on developing and then applying the coupled pest-climate-tree code to key case studies of woodland management and expansion in the UK.  

​The core goal of TP1 is to fully rewrite the existing 3PG model in a vectorised, GPU compatible, format with auto-diff for gradient computation through physiological sub-models. This will then be coupled to the existing pest/disease code using profiling tools to ensure computational efficiency and scalability. Finally, we will incorporate parameter inference using HMC and train the model on pre-existing datasets for representative tree-pest systems under current and future climates. Key focuses will be a) Sitka spruce and Elatobium aphid (using aphid monitoring data) and b) Oak and multiple pests and diseases (e.g. mildew, Acute Oak Decline, and Oak Processionary Moth).  

​We will then be able to: 

• ​Quantifying the vulnerability of new and existing woodlands to pest/pathogen outbreaks under climate stress. 
• ​Identifying data-informed planting strategies that balance carbon storage, biodiversity, and woodland resistance. 
• ​Generating fine-resolution maps of pest risk and climate suitability to guide woodland expansion planning at national and regional levels.

​This project offers direct policy relevance, producing modelling tools to support expansion strategies that are resilient to compound risks.

​Teaser Project 2 Objectives: Digital twinning 

​This project strand emphasises methodological advances in GPU-accelerated modelling and remote sensing technology to create a digital twin of woodland expansion. The objective is to integrate real time data streams from observations and sensors with different resolutions and modalities with the pest–climate–tree model, enabling real-time or near-real-time assimilation of key data sources, including citizen science initiatives. Specific goals include: 

• ​Scaling the software developed in TP1 to harness exascale facilities for high-resolution environmental modelling. 
• ​Coupling pest–climate–tree interactions to real-time multi-modal observation streams,  
• ​Integrating GPU-accelerated inference schemes to constrain free parameters. 
• ​Developing an accessible user interface to maximise the impact of the project. 

References & Further Reading

​​[1] J.J. Landsberg, R.H. Waring, Forest Ecology and Management, 95, 3 (1997) 
​[2]  K Reed, et al. Ecology and Management, 476, 118441 (2020)​ 

Overview and Background

​​The Met Office is home to the London Volcanic Ash Advisory Centre (VAAC). The role of the London VAAC is to provide advice and guidance to the aviation authorities on the presence of volcanic ash in the atmosphere. Ash forecasts are generated using the Met Office’s Numerical Atmospheric-dispersion Modelling Environment (NAME), initialised with eruption source parameters and driven by meteorological data. Currently, model simulations are conducted for specific individual volcanic events, each with unique eruption source parameters (i.e., the model inputs). This is necessary for real-time eruption response but does not allow for broader questions surrounding the interconnectivity between the eruption characteristics, the metrological conditions, and the associated impact on airspace to be addressed. The lack of these insights hinders long-term and strategic planning for the presence of volcanic ash in the atmosphere. 

Methodology and Objectives

​Suggested headings:  

1. ​Harnessing Machine Learning to inform risk-based decision making for volcanic ash clouds 
2. ​Using NAME as a digital twin to support impact assessments. 

​Methods Used: Ensemble forecasting; Monte Carlo analysis; Numerical Weather Prediction models; Lagrangian and Eulerian particle dispersion models (NAME); Parallel computing; probability mapping; Machine Learning; ORCHID GPU cluster on JASMIN; High Performance Computing.  

​Teaser Project 1 Objectives: In this teaser project, you will perform probabilistic model simulations to construct a large queryable dataset. On the order of 100,000 model runs will be conducted that vary the meteorological conditions, the source/volcano location and the eruption source parameters (e.g., mass eruption rate, plume height, particle size distribution). You will devise an accessible visualisation of this big dataset to enable end-users to explore the impact that specific source terms have on the transport and dispersal of a volcanic ash cloud, as total column mass loadings (appropriate for comparison with satellite data), and ash concentration within each flight level. Furthermore, these data will also serve as a training dataset for a machine-learning based model designed to run on GPUs in a parallel setting to address the time-sensitive nature of eruption response. Testing will occur on the ORCHID GPU cluster on JASMIN and this will build towards support for real-time eruption response conducted by the London VAAC wherein observables (e.g., satellite retrievals, plume height) can be used to query the database for likely, corresponding eruption source parameters (e.g., particle size distribution) which are difficult to measure in real-time. Without this carefully constructed training dataset proposed here, machine learning techniques have limited applicability to real-time London VAAC forecasts.   

​Teaser Project 2 Objectives: In this teaser project you will couple NAME model outputs (potentially the large suite of NAME outputs generated in teaser project 1) to several existing models/databases of critical infrastructure and services. Examples of coupling include: 

• ​airspace use models to determine the number of flights affected 
• ​building type and roof structure databases to determine the likelihood of roof collapse due to ash loading 
• ​road and public transport network layout and usage models to determine the number of people affected and delay times due to ash ground accumulation.  

​Coupling NAME with these other models, forming a ‘digital twin’, and ensuring that the result can be parallelised and run on GPUs will enable rapid, quantitative impact assessments to be made for explosive volcanic eruptions. Then, when exploiting exascale resources, the digital twin could be used to provide evidence for informed decision making at speed. Example questions that could be addressed include: is there a relationship between the eruption start time and the area of airspace affected? Do specific weather regimes cause a more severe disruption to air traffic/ public transport/ roads? How does impact scale with mass eruption rate? 

References & Further Reading

​​https://www.metoffice.gov.uk/research/approach/modelling-systems/dispersion-model 

​https://jasmin.ac.uk  

​Madankan, R., Pouget, S., Singla, P., Bursik, M., Dehn, J., Jones, M., Patra, A., Pavolonis, M., Pitman, E.B., Singh, T. and Webley, P., 2014. Computation of probabilistic hazard maps and source parameter estimation for volcanic ash transport and dispersion. Journal of Computational Physics, 271, pp.39-59. 

​Leadbetter, S.J., Jones, A.R. and Hort, M.C., 2022. Assessing the value meteorological ensembles add to dispersion modelling using hypothetical releases. Atmospheric Chemistry and Physics, 22(1), pp.577-596. 

​Capponi, A., Harvey, N.J., Dacre, H.F., Beven, K., Saint, C., Wells, C. and James, M.R., 2022. Refining an ensemble of volcanic ash forecasts using satellite retrievals: Raikoke 2019. Atmospheric Chemistry and Physics, 22(9), pp.6115-6134. 

​Beckett, F., Barsotti, S., Burton, R., Dioguardi, F., Engwell, S., Hort, M., Kristiansen, N., Loughlin, S., Muscat, A., Osborne, M. and Saint, C., 2024. Conducting volcanic ash cloud exercises: practising forecast evaluation procedures and the pull-through of scientific advice to the London VAAC. Bulletin of Volcanology, 86(7), p.63. 

​Hayes, J.L., Wilson, T.M. and Magill, C., 2015. Tephra fall clean-up in urban environments. Journal of Volcanology and Geothermal Research, 304, pp.359-377.​ 

Overview and Background

Many real-world processes, including those governing landscape evolution, can be effectively mathematically described via differential equations. These equations describe how processes, e.g. the physiography of mountainous landscapes, change with respect to other variables, e.g. time and space. Conventional approaches for performing statistical inference involve repeated numerical solving of the equations. Every time parameters of the equations are changed in a statistical optimisation or sampling procedure, the equations need to be re-solved numerically. The associated large computational cost limits advancements when scaling to more complex systems, the application of statistical inference and machine learning approaches, as well as the implementation of more holistic approaches to Earth System science. This yields to the need for an accelerated computing paradigm involving highly parallelised GPUs for the evaluation of the forward problem. 

Beyond advanced computing hardware, emulation is becoming a more popular way to tackle this issue. The idea is that first the differential equations are solved as many times as possible and then the output is interpolated using statistical techniques. Then, when inference is carried out, the emulator predictions replace the differential equation solutions. Since prediction from an emulator is very fast, this avoids the computational bottleneck. If the emulator is a good representation of the differential equation output, then parameter inference can be accurate. 

This work is highly relevant to the area of Earth Science research. By facilitating the practical use of these systems through emulation, we can gain insights into Earth’s processes, for example, predicting potential triggers for natural hazards such as landslides. Other insights include; tracing locations of erosion and/or deposition for potential critical resource identification, determining principal and potentially hidden natural or anthropogenic drivers of landscape evolution globally, and gaining further understanding as to how future climate change will affect Earth’s surface.

There are a number of applied and methodological directions this work can be generalised to. For example, since a main focus of this project involves making these systems faster and more efficient for use in practice, a natural extension would be to upscale model resolution to m-scale, allowing for a finer analysis and direct assessment of natural hazard risks. Once the principal work is complete, it also provides a framework for scaling to systems with additional modules, such as including ecological factors, advanced drainage flow routing components, and elements that model high-intensity/high-frequency storm events. Future work could also explore and adapt the improved emulation strategies we develop, making digital twinning more viable for use in Earth Science research, as well as other fields of study. 

Methodology and Objectives

Methods Used: Gaussian process interpolation (for building the emulator), Bayesian inference (for parameter inference), geomorphological analyses, surface processes modelling. 

Teaser Project 1 Objectives: GPU-accelerated differential equation solver. Geodynamic models in Earth Science are used to simulate a range of natural processes. Landscape evolution models specifically contain, amongst others, equations that describe surface processes such as erosion and sediment deposition as well as rock/surface uplift and aspects of climate change. However, the numerical solver executes consecutively, rather than generating solutions in parallel. This first teaser project will commence at the beginning of the PhD project (semester 1) and will focus on familiarising the student with parallel computing via GPUs, including the optimisation of existing landscape evolution models for GPU use. At the same time, the student will take training from ExaGEO, equivalent to 20 UoG credits, in GPU programming and Exascale principles. This teaser project will support the PhD project in developing robust, reliable and efficient emulators for landscape evolution models, utilising GPU power, which will allow for a denser training set and the inclusion of a broader variety of geomorphological scenarios. This teaser project will also give insight on possible GPU acceleration in the emulation process itself. 

Teaser Project 2 Objectives: Emulator development. The second teaser project will look at creating an emulator for a simple mathematical model describing elevation change as a function of spatial and temporal variations in surface uplift and efficiency of erosion. This will take place in semester 2 and the student will also undergo training at the same time from ExaGEO, in statistical and numerical methods in computing, complementing the students research aims at this stage. The skills the student will develop during this teaser project will set them up well, in combination with what they have attained from teaser project 1, to develop efficient emulators for more complex landscape evolution models, as the PhD project evolves.  

The student will be well supported by the supervisory team. Dr Eizenhöfer has expertise in landscape evolution modelling and Earth System science, Dr Macdonald and Dr Niu have expertise in developing statistical methodology in the area of statistical emulation and Dr Febrianto has expertise in highly parallelised architecture for scientific computing and will be able to advise on software development and design with open-source vision, as well as aspects of the GPU software development. 

References & Further Reading

Rasmussen, C.E., & Christopher K. I. Williams, C.K.I. (2006). Gaussian Processes for Machine Learning. The MIT Press. ISBN 0-262-18253-X. 

Donnelly, J., Abolfathi, S., Pearson, J., Chatrabgoun, O., & Daneshkhah, A. (2022). Gaussian process emulation of spatio-temporal outputs of a 2D inland flood model. Water Research. Volume 225. ISSN 0043-1354. 

Clark, M. K., Royden, L. H., Whipple, K. X., Burchfiel, B. C., Zhang, X., & Tang, W. (2006). Use of a regional, relict landscape to measure vertical deformation of the eastern Tibetan Plateau. Journal of Geophysical Research: Earth Surface, 111(F3). 

Eizenhöfer, P. R., McQuarrie, N., Shelef, E., & Ehlers, T. A. (2019). Landscape response to lateral advection in convergent orogens over geologic time scales. Journal of Geophysical Research: Earth Surface, 124(8), 2056-2078. 

Mutz, S. G., & Ehlers, T. A. (2019). Detection and explanation of spatiotemporal patterns in Late Cenozoic palaeoclimate change relevant to Earth surface processes. Earth Surface Dynamics, 7(3), 663-679. 

Whipple, K. X., Forte, A. M., DiBiase, R. A., Gasparini, N. M., & Ouimet, W. B. (2017). Timescales of landscape response to divide migration and drainage capture: Implications for the role of divide mobility in landscape evolution. Journal of Geophysical Research: Earth Surface, 122(1), 248-273. 

Whittaker, A. C., & Boulton, S. J. (2012). Tectonic and climatic controls on knickpoint retreat rates and landscape response times. Journal of Geophysical Research: Earth Surface, 117(F2). 

Yang, R., Willett, S. D., & Goren, L. (2015). In situ low-relief landscape formation as a result of river network disruption. Nature, 520(7548), 526-529. 

Zachos, J. C., Dickens, G. R., & Zeebe, R. E. (2008). An early Cenozoic perspective on greenhouse warming and carbon-cycle dynamics.  nature,  451(7176), 279-283.