Project institution: University of Glasgow
Project supervisor(s): Dr Andrew Elliott (University of Glasgow), Prof Lindsay Beevers (University of Edinburgh), Prof Claire Miller (University of Glasgow) and Prof Michèle Weiland (University of Edinburgh)
Overview and Background
Flood modelling is crucial for understanding flood hazards, now and in the future as a result of climate change. Modelling provides inundation extents (or flood footprints) which provide outlines of areas at risk which can help to manage our increasingly complex infrastructure network as our climate changes. Our ability to make fast, accurate predictions of fluvial inundation extents is important for disaster risk reduction. Simultaneously capturing uncertainty in forecasts or predictions is essential for efficient planning and design. Both aims require methods which are computationally efficient whilst maintaining accurate predictions. Current Navier-stokes physics-based models are computationally intensive; thus this topic would explore approaches to hybrid flood models which utilise GPU-compute and ML fused with physics-based models, as well as investigating scaling the numerical models to large-scale HPC resources.
Methodology and Objectives
Methods Used: Machine learning, statistical modelling, optimised process models using GPU computation.
Teaser Project 1
Exploring the advantages and limitations of GPU enabled approaches to flood modelling in contrast to traditional process-based flood modelling. Key considerations would be characterising the computational advantage of different ML approaches (especially physics informed machine learning models) considering both training and inference and the corresponding accuracy in comparison to the traditional process-based models. In addition, we will explore enhancing traditional process-based models by investigating the opportunities for exploiting large-scale, GPU-accelerated HPC.. Using data available from a range of sources (e.g. satellite, sensor networks as well as model outputs), different ML approaches will be explored to represent the complex hydrodynamics which a process based model would capture.
This project will naturally extend to a full PhD exploring hybrid modelling approaches with a key understanding how the level of accuracy of these models.
Teaser Project 2
Uncertainty quantification is becoming increasingly important as binary predictions give at best a limited outlook on the model and at worse can be misleading to policy makers who may not consider the implications of enforcing a binary outcome to flood forecasting models, or for adaptation development. However, with particularly slow high-fidelity models, gaining accurate and meaning uncertainty estimates via Monte Carlo, is either incredibly time consuming or indeed impossible. There are multiple solutions to this, including use surrogate/ML models (which can run the simulation faster) or improved Monte Carlo procedures (e.g. see Aitken et. al. 2024). Needless to say, while this computationally useful, it is important to understand the implications for the calibration of the uncertainty quantification of these approaches.
Thus, following from Aitken et. al. 2024, in this teaser project we will consider a large range of possible approaches, use them to obtain uncertainty quantifications and compare them to the uncertainty estimation which we can obtain from a high fidelity model, e.g. using LisfloodFP or Telemac2D. Due to the computational requirements of this approach, this is likely to require large scale compute, in both traditional and GPU compute. Comparisons will then be made between the UQ relying on large compute and those developed in this teaser project, allowing an understanding of the trade-offs between these approaches.
This teaser project naturally expands into a wider PhD designing and developing novel GPU enabled methods to obtain well calibrated uncertainty estimates via a combination of statistical and machine learning techniques to give rapid outputs to decision and policy makers.
References and Further Reading
- Aitken, G.; Beevers, L.; Christie, M.A. Advanced Uncertainty Quantification for Flood Inundation Modelling. Water 2024, 16, 1309 (click here)
- Andersson, T.R., Hosking, J.S., PĂ©rez-Ortiz, M. et al. Seasonal Arctic sea ice forecasting with probabilistic deep learning. Nat Commun 12, 5124, 2021 (click here)
- Aitken, G., Beevers, L., & Christie, M. A. (2022). Multi-level Monte Carlo models for flood inundation uncertainty quantification. Water Resources Research, 58, e2022WR032599 (click here)
- Fraehr, Niels, et al. “Assessment of surrogate models for flood inundation: The physics-guided LSG model vs. state-of-the-art machine learning models.” Water Research 252 (2024): 121202
