Project institution: University of Edinburgh
Project supervisor(s): Dr Mark Bull (University of Edinburgh), Dr Maggie Creed (University of Glasgow), Prof Simon Mudd (University of Edinburgh) and Dr Declan Valters (British Geological Survey)
Overview and Background
Flood forecasting at regional and national scale is imperative to predicting the scale and distribution of floodwaters during extreme weather events, mitigating the impact on communities most at risk from flooding. The LISFLOOD family of surface water models have proved suitable to being parallelised at scale, allowing research and forecasting communities to take advantage of the previous generation of supercomputers, such as ARCHER.
The increasing availability of high resolution topographic and meteorological data provides an opportunity to extend the capability of the LISFLOOD modelling framework to produce large-scale or high resolution flood forecasts at operational timescales – i.e., producing model runs at sufficient lead-in times to alert communities to impending flood risk from forecasted extreme weather events. GPU-based exascale HPC systems provide the technological basis to develop forecast models delivering at operational timescales.
Methodology and Objectives
LISFLOOD is a family of hydrological models based on a 2D grid simulating rainfall-runoff. The water routing across a flood basin/river catchment is based on a simplified version of the shallow water (St Venant) equations. The model is process (physics) based, and there have been several implementations (see below), usually in C or C++, using a cellular automaton approach. These have been parallelised for CPU using OpenMP and in one spin-off project, MPI (see here for more details).
The stencil-code library used in the previous CSE project, LibGeoDecomp, purports to have support for NVIDIA GPUs and CUDA (see here for more details).
Teaser Project 1:
- Implement the hydrodynamic core of the LISFLOOD model on GPU hardware to demonstrate proof-of-concept that the current CPU parallelised code is portable to GPU hardware.
- Methods for GPU parallelisation would include OpenMP offloading as initial approach to verify proof of concept. Project could then be extended to investigate CUDA bindings available in the libgeodecomp library.
Teaser Project 2:
- Profile, then optimise GPU ported code and test using case studies of UK extreme flood events, to indicate potential for near-realtime flood forecasting of GPU-enabled LISFLOOD code.
- This objective would aim to deliver a minimum-working example of a the GPU-ported flood model to deliver forecasts/re-analysis of a historic flood event in the UK within an operational timescale.
Development into a full PhD would involve further profiling and optimisation of the GPU code using either the libgeodecomp library, or another suitable GPU parallelisation framework. Delivering a proof-of-concept for a working flood forecast model at a regional scale would be a key aim of this project, demonstrating the potential to be used in operational flood forecasting systems. The full PhD may therefore look at workflow tools to integrate the various stages of forecast production such as: ingestion and pre-processing of data (i.e. from rainfall forecast/nowcasting data products), model scheduling on HPC systems, and post-processing of the outputs.

References and Further Reading
- Coulthard, T.J., Neal, J.C., Bates, P.D., Ramirez, J., de Almeida, G.A. and Hancock, G.R., 2013. Integrating the LISFLOOD‐FP 2D hydrodynamic model with the CAESAR model: implications for modelling landscape evolution. Earth Surface Processes and Landforms, 38(15), pp.1897-1906
- LISFLOOD model high level overview (click here)
- Stencil Code for LibGeoDecomp (click here)
- Open Source version of the C++ code developed by Declan Valters (click here)
- Overview of an earlier project that developed an experimental version of the code for multi (CPU) node using stencil code (click here)