PhD: GPU-Accelerated High-Fidelity Hydrodynamics Modelling for Tidal Energy Resource and Environmental Impact Assessment

University of Edinburgh

Edinburgh, UK 🇬🇧

Project institution: University of Edinburgh

Project supervisor(s): Dr Joseph O’Connor (University of Edinburgh), Dr Brian Sellar (University of Edinburgh) and Dr Athanasios Angeloudis (University of Edinburgh)

Overview and Background

Tidal energy offers a predictable and sustainable energy source, driving increased interest in its development. However, as tidal energy deployments are scaled up, this places a greater burden on the local environment/ecosystem. High-fidelity hydrodynamic modelling tools are essential for predicting and mitigating environmental impacts, while also maximising energy extraction, to ensure this limited resource is used in a responsible way. However, these models are computationally demanding. Moreover, many existing tools are developed for traditional (CPU-based) HPC systems. With the advent of large GPU-based exascale machines, there is a need to prepare existing codes for this new HPC paradigm. As well as futureproofing existing codebases, this will provide a step change in computational capacity, unlocking new types of simulations (e.g. high-fidelity multi-physics) and workflows (e.g. optimisation, uncertainty quantification, data assimilation) not currently possible with today’s methods.

Methodology and Objectives

Both projects will use advanced computational techniques to improve hydrodynamic modelling capability for tidal energy resource assessment, as well as predicting/mitigating environmental impacts. This will involve porting existing open-source hydrodynamic modelling tools to GPU to enable faster, larger, and more detailed simulations/workflows than what is currently achievable. The focus is specifically on high-fidelity models, where the 3D hydrodynamic equations are solved numerically. Initially, this project will focus on tools already used throughout the supervisory team (e.g. TELEMAC-3D, Thetis). However, a survey period is envisaged to identify the most suitable tools to take forward to exascale applications. 

Teaser Project 1: Accelerating high-fidelity hydrodynamic models for large-scale ensemble-based workflows

In most real-world applications, running a single simulation is often insufficient. Tasks such as optimisation and uncertainty quantification typically require hundreds/thousands of model evaluations. This is extremely computationally demanding, making it impractical with today’s high-fidelity methods. This project will enable these types of workflows on emerging GPU-based exascale machines by porting an existing high-fidelity hydrodynamics model to GPU. The objectives for the teaser project are:

  1. Survey and profile existing high-fidelity hydrodynamic modelling tools (starting with TELEMAC-3D and Thetis) to identify computational bottlenecks and evaluate the potential for porting to GPU.
  2. Build a proxy application replicating one of the identified bottlenecks (e.g. advection-diffusion equation) to test and evaluate different programming models/frameworks for porting to GPU (e.g. CUDA, OpenMP, SYCL).

Following the first year, if this project is selected the remaining objectives will be to:

  1. Port selected components of the chosen model to GPU to improve computational performance. This will involve implementing kernels for GPU execution and optimising memory management.
  2. Benchmark the GPU-accelerated implementation to compare the performance against the existing CPU implementation and identify areas for optimisation. This will also involve testing the scalability across multiple GPUs for large-scale exascale applications.
  3. Integrate the GPU-accelerated model within a large-scale ensemble-based framework to enable workflows that require hundreds/thousands of model evaluations. This will be demonstrated on real-world cases by performing large-scale optimisation (e.g. for marine spatial planning) and uncertainty quantification (e.g. for model reliability) campaigns.

Teaser Project 2: Accelerating data assimilation for enhanced model calibration and predictive capability

Data assimilation (DA) combines real-world observational data with numerical simulations to enhance model predictions, leading to more reliable simulations for real-world problems. However, DA is computationally intensive, requiring sophisticated large-scale modelling and data processing techniques. This project will develop a GPU-accelerated DA framework tailored for tidal resource assessment to enable these workflows on emerging GPU-based exascale machines. The objectives for the teaser project are:

  1. Survey existing methods (e.g. 3D/4D variational DA, ensemble Kalman filter) and libraries for combining observational data with high-fidelity hydrodynamic models. This should also consider the form of the observational data (e.g. satellite, acoustic doppler current profiler, etc.).
  2. Build and profile a small CPU-based example of the DA framework to identify computational bottlenecks and determine priority components for porting to GPU (e.g. model, data processing, or a mix of both).

Following the first year, if this project is selected the remaining objectives will be to:

  1. Port selected components of the DA framework to GPU to improve computational performance. This will involve selecting a suitable programming model/framework, implementing kernels for GPU execution, and optimising memory management.
  2. Benchmark and profile the GPU-accelerated implementation to compare the performance against the existing CPU implementation, as well as identifying areas for further development and optimisation.
  3. Demonstrate the new GPU-accelerated DA framework on real-world tidal energy applications (e.g. resource assessment and environmental impact). This will enable enhanced model calibration for uncertain parameters (e.g. bottom friction, turbulence parameters), as well as improve predictive accuracy by optimally combining model solutions with observational data.

References and Further Reading

  1. Almoghayer, M. A., Lam, R., Sellar, B., Old, C., & Woolf, D. K. (2024). Validation of tidal turbine wake simulations using an open regional-scale 3D model against 1MW machine and site measurements. In Ocean Engineering (Vol. 299, p. 117402). Elsevier BV (click here)
  2. Old, C., Sellar, B., & Angeloudis, A. (2024). Iterative dynamics-based mesh discretisation for multi-scale coastal ocean modelling. In Journal of Ocean Engineering and Marine Energy (Vol. 10, Issue 2, pp. 313–334). Springer Science and Business Media LLC (click here)
  3. TELEMAC-MASCARET 
  4. Thetis

2 days remaining

Apply by 17 February, 2025

POSITION TYPE

ORGANIZATION TYPE

EXPERIENCE-LEVEL

DEGREE REQUIRED

You ad could be here!