Highlights
- Use of graph neural networks provides a way to solve the matrix equation systems using GPUs
- Approximation of SIMPLE algorithm terms that disregarded in the traditional approaches lead to a considerable improvement of performance, accuracy and robustness for several canonical, benchmark and simple Industrial-typical CFD cases
- Exasimple.py code permits all CFD practitioners to reproduce the results here obtained and transfer the major findings to any CFD code.
| Keywords | Energy, Environment, Mechanical Engineering, Ocean |
| Technologies used | CFD, SIMPLE, ML, PeTSC |
Research Topic
The exaSIMPLE project developed hybrid machine learning-computational fluid dynamic algorithms to enhance the efficiency of CFD solvers. The approach builds upon the SIMPLE algorithm and integrates ML to improve key computational aspects. First, a more efficient way to solve the pressure correction equation is sought by employing graph neural networks to replace traditional solvers. The approach aims to accelerate CFD codes and enable them to more easily use GPUs. Second, improvements in the pressure-velocity coupling are achieved by lifting a key simplification made in SIMPLE and using ML to approximate a matrix inverse that is typically left out, aiming for faster convergence.
Challenge
The use of CFD has become ubiquitous throughout many industries, including the maritime and offshore renewable sectors. However, the computations remain very computationally intensive, reducing their applicability to optimisation studies or necessitating the use of simplified modelling. Thus, accelerating CFD would enable far more advanced fluid dynamic analysis to be employed by the industry. Unfortunately, the current generation of CFD codes relies on outdated software design choices that did not account for the now omnipresent existence of massively-parallel CPU and hybrid CPU-GPU hardware architectures. This has left the codes lagging behind and not benefiting from the speed-ups seen in the world of AI training, for instance. A new reimagining of the core algorithms is necessary to bring CFD back in line with hardware advancements in a future-proof manner. To fully leverage modern hardware, integrating cutting-edge machine learning techniques with CFD solvers is essential and yet remains largely unaddressed.
Solution
The adopted solutions have relied on utilising machine learning at the critical focal points of the overall CFD framework, namely the pressure-velocity coupling and linearised equation solution. For both aspects, ML has been used to infer solutions to sub-problems that can only be obtained in a computationally-burdensome way using traditional methods. For the pressure-velocity coupling, the new approach utilises ML to compute the Navier-Stokes equation solution in a much more accurate way, greatly reducing the required number of iterations and hence bringing down the net computational cost. For the linear solvers, the solution can now be offloaded entire to GPUs, promising better integration with new hardware stacks and better performance at a reduced energy cost.
| Time period of the study | March 2024-March 2025 |
| Country | Portugal Netherlands |
| Name of the scientific domain of the software | Ocean, Maritime, Naval applications |