exaSIMPLE – Leveraging Machine Learning for Exascale CFD

Summary

exaSIMPLE is a project financed by Inno4scale (European High-Performance Computing Joint Undertaking project), and it aims at preparing CFD flow solvers for the newly deployed exascale HPC systems incorporated state-of-the-art machine learning techniques in traditional SIMPLE-like flow solvers.

Budget

160.000,00 €

Budget (blueOASIS)

80.000,00 €

Introduction

The exaSIMPLE project, funded through Inno4scale under the European High-Performance Computing Joint Undertaking (EuroHPC JU), aims to enhance the computational efficiency and scalability of Computational Fluid Dynamics (CFD) solvers through the integration of machine learning techniques. Developed by blueOASIS, INESC TEC, and the Maritime Research Institute Netherlands (MARIN), the project addresses the limitations of traditional numerical algorithms in exploiting modern high-performance computing architectures, with particular emphasis on maritime, ocean, and renewable energy applications.

The developed methodologies combine PyTorch and PETSc scientific computing libraries and were implemented on the Deucalion EuroHPC supercomputer, leveraging hybrid CPU-GPU architectures. A novel GNN architecture was developed using more than 7,000 CFD test cases, covering canonical benchmarks and industrially relevant flow configurations. The results demonstrate the feasibility of integrating machine learning into CFD solvers and show that enhanced pressure-velocity coupling can reduce the number of iterations required to solve the Navier-Stokes equations.

The outcomes of exaSIMPLE represent a step towards more efficient and scalable CFD simulations on future exascale supercomputers. Although further optimisation is required to improve inference accuracy and computational performance, the developed approaches offer promising opportunities for reducing computational costs and addressing increasingly complex engineering problems. The resulting software is publicly available through the exaSIMPLE GitHub repository, supporting further research and industrial adoption.

Funded by

EuroHPC & Inno4scale