The intersection of machine learning (ML) and computational fluid dynamics (CFD) has opened new opportunities for improved performance, scalability and efficiency. This study is about exaSIMPLE, that aims to revolutionise CFD simulations by addressing the critical aspect of performance through the integration of ML methods and tailored exascale high-performance computing (HPC) hardware components.

The primary focus of exaSIMPLE is to accelerate the computational time and scalability of the SIMPLE algorithm, a fundamental numerical technique widely used in CFD simulations. The study strategically targets three key categories in the use of ML within CFD: performance, discretisation and modelling. Here we look at the performance domain, where exaSIMPLE aims to drive advances using ML techniques and specialised exascale hardware components.

In the first layer of the SIMPLE algorithm (Level 1), exaSIMPLE aims to accelerate the linear solution of the pressure Poisson equation through the application of ML-driven matrix solvers. These solvers use ML acceleration techniques to optimise the computation and significantly reduce the turnaround time. Moving to the second level (Level 2), exaSIMPLE addresses the non-linear loops of the pressure-velocity coupling, using ML models to refine classical SIMPLE approximations. This refinement not only minimises the number of iterations, but also contributes to an overall reduction in computational time.

An innovative aspect of exaSIMPLE is the use of Graphics Processing Units (GPUs) for offline and online ML training, while the core computations of the traditional SIMPLE algorithm are performed on Central Processing Units (CPUs). This strategic allocation of tasks minimises task synchronisation, thereby increasing the scalability of the parallelization. In addition, exaSIMPLE integrates seamlessly with popular HPC libraries, including PETSc, Trilinos and MAGMA, ensuring compatibility.

The results of exaSIMPLE will be CFD code agnostic, promising broad applicability. By catalysing the hybridisation of Artificial Intelligence and Computational Fluid Dynamics, exaSIMPLE aims to unlock the full potential of upcoming exascale hardware, ushering in unprecedented scalability, energy efficiency and speed-up in CFD simulations. Through its integrative approach and adherence to well-established interfaces, exaSIMPLE will have a lasting impact on the optimal use of technology in the field of CFD.