EuroHPC and FCCN

Deucalion: a new EuroHPC green supercomputer

Co-Funded by:


Combining CFD and AI for the exa-scale HPC era

Access typology:

A3 Larger Access

Cluster Facility:

Deucalion, MACC, Guimarães, Portugal

CPU time:

10 009 600 core.hours

GPU time:

4 000 gpu.hours


Deucalion – X86 cluster e GPU


12 months

Start date:



Computational Fluid Dynamics (CFD) methods, widely used for this purpose, require large amounts of computational resources, making this an area of interest for future exascale high-performance computing (HPC) clusters. However, poor scalability of the current CFD methods limits their applications. To overcome this, simplified models, coarse grids, or less accurate numerical settings are nowadays used. Enhancing scalability of engineering CFD codes would allow to overcome this barrier, greatly advancing the capabilities of industrial computational tools.

To keep increasing FLOPS without exponentially increasing supercomputer energy consumption, the industry has embraced massive parallelization and heterogeneous platforms, paired with reduced clock speeds. However, this shift in computing architecture has posed additionally scalability challenges for CFD tools, particularly in handling lower cells-per-core ratios, achieving finer-grain parallelization, and efficiently using mixed CPUs-GPUs architectures. The most widely used CFD solvers for incompressible viscous flow simulations solve the Navier-Stokes equations using pressure-based SIMPLE-like methods. In this case, the pressure equation, core of any SIMPLE variant, constitutes a major performance bottleneck, taking between 60% and 90% of the computational time. Scaling of this linear solver is also a barrier to achieve strong scaling of CFD and is mainly due to memory-bandwidth constraints and global communication patterns.

Over the last few years, many scientific fields have started applying Machine Learning (ML) algorithms to enhance the capabilities of traditional tools, and CFD is no exception. The use of ML in CFD can be divided into three categories: performance, discretization and modelling. This project proposal, exaSIMPLE, tackles the performance topic, by using ML methods and tailored HPC exascale hardware components, to speed up the turn-around computational time and scalability of the SIMPLE algorithm. exaSIMPLE will work on two different layers of the SIMPLE algorithm. First, the linear solution of the pressure Poisson equation (Level 1) will be attempted using matrix solvers utilising ML for acceleration. Secondly, the non-linear loops of the pressure-velocity coupling (Level 2) will be improved by leveraging ML models to substantially improve the classical SIMPLE approximations, therefore minimization the number of iterations and total computational time.

exaSIMPLE will contribute to the hybridization of Artificial Intelligence and Computational Fluid Dynamics fields, thereby catalysing the optimal usage of upcoming exascale hardware, all with the aim of achieving scalabilities, energy efficiency and speed-ups of CFD tools so far never achieved.