The exaSIMPLE project addresses critical challenges in high-performance computing (HPC) and computational fluid dynamics (CFD). Led by Dr. Guilherme Vaz, exaSIMPLE embeds machine learning (ML) directly into CFD algorithms based on the SIMPLE algorithm, which has remained largely unchanged for decades.
BlueOASIS, the project coordinator, is leading the development with a focus on CFD, AI, and scientific programming, ensuring smooth project management and coordination. INESC-TEC, a prominent research institute in Portugal, provides invaluable HPC resources, including access to the powerful Deucalion supercomputer, and supports the optimization of exaSIMPLE for heterogeneous computing environments. Meanwhile, MARIN, the maritime research institute from the Netherlands, brings decades of experience in CFD development and facilitates the use of their large-scale datasets to train the machine learning models.
Motivation and Vision
ExaSIMPLE was not initially chosen to be a core part of the Inno4scale initiative. However, Dr. Vaz and his team saw an opportunity to prove that a small, agile company could tackle challenges traditionally reserved for larger institutions. “We wanted to show that a small company can play with the big boys,” says Vaz, emphasizing the team’s deep expertise in CFD, machine learning, and high-performance computing. ExaSIMPLE merges these areas to solve longstanding problems in CFD, particularly related to scalability and computational expense.
At its core, the project aims to revolutionize Finite-Volume based CFD codes by improving the SIMPLE algorithm, which is commonly used to solve the Navier-Stokes equations for fluid flow. The team seeks to address two key areas: reducing the extensive communication required between computational nodes in large simulations, and eliminating outdated approximations that slow convergence rates. These improvements are critical for advancing CFD into the exascale era, in the algorithmic front, where computations span massive supercomputing resources across Europe.
Overcoming Traditional Bottlenecks in CFD
Traditional CFD solvers, particularly those based on the SIMPLE algorithm, face two major challenges when applied to exascale computing. First, the communication between thousands of nodes in large-scale simulations slows down performance. Second, the linear system solvers used in traditional methods are not optimized for the complexity and scale of modern computations. ExaSIMPLE addresses these problems by replacing traditional solvers with machine learning models that are seen to be more efficient and scalable, specially in new HPC architectures.
The SIMPLE algorithm, originally developed in the 1960s, involves solving a pressure correction equation, that is computationally intensive and challenging in large-scale simulations. ExaSIMPLE focuses on optimizing this process by applying machine learning – particularly graph neural networks or GNNs – to two levels of the algorithm: Level 1, which addresses the linear pressure correction equation, and Level 2, which focuses on the nonlinear pressure-velocity coupling.
How does ExaSIMPLE Redefine the SIMPLE Algorithm?
One of the key innovations is the use of Graph Convolutional Neural Networks (GCNNs) to solve the linear system of equations (Level 1). By training a neural network to approximate the solutions to these equations, exaSIMPLE can significantly reduce the time it takes to complete each iteration of the algorithm. The result is faster convergence with less communication between nodes, which is essential for the efficient use of exascale hardware.
At Level 2, the team is improving the pressure-velocity coupling by using machine learning to reduce the number of iterations required for convergence. The traditional SIMPLE method involves approximations that were originally made to accommodate the limited computational resources of earlier systems. However, exaSIMPLE’s machine learning models allow the algorithm to bypass these approximations, thereby reducing the number of nonlinear iterations required for convergence.
These advances are critical for industries such as maritime engineering, offshore wind energy, and aerospace, where complex fluid dynamics simulations are essential for design and optimization. For example, simulating the flow around a ship or wind turbine requires thousands of iterations, which can take days or even weeks using traditional systems. By reducing this time, exaSIMPLE allows engineers to explore more design iterations and optimize systems faster and with greater accuracy.
Technical Challenges and Breakthroughs
Despite these promising developments, the project has faced significant technical challenges, particularly in integrating machine learning into the CFD framework. One of the most difficult aspects has been the application of GCNNs to real-world CFD problems. Although previous research had shown promising results, Dr. Vaz’s team encountered difficulties reproducing those results. “We had to invent our own solutions,” says Vaz, noting that even with help from the original researchers, replicating the results proved challenging.
However, by adapting and refining their approach, the team has begun to achieve promising results. In one test case, they were able to reduce the root-mean-square errors (RMSE) to 1e-7 values for the ML-based linear sovlers, and significantly improve convergence rates in nonlinear iterations, suggesting that the project is on the right track. The ultimate goal is to open-source the exaSIMPLE framework, allowing the wider CFD, AI, and HPC communities to further develop and apply these methods to a wide range of engineering codes and challenges.
The exaSIMPLE team has secured valuable computing time on the new EuroHPC Portuguese Deucalion supercomputer, a key resource for their high-performance computational work. Deucalion features two distinct subclusters: ARM Futjitsu A64FX-based nodes, each with 48 cores at 2.0 GHz, optimized for highly vectorizable code and energy-efficient machine learning tasks, and x86 AMD EPYC 7742-based nodes, offering 128 cores at 2.25 GHz. Additionally, up to 33 nodes are equipped with Nvidia A100 GPUs, enabling powerful parallel processing for AI-driven CFD simulations. The research team is currently using this advanced infrastructure to test different versions of PyTorch (core library for the previous developments), accelerating their algorithm development for exascale readiness.
Real-World Applications and Future Directions
The real-world impact of exaSIMPLE should be soon felt in industries such as maritime engineering, where CFD simulations are being used to optimize ship hull designs and reduce fuel consumption. By integrating exaSIMPLE’s machine learning algorithms into these simulations, engineers can run faster and more accurate tests, allowing them to design more efficient vessels. Similarly, in the offshore wind energy sector, exaSIMPLE could be used to simulate the complex flow patterns around wind turbines, enabling faster optimization and more accurate designs.
Looking to the future, Dr. Vaz envisions exaSIMPLE as a critical tool for advancing CFD beyond its current limitations. The project aims to achieve Technology Readiness Level 4 (TRL4) by the end of the Inno4scale initiative, with plans to progress through contributions from the open source community and further research. “We are doing something unusual,” Dr. Vaz explains. “Traditionally, CFD generates data for ML/AI models, but in our approach we use AI to accelerate CFD processes. We believe in the synergy between these two fields, which will lay the foundation for the future of fluid dynamics and beyond.”
The exaSIMPLE project is a step forward in the field of computational fluid dynamics. By integrating advanced machine learning techniques into traditional CFD algorithms, the project is poised to have a significant impact across multiple industries. With faster simulations, reduced computational costs, and the potential for wider application, exaSIMPLE can become a key innovation in CFD and AI.