Highlights
- Developed a GPU-accelerated spectral element solver for high-fidelity simulation of hydrogen combustion with detailed chemistry
- Achieved excellent strong and weak scalability on large European HPC systems, including runs on up to 3544 A100 GPUs
- Enabled the first large-scale DNS database of turbulent hydrogen combustion across a broad range of conditions
- Demonstrated exascale readiness, leading to a successful early access application for next-generation supercomputers
- Provided valuable insights into flame dynamics and pollutant formation, supporting next-generation combustion model development
| Keywords | Energy, Environment, Climate, Weather, Mechanical Engineering |
| Technologies used | GPU |
Challenge
The transition to carbon-free energy systems presents a major scientific and engineering challenge, particularly in the development of clean combustion technologies based on hydrogen and other alternative fuels. Accurately predicting flame behavior, pollutant formation, and combustion stability under practical conditions requires the resolution of a wide range of physical and chemical processes across multiple scales. Direct Numerical Simulation (DNS) offers the highest fidelity but is computationally demanding, especially for reacting flows with detailed chemistry. Traditional CPU-based solvers face significant limitations in terms of memory, speed, and scalability, making large-scale simulations of practical relevance infeasible. Ex3S addressed the challenge by developing a disruptive GPU-accelerated spectral element solver capable of solving large sets of coupled scalar species with high efficiency and accuracy. By leveraging mixed-precision computing, advanced stabilization techniques, and asynchronous communication, the solver overcomes key bottlenecks in performance and memory usage. The resulting tool enables DNS of hydrogen combustion at unprecedented scale and fidelity, providing valuable data for model development and validation. This addresses a critical gap in the predictive capabilities needed for designing next-generation, low-emission combustion systems, and supports Europe’s broader goals in clean energy and exascale computing.
Research Topic
This project focused on the development of a high-performance GPU-accelerated solver for simulating turbulent hydrogen combustion using the spectral element method. By enabling large-scale direct numerical simulations (DNS) with detailed chemistry, the research helps to improve the understanding of flame dynamics and pollutant formation under realistic conditions, supporting the development of predictive combustion models for carbon-free fuels.
Solution
To overcome the limitations of conventional solvers for reacting flows, Ex3S developed a GPU-accelerated spectral element solver integrated into the NekRS framework via the NekCRF chemistry plugin. The solution uses a mixed-precision approach to balance accuracy and performance, solving the coupled transport equations for species and temperature with advanced iterative methods. A hybrid stabilization strategy and asynchronous communication were introduced to address challenges such as Gibbs oscillations and scalability. The solver runs entirely on GPUs, maximizing hardware efficiency, and demonstrates excellent strong and weak scaling on leading HPC systems. This solution enables high-fidelity DNS of hydrogen combustion, supporting advanced model development.
