Project
Assessment of combustion models for thermodiffusive instabilities using a newly developed GPU library
Hydrogen is widely recognized as a leading candidate among carbon-free energy carriers. However, its use as a fuel in combustion systems introduces significant challenges. In particular, lean hydrogen–air flames are prone to thermo-diffusive (TD) instabilities, which strongly influence flame behavior in both laminar and turbulent regimes. To enable predictive simulations of hydrogen combustion, it is therefore essential to develop accurate models that incorporate thermo-diffusive effects.
In parallel, high-performance reactive computational fluid dynamics (CFD) codes are critical for supporting the transition to carbon-free energy conversion technologies. Leveraging modern GPU-accelerated high-performance computing (HPC) platforms is key to improve simulation performance. A practical and efficient strategy to harness GPU capabilities, without extensive modifications to the original code, is to offload the most computationally expensive tasks using external, modular libraries.
This project is structured into two sub-projects, each addressing one aspect of the dual objective. The first sub-project focuses on evaluating the performance of LES combustion models specifically designed for thermo-diffusive instabilities. The second sub-project aims to benchmark and validate a CUDA-based library capable of efficiently computing chemical source terms on both CPUs and GPUs.
Project Details
Project term
July 18, 2024–July 17, 2025
Affiliations
RWTH Aachen University
Institute
Institute for Combustion Technology (ITV)
Principal Investigator
Methods
All the simulations are performed with the high-fidelity in-house code CIAO, which solves the reacting Navier-Stokes equations in the low-Mach number limit. CIAO is a massive-parallel, higher-order semi-implicit finite difference code that uses Crank-Nicolson time advancement and an iterative predictor-corrector scheme. To model TD instabilities, the tabulated chemistry approach developed from Berger et al. has been used.
The ODE solver implemented in the GPU library is based on a second-order backward differentiation formula (BDF) method and utilizes an adaptive time-step, variable-order, stiff-optimized approach derived from the CVODE package. Each GPU thread independently solves a single ODE system, enabling different time-step sizes across systems to accommodate stiffness variability.
Results
Sub-project 1:
Assessment of LES Combustion Models for TD Instabilities
The performance of the LES modeling approach proposed by Berger et al. was evaluated across varying Reynolds numbers and grid resolutions using both a priori and a posteriori analyses. Reference data for the study were obtained from a direct numerical simulation (DNS) dataset comprising three turbulent lean hydrogen flame cases at Reynolds numbers and . The evaluation also included an investigation into the influence of different progress variable definitions. The a priori analysis involved comparing source terms extracted from DNS snapshots with those predicted by the model’s flame-state lookup table across various filter sizes. Results were consistent across all three Reynolds numbers and different filter sizes. The lookup table provides an overall reasonable prediction of the progress variable source term. However, some discrepancies are observed in regions with high mixture fraction values for the hydrogen-based progress variable and in the zone with high mixture fraction and high water mass fraction values for the water-based progress variable, with the latter appearing more severe. Thea posteriori analysis consisted of LES at multiple resolutions and Reynolds numbers. Quantitative evaluation was based on the flame length predicted by LES, compared with the corresponding filtered DNS value . Results confirm that the hydrogen-based progress variable yields better agreement with DNS data than the water-based formulation. Nevertheless, flame length trends with increasing Reynolds number are well captured by both formulations, even at coarse LES resolution.
Sub-project 2:
Development and Benchmark of a GPU Library for Chemical Kinetics
The newly developed GPU-accelerated library was benchmarked using 2D flame simulations with multiple chemical kinetic mechanisms. Although moderate speed-ups were achieved, the current implementation does not yet justify full-scale GPU deployment. The primary limitation stems from the nonlinear solver. The library currently employs a fixed-point iteration method, which, although memory-efficient and simpler to implement on GPUs (due to the absence of linear system solves), exhibits poor convergence behavior. For moderately stiff systems, the fixed-point method required up to 10 times more internal steps than a Newton-based approach.
Discussion
The LES combustion modeling approach by Berger et al. has been successfully evaluated for its ability to capture thermo-diffusive instabilities. While the hydrogen-based progress variable formulation is found to provide slightly better accuracy than the water-based formulation, both variants yield low modeling errors and provide a suitable representation of the local flame state. The model accurately reproduces flame length trends with increasing Reynolds number and provides reliable temperature predictions, even at low LES resolutions. Further validation in a broader set of flame configurations is planned for the next phase.
Additional Project Information
DFG classification: 404 Fluid Mechanics, Technical Thermodynamics and Thermal Energy Engineering
Software: CIAO
Cluster: CLAIX