DeepFlame 2.0: A new version for fully GPU-native machine learning accelerated reacting flow simulations under low-Mach conditions

This paper presents DeepFlame v2.0, a significant computational framework upgrade designed for high-performance combustion simulations on GPU-based heterogeneous architectures. The updated version implements a comprehensive CUDA-accelerated architecture incorporating fundamental combustion modelling components, including: implicit/explicit finite volume method (FVM) discretisation schemes, chemical kinetics integrators, thermophysical property models, and subgrid-scale closures for both fluid dynamics and combustion processes. The redesigned code supports diverse boundary conditions and discretisation schemes for broad applicability across combustion configurations. Key performance optimisations integrate advanced CUDA features including data coalescing techniques, CUDA Graphs for kernel scheduling, and NCCL-based multi-GPU communication. Validation studies employing the fully-implicit low-Mach solver demonstrate two-order-of-magnitude acceleration compared to conventional CPU implementations across canonical test cases, while maintaining numerical accuracy.