An efficient GPU-accelerated adaptive mesh refinement framework for high-fidelity compressible reactive flows modeling

This paper presents a heterogeneous adaptive mesh refinement (AMR) framework for exascale simulations of non-stiff/moderately stiff chemical kinetics. The framework features an efficient time-subcycling stepping algorithm along with a specialized refluxing method, all unified in a highly parallel, scalable codebase. In addition, we develope a GPU-optimized low-storage explicit Runge–Kutta chemical integrator designed to minimize register usage, achieving higher efficiency than its implicit counterparts for detailed chemical kinetics with small mechanism size in high-speed combustion problems. A suite of benchmarks demonstrates the framework's high fidelity for both non-reactive and reactive simulations on both uniform and adaptively refined grids. By leveraging our parallelization strategy developed on top of AMReX, we demonstrate significant speedups on various problems using an NVIDIA V100 GPU compared to an Intel i9 CPU within the same codebase. In particular, for problems with complex physics and spatiotemporally distributed stiffness, such as hydrogen detonation propagation, we achieve an overall speedup of 6.49× with substantial computational throughput. Finally, this AMR framework is applied to a large-scale three-dimensional direct numerical simulation. Compared to prior CPU computations on a uniform grid, it yields a substantial reduction in total degrees of freedom involved in the calculation without compromising accuracy.