An efficient and consistent numerical simulation framework coupling lattice Boltzmann flux solvers and graphic processing unit for single- and two-phase flows

This study presents an efficient and consistent numerical simulation framework by coupling efficient lattice Boltzmann flux solvers (LBFS) with graphic processing units (GPU) for both single- and two-phase flows, which significantly improves the computational efficiency compared with the serial implementations on the central processing unit (CPU). The framework is based on the well-developed LBFS, in which the macroscopic governing equations are discretized by the finite volume method (FVM), and their fluxes at each cell interface are evaluated using the lattice Boltzmann method (LBM). Both Newtonian and non-Newtonian power-law fluids are considered in single-phase flows. For two-phase flows, the Cahn-Hilliard equation is applied to capture the phase interfaces, which is solved by using the improved scheme of third-order weighted essentially non-oscillatory with Z-indicator (WENO-Z). A consistent CPU/GPU heterogeneous parallel programming framework realizes the overall solution procedures with the single instruction multiple threads (SIMT) architecture. The main computing tasks are distributed to many threads on the GPU corresponding to the cell centers and interfaces in FVM to achieve high computational efficiency. Both 2D and 3D benchmarks are considered to examine the accuracy and efficiency of the present numerical simulation framework, such as Poiseuille flow, lid-driven cavity flow, Rayleigh-Taylor instability, and droplet-impacting solid surfaces. The results indicate that, compared to purely CPU-based serial codes, the present GPU-accelerated LBFS simulation framework achieves acceleration ratios over 2300 without sacrificing numerical accuracy. The present GPU-accelerated solvers seem to provide a promising and reliable tool for large-scale, high-performance computing of complex fluid flow problems.