gpu-accelerated high-fidelity simulation of turbulent natural convection

This paper presents the hybrid meso-macroscopic model for high-fidelity simulation of turbulent thermal flows. the two-relaxation lattice boltzmann solver with the fourth order equilibrium distribution function representation was used to compute flow fields. numerical integration of the energy equation was performed using a fourth-order runge-kutta method. in-house matlab and julia codes were developed using cuda tools. turbulent natural convection of air was considered in a rayleigh number range of 1010 ≤ ra ≤ 1011 and aspect ratio range of 0.5 ≤ ar ≤ 2. it was found that distribution of the second order statistics is mainly determined by convective plumes dynamics. the aspect ratio affects the level of heat stratification. moreover, the horizontal distribution of temperature variance and turbulent kinetic energy with ra = 1010 is altered to vertical distribution with ra = 1011 when ar = 2. the hybrid lattice boltzmann scheme predicts the mean nusselt numbers with error of no more than 1% and 4% when ra = 1010 and ra = 1011, respectively, compared to classical dns techniques. the model performance achieves 1776 mlups and 1048.58 mlups with nvidia tesla v100 when using single and double precision variables, respectively. this level of computational performance creates an opportunity for the real-time prediction of thermal turbulence characteristics.