Large-eddy simulation of particle-laden isotropic turbulence using machine-learned subgrid-scale model

Abstract

We apply a machine-learned subgrid-scale model to large-eddy simulations (LES) of heavy particles in isotropic turbulence with different Stokes numbers. The data-driven model, originally developed for high Reynolds number isotropic turbulent flows based on the gene expression programming (GEP) method, has explicit model equations and is for the first time tested in multiphase problems. The performance of the GEP model has been investigated in detail, focusing on the particle statistics including particle acceleration, velocity, and clustering. Compared with the commonly used dynamic Smagorinsky model, the GEP model provides significantly improved predictions on the particle statistics with Stokes numbers varying from 0.01 to 20, showing satisfactory agreement with the results from direct numerical simulations. The reasons for the enhanced predictions of the GEP model are further discussed. As the GEP model is less dissipative and it introduces high-order terms closely related to vorticity distribution, the fine-scale structures usually missing in LES simulations can be better recovered, which are believed to be closely related to the intermittency of particle motion and also particle clustering.