Machine-learning based defiltering method for Lagrangian particle tracking in coarse-grid simulations

We propose a method for accurate numerical simulations of Lagrangian particle tracking based on machine learning techniques. In industrial applications, numerical simulations must be conducted with limited computational resources, which requires reduced grid resolution. However, the low resolution of fluid fields tends to underestimate the fluctuation of particle velocity, leading to inaccurate estimations of particle behaviors such as accumulation and deposition. To address this issue, we introduce a novel approach to reconstruct (i.e., defilter) the fluid velocity, thereby improving particle motion in coarse-grid (i.e., filtered) simulations. The proposed method utilizes the machine-learning model trained in a supervised manner using data from the direct numerical simulation of a turbulent channel flow with a friction Reynolds number of 180. We demonstrate that the model can accurately reconstruct the fluid velocity at particle locations using local stencils around the particles in coarse-grid fields. The results of particle tracking using the proposed method show significant improvements in calculating particle trajectories, velocity fluctuations and depositions. Moreover, we show that the trained model can be applied to computational domains of different sizes and flows with different Reynolds numbers. The performance and generalizability of the proposed method indicate its potential to enable accurate prediction of particle motion in practical engineering applications.