Prediction of submicron particle dynamics in fibrous filter using deep convolutional neural networks

Abstract

This study developed a data-driven model for the prediction of fluid–particle dynamics by coupling a flow surrogate model based on the deep convolutional neural network (CNN) and a Lagrangian particle tracking model based on the discrete phase model. The applicability of the model for the prediction of the single-fiber filtration efficiency (SFFE) for elliptical- and trilobal-shaped fibers was investigated. The ground-truth training data for the CNN flow surrogate model were obtained from a validated computational fluid dynamics (CFD) model for laminar incompressible flow. Details of fluid–particle dynamics parameters, including fluid and particle velocity vectors and contribution of Brownian and hydrodynamic forces, were examined to qualitatively and quantitatively evaluate the developed data-driven model. The CNN model with the U-net architecture provided highly accurate per-pixel predictions of velocity vectors and static pressure around the fibers with a speedup of more than three orders of magnitude compared with CFD simulations. Although SFFE was accurately predicted by the data-driven model, the uncertainties in the velocity predictions by the CNN flow surrogate model in low-velocity regions near the fibers resulted in deviations in the particle dynamics predictions. These flow uncertainties contributed to the random motion of particles due to Brownian diffusion and increased the probability of particles being captured by the fiber. The findings provide guidelines for the development of data science-based models for multiphysics fluid mechanics problems encountered in fibrous systems.