Solid-liquid dual channel data-driven method for Lagrangian fluid simulation

Abstract

To solve the problems of low accuracy in long time series prediction and low generality of network parameter model in the existing data-driven Lagrangian fluid simulation, a light neural network prediction model which is physics-based multi-layer shared perceptron was proposed. Each fluid particle is standardized by searching neighbor particles through the optimized parallel processing module. The neural network is used to predict the effect of each neighbor particle on the central particle. The solid-liquid two-state differentiated aggregation operation is used to predict the acceleration of each fluid particle. The experimental results show that, compared with the existing methods, the method proposed in this paper greatly improves the prediction accuracy with less time overhead, and at the same time maintains more fluid motion details. In addition, we can enables more accurate long-term fluid motion prediction. Compared with PointRNN, PointNet++ and other single-channel data-driven methods, we can better deal with the fluid-solid coupling problem, and has wider network versatility.