Surrogate model of DEM simulation for binary-sized particle mixing and segregation

Abstract

Segregation is a well-known phenomenon that occurs during powder mixing, wherein particles with similar properties are collected from specific regions of a bulk powder. Numerical simulation using the discrete element method (DEM) is recognized as a potent tool for investigating and predicting segregation. However, DEM simulations are computationally expensive. To address the high computational overhead, in our prior work, we proposed a machine-learning-based surrogate model for DEM simulation, namely, recurrent neural network with stochastically calculated random motion (RNNSR). The model was designed to predict the local mean component of the particle behavior using the Lagrangian approach and the local variability component of the particle behavior using the Eulerian approach. However, the RNNSR was demonstrated exclusively for monodisperse particles with homogeneous properties. Hence, in the current study, we extended the RNNSR to simulate the mixing and segregation of powders with inhomogeneous properties. The dependence of the particle size on the Lagrangian and Eulerian behaviors of the particles was investigated. Based on this analysis, an extended-RNNSR was developed for binary-sized particle system by adding the particle size data for the training data. The prediction accuracy of the extended-RNNSR was evaluated in terms of the mixing degree, particle velocity, granular temperature, and computing speed. It was demonstrated that the extended-RNNSR constructed on learning data initiated from the randomly mixed initial condition could predict mixing and segregation initiated from both segregated as well as the randomly mixed initial conditions. The extended-RNNSR also demonstrated a much faster computing speed than the DEM.