(333b) Development of High-Speed Computing Method for Powder Mixing Using Machine Learning with Random Model

Powder mixing is ubiquitous in a wide variety of manufacturing sectors including chemical industries, and it plays an important role. The degree of mixing of powders (a measure of uniformity) has a significant effect on the quality of the final product obtained by processing the powders, and thus methods to predict the degree of mixing are required. Computer simulation using the Discrete Element Method (DEM) is a powerful tool for the analysis of powder mixing process. In a DEM, dynamics (position and velocity) of an individual particles are numerically solved. These calculations are performed at extremely small time-step (micro seconds order). Therefore, a large number of iterations are required to calculate the entire powder mixing process, and it is difficult to predict a practical mixing in an actual process.

Recently, some studies on the combination of DEM and machine learning (ML) have been conducted. In these studies, macroscopic bulk powder phenomena (angle of repose, kinetic energy, mixing index etc.) were predicted. However, there is no study on prediction method of an individual particle flow and mixing behavior. In this study, we developed an original machine learning model, namely recurrent neural network with stochastically calculated random motion (RNNSR), which allows long time scale powder mixing simulation with low computational cost and high accuracy. RNNSR learns individual particle dynamics from short-time DEM simulation results and predicts a powder mixing for a long time. RNNSR combines recurrent neural network with stochastic model to predict convective and diffusive mixing. RNNSR is evaluated in terms of degree of powder mixing, particle velocity, granular temperature, and computing speed.