Optimizing data-sampling period in a machine learning-based surrogate model for powder mixing simulations

The discrete element method (DEM) has been widely employed in computer simulations of various powder handling processes. However, it has limitations in treating long-term evolution owing to its high computational cost. To predict long-time-scale individual particle motions with low computational cost and high accuracy, we developed a machine learning-based surrogate model: a recurrent neural network with stochastically calculated random motion (RNNSR). The RNNSR learns and predicts individual particle trajectories with a much larger time step than the DEM, enabling the ultrafast computation of the time evolution of particle motions. However, its accuracy was evaluated under only one specific operating condition, leaving open the question of whether the RNNSR can predict powder mixing even at different powder flow velocities. In addition, the accuracy of the RNNSR depends on the time step of the learning data used to train the recurrent neural network. However, the criteria for an appropriate time step in RNNSR remain uncertain. Hence, this study investigated the criterion for an appropriate time step for the RNNSR and evaluated its performance at various powder flow velocities. The appropriate time step for the RNNSR was determined based on the mean period of the original particle trajectory data. The criterion for the appropriate time step was 4 % of the mean period of the original particle trajectory. The RNNSR with the optimized time step accurately predicted the powder mixing behavior, particle velocity, and granular temperature in a rotating drum mixer regardless of the drum rotation speed. © 2024 Elsevier B.V.