An improved machine learning approach for predicting granular flows

Granular flow is widely practised in many industry processes. The previous prediction methods of granular flow are limited in efficiency, including conventional Discrete Element Modelling (DEM) where direct computation of particle collisions is very time-consuming; and recent machine learning (ML) approach where particle positions were straight predicted and particle–particle and particle–wall collisions were not considered separately, likely compromising the prediction accuracy. In this study, an improved ML approach is developed for predicting granular flows efficiently in terms of computational speed and accuracy. In the proposed ML approach, inspired by Newton's second law's concept – from particle acceleration to calculate velocity and then position, a new continuous convolutional neural network (CNN) is established to predict the particles' accelerations first based on the particle–particle and particle–wall collisions separately, and the particle accelerations are used for calculating particle velocities and finally particle positions. The ML approach is applied to a typical granular flow - particle packing for demonstration. A dataset of 100 scenes of DEM simulations in one scenario is established for network training and examination. The results show that, in long-sequence predictions, the accuracy of the ML approach is three times higher than the previous ML approach. The effects of hyperparameters in the network are quantified. Then the ML approach with the optimized hyperparameters is used in additional three scenarios for further examining the prediction effectiveness. It is indicated that the improved ML approach can satisfactorily capture the morphology of granular flows under three new different scenarios; and the computational cost is only one-seventh compared to the DEM approach under the present conditions. The ML approach provides a simple and time-effective tool for simulating granular flows.