Prediction of random packing density and flowability for non-spherical particles by deep convolutional neural networks and Discrete Element Method simulations

Abstract

Non-spherical particles can be found in many industrial processes. Unfortunately, scalar shape parameters are often insufficient to predict their complex behavior. Therefore, this study used convolutional neural networks (CNN), which are able to interpret abstract shape features to derive the flow and packing behavior. For training data generation, experimentally validated Discrete Element Method simulations of silo filling and discharge were conducted. The subsequent training of the neural network was achieved with extensive input data augmentation. The CNN was then applied for the investigation of common shapes like ellipsoids or frustums for validation. The trained network showed good accordance to literature and high predictive robustness. Finally, the CNN assessed various food grains and tablet shapes. The network predicted the densest packing and best flowability for rounded lentils, whereas long rice kernels showed poor flow behavior. For pharmaceutical tablets, rounded and hexagonal shapes were predicted to yield denser packings than plain cylinders.