DEM Study Investigating the Effect of Particle Shape on Compaction of Realistic Non-Spherical Particles Using Convolutional Neural Network

Abstract

This study focused on investigating the deformation behavior of non-spherical particles under high-load compaction, utilizing the multi-contact discrete element method (MC-DEM). To account for the non-spherical shape of the particles, two methods were employed: the bonded multi-sphere method (BMS) and the conventional multi-sphere (CMS). The BMS approach yielded accurate results in predicting the compression behavior of a single rubber sphere, while the CMS method failed to replicate the same behavior. Building on these findings, the BMS method was utilized to study the uniaxial compaction of Avicel® PH 200, a popular choice of excipient due to its ability to enhance the stability, flowability, and compressibility of tablet formulations. The results obtained from this study showed very good agreement with experimental data. To generate realistic 3D models of particles, a novel approach was introduced, which combines 2D projections and deep learning algorithms utilizing a 3D convolutional neural network (3D-CNN) methodology. Surrogate model was used to overcome the computational cost of DEM simulations. The findings of this study offer a valuable tool for researchers and engineers to efficiently and accurately generate 3D models of particles, leading to new insights and innovations in a range of applications such as rock and mineral analysis, battery materials, pharmaceuticals, and space exploration.