Predicting packing density in packed bed reactors using convolutional neural networks trained on discrete element method simulations: Effects of pellet shape, material, and filling conditions

Accurately predicting the packing density of packed beds depending on the pellet shape and material properties under different filling conditions is crucial for optimizing packed bed processes. In this study, a Multi-Input 3D Convolutional Neural Network (CNN) is developed to predict the mean packing porosity from the pellet shape and five additional parameters, i.e., the tube-to-pellet diameter ratio, friction coefficient, restitution coefficient, Young’s modulus, and pellet fill rate. The model is trained on a dataset obtained from Discrete Element Method (DEM) simulations, covering full-body pellets and corresponding hollow variants with diverse outer shaping. The trained model accurately captures the complex interplay between pellet geometry, material properties, and filling conditions, providing consistent predictions for unseen configurations with a mean absolute percentage error (MAPE) of less than 1 %. This highlights its potential as a fast surrogate for DEM-based packing generation simulations, with broader applicability than conventional correlations, which are restricted to specific pellet shapes or materials. Moreover, an example use case of the neural network is presented, identifying pellet shapes that result in packings with high surface area and porosity, achieved with minimal computational demand. • Prediction of packing density in packed-bed reactors using deep learning. • 3D convolutional neural network for learning directly from the full pellet shape. • Discrete inputs for dimensioning, material properties, and filling conditions. • Training and validation data from DEM simulations using Ansys® Rocky®. • Identification of pellet shapes yielding high porosity and packing surface area.