Integrating Graph Neural Network-Based Surrogate Modeling with Inverse Design for Granular Flows

Abstract

Granular flows are central to a wide range of natural phenomena and industrial processes such as landslides, industrial mixing, and material handling and present intricate particle dynamics challenges. This study introduces a novel approach utilizing a Graph Neural Network-based Simulator (GNS) integrated with an inverse design for optimizing Discrete Element Method (DEM) parameters in granular flow simulations. The GNS model, trained on data sets generated from high-fidelity DEM simulations, exhibits enhanced predictive accuracy and generalization capabilities across various materials and granular collapse scenarios. Methodologically, the study contrasts the GNS approach with conventional Design of Experiment (DoE) methods, highlighting its enhanced computational efficiency and dynamic optimization capacity for complex parameter interactions in granular flows. The results demonstrate the GNS method superiority over the DoE in terms of computational speed and handling intricate parameter relationships. This work offers an advancement in computational techniques for granular flow studies, showing the potential of using differential simulations for realistic problems.

Multiscale Modeling and Artificial Intelligence for Multiphase Flow Science

“…Zhong et al developed a deep learning model with convolutional neural networks that efficiently predicts spatiotemporal distributions in bubbling fluidized-bed biomass fast pyrolysis, enabling low-cost, high-accuracy simulations. Jiang et al proposed a graph neural network-based simulator (GNS) for optimizing DEM parameters in granular flow simulations, outperforming the conventional design of experimental methods in computational efficiency and parameter optimization. Omar applied recurrent neural networks, specifically long short-term memory (LSTM), for multistep time-series predictions in complex multiphase systems such as mixers.…”