Prediction of force chains for dense granular flows using machine learning approach

Force chain networks among particles play a crucial role in understanding and modeling dense granular flows, with widespread applications ranging from civil engineering structures to assessing geophysical hazards. However, experimental measurement of microscale interparticle contact forces in dense granular flows is often impractical, especially for highly complex granular flow systems. On the other hand, discrete-based simulation approaches suffer from extremely high computational costs. Thus, this study proposes an innovative machine-learning framework aimed at accurately predicting the force chain networks in dense granular flows, using particle-scale and bulk-scale flow features, and novel topological parameters. A deep neural network was developed, achieving an excellent accuracy of 94.7%, recall of 100%, precision of 90.3%, and an f1-score of 95% for non-Bagnold type flow, where the force chains significantly affect flow characteristics. In addition, to enrich the future application of the proposed model, we introduce an experimentally accessible feature set, demonstrating effective performance in detecting force chains. More importantly, our analysis of feature importance using Shapley additive explanations values facilitates informed decision-making when identifying force chains in real-world dense granular flow experiments. The proposed machine-learning architecture will be of interest and essential for any dense granular flows where detecting force chains proves to be exceedingly challenging.