3D Visualization and Quantitative Evaluation of Graph Neural Network–Predicted Force Chains in Granular Materials

Deepak Solanki

Abstract

Force chains govern stress transmission in granular materials, yet direct measurement of inter-particle forces remains challenging. In this work, a Graph Neural Network (GNN) is implemented to predict the Normalized Particle Maximum Normal Contact Force (NPMNCF) at the particle level using data from discrete element method simulations. Granular assemblies are represented as graphs with particles and contacts modeled as nodes and edges, and an encode–process–decode architecture with iterative message passing is employed. The model is evaluated through training convergence, particle-scale correlation, and comparison of predicted and ground-truth force distributions. A persistent error analysis is further conducted to identify spatial regions where the model consistently underperforms, revealing that prediction failures concentrate at specimen boundaries due to incomplete graph neighborhoods and force chain termination gradients. Results show that the GNN captures particle-scale force heterogeneity and enables reliable identification of force-chain particles, providing a foundation for future contact-level force prediction and force-chain analysis.

The code and data generated for computational experiments are available here.

[PDF]