Single-Layer Graph Neural Networks for Discrete Element Method Simulations

Graph neural networks (GNNs) have demonstrated success as surrogate models for fluid simulations with smoothly varying dynamics, but their efficacy for discrete element method (DEM) simulations remains questionable. Despite this, multilayer Encoder-Processor-Decoder (EPD) architectures from fluid domains have been uncritically applied to DEM in limited prior work. We systematically demonstrate that these fluidoptimized architectures fail to capture the fundamental contact mechanics governing DEM particle collisions, instead encoding statistical bulk behavior that breaks down under rigorous testing. Through novel single-particle and double-particle collision tests, we identify critical shortcomings in the physical fidelity of current approaches. We introduce a single-layer message-passing GNN with physics-informed features that explicitly encode HertzMindlin contact laws and normal-tangential force decomposition. Experimental results confirm that our architecture achieves lower positional error in collision scenarios and improved energy conservation, while delivering faster inference with fewer parameters compared to multi-layer counterparts. This work establishes that physically-constrained architectural choices outperform generic deep learning approaches for contact dynamics in DEM simulations.