Accelerating granular dynamics simulations: A graph neural network surrogate for complex high-energy ball milling

Abstract

While the Discrete Element Method (DEM) provides high-fidelity insights into granular processes like high-energy ball milling, its computational cost can become prohibitive when exploring extensive parameter spaces required for scale-up. This limitation hinders the rapid design and optimization cycles crucial for emerging applications, like mechanochemistry. Surrogate modeling offers a promising path to overcome these computational barriers, yet existing approaches often struggle to accurately represent the complex, moving boundary conditions typical of milling equipment. In this work, we leverage a Signed Distance Function Graph Neural Network (SGN) surrogate tailored to the high-energy, moving-boundary regime of the Emax mill. Trained on DEM data, the SGN jointly predicts particle kinematics for recursive roll-out and a mechanochemistry-relevant global quantity, the global dissipated energy. The model exhibits strong generalization to unseen motion trajectories and moderate jar-shape edits without retraining, while operating with a timestep over 100x larger than required by DEM. In a CPU-only comparison, it achieves a minimum of 6.6× wall-clock speedup. This approach provides a powerful and promising technique for the simulation, analysis, and design optimization of high-energy ball milling equipment.