A machine learning-embedded discrete element method: contact force surrogates for resolution-efficiency balance

Particle-particle contact dominates computational cost in conventional Discrete Element Method (DEM) simulations, especially at large scales, and machine learning (ML) has recently been applied by training models using DEM data, delivering significant speed-ups but producing distorted results and weak generalisation. In this study, we embed an ML model within the DEM contact-resolution process (MLE-DEM), capable of preserving critical contact-level information while accelerating DEM simulations. To demonstrate its effectiveness in balancing resolution and efficiency, the proposed MLE-DEM method is applied in eight cases. It achieves speed-ups of 20.66% in the rotary drum case, 32.23% in the random packing case, and 27.16% in the scaled-up packing systems, while maintaining high consistency with the ground-truth DEM results. Moreover, this MLE-DEM also outperforms previous ML-DEM approaches, achieving approximately smaller errors in the prediction of granular temperature (GT error of −0.00119 versus −0.03119, i.e., about 26 times smaller) and angle of repose (AOR error of 1.33° versus 8.35°, i.e., about 6.3 times smaller). A detailed analysis of the acceleration mechanism is also presented, showing that the observed speed-up is influenced not only by the evaluation time of the single contact force, but also by the difference in the number of collisions and the overall computing load. Overall, this new MLE-DEM model offers an effective balance among simulation accuracy, robustness, and efficiency, strengthening the reliability of ML-assisted DEM frameworks for broader applications.