Physics-Informed Active Learning for Calibrating Mesoscopic Dynamic Parameters of Multiphase Concrete in DEM Simulations

The discrete element method (DEM) is widely used to simulate concrete failure, but calibrating its mesoscopic dynamic parameters is computationally expensive due to the high-dimensional parameter space. This study proposes a physics-informed active learning framework to autonomously calibrate these parameters under impact loads. An FDM-DEM coupled split Hopkinson pressure bar model is established to simulate macroscopic dynamic compressive responses. Subsequently, a Plackett–Burman experimental design reduces the parameter optimization space from 16 to 8 core dimensions. A multi-layer perceptron surrogate model is then constructed. By comparing two heuristic active sampling strategies, results indicate that a parameter priority-guided strategy incorporating physical priors significantly outperforms a mid-value exploration strategy. The proposed approach achieves coefficients of determination exceeding 0.9 for predicting multiple macroscopic dynamic indicators on an independent testing set. Building upon this forward mapping, a robust inverse parameter prediction mechanism is established, achieving a closed-loop reconstruction of 0.8662. This framework provides a reliable, data-efficient, and automated pathway for calibrating complex multiphase particulate systems.