Calibration of Mesoscopic Parameters for Lunar Soil Using a Machine Learning–Based Discrete Element Method

Accurate calibration of the mesoscopic parameters of lunar soil is essential for effectively simulating and analyzing lunar engineering problems using the discrete element method (DEM). However, conventional trial-and-error approaches are often time-consuming and inefficient, particularly when multiple parameters require calibration. This study proposes a novel calibration method for the mesoscopic parameters of lunar soil based on machine learning techniques. Initially, the predictive performance of four machine learning models—support vector machine, backpropagation neural network, random forest, and extreme gradient boosting (XGBoost)—was evaluated and compared. The best-performing model was then further optimized using four swarm intelligence algorithms: gray wolf optimizer (GWO), particle swarm optimization, satin bowerbird optimization, and the Kepler optimization algorithm. The effectiveness of the optimization was assessed using a comprehensive scoring method. The results show that the XGBoost model achieved the highest prediction accuracy and computational efficiency, with the GWO-optimized XGBoost model delivering the best overall performance. The mesoscopic parameters predicted by the optimized model yielded strength parameters that closely matched actual results, demonstrating the model’s reliability. Overall, the proposed method enables rapid and accurate calibration of mesoscopic parameters and significantly improves the efficiency of DEM-based analysis. It offers a reliable foundation for addressing practical problems and supporting future lunar surface engineering applications.