Method of soil-elastoplastic DEM parameter calibration based on recurrent neural network

In order to accurately simulate the elastoplastic macroscopic mechanical behavior of soil, the microscopic mechanical parameters in the particle contact model must be carefully calibrated. In this paper, the Discreat Element Method (DEM) parameter calibration network is constructed based on a recurrent neural network, and the soil stress-strain training and test dataset are established by DEM simulation. According to the results of the parameter sensitivity analysis of the soil compression simulation test, the evaluation criteria for model prediction accuracy are determined. Furthermore, optimize the model hyperparameters. Based on the optimized model weight and accuracy evaluation index, the difference between the traditional Design of Experiment (DoE) parameter calibration method and the Deep Learning (DL) method is compared. The relationship between simulation efficiency and model prediction accuracy is also analyzed. Finally, the actual soil sample parameters were predicted and verified by simulation. This research can provide a perspective for parameter calibration method.