A methodology for calibrating parameters in discrete element models based on machine learning surrogates

The discrete element method (DEM) is well suited for calculating the behaviour of bulk materials. However, its application is limited because of the cumbersome calibration process required. Trial and error calibration can be useful for the computation of single outputs, but is unfeasible when the aim is reproducing more complex phenomena with high accuracy. This paper describes an iterative procedure based on machine learning to automatically calibrate the parameters of DEM models for reproducing the behaviour of bulk materials. The performance of the methodology is assessed by its application to the calibration of a DEM model to compute the stress–strain evolution of a cohesive material under uniaxial compression. In this case, a random forest model is used in conjunction with the iterative calibration algorithm proposed. The results of this study show that the algorithm is accurate and flexible for the calibration of material parameters.