Identification of DEM simulation parameters by Artificial Neural Networks and bulk experiments

Abstract In Discrete Element Method (DEM) simulations, particle–particle contact laws determine the macroscopic simulation results. Particle-based contact laws, in turn, commonly rely on semi-empirical parameters which are difficult to obtain by direct microscopic measurements. In this study, we present a method for the identification of DEM simulation parameters that uses artificial neural networks to link macroscopic experimental results to microscopic numerical parameters. In the first step, a series of DEM simulations with varying simulation parameters is used to train a feed-forward artificial neural network by backward-propagation reinforcement. In the second step, this artificial neural network is used to predict the macroscopic ensemble behaviour in relation to additional sets of particle-based simulation parameters. Thus, a comprehensive database is obtained which links particle-based simulation parameters to specific macroscopic bulk behaviours of the ensemble. The trained artificial neural network is able to predict the behaviours of additional sets of input parameters accurately and highly efficiently. Furthermore, this method can be used generically to identify DEM material parameters. For each set of calibration experiments, the neural network needs to be trained only once. After the training, the neural network provides a generic link between the macroscopic experimental results and the microscopic DEM simulation parameters. Based on these experiments, the DEM simulation parameters of any given non-cohesive granular material can be identified.