The potential of a multi-fidelity residual neural network based optimizer to calibrate DEM parameters of rock-like bonded granular materials

The commonly used trial-and-error approach on selecting appropriate inter-particle parameters in DEM simulations incurs criticism such as user dependence and high computational cost. This study proposes a new framework based on a multi-fidelity residual neural network (MR-NN) as an alternative for calibrating interparticle parameters in rock-like bonded granular materials. The model is first trained using low-fidelity data (LF) to focus on capturing the main underpinning correlations between macroscopic elastic and strength parameters with inter-particle parameters of contact models, where the LF data is generated from micro-macro quantitative relations. Subsequent training on sparser high-fidelity (HF) data is then used to calibrate and refine the model, in which the HF data is generated from DEM simulations for rock. Feedforward neural network (FNN) is considered as the baseline algorithm for training models. The trained MR-NN with the same LF data is finally used to predict the inter-particle parameters of calcarenite and granite in DEM simulations to examine its feasibility and generalization ability. Furthermore, the robustness of the framework is verified by discussing the effect of LF data on the performance of MR-NN. All results demonstrate that the proposed method can provide a fast and accurate determination of inter-particle parameters for DEM simulation.