A multi-fidelity residual neural network based surrogate model for mechanical behaviour of structured sand

The structured sand presents significant interparticle bonding and anisotropy, resulting in significant differences in the physical and mechanical properties from the pure sand. This study proposes a new surrogate model based on the concept of multi-fidelity residual neural network (MR-NN) as an alternative to DEM simulation for predicting the mechanical behaviours of structured sand with different initial anisotropy and saving largely computational costs. The model is initially trained using low-fidelity (LF) data to focus on capturing the main underpinning correlations between macroscopic mechanical parameters with inter-particle properties and anisotropic state variables (i.e., microscopic fabric and tilting angle of sand), where the LF data are generated from previously proposed anisotropic macro-micro quantitative correlation. Subsequent training on sparser high-fidelity (HF) data is used to calibrate and refine the model with HF data generated from DEM simulations for anisotropic structured sand. Feedforward neural network (FNN) is adopted as the baseline algorithm for training models. The macroscopic anisotropic parameters of structured sand predicted by the surrogate model are compared with DEM simulations to examine its feasibility and generalization ability. Furthermore, the robustness of the surrogate model is examined by discussing the effect of LF data on the performance of MR-NN. The superiority of the MR-NN is further verified by comparing the performance of the trained MR-NN with the one-shot trained FNN based on the same HF data. All results demonstrate that the proposed surrogate model can provide a fast and accurate simulation of the anisotropic parameters of structured sand.