CNN-based calibration of discrete element method parameters for calcareous sand

The selection of appropriate inter-particle parameters in discrete element method (DEM) simulations is crucial, and the commonly used trial-and-error method has been criticised for its uncontrollability and high computational cost. Therefore, this study proposes a new framework based on convolutional neural networks (CNN) as an alternative method for calibrating inter-particle parameters in calcareous sand materials. Firstly, a biaxial test dataset for calcareous sand was generated using DEM simulations. This data set was then used to train a CNN to capture the primary underlying correlation between macroscopic mechanical properties and the inter-particle parameters of the contact model. To demonstrate the powerful performance of CNN, this paper also established a back-propagation neural network (BP) and gated recurrent units (GRUs) as control experiments. The results showed that the CNN had higher prediction accuracy compared to the BP and GRU models. After DEM simulation using the parameters predicted by the CNN, it was found that the stress-strain curves and failure patterns closely matched the results of the laboratory tests. This confirms that the CNN can quickly and accurately determine the inter-particle parameters for DEM simulation and verifies the robustness of the CNN model in predicting laboratory test results, this method provides a reference for the calibration of DEM parameters for calcareous sand, thereby offering strong support for the use of calcareous sand in marine development and construction projects.