On deep learning calibration for DEM simulation of sedimentary and igneous rocks

We propose a Deep Learning (DL)-based calibration procedure of particle-based material model parameters for rock materials simulated by the Discrete Element Method (DEM). We create Discrete Element models using Bond-Particle Models to simulate Uniaxial Compression tests and Brazilian tests commonly used for experimental determination of material parameters. After simulation, the microscopic parameters of the model are fed into a constructed DL system based on the Multi-Layer Perceptron regressor as input data and the functional depending on macroscopic critical stresses and strains as the output. As a result, a set of optimal microscopic parameters has been found and tested in a simulation using the same DEM Model. Then, we compared the resulting stress–strain curve with the experimental data used as a reference for calibration. Additionally, the stability of the obtained model in terms of reproducibility of the stress–strain curve has been tested by simulating Uniaxial Compression for DEM models with different particle packing. The discrete element models obtained with material parameters have showcased excellent stability and consistency with experimental data, especially for hard rocks. For soft rocks, the model showcases nonlinear elastic behavior upon loading. The procedure can be used to simulate rocks in geomechanics and geoengineering problems demanding discontinuous rock description, such as multiple crack formation, drilling, boring, and rock blasts. • Deep learning can be effectively used to calibrate DEM bond-particle models. • Uniaxial compression and Brazilian disk tests are enough for model calibration. • The strain-stress curve for granite is well-approximated by the bond-particle model. • Bond-particle model better describes hard sedative rocks than soft ones.