AI calibration of modelling parameters in UDEC

The Universal Distinct Element Code (UDEC) based on Discrete Element Method (DEM) has gained widespread prevalence in simulating multiscale rock failure in varied branches of geotechnics. Simulation performance demonstrates a heavy reliance on modelling parameters. The trial-and-error approach and parametric sensitivity analysis have long been the primary method employed in the parametric calibration in UDEC. However, they share the drawbacks of excessive computational resources, high dependence on human subjectivity, and the challenge of handling high-dimensional and nonlinear complex parameter spaces. To address this issue, we employed artificial intelligence (AI) to handle multidimensional data and higher-order nonlinear relationships between modelling parameters and macroscopic responses of numerical models. A wide range of preset gradient-based modelling database was established to pre-train the machine learning model to map the parametric relationships. Then, this pre-trained model was combined with an experimental database with various lithologies to conduct an inverse search of the input parameters in UDEC. To further improve the estimates, a gradient-based hyperparameter optimisation, implemented via GridSearch, was applied to identify the optimal parameter set by minimising the loss function. The calculated modelling parameters were subsequently input into UDEC for simulation and validation. Hundreds of comparisons reveal that the simulated results by UDEC align closely with those from the experimental database, demonstrating the feasibility of our model. This research provides a substantive solution to the parametric calibration in UDEC, significantly improving both the reliability and convenience of UDEC simulations.