An adaptive UDEC calibration tool for modelling intact rocks based on machine learning

Determining modelling parameters in the universal distinct element code (UDEC) for modelling intact rocks has been a fundamental issue. Despite the promise of machine learning (ML) in addressing this issue compared to traditional trial-and-error and sensitivity analyses, existing attempts have been limited to specific experimental inputs or ML models. To overcome this limitation, we developed an open-source adaptive ML-based UDEC calibration tool, named UPCal, featuring compatibility across experimental inputs and ML models. In UPCal, an existing or user-defined numerical database is fed into optional ML models to establish the mapping from modelling parameters to modelled properties. Given complete or incomplete experimental data (target modelled properties), UPCal automatically outputs the corresponding UDEC modelling parameters according to the established mapping integrated with grid search. For validation, a comprehensive experimental database was input into UPCal to derive the corresponding UDEC modelling parameters, followed by UDEC simulations using these UPCal-calibrated modelling parameters. Comparisons between the modelled properties and experimental data demonstrated the feasibility and robustness of UPCal. We also found that the calibration accuracy varied with both experimental inputs at different completeness levels and ML models, upon which an optimal ML calibration strategy was proposed. The potential impacts of lithology and block size were also observed. As a powerful calibration tool, UPCal provides valid UDEC parameters for modelling intact rocks. With field conditions integrated, UPCal enables credible UDEC simulations in rock mass engineering.