Parametric calibration in bonded block models for simulating mechanical behaviours of intact rocks using machine learning

Despite widespread adoption of the bonded block model (BBM) in modelling intact rocks, the calibration of BBM modelling parameters remains a significant challenge, undermining the trustworthiness of BBM-simulated results. Existing trial-and-error and sensitivity analyses for calibration suffer from inefficiency, subjectivity, and difficulty in establishing the high-dimensional and nonlinear complex mapping from modelling parameters to modelled properties in BBM. To address this issue, built on BBM-based universal distinct element code (UDEC), we employed machine learning to clarify this complex mapping. A comprehensive numerical database with 3456 UDEC simulations was constructed for training machine learning models, followed by the selection of the optimal machine learning models by comparing their predictive performances. Subsequently, we collected experimental data from 99 rock types that served as modelled properties to be input into the selected trained machine learning models. Through an inversion by integrating grid search, the corresponding modelling parameters could be output, that is, the machine learning–calibrated modelling parameters. They were further imported into UDEC to perform another 1485 simulations to validate their reliability and robustness. It was also found that both lithology and block size affect calibration accuracy differently across modelled properties. In applying the framework, specific rock model configuration may be considered when establishing the numerical database, including the constitutive laws of blocks and contacts and specific rock structure. This study provides an effective solution for parametric calibration in BBM, advancing more reliable use of BBM in scientific and engineering contexts.