Impact of rock discontinuity geometry on uniaxial compressive strength: Integrating the finite–discrete element method with data-augmented interpretable machine learning

Understanding how discontinuity geometric parameters affect the uniaxial compressive strength of complex rock masses (UCS RM ) is critical for rock engineering, but traditional methods face challenges in multiparameter analysis due to high costs and inefficiency. This study introduces an integrated framework of finite-discrete element method (FDEM), Mixup data augmentation, and interpretable machine learning to predict UCS RM . Expanding the dataset with Mixup, the gradient boosting regressor model achieves a notable accuracy improvement, with the coefficient of determination increasing from 0.813 to 0.899. The model predicts 1406 samples in 3 s—markedly faster than FDEM simulations, which require 20 days for the same task—demonstrating superior computational efficiency. Shapley Additive exPlanations analysis reveals the parameter importance ranking: bedding dip angle ( α ) > joint dip angle ( β ) > joint length ( l ) > joint intensity ( i ) > bedding thickness ( t ). UCS RM shows a U-shaped trend with increasing α or β , increases with t , and decreases with l or i . Parameter interactions indicate α governs the effects of β and t: β dominates UCS RM at α = 0°–30° or 75°–90°, while t’s influence strengthens at α = 30°–75°. Conversely, β controls l/i’s negative effects, most pronounced at β = 30°–75°. These findings offer a data-driven approach for efficient rock mechanical parameter prediction, guiding key parameter screening and numerical model simplification in tunnel and slope engineering. The framework bridges computational simulations and machine learning, providing an innovative solution for understanding complex rock mass behavior by integrating numerical modeling with interpretable artificial intelligence. • Established a multi-parameter FDEM dataset for discontinuity geometry and rock strength, enabling coupled effect analysis. • Developed a Mixup-augmented GBR model for UCS prediction, effectively improving computational efficiency over FDEM simulations. • The influence of discontinuity geometric parameters on UCS was revealed.