A machine learning-driven model for predicting macro- and micro-mechanical responses of rockfill materials considering particle breakage

Conventional constitutive models encounter challenges in comprehensively capturing the nonlinear behavior and particle breakage effects of rockfill materials subjected to large deformation and multiaxial loading conditions. To address this issue, this study proposes a constitutive model based on machine learning (ML) that considers particle breakage in rockfill materials. By learning the underlying patterns from a large dataset of experimental data, this model can effectively reconstruct the nonlinear and high-dimensional characteristics of the material while also accounting for its loading history and stress path dependence. The model shows outstanding predictive performance on the test set, with relative prediction errors confined to within ± 5 %. Utilizing this model, the macro- and micro-mechanical responses of rockfill materials are systematically investigated, revealing the mechanisms and quantitative relationships between particle gradation, intermediate principal stress coefficient, and ellipsoidal axis ratio on particle breakage behavior. Additionally, the ML model exhibits robust interpolation within its training range (mean absolute percentage error, MAPE < 3.5 %), yet its extrapolation performance varies outside this scope, maintaining high accuracy for unbreakable particles and cyclic loading (MAPE < 2.7 %) but declining for extreme shapes (e.g., MAPE > 10.0 % for breakage rate), highlighting data dependency as a key limitation. This study brings fresh insights and establishes a novel theoretical framework for understanding the mechanical behavior of rockfill materials while also highlighting potential avenues for future model optimization.