Physics-Informed Machine Learning for Predicting Stress Wave Transmission Across Realistic Rock Joints
Qiong Nie, Jicheng Zhang, Yingang Wang, Dingjian Wang, Qianyun Wang, Yuxuan Shi, Chonglang Wang
Abstract
Machine learning (ML) models are increasingly applied in rock mechanics. However, purely data-driven approaches trained on limited experimental data can often exhibit physically unrealistic results, such as violating established physical laws, which undermines their reliability. To address this, we introduce a Physics-Informed Machine Learning (PIML) approach utilizing XGBoost to predict the stress wave transmission coefficient (T) across realistic rock joints. Our model was developed using a dataset from 279 Split Hopkinson Pressure Bar (SHPB) tests conducted on granite samples replicating 80 unique natural joint morphologies. We integrated known physical constraints directly into the modeling process: (1) boundedness of T within [0, 1] enforced through logit and sigmoid transformations, and (2) hard monotonicity constraints ensuring T consistently increases with joint matching coefficient (JMC) and decreases with aperture (e). To ensure the model generalizes to unseen joint morphologies, we employed a rigorous Nested Leave-One-Group-Out (LOGO) cross validation strategy. The con-strained XGBoost model achieved R2 of 0.761 and demonstrated a statistically significant 20.55% reduction in Mean Absolute Error (MAE) compared to a Ridge Regression baseline. Post-hoc analyses using Partial Dependence and Individual Conditional Expectation plots confirmed that the model’s predictions are not only accurate but also physically realistic and interpretable. This PIML approach successfully guarantees physically consistent predictions and reaffirms the dominant influence of contact area parameters on stress wave transmission.