Physics-Consistent Monotonic Machine Learning for RMR Prediction from Q-System Parameters Under Duplicate-Aware Validation

Abstract The Q-system and rock mass rating (RMR) are widely used for rock mass classification in tunnels and underground excavations. Reliable prediction of RMR from Q-system parameters can support cross-checking between different classification systems. However, empirical Q–RMR correlations may not fully describe nonlinear effects among individual Q-system parameters. Meanwhile, flexible machine learning models may produce optimistic accuracy and engineering-inconsistent responses when repeated-input states are shared between training and testing data. To address these issues, this study develops a physics-consistent and duplicate-aware machine learning framework for continuous RMR prediction from six Q-system parameters, namely RQD, Jn, Jr, Ja, Jw, and SRF. The framework integrates monotonic gradient boosting, duplicate-aware grouped validation, boundary-sensitive diagnosis around the Poor–Fair transition threshold, and interpretable analysis using SHAP and accumulated local effects. A published field-derived dataset with 356 observations was used for model development. The response distribution was strongly imbalanced: 345 observations (96.9%) were within the Fair range, whereas only 11 observations (3.1%) had RMR ≤ 40. The observed RMR range was 36.00–59.43, and no Very Poor, Good, or Very Good rock mass conditions were represented. The data audit showed a strong repeated-input structure, with only 184 unique six-parameter combinations. Under conventional random splitting, boosting models showed high apparent accuracy, but their performance decreased under duplicate-aware grouped validation. This indicates that repeated Q-system states can lead to optimistic performance estimates if they are not considered during validation. Under duplicate-aware validation, CatBoost achieved the best overall numerical accuracy, while Monotonic XGBoost provided a better balance between prediction performance and engineering consistency. The selected monotonic model eliminated all observed monotonic violations and reduced the boundary-crossing error rate from 53.85% to 38.46% compared with the best unconstrained model. External validation on an independent high-geostress tunnel dataset further showed acceptable transferability, with an R 2 of 0.8650. SHAP and ALE analyses identified Ja and SRF as the dominant contributors and confirmed that the learned response directions were consistent with rock engineering knowledge. The results demonstrate that validation reliability and engineering consistency are as important as predictive accuracy in machine learning-based rock mass classification. The proposed framework provides a transparent and physically meaningful tool for supporting RMR estimation from Q-system data, especially within the investigated Fair-dominated range and near the Poor–Fair transition boundary. Accordingly, the proposed model should not be interpreted as an all-class RMR predictor. Its validated scope is limited mainly to continuous RMR estimation within the investigated Fair-dominated range and exploratory diagnosis near the Poor–Fair transition.