Abstract
Accurate rock mass classification is a fundamental requirement in road tunnel design and construction. The Rock Mass Rating (RMR) system, developed by Bieniawski (1989), remains one of the most widely adopted frameworks for evaluating rock mass quality; however, its application relies heavily on subjective engineering judgment, introducing variability and inconsistency in assessment outcomes. This study presents a comparative investigation of five supervised machine learning (ML) models — Random Forest (RF), Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), Multilayer Perceptron Artificial Neural Network (MLP-ANN), and Gradient Boosting (GB) — for the prediction of RMR classes in road tunnel construction scenarios. A synthetic geotechnical dataset of 2,000 samples was generated using the standard Bieniawski (1989) RMR formulation, incorporating six input parameters: Uniaxial Compressive Strength (UCS), Rock Quality Designation (RQD), joint spacing, joint condition, groundwater condition, and discontinuity orientation adjustment. Rock physics modeling principles, consistent with RokDoc software workflows, informed the parametric distributions used for data generation. Model performance was evaluated using test accuracy, five-fold cross-validation, confusion matrices, and classification reports. The RF model achieved the highest test accuracy of 98.7%, followed by XGBoost at 98.2%, GB at 97.9%, MLP-ANN at 96.4%, and SVM at 95.1%. SHapley Additive exPlanations (SHAP) analysis identified joint condition and RQD as the most influential predictors of RMR class. The findings demonstrate the viability of ML-based approaches for objective, rapid, and consistent rock mass classification, with significant implications for tunnel face assessment and geotechnical risk management in road tunnel projects.
Academians Publishers