This study develops a series of statistically rigorous, interpretable multivariate regression models to predict the Factor of Safety (FoS) of adhesive rock bolts in underground excavations. Recognising the limitations of black-box machine learning models and overly simplified empirical methods, the research emphasises transparency, accuracy, and engineering applicability. Using synthetically generated datasets grounded in empirical ranges from literature, four progressively refined regression models were constructed, culminating in a final model (Model 4) featuring five key predictors: Installation Time After Blasting, Load-Carrying Capacity, Epoxy Peak Strength, Excavation Width, and Excavation Height. Model 4 achieved a near-perfect R² value (> 0.999) and minimal residual error, underscoring the internal dataset consistency and model fit. Backward multivariate regression and Elastic-Net methods further supported model parsimony and reduced multicollinearity. Residual diagnostics, variance inflation factors, and normality assessments validated the robustness of the models. A comparative review of existing literature, spanning machine learning models to experimental pullout tests, revealed that while many studies focus on prediction, few integrate model transparency, variable sensitivity, and diagnostics to this extent. This study contributes to bridging that gap, advocating for robust yet interpretable models in geotechnical engineering. The inclusion of continuous geological variables, such as the Geological Strength Index (GSI), improves sensitivity and predictive resolution. Recommendations emphasise real-world validation, adoption in design practices, and integration with diagnostic feedback loops for continuous model refinement. These findings lay the groundwork for safer, more reliable tunnel reinforcement design, especially in data-scarce environments or where black-box models are impractical.
Mollo et al. (Thu,) studied this question.