Introduction
Rock shear strength parameters are critical for assessing slope stability, excavation safety, and tunnel design. Traditional approaches often rely on empirical correlations or single-model predictions that struggle with complex, nonlinear relationships in geological data. This article presents a novel prediction framework that integrates data-driven modeling with interpretability analyses to predict rock shear strength parameters more accurately and transparently. By addressing the limitations of single models and the randomness in hyperparameter selection, this framework enhances reliability for real-world geotechnical applications.
Challenges in Predicting Rock Shear Strength Parameters
Geotechnical data are typically heterogeneous, featuring nonlinear interactions between mineralogy, grain size, porosity, confining pressure, moisture, and rock type. These factors jointly influence shear strength in ways that single models may fail to capture. Moreover, hyperparameter randomization in many machine learning pipelines can lead to inconsistent results across runs, undermining confidence in the predictions. The proposed method tackles these challenges by combining data-driven learning with robust interpretability techniques to reveal how input features drive the predicted shear strength parameters.
Framework Overview
The core of the framework consists of three interconnected layers: data-driven modeling, hyperparameter stabilization, and interpretation. Each layer is designed to improve accuracy while maintaining clarity about how predictions are formed.
Data-Driven Modeling Layer
We employ a suite of complementary algorithms capable of capturing nonlinearities and interactions among features. Ensemble methods (such as gradient boosting and random forests) are used alongside neural-network-inspired models when data volume permits. The model selection process emphasizes robustness, cross-validation, and reproducibility to mitigate random hyperparameter effects. Importantly, the framework emphasizes feature engineering tailored to rock mechanics, including effective stress, consolidation state, mineral content, and saturation level.
Hyperparameter Stabilization
To address the hyperparameter random-selection problem, the framework utilizes consensus techniques and stability metrics. Instead of relying on a single best model, multiple models with varied hyperparameters are trained, and their predictions are aggregated. This ensemble approach reduces variance and provides more stable estimates of shear strength parameters, while preserving interpretability through post-hoc analysis.
Interpretability and Explainability
Interpretability is integral to the framework, ensuring the results can be trusted by engineers. Techniques such as SHAP (Shapley Additive Explanations) values, feature importance analysis, and partial dependence plots help elucidate how each input feature contributes to the predicted shear strength. This transparency supports diagnostic checks, feature selection, and domain-informed adjustments without sacrificing performance.
Data and Feature Engineering
Geotechnical datasets used for training combine laboratory test results, field measurements, and material properties. Features include effective stress, mean effective normal stress, cohesion, friction angle proxies, rock type indicators, porosity, moisture content, and confining pressure. Feature engineering focuses on capturing the physics of shear resistance, such as the interaction between mineral rigidity and pore pressure, and the role of saturation on shear response.
Results and Practical Implications
Initial experiments demonstrate that the integrated framework achieves higher predictive accuracy and lower error variance than any single-model baseline. The interpretability layer provides actionable insights, revealing which factors most strongly influence rock shear strength in specific geological contexts. Practically, engineers can use these predictions to optimize excavation schedules, assess slope stability, and design supports with quantifiable confidence.
Limitations and Future Work
While promising, the framework depends on high-quality, diverse datasets that cover a wide range of rock types and environmental conditions. Future work may involve incorporating physics-informed constraints, expanding interpretability to real-time monitoring data, and exploring transfer learning across mine sites and geological formations.
Conclusion
The proposed data-driven and interpretability-focused framework offers a robust solution for predicting rock shear strength parameters. By combining ensemble learning with transparent explanations and stabilizing hyperparameters through consensus, the approach addresses both accuracy and reliability, delivering practical value for geotechnical engineering projects.
