Development of a reliable rock slope stability model utilizing field and analytical data – An integration of FE-ML approaches

Slope instability in hilly regions is a highly complex phenomenon, with triggering factors ranging from natural events to anthropogenic activities. Such failures hit disastrous losses both in terms of material as well as life. It is necessary to comprehend the mechanism of these failures to mitigate such events and also to predict their vulnerability for better preparedness. Significant advancements have already been done in the area of slope stability analysis, and scores of valued tools and techniques have been developed, such as limit equilibrium methods, finite element and finite difference methods, stochastic methods, and several of their combinations. In this study, an attempt has been made to capitalize on machine learning tools to predict the factor of safety of rock slope stability in hilly regions. Three road-cut slopes have been considered and their stability is determined using both finite element (FE) and machine learning (ML) techniques. The idea to intertwine these approaches is to supplement each other and enhance the reliability of the results. The geotechnical data was acquired through field investigation trips to the adopted mountainous sites. Since the slopes at the site are rocky, in the FE model, the Generalized Hoek Brown (GHB) material model with shear strength reduction technique have been used. In the implementation of ML models, Random Forest (RF) and Gradient Boosting Machine (GBM) models have been used. For the training of the ML model, ample published data has been utilized, while for testing the ML model, the data from the current slope site is used. The analysis in ML model is carried out in three stages: a) without Hyperparameter tuning, b) with Hyperparameter tuning using GridSearchCV, and c) Pipeline incorporating Recursive Feature Elimination (RFE). Performance metrics, including Mean Absolute Error (MAE), Mean Squared Error (MSE), and R2 score, were evaluated to assess the accuracy of the model. A slight discrepancy within a range of 10 percent has been found, which is rather expected due to factors such as grid refinement and, data volume and variability. Overall, the proposed ML model demonstrates excellent compatibility with the FE model results. This study is an attempt to pick relevant ML techniques to develop a purpose-built framework that has the potential to validate the rock slope stability obtained using the traditional methods.