Supervised intelligent prediction of shear strength of rockfill materials based on data driven and a case study

Abstract

The rockfill materials (RFM) are emerging and regarded waste reuse product in construction and mining engineering. In this paper, six distinctive supervised machine learning (SML) models, namely artificial neural network (ANN), extreme learning machine (ELM), random forest (RF), relevance vector regression (RVR), support vector regression (SVR), and extreme gradient boosting (XGBoost), are adopted to predict the RFM shear strength using 165 data cases with 13 features. To improve model performance, an improved physics meta-heuristic algorithm, named snow ablation optimizer with Logistic mapping (LogSAO), is utilized to select the optimal hyperparameter combination of the proposed models. Four popular statistical indices, including coefficient of determination (R2), root mean squared error (RMSE), Willmott’s index (WI), and Scatter index (SI) are employed to quantify the model performance. The model evaluation results show that the LogSAO-SVR model is the best prediction model with the highest values of R2 and WI while the lowest values of RMSE and SI in both training and testing phases (0.9985, 0.0257, 0.9996, and 0.0380; 0.9802, 0.0812, 0.9950, and 0.1335). Normal stress (Ns) is the most important feature with the highest importance score of 0.925 in predicting the RFM shear strength. Interestingly, the negative contribution of Ns to the RFM shear strength prediction will be transformed into the positive contribution when coefficients of curvature (Cu) values are small. The validation results demonstrate that the predicted values of the proposed LogSAO-SVR model are highly consistent with the measured values from the direct shear tests based on a case study.