This study investigates the applicability of four distinct machine learning (ML) models – Random Forest (RF), Relevance Vector Machine (RVM), Support Vector Machine (SVM), and Artificial Neural Network (ANN) - for predicting the unconfined compressive strength (UCS) values of Precambrian basement rocks. The models employ input parameters such as rock aggregate index (Aggregate Crushing Value (ACV), Aggregate Impact Value (AIV), and Los Angeles Aggregate Value (LAAV)) and mechanical test index values (Schmidt Hammer Rebound Number (SHR) and Point Load Strength Index (IS(50)) for prediction purposes. Additionally, a multilinear regression (MLR) model was implemented to compare the predictive performance of the ML models against the traditional linear regression model. The developed models were validated using performance metrics such as Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Adjusted R-Squared (Adj.R2). The Friedman statistical test was also employed to investigate the significance of the differences in the performance of the models. The results indicate that all the models performed well and showed better predictive performance compared to the MLR model during both the training and testing phases, except for the ANN model, which exhibited slightly poorer performance than the MLR model in the testing phase. However, the SVM model demonstrated the best predictive accuracy among all the models with MAE, RMSE, and Adj.R2 scores of 0.76 and 1.45; 1.16 and 1.8; 0.98 and 0.90, respectively, for the training and testing phases. The difference in the performance of the predictive models was statistically significant, as revealed by the Friedman statistical test. A feature importance analysis was conducted to verify the significance of each input parameter, revealing that all input features had a positive effect on the output of all the models. However, IS(50) had the most significant impact on the models.
Read full abstract