Advanced machine learning approaches for uniaxial compressive strength prediction of Indian rocks using petrographic properties

To design and plan effectively, understanding the engineering properties of rocks, such as the Uniaxial Compressive Strength (UCS), is crucial. While direct laboratory testing is a common method for determining UCS, it is often time-consuming, costly, and challenging, especially for soft or highly fractured rocks. Numerous past studies have explored indirect methods using basic rock properties for UCS prediction, and many correlations have been established. This study introduces an innovative indirect approach using machine learning models, including Multiple Linear Regression, Support Vector Regression, Bi-directional Long Short-Term Memory, and two hybrid models’ i.e., artificial neural networks with teaching learning-based optimization (ANN-TLBO) and particle size optimization (ANN-PSO) using Petrographic properties. This study utilized 54 rock samples from three different geological locations in India, with corresponding thin sections prepared and tested in the laboratory. The prediction of UCS was based on unit weight and petrographic analysis, specifically mineral compositions. Statistical analysis showed that the ANN-PSO model outperformed others, achieving R² and TIC values of 0.9911 and 0.0134 during training, and R² = 0.9868 and TIC = 0.0154 during testing. We conducted various statistical analyses, including a heatmap matrix, Taylor plot, REC curve and uncertainty analysis to evaluate the model performance. Our results indicate that machine learning models, particularly ANN-PSO, offer a reliable and efficient alternative to traditional methods for predicting UCS. Furthermore, Sensitivity analysis was investigated to check the importance of input parameters to the output parameters. Results identified quartz as the most significant parameter influencing UCS prediction.