Leveraging machine learning for precision prediction of geomechanical properties of granitic rocks: a comparative analysis of MLR, ANN, and ANFIS models

Geomechanical properties such as uniaxial compressive strength (UCS) and modulus of elasticity (E) are essential for evaluating the mechanical behavior of rocks and crucial for rock engineering projects. Researchers frequently estimate these values indirectly because it is not always feasible to get high-quality core samples. Four machine learning models including multiple linear regression (MLR), multiple nonlinear regression (MNLR), artificial neural networks (ANN) and Adaptive Neuro-Fuzzy Inference System (ANFIS) were introduced and employed to examine the prediction of UCS and E in 221 granitic rock samples from Bataapáti nuclear reservoir in Hungary. The variables to calculate UCS and E are: density, P-wave velocity, shear strength and Brazilian tensile strength. The study compared the performance of the proposed machine learning models and found that the ANFIS model had obtained the most satisfactory prediction accuracy. Based on the sensitive analysis, the BTS is the most important variable for predicting UCS and Vp for E. Six popular performance indices, including the root-mean-square error (RMSE), mean absolute error (MAE), the coefficient of determination (R2), mean absolute percent error (MAPE), variance accounts for (VAF) and a20-index were adopted to evaluate all models. The ANFIS model demonstrates outstanding predictive performance for UCS, achieving an R²=0.99, RMSE = 2.445, MAE = 1.489, MAPE = 1.3%, VAF = 99.16%, and an a20-index of 1 for the training datasets. For the testing datasets, it attains an R²=0.97, RMSE = 5.544, MAE = 3.502, MAPE = 3.1%, VAF = 96.43%, and an a20-index of 1. Similarly, the model excels in predicting E with an R²=0.93, RMSE = 3.914, MAE = 2.91, MAPE = 5.2%, VAF = 92.23%, and an a20-index of 1 for training datasets, and an R²=0.92, RMSE = 3.195, MAE = 2.613, MAPE = 4.4%, VAF = 92.22%, and an a20-index of 1 for testing datasets. Also, a sequence of Taylor diagrams was generated to assess the performance of the models. The findings suggest that the ANFIS model is more effective in predicting UCS and E of rocks compared to other methods. However, its application should be approached cautiously, especially at the initial design phase, and limited to specific rock types. © The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature 2024.