Abstract:
Uniaxial compressive strength (UCS) is substantially used mechanical parameters to observe and classification of rocks, but this test is subsersive, taking a long time and required well equipped laboratory conditions and properly prepared samples. Therefore it is important to estimate this parameter from other physico-mechanical rock parameters that are nondestructive, easy to prepare samples and required less time. Machine learning methods which are among these methods and increase their importance and validty are Multilayer Perceptron Neural Network (MLPNN), M5 Model Tree (M5MT), Extreme Learning Machine (ELM) methods. In this study, Brazilian tensile strength, ultrasonic P-wave velocity, shore hardness tests of different rock types (Basalt, limestone, dolostone) were performed. The results were used for estimating UCS using MLPNN, M5MT, ELM methods. The validation of models were checked root mean squared error (RMSE), mean absolute error (MAE), variance account for (VAF) and coefficient of determination (R2) and a10-index. Weights and bias values for MLPNN and ELM approaches and the tree structure for the M5MT method are presented. The result indicated MLPNN model outperforms the other models. Based on the result of predictive models with RMSE, MAE, VAF and R2 equal to RMSE: 1.3421, MAE: 0.7985, VAF: 99.7409, R2: 0.9982%. © 2020 Elsevier Ltd