Three machine learning techniques comparison for the prediction of the uniaxial compressive strength of carbonate rocks

In engineering practices, it is critical and necessary to either measure or estimate the uniaxial compressive strength (UCS) of the rock. Measuring the UCS of rocks requires comprehensive studies in the field and in the laboratory for the rock block sampling, coring and testing. These studies are time-consuming, expensive and go through difficult processes. Alternatively, the UCS can either be estimated by empirical relationships or predictive models with various measured mechanical and physical parameters of the rocks. Previous studies used different methods to predict UCS, including least squares regression techniques, adaptive neuro-fuzzy inference system, artificial neuron networks and others. This study is intended to estimate the UCS of the carbonate rock by using a simple, measured Schmidt Hammer (SHV C ) test on core sample and a unit weight (γ n ) of carbonate rock. Principal components regression, multiple regression and artificial neural networks are employed to predict the UCS. We are not aware any study compared the performance of these methods for the prediction of the UCS values. The results of those models are very close; however, based on both roots mean square error (RMSE) and mean absolute error (MAE), artificial neural networks (ANN) have a slight advantage against the other two models.