Estimating UCS of South China sandstones using mineralogical and machine learning approaches

Slope stability analysis, rock mass classification, and foundation modeling necessitate measuring rocks' uniaxial compressive strength (UCS). Direct measurement is costly and time-consuming, prompting researchers to seek indirect methods. This research aimed to predict the UCS of sandstone samples using the quartz ratio and index properties. Models—including Feed-Forward Artificial Neural Network (FANN), Adaptive Neuro-Fuzzy Inference System (ANFIS), Support Vector Regression (SVR), K-Nearest Neighbor (KNN), and Multivariate Linear Regression (MLR)—were tested with varying input quantities and evaluated using Taylor's diagram, error level, A20 index, agreement index, and Calculated Performance Index (CPI). Petrography classified the sandstones as arenite, litharenite, and feldspathic litharenite; based on the results, the latter showed higher UCS, and fracture modes shifted from axial to multiple types as strength increased. Modeling revealed that KNN and FANN performance varied with distance metrics and training algorithms. Increasing inputs improved KNN and MLR accuracy but reduced SVR, ANFIS, and FANN accuracy. Additionally, the MLR's sensitivity to changes in inputs was greater than that of other methods. Comparing modeling results showed that the SVR based on the radial basis function with, CPI of 1.98, mean absolute percentage error of 0.75, A20 index of 1.00, and agreement index of 1.00, displayed the highest performance in UCS prediction. • Studied effects of physical and petrographic properties on sandstone geomechanics. • Predicted UCS using statistical, SVR, KNN, ANFIS, and FANN algorithms. • Assessed input count, KNN metrics, SVM kernels, and ANN training algorithms. • Increasing inputs improved KNN and MLR accuracy; reduced SVR, ANFIS, FANN accuracy. • SVR-RBF with 4 inputs achieved highest accuracy in UCS estimation.