Machine learning approaches to predicting the uniaxial compressive strength of granite from image-derived mineralogical features

This study uses machine learning (ML) models to predict the numerically simulated uniaxial compressive strength (UCS) of granite directly from digital images. Specifically, the images were processed using an in-house Digital Image Processing (DIP) tool to estimate mineralogical features (used as input features for the ML models), including mineral content, grain size, and spatial distribution. Mineral content and distribution were quantified using m-harmonic Fourier series equations, whereas mineral grain size was determined using the 4-connectivity method. The target UCS values were derived from the 2D physically informed Subspring Network Breakable Voronoi (SNBV) microstructural models, replicating the mineralogical features observed in the granite images. Extreme Gradient Boosting (XGBoost) models with different input combinations and hyperparameter optimization methods were trained and evaluated on 126 granite images using a single train/test split and repeated 5-fold cross-validation. Results indicate that the input combination of mineral content, grain size, and spatial distribution parameters from m-harmonic Fourier series combined with SHapley Additive Explanations (SHAP)-based feature selection, yield the best and robust performance, whereas increasing the harmonic order has a limited effect on accuracy. Among the tested optimization methods, the Optuna–XGBoost model achieved the best performance. In addition, UCS prediction is controlled mainly by the content and grain size of biotite and plagioclase, while the corresponding attributes of quartz/K-feldspar, as well as overall mineral distribution play a comparatively minor role.