Enhancing sedimentary rock brittleness prediction with interpretable machine learning and well log data

This study focuses on developing an interpretable intelligent framework for feature selection and an efficient model for brittleness index (BI) prediction of sedimentary rocks. The intelligent models are developed by coupling geophysical logging data with five machine learning algorithms. These models are evaluated with post-cross validation using statistical performance metrics, with hyperparameters optimised through a grid search procedure. Computationally resilient model interpretability is established through Shapley Additive Explanations (SHAP), enabling the identification of key geophysical features and ensuring consistency with established geomechanical principles. In contrast, the least square support vector machine (LSSVM) exhibited superior predictive performance, achieving an exceptionally high accuracy along with the lowest statistical error metrics on the cross-validated datasets. The LSSVM-SHAP analysis indicated that acoustic compressional and shear velocities are the most influential predictors for BI estimation, whereas rock bulk density, resistivity, gamma ray and neutron porosity exhibit relatively lower significance. The study demonstrated that the LSSVM interpretable intelligent framework not only improves prediction reliability for rock brittleness using well log data but also provides meaningful geophysical insights, thereby bridging the gap between data-driven techniques and domain knowledge. This interpretable and scalable methodology offers significant potential for application in complex drilling optimisation across geomechanics and geoengineering fields.