Abstract: Currently, there are numerous methods for evaluating rock brittleness index, which are mainly based on mineral composition or
rock mechanical properties, but most evaluation indicators are costly and time-consuming to obtain. By utilizing machine learning technique, a rock brittleness index prediction method based on Stacking ensemble learning concept has been proposed, which involves parallel training of Gradient Boosting Decision Tree model (GBDT), Random Forest model (RF), Naive Decision Tree model (DT), Support Vector Regression model (SVR), and LightGBM model. After hyperparameter optimization using a tree-structured Parzen estimator for each model, the XGBoost model has sequentially been used to fuse training results of the base models to achieve rapid parameter optimization and prediction of rock brittleness index. The study results indicate that the prediction results using two-layer Stacking model optimized by tree-structured Parzen estimator show significant advantages compared to those by the base models, with the explained variance score (EVS) reaching up to 0.97, and the coefficient of determination (R²) reaching a maximum of 0.967. Given the same dataset, this model obtains the lowest Mean Absolute Error (MAE) and Root Mean Square Error (RMSE), indicating that it can effectively fit the variation pattern of rock brittleness index in the technical context of supervision and learning, which verifies its practical value in predicting rock brittleness index.

Key words: rock brittleness index, Stacking model, ensemble learning, tree-structured Parzen estimator