Mean Block Size Prediction in Rock Blast Fragmentation Using TPE-Tree-Based Model Approach with SHapley Additive exPlanations

The optimum fragmentation size measures the quality of a blasting operation. Boulders or large fragments can result in more costs because they need secondary blasting, while small fragments can result in ore loss and dilution. Therefore, it is very significant to accurately predict the mean block size to reduce production costs and enhance efficiency. Due to the shortfalls of the empirical models, scholars have been inclined toward artificial intelligence (AI) techniques for fragmentation size prediction over the decades. Firstly, in this study, three tree-based models, i.e., the random forest (RF), extra-trees (ET), and CatBoost (CB), are employed for basic prediction. The model uses eight parameters, seven input parameters, and mean block size (MBS) as the output parameter. Secondly, their performance and hyper-parameters were fine-tuned using Bayesian optimization: tree-structured Parzen estimators (TPE) algorithm using Optuna. Among the three models, the TPE-ET model showed superior performance with the following metric scores on the training dataset: 0.9896, 0.0184, and 0.0003, and on the test dataset with the following metric scores: 0.9463, 0.0415, and 0.0017, i.e., R2, RMSE, and MSE, respectively. In conclusion, analysis by the SHapley Additive ExPlanations approach shows that elastic modulus significantly impacts the model's prediction of rock fragmentation.