Pioneering Coke Reactivity Index via Optimized Gradient Boosting Decision Tree‐Based Algorithms

Accurate prediction of the coke reactivity index (CRI) is vital for controlling the performance and efficiency of metallurgical coke production. This study aims to establish an advanced predictive framework based on gradient boosting machine (GBM) algorithm optimized through four metaheuristic strategies. A total of 636 coal samples obtained from an operating coke plant were analyzed, with eight physicochemical predictors measured according to ASTM standards. Data integrity was confirmed using the Isolation Forest technique, and model accuracy was assessed through five‐fold cross‐validation employing R 2 , MSE, RE%, and AARE% metrics. The results demonstrated that GBM models optimized via Bayesian probability improvement (BPI) and Gaussian process optimization (GPO) achieved superior prediction accuracy and generalization ( R 2 = 0.853 and total MSE ≈ 0.47) compared to other optimization methods. SHAP interpretation revealed volatile matter, sulfur, and ash as dominant features governing coke reactivity, while MMR, moisture, and fluidity exhibited inverse contributions. The developed hybrid optimization‐based gradient boosting decision tree (GBDT) methodology provides a robust and transparent modeling framework for industrial coke quality assessment and can be extended for real‐time evaluation within process control environments.