Interpretable Machine-Learning-Assisted Analysis of Caustic Leaching for Purification of Spent Cathode Carbon

Spent cathode carbon (SCC), a hazardous by-product generated during aluminum electrolysis, contains significant amounts of fluorides, cyanides, and other impurity phases that require effective purification before recycling or reuse. Caustic leaching is an important step in the purification of spent cathode carbon (SCC), but the nonlinear relationships between operating conditions and the carbon content of the caustic-leaching residue remain insufficiently quantified. In this study, an interpretable machine-learning (ML) framework was developed to quantitatively investigate the caustic leaching purification behaviour of SCC and evaluate the influence of key process variables on the carbon content of leaching residues. A dataset of 42 experimental observations was reconstructed from the published single-factor and Box–Behnken experiments reported by Yuan et al. The dataset describes the NaOH-based caustic-leaching stage and includes temperature, leaching time, liquid-to-solid ratio, alkali-related operating conditions, and stirring rate as input variables. Six ML algorithms, including GBDT, Random Forest, XGBoost, LightGBM, GBR, and CatBoost, were systematically evaluated. Among the evaluated models, CatBoost showed comparatively better predictive consistency in terms of testing-set R2 (0.9604) and RMSE (0.4087). However, given the limited dataset size (n = 42), five-fold cross-validation revealed substantial performance variability (mean R2 = 0.29 ± 0.59), underscoring the need for cautious interpretation. Therefore, the primary value of this study lies not in establishing a universally predictive model, but in demonstrating how SHAP analysis, when coupled with strict cross-validation, can extract meaningful process insights—specifically, the dominant roles of alkali concentration and liquid-to-solid ratio—even under small-data conditions. Overall, this study demonstrates the potential of interpretable ML as a complementary data-reanalysis tool for identifying model-attributed variable–response associations and prioritizing experimental variables for further validation in small-sample hydrometallurgical systems.