Machine learning for prediction and experimental validation of hydrochar characteristics via co-hydrothermal carbonization of sewage sludge and pomelo peel

Optimizing the co-hydrothermal carbonization (co-HTC) of sludge and biomass requires understanding complex nonlinear relationships among feedstock characteristics, process parameters, and hydrochar properties. This study constructs an integrated machine learning framework encompassing prediction, interpretation, and experimental validation. Ridge Regression was introduced as a linear baseline, while GBR and XGB were used as complementary models to reduce dependence on a single algorithmic structure. Based on the revised feature design, the average test R2 values across the retained prediction targets were 0.8229, 0.8211, 0.8188, and 0.7968 for GBR, XGB, ET, and RF, respectively. GBR-based SHAP, PDP, and ALE analyses identified feedstock ash content, carbon content, and sludge mixing ratio as key predictive variables associated with HHV_char. These trends were interpreted as predictive associations rather than direct causal mechanisms. The model suggested a candidate operating window near a 20% sludge mixing ratio. Targeted triplicate experiments at sludge mixing ratios of 10%–30% further examined this window. Under the 1 h condition, the relative error for HHV_char was 4.4%. These results show that interpretable machine learning combined with targeted validation can support data-driven screening of co-HTC conditions for sustainable waste-to-resource conversion.