Interpretable machine learning reveals dominant controls and operational windows for humic-associated phosphorus formation during aerobic composting

Humic-associated phosphorus (NaOH-P) is an important stabilized phosphorus fraction formed during aerobic composting, yet its dominant controls, threshold responses, and interaction-defined regulation windows remain poorly quantified across heterogeneous composting systems. In this study, an interpretable machine-learning framework was developed to predict NaOH-P formation and identify the key factors governing its accumulation. A literature-derived dataset was analyzed using nine machine-learning algorithms before and after preprocessing. After preprocessing, CatBoost showed the best balance between predictive accuracy and train-test stability, with a test-set R² of 0.913, RMSE of 0.179, and MAE of 0.129. SHAP analysis revealed a highly concentrated attribution structure, with TP, N/P, MC, and pH showing the largest contributions to model prediction. TP showed predominantly positive associations with predicted NaOH-P, N/P showed predominantly negative associations, and MC and pH exhibited nonlinear response patterns. Partial dependence analysis further showed that NaOH-P regulation was threshold-controlled and interaction-dependent rather than driven by independent single-factor effects. Favorable NaOH-P formation was generally associated with TP > 20 g·kg^-1, N/P below approximately 1.0, MC > 40 %, and pH 6.0-7.5, particularly after the initial 15-20 d of composting. Laboratory-scale validation showed close agreement between predicted and measured NaOH-P values under the tested conditions, yielding an R² of 0.968, RMSE of 0.037, and MAE of 0.028. Overall, this study shows that interpretable machine learning can convert heterogeneous composting data into quantitative regulation insights for targeted phosphorus stabilization.