Cross-study interpretable machine learning for prediction of NH3 and H2S emissions and identification of operational windows during composting.

Accurate prediction of NH3 and H2S emissions is essential for odor-risk control and process management during composting, but remains challenging because emissions are driven by nonlinear, stage-dependent, and interacting operating conditions. In this study, we developed an interpretable cross-study machine-learning framework for NH3 and H2S prediction using routinely monitored composting variables. A harmonized dataset from 46 published studies provided 542 NH3 and 496 H2S observations after quality control and predictor screening. Five algorithms were benchmarked under study-wise grouped validation to reduce cross-study data leakage and evaluate out-of-study transferability. Tree-based ensemble models outperformed the linear baseline, and the final XGBoost models achieved out-of-study test R2 values of 0.8759 for NH3 and 0.9232 for H2S. SHAP analysis showed that NH3 predictions were mainly associated with temperature, moisture, and nitrogen-related variables, whereas H2S predictions were more strongly linked to EC, pH, and redox-sensitive process conditions. Nonlinear dependence and interaction analyses further identified operational warning windows rather than universal single-factor thresholds, highlighting thermophilic temperature-moisture conditions for NH3 risk and high-moisture/low-aeration conditions for localized H2S risk. External validation with an independent composting run reproduced the major early-stage emission peaks and subsequent declines, although peak magnitudes were moderately underestimated. Overall, the framework provides an uncertainty-aware tool for NH3/H2S-specific risk screening, operational-window identification, and stage-specific composting management, but should not be used as a substitute for exact peak-flux design or comprehensive odor-unit assessment.