A scalable and interpretable machine learning framework for optimizing methane yield and process reliability in psychrophilic anaerobic digestion

Anaerobic digestion (AD) is an established technology for organic waste stabilization and methane recovery; however, process optimization remains challenging due to complex nonlinear interactions among biochemical and operational factors, particularly under psychrophilic (20–25 °C) conditions. In this study, we present a scalable, data-centric machine learning framework to optimize methane yield and process reliability in psychrophilic anaerobic co-digestion systems. A benchmarking analysis was conducted across 11 computational architectures, including artificial neural networks (ANN), Random Forest, XGBoost, and support vector machines, using a 38-month longitudinal dataset from swine carcass–manure co-digestion. To ensure methodological rigor, the framework incorporates a strict zero-leakage feature policy and reactor-independent validation through a Leave-One-Reactor-Out (LORO) strategy based solely on physicochemical state variables. Model robustness was further assessed using 1,000-iteration bootstrap resampling and uncertainty quantification via Monte Carlo dropout and probabilistic modeling. Results show that ANN outperformed all baseline models (p < 0.001), achieving a mean coefficient of determination (R 2 ) of 0.854 (95% CI: 0.752–0.933) and a 70.2% operational success rate on unseen reactors. This corresponds to a 12.2% higher R 2 value compared to the second-best non-evolutionary baseline model (Random Forest, R 2 = 0.7615; 95% CI: 0.611–0.876). SHAP analysis identified hydraulic retention time (HRT) and organic loading rate (OLR) as dominant drivers, while deviations in the FAN–TAN ratio and volatile fatty acids (VFAs) were consistent indicators of process instability. This study introduces a leakage-free, reactor-independent, and uncertainty-aware framework for industrial AD, enabling robust forecasting and confidence-informed decision-making. The proposed approach supports scalable optimization of methane production and enhances process resilience under low-temperature conditions.