Anaerobic co-digestion of wastewater sludge and food waste: A machine learning approach to process modeling and optimization

Due to the complex microbial interactions in anaerobic co-digestion (An-CoD), predicting process outcomes remains challenging. This study applies machine learning (ML) to model and optimize the co-digestion of wastewater sludge and food waste (FW). A generalized, transferable framework is developed that reflects real-world operational constraints by isolating the mixing ratio as the primary input variable under fixed digester conditions, an approach aligned with full-scale facility limitations. Unlike prior studies limited to data from single facilities, this model is trained on a diverse dataset compiled from multiple literature sources, enhancing its applicability across varying sludge and FW characteristics. Three ML models, random forest (RF), XGBoost, and artificial neural networks (ANN), were trained to predict methane yield and assess the influence of feedstock and operational parameters. Regularized ANN outperformed other models, reducing overfitting, and achieved an R² of 0.86 and NRMSE of 0.31. The trained model was then coupled with a global optimization algorithm (dual annealing) to identify mixing ratios that maximize methane yield based on feedstock properties. Optimization across seven candidate points, spanning a range of solid contents, revealed that methane yield scales with feedstock quality and process intensity. High-VS sludge and FW mixtures under mesophilic continuous operation yielded up to 510 mL/g VS, while inorganic-rich, low-VS sludge limited yields to ∼130 mL/g VS. Medium-quality mixtures reproduced experimental trends, exceeding 300 mL/g VS. Predicted yields ranged from 134 to 510 mL/g VS, with optimal mixing ratios varying from 0.99 % to 88 % VS across scenarios.