Introduction: Anaerobic digestion (AD) is a widely employed biological process for converting organic waste into biogas, primarily methane (CH4) and carbon dioxide (CO2). Despite its role in renewable energy recovery and wastewater management, accurate prediction of biogas yields across diverse feedstocks remains a challenge. Mechanistic models such as the Anaerobic Digestion Model No. 1 (ADM1) provide simulation frameworks but are constrained by extensive parameterization requirements for each feedstock. Data-driven machine learning (ML) methods offer complementary speed and pattern recognition capability.
Methods: A hybrid simulation ML framework was developed by integrating ADM1 as a structured data generator with four supervised ML algorithms: Random Forest (RF), LightGBM, Support Vector Regression (SVR), and a Stacking ensemble. ADM1 was implemented for six feedstocks: Biowaste, Feed Residues, Maize and Cattle Manure Co-digestion, Pig Manure, Vinasses, and Wastewater Treatment Sludge under dynamic mesophilic conditions over 100-day simulations. ADM1 state variables were used as ML inputs, with daily CH4 and CO2 production rates as target outputs.
Results: SVR showed the strongest performance, especially for Biowaste, while LightGBM provided relatively stable performance across most feedstocks; Stacking was highly competitive for Food Residue and Maize-Cattle Manure Co-digestion, whereas RF showed good performance in some cases but was less consistent overall. Among the feedstocks, Food Residue was the easiest to predict because all models achieved high R2 values for both CH4 and CO2, while Biowaste also showed strong prediction accuracy. In contrast, Maize and Cattle Manure Co-digestion was the most difficult prediction , showing the lowest R2 values, particularly for CH4 prediction.
Conclusions: This study demonstrates that simulation-informed ML, using ADM1 as a mechanistic data generator, constitutes an approach for predicting biogas production and composition across diverse organic substrates without reliance on extensive experimental measurements. The proposed framework offers significant potential for process optimization and operational decision-support in multi-feedstock anaerobic digestion facilities.