Experimental analysis and machine learning prediction of biomethane potential from organic food waste: lab-scale and semi pilot-scale approaches

Machine learning (ML) offers powerful tools for predicting renewable energy yields from complex biological processes. In this study, ML was explicitly applied to model and predict methane yield (CH₄) in lab-scale experiments and total biogas production in semi-pilot anaerobic digestion (AD) systems treating food waste (FW) and wastewater (WW). Four algorithms Stepwise Linear Regression (SLR), Quadratic Support Vector Machine (QSVM), Boosted Trees (BT), and Matern 5/2 Gaussian Process Regression (M-GPR) were trained using operational and physicochemical input variables including inoculum-to-FW (In:FW) ratio, experimental run, digestion days, temperature, pH, TS, VS, C/N ratio, and COD. Under lab-scale conditions, the optimal 1:2.5 In:FW ratio achieved the highest Biochemical Methane Potential (BMP) of 82.96 ml/g VS added. In semi-pilot scale experiments, adjusting the solids content to 8%, 10%, and 12%, the best results were obtained at 10% TS, yielding the highest gas volume of 3174.6 ml and a BMP of 30.56 ml/g VS. Among the ML models, M-GPR consistently achieved the highest predictive accuracy (R2 up to 0.999 for lab-scale CH₄ and 0.996 for semi-pilot biogas), with the lowest RMSE (7.70–9.69) and MAE (4.69–6.99). SLR also performed strongly, QSVM achieved moderate accuracy, while BT underperformed. These findings demonstrate that ML can reliably predict methane and biogas production across scales, supporting optimised and sustainable biomethane generation.