Mechanism-informed machine learning for optimizing herbal residue-assisted anaerobic co-digestion using...

Agricultural-residue co-digestion is constrained by slow lignocellulose hydrolysis and incomplete conversion of volatile fatty acids. This study used medicinal herb residue as an auxiliary co-substrate and polydopamine-modified iron-nickel foam (PDA-FN) as a recoverable conductive additive to enhance methane production from straw-manure co-digestion. We further developed a process optimization framework coupling mechanism-informed Gaussian process regression (MI-GPR) with Bayesian optimization. MI-GPR incorporated quadratic, interaction, and mechanism-informed contact and mass-transfer features and was trained using 47 of 59 experimental conditions. On the 12-condition test set, it achieved an R² of 0.9477, outperforming conventional GPR, XGBoost, and Random Forest. SHAP analysis identified the straw-manure/medicinal herb residue ratio as the dominant predictor, followed by PDA-FN dosage, PDA concentration, and Fe/Ni ratio. Experimental validation at the model-recommended condition produced 382.6 ± 11.0 mL/g VS, within 0.6 % of the predicted value and 31.8 % and 15.1 % above the Control and herb-residue-only treatments, respectively. Mechanistic analyses associated PDA-FN with community-wide restructuring of hydrolytic-acidogenic bacteria, syntrophic partners, and methanogenic archaea, accompanied by faster VFA turnover, elevated c-type cytochrome content, Methanosarcina enrichment, and higher PICRUSt2-predicted potential for electron transfer, DNRA, and ammonium assimilation. Overall, herb-residue-assisted PDA-FN jointly improved methane production from straw-manure co-digestion through substrate-nutrient regulation, enhanced electron transfer, and nitrogen-cycle modulation. MI-GPR provides an interpretable tool for mechanistically reasonable optimization of small-sample, strongly coupled, multifactor anaerobic fermentation systems.