Beyond Linearity: Transfer Learning Reveals Optimal Biochar Surface Area Window and Suggests Potential Syntrophic Mechanisms in Food Waste Co-Digestion.

UNLABELLED: Anaerobic co-digestion of food waste and paper mill wastewater offers a sustainable waste-to-energy solution, but performance instability limits its efficiency. While biochar (BC) is a promising additive, the specific surface area (SSA) required for optimal performance remains poorly contentious. This study developed a BC-physicochemical physics-informed residual transfer learning model (BP-PIRTL) framework. By integrating 1,971 literature samples and 264 experimental datasets, the model achieved high-precision cross-domain prediction (R2 = 0.99). Guided by model predictions, corncob-derived BCs with a wide SSA gradient (26 ∼ 2040 m2/g) were tested experimentally. A distinct nonlinear relationship emerged: moderate SSA (500 ∼ 800 m2/g) enhanced methane yield by up to 55.18%. Conversely, excessively high SSA (>1000 m2/g) exhibited inhibitory effects. SHAP analysis decoupled mechanistic contributions, suggesting that lag phase reduction may be associated with enhanced microbial attachment, while maximum production rate improvement potentially relates to direct interspecies electron transfer (DIET) pathway activation. Microbial profiling unveiled the underlying cause: optimal SSA BC selectively enriched syntrophic consortia (Clostridium, Syntrophomonadaceae, Methanobacterium). In contrast, excessive SSA triggered metabolic dysbiosis, likely through competitive substrate adsorption. These results suggest SSA is as a potential key factor influencing microbial syntrophy, providing hypothesis-driven guidelines for rational BC design in waste-to-energy processes. RESEARCH BACKGROUND: Biochar (BC) amendment stabilizes anaerobic co-digestion (AcoD). However, contradictory literature regarding specific surface area (SSA) effects reveals critical knowledge gaps. Unified frameworks to predict optimal SSA ranges and elucidate microbial mechanisms remain lacking. ANALYTICAL METHODS: A BC-physicochemical physics-informed residual transfer learning (BP-PIRTL) framework integrated 1,971 literature samples and 264 experimental datasets. XGBoost pre-training followed by gradient boosting regressor residual correction achieved cross-domain prediction. Model-guided experiments tested corncob-derived BCs spanning 26 ∼ 2040 m2/g SSA in mesophilic batch AcoD (F/S = 3:1, 35°C). SHAP analysis decoupled SSA contributions to kinetic parameters, while 16S rRNA sequencing elucidated microbial mechanisms. PRINCIPAL CONCLUSIONS: BP-PIRTL achieved high-precision prediction (R2 = 0.99). Moderate SSA (500 ∼ 800 m2/g) enhanced methane yield by up to 55.18%, whereas excessive SSA (>1000 m2/g) reduced yield by 29.4%. SHAP revealed that lag phase reduction may be associated with microbial attachment, and rate improvements potentially relate to DIET activation. Optimal SSA enriched syntrophic consortia, whereas excessive SSA triggered dysbiosis. NEW ASPECTS: This study pioneers transfer learning in BC-enhanced AD. By introducing baseline methane yield as an anchor feature, the model quantitatively decouples SSA contributions. The optimal SSA window (500 ∼ 800 m2/g) provides hypothesis-driven design guidelines that warrant further mechanistic investigation.