Predicting methane production in biochar-assisted anaerobic digestion using random forest and SHAP analysis of multi-study experimental data.

Anaerobic digestion (AD) optimization for enhanced biogas production remains challenging because methane production is governed by complex interactions among substrate characteristics, biochar (BC) properties, and operating conditions. In this study, a literature-derived dataset comprising 623 experimental conditions from 107 peer-reviewed studies was compiled to develop an interpretable machine learning framework for predicting methane production in BC-assisted AD systems. Principal component analysis showed that the first two principal components explained 34.5 % of the total variance (PC1: 18.4 %; PC2: 16.1 %), highlighting the intrinsic heterogeneity of multi-study datasets and supporting the application of nonlinear predictive models. A Random Forest model achieved R2 = 0.683, RMSE = 159.75, and MAE = 82.41  mL CH4/gVS under conventional validation, whereas Group K-Fold cross-validation yielded a more conservative estimate of model generalization (R2 = 0.36) by preventing study-level data leakage. Feature importance analysis identified BC specific surface area, pyrolysis temperature, and BC dosage as the strongest structural predictors, while SHAP analysis indicated that substrate composition, particularly food waste, exhibited the greatest positive statistical contribution to methane production predictions, whereas sewage sludge showed a negative contribution. These statistical associations were consistent with previously reported experimental observations but should not be interpreted as direct biological causality. Overall, the proposed framework provides a interpretable approach for integrating heterogeneous literature-derived datasets, quantifying the relative contribution of intrinsic variables, and supporting hypothesis generation and future model development for BC-assisted AD systems.