Explainable machine learning deciphers and quantifies material-microbe-methane pathways in zero-valent iron-enhanced...

Zero-valent iron (ZVI) has been widely applied to enhance methane recovery in anaerobic digestion (AD), but how ZVI material properties and operating conditions are transmitted through microbial communities to alter methane output remains poorly quantified, limiting rational selection and optimization across substrates. Here, we constructed a cross-study database integrating ZVI material descriptors, operating parameters, bacterial and archaeal community features, and cumulative methane yield (CMY), and developed a pathway-aware machine learning framework to resolve effect transmission and conditional dependence along the material/operation-microbe-methane chain. Among seven algorithms, XGBoost showed the best performance (test R² = 0.916; external validation R² = 0.817). Model interpretation identified methylotrophic and hydrogenotrophic methanogens, Synergistetes, and Proteobacteria as key microbial nodes driving CMY. Further mediation and pathway-decomposition analyses showed that approximately 72% of the total effect of controllable material and operational factors was transmitted through microbial groups, with ZVI specific surface area and initial COD mainly affecting methane production by reshaping bacterial syntrophic modules and methanogenic pathway balance. This transmission mechanism explains the substrate dependence of ZVI optimization: nano-scale, high-SSA materials are more suitable for simple synthetic substrates, whereas sub-millimeter, higher-dose ZVI is more appropriate for complex feedstocks. The resulting two-stage model enables scenario prediction using only controllable inputs when sequencing data are unavailable. This study advances ZVI-enhanced AD optimization from empirical dosing toward mechanism-guided material selection and process regulation, and provides a transferable framework for interpreting material-microbe-function relationships in engineered ecosystems.