Integrated Machine-Learning Framework for Balanced Performance-Safety Design of Iron-Based Remediation Materials

Optimizing iron-based materials via surface coatings and lattice engineering has been a major research focus for environmental remediation, yet higher reactivity to contaminants often elevates risks to organisms. To address this trade-off, we develop an integrated machine-learning framework for the quantitative coassessment of performance and safety. The model was trained on a literature-derived data set of 1007 cases, encompassing 80 iron-based materials, 136 target pollutants, and 50 test organisms. Using AutoGluon, a binary QSAR model was constructed to screen reaction performance and predict the direction of material-pollutant joint toxicity. Inputting molecular descriptors of materials and pollutants alongside experimental conditions, this model achieved robust remediation performance and joint toxicity predictions with balanced accuracy of 0.88 and 0.90, respectively. Feature attribution analysis predicted that material heterogeneity was a key driver of performance, while organismal attributes primarily governed toxicity outcomes, highlighting niche-targeting material doping and hybrid nanobio remediation strategies. A dominance-aware dynamic weighting scheme that integrates calibrated probabilities with an upper-zone attenuation and low-zone penalty was introduced to rank material candidates, plus an applicability-domain filter to limit overconfident extrapolation. These findings support a practical prescreening framework that generates prioritized candidate materials to inform the rational application of iron-based remediation systems.