Attribution Fingerprinting by Interpretable Machine Learning Reveals Category-Specific Additive Design Rules for Zinc Anodes

Electrolyte additives are one of the most direct remedy for interfacial instability in aqueous zinc-ion batteries, yet their development remains empirical. Existing machine-learning (ML) screening strategies presume a single global structure-property relationship shared by all additives. Here, an interpretable ML framework predicts cycle life from molecular descriptors and test conditions, and molecule-specific SHapley Additive exPlanations (SHAP) profiles serve as attribution fingerprints. Unsupervised clustering resolves the additives into three attribution-based categories, governed respectively by adsorption energetics, charge-transfer gap and hydrogen-bond-network regulation, indicating that one universal descriptor set may not hold across chemically distinct additives. Representative additives from each category, none present in the training set, agree with experiment to within 11%. Category I screening identifies 2-fluoropyridine-5-carbaldehyde (2FPD), predicted at 3218 h, and Zn//Zn cells deliver 3010 h at 1 mA cm^-2 and 1 mAh cm^-2, within 6.5% of prediction. Structural controls trace the improvement to the aldehyde: retaining it alone gives 900 h against an 89 h baseline, whereas fluorine alone gives 90 h. 2FPD constructs a graded organic-inorganic interphase that limits water access to the zinc surface and accelerates Zn²⁺ desolvation, and Zn//MnO₂ full cells retain 96% capacity over 4000 cycles. This workflow offers a route towards interpretability-guided additive design.