Decoding Perovskite Interfaces: Machine Learning Uncovers Geometry-Matched Dual-Site Hydrogen Bonding.

Molecular passivation is central to efficient and stable perovskite photovoltaics, yet interfacial molecules are still largely selected through chemical intuition and iterative experimentation. Here, an interpretable machine-learning framework is developed that extracts chemically meaningful structure-performance relationships from literature data and uses them to guide molecular screening. Using approximately 100 data points, the model quantitatively relates six molecular and material features to photovoltaic performance, while SHapley Additive exPlanations (SHAP) analysis ranks their relative contributions and identifies hydrogen-bonding capacity and adsorption strength as the two most influential factors. Guided by these data-derived relationships, nine previously unexplored molecules were screened, with 2-amino-thiazole-4-carboxamide (ATC) and 1-(2-aminoethyl)-1H-pyrrole-2,5-dione (MPD) emerging as leading candidates. Theoretical and experimental investigations show that MPD exhibits stronger Pb2 + coordination, whereas ATC forms stronger geometry-matched multisite N─H···I interactions, increasing the calculated iodide-migration barrier from 0.39 to 0.67 eV and suppressing interfacial iodide redistribution. ATC-modified cells achieve a PCE of 27.13% and a certified stabilized efficiency of 26.59%, while retaining 90% of their initial efficiency after 1460 h of unencapsulated operation. The strategy further delivers efficiencies of 24.01% and 21.14% in 20.98 cm2 rigid and 21.0 cm2 flexible modules, respectively, providing a data-to-mechanism route for rational and scalable perovskite interface engineering.