Paper2CIF: An AI Agent for Reconstructing Atomistic Models to Discover Structure-Reactivity Relationships
Year: 2026
Type: preprint
Abstract: Reinterpreting published chemical findings requires explicit atomic structures that connect experimental observations with reactivity. Here we report Paper2CIF, an artificial intelligence (AI) agent that reconstructs sample-specific atomic structures from published chemical evidence for structure-reactivity relationship investigations. Across 171 interdisciplinary articles, Paper2CIF achieves 97.08% phase identification accuracy and 91.23% final structural accuracy, with correct structures satisfying all applicable reported constraints, including composition, site occupancy, and local coordination. Calculations on reconstructed BiVO4 structures reveal distortion-dependent spatial contributions to the transition dipole, guiding targeted synthesis toward enhanced near-edge photon conversion. The resulting BiVO4-based Z-scheme pair exhibits a 3.47-fold increase in apparent quantum yield from 4.51% to 15.64% at 475 nm, compared with a 1.03-fold increase from 18.51% to 19.02% at 365 nm. The solar-to-hydrogen efficiency increases from 0.969% to 1.448% for overall water splitting. Paper2CIF establishes atomic reconstruction as a common basis for deriving chemical mechanisms from published evidence and translating them into experimentally testable principles for materials synthesis.
Source: ChemRxiv
Language: English
Subfield: Materials Chemistry
Field: Materials Science
Domain: Physical Sciences
Open Access status: green
Award: G1991018553
OpenAlex ID: W7221196749