Spectroscopy-Informed XANES–PXRD Framework for Multi-Property Prediction and Structure Inference

Abstract

Accelerating functional material characterization and rational design faces a fundamental circular bottleneck in which all mainstream computational screening methods rely entirely on a priori crystallographic knowledge, information inherently unavailable for novel, uncharacterized samples. To bypass this bottleneck, we utilize spectroscopy as a predictive driver. While X-ray absorption near-edge structure (XANES) captures element-specific local states, powder X-ray diffraction (PXRD) resolves global long-range order. Here, we integrate XANES and PXRD into a unified spectroscopic representation to address the challenge of structure inference from spectral data. Trained on 34,929 inorganic compounds with over 100,000 simulated spectra, our framework jointly predicts key physical properties (e.g., band gap, magnetism, density) while inferring oxidation states, coordination numbers, and crystal systems. Notably, a composition-aware partial-measurement strategy utilizing only transition-metal edges matches the accuracy of all-element models, significantly reducing experimental burden. Interpretability analyses reveal that transition-metal and non-transition-metal features cooperatively encode the correlations linking local electronic motifs to global symmetry. Crucially, the framework goes beyond property mapping to reconstruct charge-balanced formulas and retrieve structural templates from existing databases, enabling structure mining without prior structural knowledge. This approach achieves a top-1 accuracy exceeding 0.80 across binary to quinary systems, validated experimentally on eight representative samples including single-phase compounds and heterogeneous composites. These results establish spectroscopy as a quantitative, interpretable medium for decoding structure-property relationships, offering a practical pathway for spectroscopy‑informed materials characterization and structure mining.