X-ray absorption spectroscopy (XAS) reveals local structures in catalysts and materials under operando conditions but faces ill-posed challenges, with high computational costs and ambiguous solutions. DeepFit pioneers a universal deep learning remedy, deploying a differentiable minimization of both a spectral mismatch and quantum-chemical energy. The model is based on an E(3)-equivariant neural network trained on the original database of 67,000 theoretical K-edge spectra from diverse 3d/4d transition metal complexes. The approach ensures physically equivariant and chemically plausible refinement of 3D structure and is validated on the experimental case studies for homogeneous catalysts. The benchmark for the series of 3d compounds shows agreement of DeepFit refined bond lengths with respect to XRD/EXAFS (MAE 0.06 Å) and 0.73 ROC-AUC selectivity over local motifs. By merging machine learning spectra approximation and quantum chemical constraints, DeepFit transforms quantitative XAS analysis toward a black-box, routine tool.
Kulaev et al. (Thu,) studied this question.