Universal rapid machine learning models for predicting unconvoluted and convoluted X-ray Absorption Spectra

TL;DR

The paper addresses the need for fast, quantitative XANES predictions from 3D structures across multiple elements and edges. It introduces XAS3D, a universal absorber-centered 3D GNN that processes a local atomic cluster with species embeddings and geometry-derived edge weights, producing either unconvoluted or convoluted XANES spectra and enabling multi-element predictions with a single model. Using a dataset built from CCDC structures and XANES spectra simulated by FDMNES within a 5 Å cluster, the authors demonstrate state-of-the-art accuracy across Ni, S, and Ru K-edges and show robustness to hyperparameter choices. They also present a practical XANES-fitting workflow that refines 3D structures against experimental spectra, highlighting potential for online beamline data analysis and real-time structure validation in materials science and physics. Overall, the work provides a scalable, robust framework for rapid XANES analysis and 3D-structure–spectra inference that can adapt to different scattering methodologies.

Abstract

X-ray absorption near edge structure (XANES) is an essential tool for elucidating the atomic-scale, local three-dimensional (3D) structure of given materials and molecules. The rapid computation of XANES based on molecular 3D structures constitutes a vital element of quantitative XANES analysis. Here, we present an XANES prediction model. It takes 3D structures as input and generates either unconvoluted XANES or convoluted spectra as output, demonstrating excellent generalizability across diverse instrumental broadening. This model has validated its predictive capability for both hard X-ray XAS (exemplified by K-edges of 3d 4d metals and lanthanides) and soft X-ray XAS (using S K-edge as examples). Adopting the model, XANES spectra of multiple elements can be predicted using a single unified model. A highly efficient 3D structure fitting algorithm based on this unconvoluted XANES prediction model, aiming to serve as an online data analysis method suitable for XAS beamlines.

Figures (4)

• Figure 1: (A):Boxplots of performance metric (MAE) of XAS3D, GraphNet and SGN GNN models using different hyperparameters in Ni K-edge XANES prediction. (B):Prediction results generated by the GraphNet model for Ni K-edge XANES prediction at different quantile levels from q10 through q90. (C):Prediction results generated by the SGN model for Ni K-edge XANES prediction at different quantile levels from q10 through q90.(D):Prediction results generated by the XAS3D model for S K-edge XANES prediction at different quantile levels from q10 through q90. (E):Prediction results generated by the XAS3D model for Ni K-edge XANES prediction at different quantile levels from q10 through q90. (F):Prediction results generated by the XAS3D model for Ru K-edge XANES prediction at different quantile levels from q10 through q90. In the above subplots, XANES simulated with multiple scattering framework(solid lines) and with machine learning model(dashed lines) on validation dataset.
• Figure 2: (A):MAE of convoluted XANES predictions for 3d transition metals versus MAE calculated after applying convolution to results of unconvoluted XANES prediction models. (B):Prediction results generated by the XAS3D model for Sc K-edge XANES prediction at different quantile levels from q10 through q90. (C):Prediction results generated by the XAS3D model for Fe K-edge XANES prediction at different quantile levels from q10 through q90. (D):Prediction results generated by the XAS3D model for Zn K-edge XANES prediction at different quantile levels from q10 through q90. (E):The XAS3D model first predicted unconvoluted Sc K-edge XANES, then applies convolution to unconvoluted spectrum. MAE were calculated between predicted and theoretical spectra both after convolution. The performance across q10 through q90 quantiles was presented. (F):The XAS3D model first predicted unconvoluted Fe K-edge XANES, then applies convolution to unconvoluted spectrum. MAE were calculated between predicted and theoretical spectra both after convolution. The performance across q10 through q90 quantiles was presented. (G):The XAS3D model first predicted unconvoluted Zn K-edge XANES, then applies convolution to unconvoluted spectrum. MAE were calculated between predicted and theoretical spectra both after convolution. The performance across q10 through q90 quantiles was presented. In the above subplots, XANES simulated with multiple scattering framework(solid lines) and with machine learning model(dashed lines) on validation dataset.
• Figure 3: The flowchart of XANES fit combined unconvoluted XANES prediction model and optimization algorithm.
• Figure 4: (A)Comparison of the experimental spectra, the model predicted spectra and multiple scattering calculated one based on the fitted 3D structure of $Fe_2O_3$ material.(B)Comparison between the model-predicted unconvoluted spectra and the unconvoluted spectra obtained from multiple scattering calculations.