Universal and Experiment-calibrated Prediction of XANES through Crystal Graph Neural Network and Transfer Learning Strategy

TL;DR

This work tackles the challenge of fast, accurate XANES interpretation by developing CGXAS, a crystal-graph neural network trained on a large simulated XANES dataset to achieve universal predictions across 48 elements. It demonstrates that a single universal model outperforms element-specific variants and reveals interpretable latent representations via t-SNE, linking absorption element and local environment to the prediction. To bridge the gap to experiments, the authors apply transfer learning using a small experimental XANES set, achieving about a 55% reduction in edge-energy misalignment and sizable drops in prediction loss, illustrating practical feasibility for experiment-calibrated XANES prediction. The approach promises fast, universal, and experimentally calibrated XANES predictions suitable for high-throughput and in-situ analysis, with the potential to leverage more high-fidelity data to further improve accuracy in challenging spectral regions.

Abstract

Theoretical simulation is helpful for accurate interpretation of experimental X-ray absorption near-edge structure (XANES) spectra that contain rich atomic and electronic structure information of materials. However, current simulation methods are usually too complex to give the needed accuracy and timeliness when a large amount of data need to be analyzed, such as for in-situ characterization of battery materials. To address these problems, artificial intelligence (AI) models have been developed for XANES prediction. However, instead of using experimental XANES data, the existing models are trained using simulated data, resulting in significant discrepancies between the predicted and experimental spectra. Also, the universality across different elements has not been well studied for such models. In this work, we firstly establish a crystal graph neural network, pre-trained on simulated XANES data covering 48 elements, to achieve universal XANES prediction with a low average relative square error of 0.020223; and then utilize transfer learning to calibrate the model using a small experimental XANES dataset. After calibration, the edge energy misalignment error of the predicted S, Ti and Fe K edge XANES is significantly reduced by about 55%. The method demonstrated in this work opens up a new way to achieve fast, universal, and experiment-calibrated XANES prediction.

Figures (6)

• Figure 1: Framework of CGXAS model and flowchart of training.
• Figure 2: Performance of CGXAS_Uni on the prediction of XANES. (a) Evolution of training loss and validation loss at each epoch in the training process, where the loss is measured by relative square error (RSE). (b) Distribution of prediction error in the testing set. (c)-(f) Examples of the predicted XANES with (c) low, (d) median, (e) high, and (f) ultra-high prediction error compared with simulated XANES.
• Figure 3: Universality and interpretability analysis of CGXAS. (a) Comparison of XANES prediction accuracy (measured by Pearson correlation coefficient) of CGXAS_Uni, CGXAS_Spec_X, and CGXAS_FT_X in different elements. (b) Distribution of latent site vectors of different absorption sites, where the site vectors are reduced into 2 t-SNE features and colored according to their absorption elements. (c) Distribution of O latent site vectors obtained from Li, Mg, Al, Ti, Fe, and Cu metal oxides, where the site vectors are reduced into 2 t-SNE features and colored according to their neighboring elements. The latent site vectors in (b) and (c) are derived from CGXAS_Uni model.
• Figure 4: Distribution of the samples from the simulated dataset and experimental dataset: (a) S K edge XANES, (b) Ti K edge XANES, (c) Fe K edge XANES. Here, the samples are represented by their spectra predicted by CGXAS_Uni and reduced into 2 features by t-SNE. The blue dot represents the samples from the simulated dataset, and the red star represents the samples from the experimental dataset.
• Figure 5: Performance of the CGXAS model on the prediction of experimental XANES after transfer learning. (a) Prediction loss of S, Ti, and Fe K edge XANES (measured by RSE) before and after transfer learning in training set, validation set, and testing set. (b) Error of edge energy compared with experimental XANES before and after transfer learning in training set, validation set, and testing set. Here, the column height and error bar represent the average value and standard deviation of the prediction losses or energy errors calculated over 20 models trained with different data separations respectively. (a) and (b) share the same legend. (c)-(d) Examples of S K edge XANES predicted by CGXAS_Exp_S for the samples in testing sets with different RSE values, together with the spectrum predicted by CGXAS_Uni and the ground truth experimental XANES. (e)-(f) Examples of Ti K edge XANES predicted by CGXAS_Exp_Ti for the samples in testing sets. (g)-(h) Examples of Fe K edge XANES predicted by CGXAS_Exp_Fe for the samples in testing sets.