Demonstration of an AI-driven workflow for dynamic x-ray spectroscopy

TL;DR

The paper tackles the bottleneck of slow XANES data acquisition by introducing a knowledge-infused Bayesian optimization framework that actively samples energies based on domain knowledge of XANES spectral structure. By integrating a gradient-and-residue–aware acquisition function and edge-aware reweighting into a Gaussian process Bayesian-optimization backbone, the method achieves accurate spectrum reconstruction with only 15–20% of the points used in conventional scans, and precise edge and white-line energy estimates within sub-eV to tens-of-meV accuracy. Its effectiveness is demonstrated across simulated single-spectrum and dynamic XANES scenarios as well as a live beamline test on battery and catalyst materials, including an autonomous NMC111 measurement campaign that significantly speeds up data collection. The approach reduces under- or over-sampling near absorption edges, improves time resolution for dynamic experiments, and enables more automated, autonomous XANES workflows with explicit uncertainty quantification. This work thus advances efficient, adaptive, and automated XANES experiments at modern synchrotron facilities.

Abstract

X-ray absorption near edge structure (XANES) spectroscopy is a powerful technique for characterizing the chemical state and symmetry of individual elements within materials, but requires collecting data at many energy points which can be time-consuming. While adaptive sampling methods exist for efficiently collecting spectroscopic data, they often lack domain-specific knowledge about XANES spectra structure. Here we demonstrate a knowledge-injected Bayesian optimization approach for adaptive XANES data collection that incorporates understanding of spectral features like absorption edges and pre-edge peaks. We show this method accurately reconstructs the absorption edge of XANES spectra using only 15-20% of the measurement points typically needed for conventional sampling, while maintaining the ability to determine the x-ray energy of the sharp peak after absorption edge with errors less than 0.03 eV, the absorption edge with errors less than 0.1 eV; and overall root-mean-square errors less than 0.005 compared to compared to traditionally sampled spectra. Our experiments on battery materials and catalysts demonstrate the method's effectiveness for both static and dynamic XANES measurements, improving data collection efficiency and enabling better time resolution for tracking chemical changes. This approach advances the degree of automation in XANES experiments reducing the common errors of under- or over-sampling points in near the absorption edge and enabling dynamic experiments that require high temporal resolution or limited measurement time.

Figures (7)

• Figure 1: Diagram showing the workflow of our adaptive sampling method, illustrated using the real-world battery cell experiment demonstrated in this paper. The backend algorithm receives data from the x-ray detector and sends data to the monochromator through the instrument API powered by Bluesky allan_synchrotron_rad_news_2019. With each measured energy, the algorithm updates its internal state, computes the comprehensive acquisition function and directs the monochromator to measure at a new energy. This cycle is repeated until the stopping conditions are met.
• Figure 2: Intermediate reconstructed spectra, posterior standard deviation, measured data points, and true spectrum for the YBCO data. Data are plotted without normalization. The posterior standard deviation at each point is represented by half of the vertical length of the shaded area.
• Figure 3: Results of single-spectrum sampling of the YBCO sample. Legend at the bottom right corner applies to both sub-figures. (a) Convergence of RMS error with our comprehensive acquisition function and acquisition reweighting, and its comparison with cases (i) with comprehensive acquisition but without reweighting, (ii) with posterior uncertainty-only acquisition function, and (iii) with uniform sampling. The inset bar chart on the side shows the areas under the curve before the 50th measurement for each case with the same color coding to the RMS error curves. The heights of the bars and the numbers indicate the averages of the AUCs over 5 repeated runs with different initial points, and the error bars represent their standard deviations. (b) Reconstructed spectra using these methods after 30 points are sampled (initial measurements included). The inset shows the magnified spectra within the energy range between 9007 and 9017 eV. This offers a direct comparison between the comprehensive acquisition method with and without reweighting, with the latter exhibiting larger errors in the plotted region.
• Figure 4: Results of applying our adaptive sampling method in a dynamic XANES experiment that tracks the phase transition progress of the LTO sample. (a) The stack of spectra at 50$^\circ$C sampled using our method. Data plotted and normalized. (b) The RMS error of normalized spectra plotted against spectrum index. (c) The phase transition percentages calculated through linear fitting with data collected. Percentages were calculated using normalized data. (b) and (c) are aligned with (a) in the horizontal axis. The gray dashed lines marks the beginning, end of the experiment and the onset of the phase transition.
• Figure 5: Dynamic experiment results of the Pt sample. (a) All normalized spectra during the reduction process, sampled using our method. (b) The transition percentages calculated using normalized data sampled with our method. (c) The energies of the maxima of the white lines (bold lines), as well as those of the first derivative of the absorption edge (thin lines), plotted for all spectra over the reduction process.