Towards a new paradigm for machine learning-driven analysis and surrogate modeling for X-ray absorption spectroscopy*

Spectroscopy is a foundational experimental characterization technique based on the light-matter interaction. Typically, experimental spectra are an abstract representation of the atomistic or the electronic structure of the sample, and spectral analysis often relies on computational simulation to uncover the underlying structure-spectrum relationship. However, the high computational cost of simulations can become a bottleneck during analysis, especially for complex systems with many electrons in the unit cell. Machine learning (ML) techniques, which have been deployed with unnerving effectiveness in the technology sector, may offer a solution to accelerating qualitative and even quantitative spectral analysis with, e.g., surrogate modeling. ML methods are especially good at interpolating in and exploring high-dimensional spaces, making them an excellent fit for a broad range of spectral analysis tasks. In this talk, we focus on X-ray absorption spectroscopy (XAS), which is a premier tool for probing the local chemical environments of absorbing sites. We explore multiple instances in which ML has been applied to extract physical descriptors of materials and molecules from their XAS and predict material and molecular XAS, avoiding expensive simulations. We highlight that ML is a powerful tool for XAS interpretation, especially for high-throughput studies and real-time analysis.