Physics Informed Disentangled Latent Space of X-Ray Absorption Spectra

X-ray absorption spectra provide insights into the atomic scale local structures of a wide range of materials with nanoscale dimensions. This information is often inaccessible to experimentalists through traditional structural analysis techniques. To analyze XAS spectra, theory-driven structural models are needed, particularly for fitting the extended part of the spectra (EXAFS). While the quantitative analysis of the near-edge part of XAS (XANES) remains an active research area, neural networks have proven their efficacy as universal approximators. They efficiently learn the mapping between theory generated spectra and structural descriptors in various cases of mono and bimetallic metallic nanocatalysts. However, challenges arise in uncertainty quantification and out-of-distribution performance when these networks are applied to experimental data. Recent studies have highlighted that low-dimensional embeddings of XANES, obtained via autoencoder frameworks, encapsulate these structural descriptors. To enhance the interpretability and utilize the generative capability of the autoencoder framework, a structured and continuous latent space is essential.<br/><br/>In our work, we introduce a physics-informed robust analysis of XANES spectra, striking a balance between reconstruction and descriptor-specific information. Our method, the Multitasking Algorithm for Variational Auto Encoder (MAVEN), achieves a disentangled, interpretable latent space through multi-objective optimization, focusing on reconstruction, denoising, autocompletion, and descriptor mapping. MAVEN stands out for its ability to disentangle and interpret latent space variables, offering a unique generative capability within the autoencoder framework. We’ve also developed a methodology to quantify the level of disentanglement achieved, rooted in information theory principles. We show the utility of MAVEN by interpreting experimental XANES of palladium and palladium hydride nanoparticles. These are renowned for their lattice expansion, which MAVEN distinguishes from particle size effects via coordination number analysis. The comparison of trends from the MAVEN algorithm to those obtained from extended X-ray absorption fine structure (EXAFS) spectral analysis using Demeter software indicates MAVEN’s potential for high-fidelity real-time analysis of local structures. Our discussions will explore the applications of this technique in bimetallic nanoparticles systems and assess its effectiveness on experimental XANES spectra. The promising results suggest that MAVEN could be a vital tool for real-time analysis of local structures, especially in conjunction with experimental spectroscopy of functional nanomaterials like catalysts, batteries, and fuel cells.