Robust Machine Learning Inference from X-Ray Absorption Near Edge Spectra through Featurization

Machine learning (ML), used in conjunction with materials modeling, is reshaping the way that researchers analyze and interpret materials characterization data by greatly accelerating the process and providing underlying physics [https://link.springer.com/article/10.1557/s43577-022-00446-8]. One notable application is the extraction of essential materials properties, such as oxidation states and structural information, from X-ray absorption spectroscopy (XAS) data. Traditional ML models typically utilize raw spectral intensities as the model input, with limited exploration on transforming the spectra intensities to potentially boost model performance. In this presentation, we will compare and assess the effectiveness of both reduced-dimensional features and overcomplete representations to discover the optimal representation of x-ray absorption near edge structure (XANES) data. Our system of interest is LiNi_xMn_yCo_zO₂(NMC), a typical cathode material for Li-ion batteries. This material presents challenges in studying detailed changes during electrochemical cycling due to the complexity arising from transition metal mixing. We will evaluate various input transformations for XAS through regression and classification tasks to demonstrating how such feature engineering improves prediction accuracy and interpretability of ML models. Furthermore, we will discuss model validation using unseen experimental datasets, illustrating the transferability and robustness of the feature. A thorough explanation will also be provided to elucidate why certain features outperform others, aiding in the data analysis of experimental spectra [arXiv preprint arXiv:2310.07049].<br/><br/><br/>Acknowledgement<br/>This work is supported by the U.S. Department of Energy (DOE) Office of Science Scientific User Facilities project titled “Integrated Platform for Multimodal Data Capture, Exploration and Discovery Driven by AI Tools”. M.K.Y.C. acknowledges the support from the BES SUFD Early Career award. Work performed at the Center for Nanoscale Materials and Advanced Photon Source, U.S. Department of Energy Office of Science User Facilities, was supported by the U.S. DOE, Office of Basic Energy Sciences, under Contract No. DE-AC02-06CH11357. We also acknowledge the support provided the Data Infrastructure Building Blocks (DIBBS) Local Spectroscopy Data Infrastructure (LSDI) project funded by National Science Foundation (NSF), under Award Number 1640899. MJD was supported by the U. S. Department of Energy, Office of Basic Energy Sciences, Division of Chemical Sciences, Geosciences, and Biosciences operating under Contract Number DE-AC02-06CH11357.