Machine Learning for X-Ray Spectroscopic Data Analysis

Abstract:

X-ray absorption spectroscopy (XAS) is a powerful diagnostic tool for uncovering the local structural and electronic properties of materials, which is essential for advancements in fields like catalysis and nanotechnology. Despite its utility, the traditional data analysis process is notoriously slow, often relying on manual visual comparisons against established reference spectra.

The thesis explores how machine learning (ML) can automate and accelerate this bottleneck. By utilizing models such as random forests and convolutional neural networks, the research demonstrates how complex patterns within iron K-edge spectra can be decoded instantaneously to predict coordination environments and oxidation states. To bridge the gap between theoretical training and practical application, the study meticulously addresses the challenges of using simulated data to analyze real-world experimental spectra. Significant emphasis is placed on model resilience, quantifying how errors like spectral shifts, shot noise, and self-absorption impact prediction quality. Furthermore, the research employs oversampling techniques (such as SMOTE) to handle class imbalances in mineral data and utilizes SHAP analysis to provide transparency into the decision-making processes. These ML frameworks were successfully used for online analysis at the ESRF beamline, proving their efficacy in identifying chemical phases in complex, disordered materials like ceramic glazes.