Hierarchical Bayesian Data Analysis for Challenging Structural Materials Characterization Problems

Automated quantitative analysis for structural materials characterization is a challenging task primarily because of the difficulty of physical model specification and optimization. However, potential development of robust online automated quantitative analysis would significantly expand the scope and impact of autonomous platforms for materials development.<br/> <br/>Our approach of hierarchical Bayesian analysis of high throughput structural characterization data blurs the lines between machine learning based approaches and conventional nonlinear least squares analysis. This approach allows models to pool information across related samples, and also allows integration of flexible non-parametric models such as Gaussian Process priors to model functional dependencies between parameters.<br/> <br/>We demonstrate how this modeling strategy can provide higher sensitivity for analysis of minority or trace phases in challenging high throughput X-ray diffraction (XRD) data. We will also show how to apply Bayesian modeling workflow to extended X-ray absorption fine structure (EXAFS), a spectroscopic technique that provides insight into local structural and chemistry.<br/> <br/>Finally, we will discuss the current limitations of the hierarchical Bayesian modeling approach, and how it may be complemented by other contemporary machine learning based methods.