| About this Abstract |
| Meeting |
MS&T24: Materials Science & Technology
|
| Symposium
|
Materials Informatics for Images and Multi-dimensional Datasets
|
| Presentation Title |
Hierarchical Bayesian Models for Automating Structural Materials Characterization |
| Author(s) |
Brian DeCost, Howie Joress |
| On-Site Speaker (Planned) |
Brian DeCost |
| Abstract Scope |
Machine learning systems are being widely deployed to accelerate measurements of the structure,
properties, and performance of materials. The black-box nature of the models used by these systems can limit the ability to decouple the effects of competing physical phenomena, particularly in small data or active learning settings. Our approach blends non-parametric machine learning models and Bayesian inference in physical models. Two challenging structural materials characterization tasks highlight this approach: quantitative analysis of multiphase x-ray diffraction (XRD) data and quantification of chemical short range order in multicomponent alloys via EXAFS. We show how to incorporate physical intuition into hierarchical priors, and how to incorporate flexible Gaussian Process modeling components for features without concrete physical models. Our long term goal is automated online analysis that can drive adaptive measurement selection with the aim of enabling comprehensive understanding of the relationship between structure and properties. |