Generalizability, Data Diversity and Representation in Microstructure-Property Mapping

Mapping microstructure-sensitive properties with microstructure representation is invariably challenging due to the mismatch between the high dimensionality of microstructural information (e.g., via microscopy or simulations) and the principal degrees of freedom (or salient features) governing the properties. This is because microstructural imaging aims to provide detailed, high-resolution maps. Hence, imaging techniques inevitably produce high-dimensional representations of microstructure, while the goal of establishing practically useful structure-property models is to identify the smallest set of features that can successfully predict the effective properties exhibited by the material. Often, this set is not known a priori, especially for complex multi-physics phenomena governing the material properties.

Data-driven approaches become the integral approach to establishing reliable microstructure-property mappings. However, materials science datasets are typically small, or the property evaluations are computationally or experimentally demanding, and this requires approaches that integrate the small datasets or seek smart sampling strategies. One critical aspect of data-driven approaches is the ability to represent materials microstructure in machine-friendly formats. This talk presents three methods: statistical descriptors (e.g., 2-point correlations), vector of physically meaningful descriptors, and latent space learned using autoencoder. Using the combination of different representations, we explore three questions: Given a few datasets with distinct microstructure annotated with the property of interest: 1) Can a small subset of features be selected to train a robust microstructure-property predictive model? And is this subset agnostic to the choice of feature selection algorithm? 2) Can the addition of expert-identified features improve model performance? 3) Can the generalizable model be trained for independent microstructure datasets (different microstructure types)? The questions are essential for any microstructure-sensitive properties. Still, in this talk, we will utilize the problem of constructing structure-property models for organic photovoltaics applications (OPV) to understand data-driven SP models.