Dr. Olga Wodo, Associate Professor
Materials Design and Innovation Department (MDI)
University of Buffalo, USA
Microstructure informatics: bridging materials and data science for accelerated design
The holy grail of materials science is to establish quantitative process-structure-property relationships (PSP). Defining these relationships is rarely a straightforward task, mainly due to the mismatch between (micro)structural information observed (e.g., via microscopy or simulations) and the principal degrees of freedom governing the PSP. The mismatch exists because microstructural imaging aims to provide detailed, high-resolution maps, while the purpose of establishing quantitative PSPs is to derive the smallest set of variables/descriptors that explain most of the variability in the data. To close this gap, the microstructural data sets should be reduced to meaningful descriptors (or other low-dimensional representation) to establish PSP to accelerate materials design. At the atomistic level, descriptors played a crucial role in materials design for photovoltaics, batteries, or catalysts. More than a dozen software tools are available to calculate descriptors at the electronic and atomistic levels. Descriptors at the next scale – microstructure – are relatively less explored and consist of various application-specific and disparate clusters of descriptors. Most frequently, descriptors are tailored to characterize specific mechanisms and properties.
In this talk, we will present various approaches to handle the mismatch using tools of microstructure informatics. First, we will introduce our unified framework to compute a library of generic microstructural descriptors. Our microstructure representation is derived from the graph theory and enables microstructure featurization in terms of shape (i.e., morphology), geometry, and connectedness (i.e., topology). In the simplest form, the framework can be used to quantify the microstructures. When combined with high throughput workflows, it can be harnessed to represent materials microstructure in a machine learning-friendly format for data-driven PSP map construction. The ability to represent materials microstructure in machine-friendly formats is one critical aspect of data-driven approaches. Using the combination of different representations (descriptors, two-point correlation, and autoencoder learned latent space), 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.
Short bio: Olga Wodo is an Associate Professor in Materials Design and Innovation Department at University at Buffalo. Olga received her PhD degrees in Mechanical Engineering from Czestochowa University of Technology (Poland). Her research interests are focused on microstructure informatics, computational materials science, multiscale modeling, and high-performance computing. Her research areas cover heterogeneous materials, additive manufacturing and more recently mycelium-based materials. During pandemic she developed interest in chocolate making following bean-to-bar movement, and gained practical knowledge about the importance of processing conditions on crystallization.