Graph-Based Strategy for Microstructure Similarity in Large Datasets

Data-driven approaches have been recognized as a new paradigm to establish and explore the microstructure-property relationships. However, typical exploration methods deliver high-dimensional microstructures that pose the challenge of extracting the key features and patterns that could guide the design of processing. At the core of many data-driven approaches is the similarity measure that allows to organize the high-dimensional microstructures and perform meaningful data analytics. However, the currently available approaches either take a simplified view of the morphology, e.g., focusing on pixels in the morphology images, or apply transformations that average out structural descriptors of morphologies. To address these shortcomings, we propose a new computationally efficient and configurable similarity measure that is based on graph abstraction. Our main idea is to simplify complex morphologies by representing them as graphs, that are weighted with domain-specific information, and then express similarity as a distance between morphology graphs. Because both morphology graph structure and its weights have clear interpretation, our similarity measure can be easily tailored to the specific applications. Our results demonstrate the superior performance of the proposed approach on data from simulation and synthetic data, including in real-world applications (like morphologies clustering, and optimization).