Assessing Unsupervised Learning Models for Synthetic Process Specification Analyses

In recent years, computer vision-based image analysis of microstructures captured in scanning electron micrographs (SEM) has shown to be a useful analytical tool for material property characterization, synthetic process specification prediction, and many other tasks. Prior works focused on developing supervised classification or regression models for predicting material properties. Recently, the unsupervised learning paradigm, which learns to encode descriptive representations of microstructures without using human annotations, has shown success relative to its supervision counterpart in terms of generalizability, training effort, and data parsimony when training downstream tasks such as classification. Recent unsupervised learning works have briefly explored training convolutional neural networks (CNNs) with reconstruction loss and separately with contrastive loss. In this work, we present a comprehensive evaluation of models, trained with unsupervised training paradigm, for determining synthetic process specification of uranium ore concentrates and plutonium oxides. Concretely, we train an array of computer vision-based architectures (from CNNs to vision transformers) via different unsupervised objective functions. We compare the training effort and performance on downstream classification, regression, and retrieval tasks using representations from these models. Even though the evaluation heavily emphasizes on synthetic process specification of two materials, we believe the results and conclusions herein will generalize readily to many other materials.