Machine Learning Models Pre-trained on MicroNet Show Improved Segmentation and Microstructure Quantification

Neural network encoders were pre-trained on MicroNet, a dataset of over 100,000 micrographs from NASA and the literature to learn robust microstructure representations. MicroNet contains images from 54 materials classes including metals, composites, and polymers, and includes SEM, TEM, CT, and optical images. The improved performance of these pre-trained encoders on downstream segmentation and direct regression tasks will be shown. The pre-trained encoders were applied through transfer learning to segment and extract features from micrographs of different materials. The extracted features were then linked to processing and/or property data in order to establish quantitative processing-structure-property relationships. The presentation will demonstrate the technique on several materials including: Ni-superalloys where precipitate morphology and matrix channel width are quantified, and environmental barrier coatings where a thermally grown oxide is segmented and its roughness, thickness, porosity, and inter-crack spacing are quantified. Software to use the encoders and perform automatic image analysis are publicly available.