Segmentation of Experimental Datasets Via Convolutional Neural Networks Trained on Phase Field Simulations

Abstract The ability to quickly analyze large imaging datasets is vital to the widespread adoption of modern materials characterization tools, and thus the development of new materials. Image segmentation can be the most subjective and time-consuming step in the data analysis workflow. A promising approach to segmentation of large materials datasets is the use of convolutional neural networks (CNNs). However, a major challenge is to obtain the images and segmentations needed for CNN training, since this requires segmentations performed by humans. We show that it is possible to segment experimental materials science data using a SegNet-based CNN that was trained only using simple phase field simulations. A test image from an in-situ solidification experiment of an Al-Zn alloy was used to parameterize the phase field simulations. The most important microstructural features required for the best CNN to “understand” the contents of the image are ranked as: (1) having training images with diffuse particle-background interfaces, (2) modifying the images by adding noise, (3) removing particles at the image edges, and (4) adding sub-images to the particles to account for the feint bands present on some dendrites. The CNN trained on phase field images segmented the experimental test image with 99.3% accuracy, comparable to CNNs trained on experimental data. This approach of using computationally generated images to train CNNs capable of segmenting experiments will accelerate the rate of materials design and discovery.