Computer vision and deep learning for microstructural modeling and automated characterization of materials

Deep learning has demonstrated impressive results for a variety of applications in image analysis, but the use of this powerful technique in materials science has not fully been explored. This thesis presents a survey in which deep learning is applied to challenges in modeling and automated characterization of various materials and processes. The results demonstrate how neural networks can be used for a variety of tasks in image analysis. Convolutional neural networks and transfer learning are applied to simple image classification and more advanced instance segmentation applications. Transfer learning is applied to develop high-performance models without requiring extensive data collection and labeling efforts. Graph neural networks are used to predict the occurrence of abnormal grain growth in simulations of microstructural evolution. The results demonstrate how deep learning can enable new approaches to materials characterization and modeling, providing benefits for a wide range of applications