- Advisor: Prof. Bülent Yener (Computer Science, RPI), Prof. Dan Lewis (Materials Science and Engineering, RPI)
- Collaboration: Pacific Northwest National Laboratory
We aim to characterize the microstructures of a U-Mo system in a quantative way. By segmenting the regions of different phases of microstructures, we managed to extract several interpretable features, such as the area fraction of phases and the spatial distribution of microstructures. A kinetic model is built to visualize the changes of microstructures over 10 classes.
We have submitted our first publication to Materials Characterization for peer review. It is also pre-printed on arXiv. Currently, we are working on another publication which covers our work since May 2019.
Generative adversarial networks
We aim to build a generative adversarial network to learn the underlying distributions of the microstructures and to synthetize new images from the same distribution in our dataset. Multiple GANs of differernt resolutions are trained, from 1024 by 1024 to 256x256, which serve for different purposes.
We also make some attempts in image-to-image translation, where a "label map" of microstructures are given and a synthetic microstructure image is generated. This GAN focuses on the textures and visual effects of the microstructure images and leaves the distributions and properties of microstructures to the label map.
Papers and Presentations
An image-driven machine learning approach to kinetic modeling of a discontinuous precipitation reaction
Elizabeth Kautz*, Wufei Ma*, Saumyadeep Jana, Arun Devaraj, Vineet Joshi, Bülent Yener, Daniel Lewis (* equal contribution)
Submitted to Materials Characterization for peer review. Also pre-printed on arXiv.
Adversarial Networks for Microstructure Generation and Modeling Phase
Presentation at TMS 2020. Abstract accepted for publication.
Microstructure Characterization with Computer Vision
Selected for presentation at RPI open house event on ML/AI research.