DS Ph.D. Dissertation Defense | Biao Yin | Thursday, April 18th @ 2:00PM - Gordon Library | "Facilitating Scientific Material Discovery via Deep Learning on Small Image Datasets"

DATA SCIENCE  

Ph.D. Dissertation Defense 

Biao Yin, Ph.D. Candidate 

Thursday, April 18th, 2024 | 2:00PM – 3:00PM 

Location: | Conference Room 303, Gordon Library 

Dissertation Committee: 

Dr. Elke Rundensteiner, Professor, WPI (Advisor)​ 

Dr. Ziming Zhang, Assistant Professor, WPI​ (Co-Advisor) 

Dr. Jian (Frank) Zou, Associate Professor, WPI ​ 

Dr. Robert Jensen, Team Lead of Materials Data Science, DEVCOM Army Research Lab​ 

Title: 

Facilitating Scientific Material Discovery via Deep Learning on Small Image Datasets 

Abstract: 

Scientific material discovery, important for economic prosperity and well-being from transportation, security, to healthcare, has traditionally been tackled by intensive mathematical modeling and physical experimentation. Recently, popular deep learning models have emerged as a promising solution approach; however, challenges remaining include overfitting due to small sized datasets, tiny but critical pixels, strong false positives, lack of domain knowledge, etc.  

I tackle four directions of research related to deep learning models on experimental complex data such as images from real-world projects for scientific material discovery. First, we developed the first open-source corrosion image dataset annotated for automation of scientific corrosion assessment using expert labeling. We also built an AI platform, incorporating our published deep learning model, for real-world anti-corrosive material discovery rating and analytics. Second, we focused on supervised deep-learning models for image-based scientific corrosion assessment on cast alloys. Techniques like augmentation, transfer learning, contrastive learning, and generative self-supervision were incorporated into the solution. Third, we innovated DeepSC-Edge, a science-informed deep learning model, which prioritizes submodule guidance enabling high-level edge shapes with a unique loss function to avoid segmentation overfitting. Our model also incorporates a class-balanced loss, especially for challenging edges. Fourth, we created the AlloyGAN model to generate microstructure images of uncast alloys promptable by their chemical composition and manufacturing parameters. Injecting domain knowledge, our model empowers material scientists to simulate hypothetical alloys, providing a faster yet precise alternative to traditional methods in material science. Our approach demonstrates the potential of GAN-based models to advance scientific exploration in materials discovery. 

This work is based on a collaboration with material scientists at the DEVCOM Army Research Laboratory (ARL) - with the later testing and working with the resulting technology. In a collaboration between ARL, WPI, and ASM, the technology is being transitioned into practice for marketing and release by ASM. We thank ARL, NSF, and DoE for sponsoring this research.