Facilitating Scientific Material Discovery via Deep Learning on Small Image Datasets

Yin, Biao

Etd

Facilitating Scientific Material Discovery via Deep Learning on Small Image Datasets

Public Deposited

Scientific material discovery, important for economic prosperity and well-being from transportation, construction, and security, to healthcare; has traditionally been tackled both by intensive mathematical modeling and extensive physical experimentation. Recently, popular deep learning models have emerged as a promising solution approach. However, challenges remaining include overfitting due to the typically small size of datasets in this domain, tiny but critical pixels, strong false positives, lack of domain knowledge, etc. In this dissertation, I tackle four particular directions of research related to deep learning models on experimental complex data such as images derived from real-world projects for scientific material discovery. In the first part, we designed and developed the first open-source corrosion image dataset, annotated for data-driven automation in scientific corrosion assessment using expert labeling. Using this dataset, we built an AI platform, incorporating our published deep learning model, for real-world anti-corrosive material discovery rating via automatic data collection, exchange, and visual analytics embedded with our published deep learning models. In the second part, we focused on deep learning models in image-based scientific corrosion assessment for existing alloys. Techniques like augmentation, transfer learning, contrastive learning, as well as generative self-supervised learning were incorporated into the solution to improve its effectiveness. In the third part, we innovated a science-informed deep learning model named DeepSC-Edge, enhanced by a novel edge guidance submodel. This submodel focuses attention on high-level edge shapes while utilizing our unique loss function to prevent overfitting to edges. Additionally, our model incorporates a class-balanced loss, improving segmentation, particularly with challenging yet essential edges crucial for scientific corrosion assessment. In the fourth part, we created a domain-promptable AlloyGAN model aimed at producing microstructure images for alloys that have not previously existed in the world, based on their chemical composition and manufacturing parameters. By integrating domain knowledge into the model, my research empowers material scientists to effectively handle hypothetical alloys through instant and scientifically validated material simulation and evaluation. This approach represents a quicker and equally precise alternative to conventional methods in material science for evaluating alloy microstructures, while also showcasing the potential of GAN-based models in advancing scientific exploration in the realm of material 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. In general, the application of AI techniques to material science challenges promises to save time and effort for scientific material discovery.

Creator