| About this Abstract |
| Meeting |
2025 TMS Annual Meeting & Exhibition
|
| Symposium
|
Novel Strategies for Rapid Acquisition and Processing of Large Datasets from Advanced Characterization Techniques
|
| Presentation Title |
Deep Learning Conditional Diffusion Models to Recreate Scanning Electron Microscopy Using Light Optical Microscopy Priors |
| Author(s) |
Nicholas Amano, Bo Lei, Martin Müller, Dominik Britz, Elizabeth A Holm |
| On-Site Speaker (Planned) |
Nicholas Amano |
| Abstract Scope |
Observation and analysis of microstructure is fundamental to metallurgical science, hence significant resources are allocated towards preparing and imaging structured materials. We present a deep learning based method of recreating scanning electron microscopy(SEM) images using light optical microscopy(LOM), a cheaper and more scalable imaging technique, using conditional diffusion models. This work is made possible by a unique dataset of correlated LOM and SEM micrographs collected at identical sample locations and length scales. Diffusion models are the current state of the art for image generation and we have shown their capacity to recreate SEM textures from LOM images for simple dual phase steel. We extend this work to more complex and ambiguous quenched and tempered steel micrographs, where there is no clear separation of phases. We demonstrate that diffusion models are able to produce plausible SEM microstructures but struggle with subtle features. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Computational Materials Science & Engineering, Machine Learning, Characterization |