| About this Abstract |
| Meeting |
MS&T24: Materials Science & Technology
|
| Symposium
|
Materials Informatics for Images and Multi-dimensional Datasets
|
| Presentation Title |
Synthetic 3D Microstructure Generation of Solid Oxide Cell Electrodes Using Denoising Diffusion Models |
| Author(s) |
Rochan Bajpai, William Kent, William Epting, Harry Abernathy, Paul Salvador, Rachel Kurchin |
| On-Site Speaker (Planned) |
Rochan Bajpai |
| Abstract Scope |
Detailed representations of 3D microstructures are crucial for materials simulation and design, but high-fidelity experimental measurement is costly. To address this, a denoising diffusion generative model was developed and trained on solid oxide cell anode microstructures acquired through plasma focused ion beam milling and SEM. The model matches, and often exceeds, performance of existing methods such as generative adversarial networks (GAN’s) or DREAM.3D in generating realistic microstructures, as assessed by distributions of functionally relevant properties such as phase fraction, tortuosity, and triple phase boundary density. Diffusion models also offer compelling advantages over GAN’s in data efficiency and stability/ease of training, as well as several intriguing possibilities for extension, building on algorithmic developments in related 2D models. One example that we will explore is conditional generation (e.g. targeting particular properties) using an existing pre-trained model by introducing targeted biases in the sampling process. |