About this Abstract |
| Meeting |
2026 TMS Annual Meeting & Exhibition
|
| Symposium
|
Algorithms Development in Materials Science and Engineering
|
| Presentation Title |
Label-Aware Microstructure Generation for Manufacturing Processes Using Generative Networks |
| Author(s) |
Zekeriya Ender Eger, Waris Khan, Pinar Acar |
| On-Site Speaker (Planned) |
Zekeriya Ender Eger |
| Abstract Scope |
The generation of synthetic microstructures has become important in materials science, especially in data-scarce regimes where experimental imaging is limited or costly. High-fidelity synthetic data supports robust structure-property models, surrogate modeling, and inverse design. This study presents a conditional Denoising Diffusion Probabilistic Model (DDPM) to generate microstructures based on inputs like processing conditions or morphological descriptors. Conditional DDPMs allow controllable, probabilistic generation, capturing the complexity and variability of real microstructures. The generated samples are assessed through a three-stage comparison. First, statistical image similarity metrics such as structural similarity index (SSIM), histogram comparison, and power spectral density are used to assess visual and statistical fidelity. Second, domain-specific features like grain size, shape, and orientation are compared to real data. Finally, the physical relevance of the generated samples is evaluated by simulating their effective properties using crystal plasticity simulations, enabling a direct comparison of mechanical responses between real and synthetic microstructures. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Machine Learning, Modeling and Simulation, Characterization |