| About this Abstract |
| Meeting |
MS&T24: Materials Science & Technology
|
| Symposium
|
Materials Informatics for Images and Multi-dimensional Datasets
|
| Presentation Title |
Aligning Grains in Time-Series Laboratory Diffraction Contrast Tomography (LabDCT) Data for Machine Learning of Microstructure Evolution |
| Author(s) |
Woohyun Eum, Yi Wang, Kang Yang, Vishal Yadav, Michael R. Tonks, Amanda R. Krause, Joel B. Harley |
| On-Site Speaker (Planned) |
Woohyun Eum |
| Abstract Scope |
Laboratory diffraction contrast tomography (LabDCT) is a laboratory-scale x-ray tomography for non-destructively imaging grains. Its combination with machine learning, such as the recent Physics-Regularized Interpretable Machine Learning Microstructure Evolution (PRIMME) framework, has the potential to elucidate mechanistic relationships within microstructural evolution. However, appropriate data preprocessing and cleaning is essential for machine learning, and common preprocessing pipelines for segmenting and matching grains between consecutive LabDCT measurements is often insufficient, sometimes failing to track up to 10% of grains. This study employs strategies to better identify and match grains across time by minimizing differences in grain centroids and orientation through the Hungarian algorithm. We validate our approach by observing its performance with simulated microstructure data as well as the error distributions for matched grain centroids and orientations in experimental data. Results show our approach to mitigate experimental noise and improve the PRIMME framework's ability to learn from experimental data. |