About this Abstract |
| Meeting |
2027 TMS Annual Meeting & Exhibition
|
| Symposium
|
Modeling, AI Applications, and Method Development in Nuclear Materials
|
| Presentation Title |
Deep Learning–Enabled Automated Tracking Pipeline for Helium Cavity Dynamics in In-Situ Transmission Electron Microscopy |
| Author(s) |
Chun Yin (Tommy) Wong, Quanyou Zhang, Caitlin A. Kohnert, Elijah M. Davis, Maxim A. Ziatdinov, Sergei V. Kalinin, Haixuan Xu, Khalid Hattar |
| On-Site Speaker (Planned) |
Chun Yin (Tommy) Wong |
| Abstract Scope |
In-situ transmission electron microscopy (TEM) annealing and ion irradiation experiments of helium cavities in prototypical FCC alloy systems have resulted in a large video dataset of He bubble to faceted cavity dynamics. To that end, a lightweight deep learning–enabled automated tracking pipeline was created to analyze the videos accurately and expediently. Compiled and trained on the Google Colab cloud-based platform, the Mask Recurrent-Convolutional Neural Network (R-CNN) achieved an intersection over union of 0.88 using only a few micrographs for training. The Mask R-CNN model was benchmarked against other contemporary CNNs. Individual overlapping cavities were segmented and sized. Lightweight drift correction and tracking algorithms were used to analyze changes in cavity morphology and trajectory—giving provisions for online tracking. Grain boundary motion was also analyzed using a similar pipeline. The proposed pipeline gives credence to complementary atomistic and molecular dynamics simulations, enabling predicative computational and experimental modeling. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Nuclear Materials, Characterization, Machine Learning |