| 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 |
Automated Real-Time 3D Stereo-Reconstructions Through Machine-Learning based Tracking |
| Author(s) |
Hangyu Li, Benjamin Eftink, Kevin Field |
| On-Site Speaker (Planned) |
Hangyu Li |
| Abstract Scope |
In-situ experiments offer unparalleled insights into the microstructural evolution of materials, yet analysis of the vast datasets they generate is often time- and labor-intensive. This work presents a novel strategy to accelerate in-situ irradiation-induced defect characterization through an automated, real-time 3D stereo-reconstruction workflow. Utilizing YOLOv8/YOLOv10 for defect detection and BoTSORT/ByteTRACK for tracking, we can rapidly extract features (e.g., loops, cavities) from stereo-image pairs. This approach captured 77.5% of defects across a 5° tilt series, with missed features primarily exhibiting low contrast. The modified Obtain3D code enabled 3D reconstructions in <0.1 seconds, demonstrating real-time potential.
This technique has broad applicability to in-situ experiments and promises to transform the analysis of in-situ irradiation experiments, enabling a deeper understanding of defect dynamics and ultimately accelerating the development of radiation-resistant materials. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Machine Learning, Other, |