| Abstract Scope |
Current in situ TEM ion‑irradiation experiments record unprecedented, time‑resolved projections of defect evolution; a single experiment can yield terabytes of images. Although these projected views reveal defect mobility and spatial correlations, they lack the depth fidelity of full 3D reconstructions. Achieving true 3D visualization would require at least two images per time step, further multiplying the data volume. Moreover, converting each stereopair into quantitative 3D information demands labor‑intensive labeling, tracking, and depth calculations that can take days, so most studies still rely on 2D projections.
We propose a machine‑learning‑assisted Small‑Tilt Automated Reconstruction (STAR) technique that removes these barriers. The workflow sequentially detects dislocation loops and cavities with a YOLO11‑OBB network, links corresponding defects between two tilted views using BoT‑SORT, and applies stereoscopic geometry to triangulate their 3D positions, shrinking post‑processing from days to seconds. Benchmarking STAR on in situ TEM ion‑irradiations of high‑purity Ni shows that it processes the entire tilt series, automatically reconstructs hundreds of defects, and delivers accurate number‑density, size‑distribution, and habit‑plane‑orientation metrics that previously required exhaustive manual effort. These results confirm STAR’s reliability under practical imaging conditions and pave the way for high‑temporal‑ and spatial‑fidelity 3D reconstruction in future in situ irradiation studies and other tilt‑series experiments. |