Small‐Tilt Automated Reconstruction (STAR): A Machine‐Learning‐Based Stereo Method for Realization of Real‐Time 3D Microscopy Characterization
Year: 2026
Type: article
Abstract: In situ scanning/transmission electron microscopy (S/TEM) typically records two-dimensional projections, limiting interpretation of three-dimensional defect positions and orientations. We present Small-Tilt Automated Reconstruction (STAR), a machine-learning-enabled stereo framework for automated reconstruction of discrete features from views separated by small angular tilts (<5°). STAR integrates automated defect localization, tracking-based angular correspondence, video-frame interpolation, and small-angle double-tilt geometric reconstruction. Controlled benchmarks yielded a mean 3D position error of ∼0.51% of image width for synthetic loops and an average reconstruction error of ∼0.73% for latex spheres. Under assumed irradiated face-centered cubic crystallography, STAR correctly assigned habit-plane families for 97% of synthetic loops, with failures confined to the smallest {111} loops. Applied to post-irradiation images, STAR reconstructed cavity and dislocation-loop populations at computational throughputs of 0.50 and 1.18 reconstructions/s, respectively, providing depth-resolved statistics and loop-orientation distributions. STAR provides rapid, feature-level 3D reconstruction for discrete objects with prescribed geometries. Coupled with rapid stereo-pair acquisition, this computational framework supports time-resolved in situ and in operando defect characterization.
Source: Small
DOI: 10.1002/smll.76645
Institutions: University of Michigan, Los Alamos National Laboratory
Language: English
Subfield: Structural Biology
Domain: Life Sciences
Open Access status: hybrid
Estimated APC paid: $5,790
OpenAlex ID: W7236116751
