researchsquare.comSource publication:
Loading content…
Limited-angle atomic electron tomography leaves a missing wedge because specimens cannot be tilted far enough, degrading depth resolution and atomic localization. RESIRE-Net uses only the input dataset: it rotates the reconstructed volume, projects it at the experimental tilt angles, reconstructs each synthetic tilt series with RESIRE, and trains a 3D U-Net on the resulting pairs to approximate the inverse of the missing-wedge operator. On multislice-simulated metallic-glass and medium-entropy-alloy nanoparticles, 3D positional RMSD falls from 21 to 14 picometers and from 24 to 15 picometers, with atom recovery and elemental classification improving as well.
Depth-direction localization in limited-angle atomic electron tomography improves, so atomic coordinates are no longer elongated and streaked by the missing wedge. Prior reconstructions relied on iterative algorithms such as RESIRE, which still leave an unsampled wedge under the roughly ±60° to ±75° tilt range and produce anisotropic depth resolution. On a multislice-simulated metallic-glass nanoparticle, 3D positional RMSD drops from 21 to 14 picometers, a 33% reduction dominated by depth, where the average z deviation falls from 15 to 10 picometers.
Better reconstruction carries through to atomic-model extraction: more atoms recovered, fewer spurious detections, and more accurate elemental classification. The same tracing and classification pipeline previously missed some atoms on RESIRE reconstructions and mislabeled noise as atoms. In the metallic-glass simulation, ground-truth atom recovery rises from 97.99% to 99.91%, spurious atoms fall from 494 to 132, overall classification accuracy rises from 97.35% to 99.81%, and the weakly scattering Ni class rises from 94.87% to 99.74%.
The method needs neither full-angle ground truth nor an external training set, because training pairs come from the data itself. Supervised deep learning previously depended on training data that closely matched specimen morphology, tilt geometry, noise, and resolution, so transfer to another material or acquisition setting could cause domain shift. Training pairs come from 20 rotated variants of the RESIRE reconstruction under the cubic symmetry group, each projected at the experimental tilt angles and reconstructed with RESIRE, so the network maps a twice-degraded volume back to a once-degraded one.
On an experimentally acquired medium-entropy-alloy nanoparticle, atomic peaks sharpen along the missing-wedge direction while twin boundaries and chemical short-range order are preserved. The prior RESIRE reconstruction of this experimental data was elongated along the least-sampled direction, with the first PDF minimum at 3.44 angstroms. The RESIRE-Net reconstruction moves the first minimum to 3.41 angstroms, closer to ideal FCC, lowers the mean deviation from a locally fitted FCC lattice from 0.26±0.12 to 0.23±0.11 angstroms, and keeps the dominant signs of αNiPt and αPdPt unchanged.
The demonstrated scope covers multislice-simulated metallic-glass and medium-entropy-alloy nanoparticles plus one experimentally acquired medium-entropy-alloy tilt series; analysis reuses the existing RESIRE atomic-model extraction pipeline with only the reconstruction algorithm replaced. The method targets researchers using atomic electron tomography, and its output is a 3D atomic model compatible with existing workflows. The authors propose future work on generative priors, physics-informed constraints, and implicit neural representations to further improve fidelity and robustness.
The method relies on rotational augmentation and the approximate isotropy assumption behind it; for objects with strong intrinsic anisotropy or limited symmetry, it may read genuine structural anisotropy as missing-wedge artifact and overcorrect. Probe polarization, electron channeling, detector non-uniformity, noise, tilt-angle miscalibration, mechanical drift, and alignment errors all bias the operator Tθ and can further limit reconstruction quality.