Feed-Forward Atomic Electron Tomography with Calibrated Missing-Wedge Uncertainty

Many of the catalysts that could replace platinum and iridium are made from cheap metals such as nickel, cobalt, iron and molybdenum. Their activity often comes from atomic structure that is messy rather than ideal: boron sitting between nickel atoms, asymmetric cobalt sites left behind by acid etching, sponge-like nickel left after zinc is leached out. The one technique that can see such structure atom by atom in 3D is atomic electron tomography (AET). AET reconstructs a nanoparticle from a "tilt series", a set of 2D electron-microscope projections taken as the sample is rotated. The microscope stage can only tilt to about ±70°, so a wedge of viewing directions is never measured. This "missing wedge" smears atoms along the beam axis. It is the tomographic version of the depth ambiguity in single-image 3D vision: the data cannot tell you where along the ray an atom sits.

The best current fix fits a set of 3D Gaussian blobs, one per atom, to each nanoparticle by gradient descent. It uses hand-written rules to keep the blobs round and apart. We propose to replace this with a network trained on large numbers of simulated nanoparticles. In one forward pass, plus a few learned refinement steps, the network maps a tilt series to a set of atoms. Each atom comes with a calibrated probability that it exists and a covariance describing how uncertain its position is along the unmeasured directions. Speed is not the main claim. The claim is that a learned structural prior fills the missing wedge better than hand-written regularizers, and that uncertainty expressed as competing candidate atoms captures the "is the atom here or one lattice spacing deeper" ambiguity. A single unimodal Gaussian per point cannot represent that kind of either-or ambiguity.

Core Claim and What Is New

A Gaussian Parameterization for Direct Atomic Structure Identification in Electron Tomography motivates this work. It fits position, amplitude and size of spherical Gaussians to each tilt series, using an isotropy constraint and a minimum-distance merging rule to suppress streaks. It improves atom identification over filtered backprojection and SART, and it separates chemical species by amplitude. It is also a per-scene optimizer with priors fixed by hand. In computer vision, feed-forward reconstructors such as GS-LRM and DepthSplat regress 3D Gaussians directly from images in under a second. ReSplat adds a few recurrent steps that render the current guess, compare it to the input, and feed the error back without test-time gradients. These systems supply our method, but they all assume pinhole cameras, opaque surfaces and one Gaussian per pixel. Trust3R supplies our uncertainty evaluation protocol and also defines the limitation we target. It gives feed-forward 3D models calibrated per-point Student-t uncertainty, reducing AUSE by 41% on ScanNet++, but it is unimodal by construction.

Our model combines four choices that none of these papers make together. It is trained across many specimens rather than fit to one. Its outputs are a discrete set of atoms rather than pixel-aligned or voxel densities. Its projection model is a linear, transmission-style one in which the signal adds up along the beam, rather than alpha compositing. Its uncertainty is represented as mutually exclusive candidate atoms, each with an existence probability and an anisotropic position covariance, which handles discrete depth ambiguity. Space-Time Implicit Neural Representations for Atomic Electron Tomography on Dynamic Samples and NeAT show that neural tomography is feasible. Both, however, are per-scene optimizers over continuous fields. NeAT needs about 45 minutes on an A100. Neither gives per-atom uncertainty.

Research Program

The training data is simulated. We carve nanoparticles of 2–8 nm from crystal structures spanning the element range of OC20 and add realistic disorder: grain boundaries, vacancies, interstitial light elements such as B and P, amorphized regions, and dealloyed surfaces. Each particle is projected with an incoherent HAADF-like forward model, in which each atom contributes a blurred blob whose brightness grows with atomic number. We apply tilt ranges sampled between ±50° and ±80°, Poisson dose noise, small residual misalignments, and a thin amorphous-carbon background. The test families are held out by structure type and match the Au, FePt grain-boundary, amorphous Pd and high-entropy-alloy particles used by Singh et al.

The network has three stages. First, each projection is encoded in 2D and the features are backprojected into a coarse 3D feature volume, so the network starts from the physics rather than learning tomography from scratch. Second, a transformer over sparse anchor points in that volume proposes candidate atoms. Each proposal has a 3D offset, a species distribution, an existence logit, and a full 3×3 position covariance. Third, over T recurrent steps in the style of ReSplat, the current atom set is rendered into all measured tilts. The residuals are backprojected, and the transformer updates the proposals. This step never requires gradients at test time.

Training matches proposals to ground-truth atoms with Hungarian matching and uses three losses. A Gaussian negative log-likelihood on matched positions trains the covariance. A cross-entropy loss trains existence and species. A projection-consistency loss requires that the expected rendering, with each atom weighted by its existence probability, reproduces the measured tilts.

The multimodal behaviour comes from how the outputs are read. When the data cannot decide between two depths along the beam, the calibrated answer is two proposals at about 0.5 existence each, which the matching loss rewards in expectation. Downstream users sample whole structures by drawing atoms according to their existence probabilities, subject to a minimum-separation exclusion.

The catalysis direction is the application regime and a downstream test. It is not part of the core mechanism. Adsorption energies are necessary but not sufficient to identify good catalysts shows that screening already carries about 0.3 eV of DFT uncertainty for hydrogen-evolution descriptors. We therefore ask a quantitative question: is the error introduced by reconstructing surface atoms from tomography small or large compared with that 0.3 eV? We will sample structures from our posterior for simulated Ni-based and Mo-P particles. For each sample we estimate hydrogen adsorption energies at surface sites with a surrogate force model trained on OC20, then compare the spread with the DFT uncertainty band. One caveat applies: OC20-trained surrogates are built for slabs, so this is a relative comparison, not an absolute prediction.

Baselines

The naive reference is filtered backprojection followed by standard peak-finding atom tracing, with SART plus tracing as a second classical reference. These are what most AET pipelines run today. We tune the tracing thresholds on the validation split for each structure family.

The strongest existing method is the per-scene Gaussian optimization of Singh et al.. The file reports only qualitative gains for it ("no reported numerical result found on Lacuna" for atom-identification F1 or position error). We will therefore reimplement it with its isotropy and interatomic-distance regularizers. For each structure family we tune its regularization weights and iteration count on the validation particles, and report it on identical simulated tilt series. Our model beats it if it achieves higher atom-identification F1 and lower matched-position error. The F1 comparison is at a 0.5 Å match radius on held-out families at ±70°, with paired bootstrap confidence intervals over particles. The margin should grow as the wedge widens. We also report wall-clock time on the same GPU.

The strongest learned competitor is a 3D U-Net that maps filtered-backprojection volumes to clean density, followed by the same atom tracing. It is trained on exactly our simulated data with matched parameter count and training steps.

A hybrid baseline initializes Singh et al.'s optimizer from our feed-forward prediction. Three ablations isolate the key ingredients. The first removes the recurrent residual steps (T = 0). The second replaces the learned anchor transformer with the residual-update loop alone, starting from traced filtered-backprojection atoms. The third replaces candidate-set uncertainty with a single unimodal Student-t per atom in the style of Trust3R.

Why not just…?

Why not just run the per-scene optimizer on a GPU and accept a few minutes per particle? Reconstruction time is not the bottleneck in AET, since acquiring the tilt series takes far longer. So speed alone would not justify this work. Our claim stands or falls on accuracy and uncertainty under a wide missing wedge, where hand-written isotropy and spacing rules carry no knowledge of which local structures are physically plausible. The tilt-range sweep settles this. If the per-scene optimizer matches our F1 at ±60° and ±50°, the learned-prior claim fails, and what remains is amortized throughput for high-throughput atlases.

Why not just denoise the backprojected volume with a 3D U-Net and trace atoms? This is a real threat, because it also learns a prior from the same data. Its weakness is that tracing a density volume gives no calibrated probability that an atom exists. It also tends to merge or split atoms where the wedge stretches them. The matched-capacity comparison measures both effects, and we report it even if the U-Net wins on F1.

Why not just use our network as an initializer for the optimizer? This hybrid may well be the most accurate option, and we treat it as a likely outcome rather than an objection. If it beats pure feed-forward inference on F1, the practical recommendation becomes prediction followed by a short refinement. The uncertainty outputs remain the contribution, because the optimizer produces no calibrated posterior.

Why not just attach Trust3R's unimodal evidential head? The unimodal ablation tests this directly. On particles where we deliberately place atoms in ambiguous columns, a single Gaussian must sit between the two possible depths with a large variance. The candidate-set model can instead put probability mass at both depths. Ambiguity-region coverage and negative log-likelihood will show which representation is better calibrated.

Proposed Experiments

The main benchmark uses the held-out simulated Au, FePt, amorphous Pd and high-entropy-alloy families from Singh et al. We report atom precision, recall and F1; matched position error, split into in-plane and beam-axis components; species accuracy; and runtime. We sweep tilt range from ±80° down to ±50° and dose across three levels.

For uncertainty we report existence expected calibration error and covariance coverage at 68% and 95%. We also report AURC and AUSE computed as in Trust3R, which ranks atoms by predicted uncertainty and measures how error falls as the most uncertain atoms are removed. A model that understands the wedge should put the long axis of each predicted covariance along the unsampled direction. We measure this as the angle between the covariance's major axis and the missing-wedge axis.

Transfer to real data is tested on the experimental Zr-Te carbon nanotube used by Singh et al. There is no ground truth, so we measure three things: projection residuals on held-out tilts, agreement with the per-scene reconstruction, and whether amplitude-based separation of C and Te survives.

The catalysis check compares the spread of surrogate-predicted hydrogen adsorption energies across posterior samples with the 0.3 eV DFT band.

Two results would decide the project. The first is whether the F1 gap over the per-scene optimizer widens as the wedge grows, which tests the learned-prior claim directly. The second is whether held-out tilts that our model marks as uncertain actually contain larger errors, which tests whether the uncertainty is real or decorative.