{"id":"57a78c2d-6da6-48fe-9910-e53cba8f6b04","arxiv_id":"2506.16104","paper_version":1,"verdict":"UNVERDICTED","confidence":"MODERATE","novelty_score":0.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"A review of neural network-based atomic electron tomography, covering deep learning methods for missing-wedge artifact removal, surface atomic structure determination, and an outlook toward light-element and dynamic imaging.","lead":"This preprint is a review of recent work that combines neural networks with atomic electron tomography (AET) to fix image artifacts caused by missing tilt angles. It summarizes the methods, results, and future directions for using deep learning to see 3D atomic structures of nanoparticles more accurately.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Key precision/detection gains are simulation metrics; experimental support rests on R-factor, which can reward the atomicity prior rather than recovered truth.","rationale":"The reader correctly identified the weakest premise as the possibility that neural-network-refined tomograms impose the learned atomicity prior rather than recover true atomic positions. My concern overlaps but is more specific: the two quantitative gains most likely to persuade a reader (95.6 to 98.8% detection, 26.1 to 15.1 pm precision) appear to come from simulated data with known ground truth, while the only experimental metric, the R-factor, is a consistency check that can improve simply because the network makes the volume more atom-like. The review's own limitation paragraph about surface diffusion and beam effects further weakens the assumption of a static ground truth. A leave-one-tilt-out test would directly probe whether the network is recovering missing angular information or merely imposing a prior. I recommend CONDITIONAL rather than REJECT because this is a survey, the underlying methods are credible, and the issue is mainly one of how the evidence is labeled and interpreted. The condition is that the review must explicitly distinguish simulation-derived accuracy metrics from experimental consistency metrics and either add cross-validation evidence or soften the claim that experimental positional accuracy is improved. This is not an ad hominem judgment; it is a request for evidentiary precision.","tokens_in":14180,"tokens_out":5120,"duration_ms":54092,"concrete_test":"Perform leave-one-tilt-out cross-validation on the experimental Pt tilt series from the showcase study (Ref. 108): reconstruct from a subset of tilts, run the NN augmentation, project the refined volume at the held-out tilts, and compare against the measured projections with a non-atomicity-based metric (e.g., normalized cross-correlation or Fourier ring correlation). If the augmented volume predicts held-out projections clearly better than the raw tomogram, the R-factor gain reflects recovered missing information; if not, it is consistent with the atomicity prior rather than recovered truth. Also confirm from Ref. 108 whether the 96.5/98.8% detection and 26.1/15.1 pm precision figures are simulation benchmarks; if so, the review should label them accordingly.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"In the section 'Precise determination of Pt nanoparticle surfaces and interface structures', the review reports atom detection rising from 96.5% to 98.8% and coordinate RMSD falling from 26.1 to 15.1 pm. The associated Fig. 3 panels (a-f) are described as 'from simulation of AET process for a Pt nanoparticle' with known ground truth, so these headline numbers are simulation benchmarks, not measurements on the experimental particle. The only experimental validation offered is the R-factor improving from 19.2% to 17.4%, computed by comparing the experimental tilt series to projections of the reconstructed atomic model. That comparison is not an independent ground truth: the forward projection assumes discrete atomic potentials, and the network was explicitly trained to impose the atomicity prior. Sharper, more atom-like peaks will generally lower this R-factor even if individual atoms are placed incorrectly. The review itself concedes that surface atoms diffuse on 10^-4 to 10^-7 s timescales and the dose can perturb the surface, so a static atomic model is a moving target. Consequently, the central claim that neural-network augmentation reliably improves true atomic positional accuracy in experiments is not secured by the cited evidence; the strongest quantitative support is simulation-only, and the experimental support is a consistency metric that the prior can satisfy.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This review article surveys recent applications of convolutional neural networks to atomic electron tomography (AET). It covers three broad strategies: deep-learning-based recovery of missing-wedge data, neural-network augmentation of tomograms under an 'atomicity' prior, and CNN-based image inpainting of support signals. The authors report quantitative gains from the primary literature, including atom detection improving from 96.5% to 98.8%, coordinate RMSD decreasing from 26.1 pm to 15.1 pm, R-factor improving from 19.2% to 17.4%, a 5.5% RMSE reduction for an ensemble transformer model, and 0.7 Å resolution for a deep-learning-aided reconstruction. The showcase applications are Pt and Pd nanoparticles, with emphasis on surface structure, strain, and catalytic activity. The review also discusses future directions in 4D-STEM ptychography, multislice tomography, low-dose imaging, and uncertainty quantification.","tokens_in":14418,"tokens_out":8137,"duration_ms":96089,"significance":"If the central claims are taken at face value, the review provides a useful and well-organized synthesis of an active area, and the authors are well placed to write it given their direct contributions to several of the key papers. The manuscript is readable and mostly accurate in reporting numbers from the cited studies, and it explicitly acknowledges limitations such as surface atom mobility and the lack of universally validated ground truth. The main weakness is that the most impressive quantitative claims—pm-level precision, near-perfect detection rates—are simulation benchmarks, while the experimental validation is a single consistency metric (R-factor) that can be partly satisfied by the very prior the network is trained to impose. The review should make this distinction explicit. With that clarification, it would be a valuable resource for newcomers and practitioners alike.","major_comments":[{"comment":"The quantitative gains that anchor the showcase—atom detection rising from 96.5% to 98.8% and coordinate RMSD decreasing from 26.1 pm to 15.1 pm—come from a simulation with known ground truth, as stated in the Fig. 3 caption, but the body text presents them without this qualification. The only experimental validation offered is the R-factor improvement from 19.2% to 17.4%, computed by comparing the experimental tilt series to projections of the reconstructed atomic model. This metric is not an independent check on atomic positions: the network was trained to impose the atomicity prior, so sharper, more atom-like peaks can lower the R-factor even if individual atoms are misplaced. Please state explicitly that the precision and detection figures are simulation benchmarks, and add a sentence explaining what the R-factor does and does not validate, or cite additional experimental validation from the primary sources.","section":"Fig. 3 and §'Precise determination of Pt nanoparticle surfaces and interface structures'"},{"comment":"The statement that these methods 'generalize across diverse structural types' is stronger than the evidence assembled in the review. The experimental cases shown are predominantly Pt and Pd nanoparticles, plus one nanoporous-gold example, and most of the quantitative generalization evidence is simulation-based. Please qualify this claim, or provide a systematic cross-structure benchmark or reference to one, so that readers do not over-interpret the breadth of demonstrated experimental applicability.","section":"Summary and Outlook"}],"minor_comments":[{"comment":"The version under review omits the actual figure images while the text refers to specific panels (e.g., Fig. 3a–i). Since the figures are central to a review of this type, please ensure the published version includes all figures and panels, or clearly indicate how readers can access the complete figure set.","section":"General / figure availability"},{"comment":"Please define the R-factor or provide a reference for its computation and expected range, so that the numerical change from 19.2% to 17.4% can be interpreted by readers who are not specialists in electron tomography.","section":"Fig. 3 / R-factor"},{"comment":"Use superscript notation for the timescale and dose expressions: '10^-4–10^-7 s' and '10^5 e Å^-2'. Also fix the spacing glitches in the abstract ('st rain' → 'strain', 're solution' → 'resolution').","section":"Abstract and 'Precise determination...' paragraph"},{"comment":"Reference 130 (UsiNet) is cited but not discussed in the text; a brief sentence describing this unsupervised sinogram-inpainting approach would make the survey more complete and better balanced.","section":"Deep learning for missing wedge recovery"},{"comment":"The reference list has formatting inconsistencies (e.g., refs. 49 and 56) and includes a preprint (ref. 161); please check the journal style for these entries.","section":"References"}],"recommendation":"minor_revision","confidential_remarks":"The review draws heavily on the authors' own publications (Refs. 60, 108, 134, 135, 156, 161, 162), which is acceptable for a review but worth keeping in mind. The editor may wish to encourage the authors to add a short comparative discussion of independently developed methods (e.g., UsiNet, Ref. 130) and any independent validation studies. The main required revision is the explicit separation of simulation benchmarks from experimental evidence, which is a text-level fix rather than a change of scientific direction."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a review of neural network methods for atomic electron tomography, written by two people who did much of the primary work. It is a fair, accurate survey and deserves peer review. The main thing to know: the headline precision numbers (96.5% to 98.8% detection, 26.1 to 15.1 pm RMSD) come from simulations with known ground truth, and the experimental validation is an R-factor that the network's atomicity prior can partly satisfy. The authors don't hide this—the figure caption says 'from simulation' and the text concedes surface atoms diffuse—but the prose elsewhere sometimes lets the simulation numbers stand in for experimental truth.\n\nWhat it does well: the review gives a clear taxonomy of approaches: GAN-based sinogram inpainting (Ding et al.), U-Net tomogram augmentation (Lee et al.), transformer hybrids (Yu et al.), and CNN inpainting of support background (Iwai et al.). It places them in the context of the missing-wedge problem and explains why atomicity is a powerful prior. The quantitative details (0.7 Å resolution, 5.5% RMSE reduction, 19.2% to 17.4% R-factor) are consistent with the cited papers. The limitations paragraph about surface mobility and beam effects is honest and well placed.\n\nSoft spots: the review is heavily self-referential—most showcase examples are the authors' own prior publications (Refs 60 and 108). That is not circularity, but it is a selection bias. The deeper issue is the evidence base: the strongest accuracy numbers are simulation-only, and the experimental evidence is a consistency metric, not ground truth. The R-factor improvement can occur simply because the network makes peaks sharper and more atom-like, even if atoms are individually misplaced. The review acknowledges the risk of artifacts but does not seriously address it beyond one paragraph. A referee should ask the authors to state explicitly in the main text which gains are simulation-derived and which are experimental.\n\nWho this is for: practitioners entering AET who want a map of the neural-network toolbox, and instructors who need a compact overview. It will not settle whether the learned atomicity prior invents structure.\n\nRecommendation: accept for peer review as a useful review, with minor revisions asking for a clearer simulation/experiment distinction.","headline":"A useful and mostly faithful review of neural network AET, but the headline precision numbers are simulation benchmarks and the experimental validation is a consistency metric.","tokens_in":14871,"tokens_out":2775,"would_cite":true,"duration_ms":30831,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This review claims that convolutional neural networks guided by the atomicity prior recover missing tomographic data and sharpen atomic electron tomography to picometer-level precision, making surface atom positions in nanoparticles…","keywords":["atomic electron tomography","convolutional neural networks","missing wedge","atomicity prior","3D U-Net","image inpainting","nanoparticle surface structure","ptychography"],"falsifier":"Take a tomogram of a known amorphous or highly disordered structure, such as a simulated metallic glass, run the trained U-Net augmentation, and count whether refined volumes show new well-resolved atomic peaks where the ground-truth density is continuous; a nonzero false-peak rate, or an R-factor that improves while the coordinates diverge from a simultaneously obtained ptychographic reconstruction, would indicate the network is imposing atomicity rather than recovering structure.","tokens_in":13999,"feed_emoji":"🔬","tokens_out":6501,"duration_ms":65181,"temperature":0.7,"pith_summary":"Atomic electron tomography (AET) reconstructs 3D atomic positions from a series of 2D electron-microscope images, but geometric tilt limits create a missing wedge of data that smears and distorts the result, especially at surfaces. This review argues that convolutional neural networks inserted into the reconstruction pipeline, whether as sinogram inpainting, tomogram refinement, or pre-reconstruction background removal, can recover much of that missing information and sharpen atomic maps to picometer precision. The payoff is concrete: in the showcase study, atom identification in a platinum nanoparticle rose from about 96.5% to 98.8%, coordinate precision improved from 26.1 pm to 15.1 pm, and the R-factor dropped from 19.2% to 17.4%. Because surface atomic arrangements control catalytic activity, adhesion, and corrosion resistance, making surface 3D structure reliably measurable would directly serve rational nanomaterial design.","feed_headline":"Neural nets fix the missing-wedge flaw in atomic tomography","feed_subtitle":"CNN refinement lifts atom detection to 98.8% and pinpoints surface atoms to ~15 pm in nanoparticles.","key_machinery":"The load-bearing mechanism is the atomicity prior encoded in a 3D U-Net: a trained network that maps blurred density distributions to well-separated Gaussian-like atomic peaks, acting as a post-reconstruction filter on tomograms produced by iterative algorithms such as GENFIRE. Alongside it, the review presents two complementary machinery pieces: a generative adversarial network that inpaints missing tilt angles directly in the sinogram domain to fill the missing wedge before reconstruction, and an ensemble cross U-Net transformer with attention in both encoder and decoder to capture long-range spatial dependencies and suppress residual artifacts. For supported nanoparticles, a CNN-based image inpainting step isolates the particle signal from the support background before reconstruction. Together these components carry the argument that the information lost to geometric and dose constraints is recoverable through learned priors.","core_discovery":"On the paper's own terms, the central claim is that a neural network trained with the atomicity constraint, the assumption that a sample is composed only of discrete atomic potentials, can transform a blurred, missing-wedge-distorted tomogram into a volume of well-isolated atomic peaks, and that it does so even for structures entirely different from its training set. Applying a 3D U-Net as a post-reconstruction augmentation step recovers low-coordination surface atoms that standard iterative reconstruction such as GENFIRE misses, and the resulting coordinates pass a consistency check via R-factor comparison with the original experimental tilt series. The same logic extends to the tilt series itself: a GAN-based two-step model fills the missing-wedge sinogram before reconstruction, coping with more than 80 percent missing tilt range, and a transformer-based ensemble model (EC-UNETR) further reduces root-mean-square error by 5.5 percent. In a separate move, CNN inpainting removes the support-material background from tilt series, enabling 3D atomic reconstruction of a supported palladium nanoparticle and revealing facet-dependent strain and disorder at the oxide interface. These advances are the basis for the review's conclusion that neural-network-assisted AET is becoming a data-driven platform for atomic-scale materials characterization.","pith_inferences":["The review does not test whether the atomicity prior invents peaks in genuinely amorphous samples; a check would be to run the same trained U-Net on simulated amorphous tomograms and count newly resolved atomic peaks where the ground-truth density is continuous.","The reported precision gain from 26.1 to 15.1 pm is measured against a known ground truth in simulation; an experimental cross-validation against an independent technique such as 4D-STEM ptychography on the same particle would separate true recovery from learned polishing.","Because the review notes surface atoms diffuse on $10^{-4}$ to $10^{-7}$ second timescales, the same networks could be repurposed as uncertainty quantifiers, flagging atoms whose positions vary between repeated tilt-series passes rather than reporting a single static structure.","If network refinement systematically biases toward Gaussian peak shapes, subtle anharmonic or delocalized electron density at defects may be smoothed away; comparing refined tomograms against multislice simulations of known defect configurations would reveal the bias."],"forward_implications":["Surface 3D atomic structures of nanoparticles, previously the weak point of AET due to missing-wedge elongation, can be determined at single-atom level, enabling facet-resolved strain mapping and interface analysis.","Neural-network-refined coordinates can be fed directly into density functional theory calculations, linking observed strain to catalytic activity, as demonstrated for oxygen reduction on strained platinum facets.","The missing-wedge problem loses much of its sting: with sinogram inpainting, reconstruction remains high-fidelity even when more than 80% of the tilt range is unavailable.","Supported catalysts become tractable, as CNN inpainting removes the support signal that otherwise swamps the nanoparticle.","Combined with 4D-STEM ptychography and multislice methods, the same neural-network tools are expected to extend atomic-resolution 3D imaging to light elements such as oxygen, carbon, and nitrogen."],"supporting_citations":[{"why":"Supplies the core demonstration: 3D U-Net augmentation of AET tomograms of a Pt nanoparticle, lifting atom detection from 96.5% to 98.8% and coordinate precision from 26.1 to 15.1 pm, plus the R-factor check.","marker":"[108]"},{"why":"Introduces the two-step GAN + U-Net pipeline that inpaints missing-wedge sinogram data and suppresses artifacts, the foundation for the missing-wedge recovery claim.","marker":"[132]"},{"why":"Extends the inpainting approach to 0.7 Å resolution electron tomography of nanoporous gold, supporting the sub-angstrom resolution claim.","marker":"[133]"},{"why":"Provides the EC-UNETR transformer model that reduces reconstruction RMSE by 5.5%, supporting the claim of further artifact reduction.","marker":"[134]"},{"why":"Demonstrates CNN image inpainting to remove support background, enabling 3D atomic reconstruction of a supported Pd nanoparticle.","marker":"[135]"},{"why":"Supplies GENFIRE, the iterative reconstruction algorithm whose raw tomograms the U-Net augmentation starts from.","marker":"[139]"},{"why":"Establishes atomic electron tomography at 19-pm precision without symmetry, the baseline that neural network methods advance.","marker":"[93]"},{"why":"Defines the 3D U-Net architecture used for tomogram augmentation.","marker":"[138]"}],"fun_headline_variants":["Neural nets conquer missing-wedge limits in atomic tomography","AI sharpens atomic electron tomography to picometer accuracy","Deep learning recovers surface atoms that standard tomography misses","Neural networks fix missing-wedge artifacts in atomic tomography"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The weakest load-bearing assumption is that the neural network's refined tomograms recover true atomic positions rather than imposing its learned atomicity picture, so the reported precision and R-factor gains reflect the real structure of a sample whose surface atoms may be diffusing on a $10^{-4}$ to $10^{-7}$ second timescale under an intense electron beam.","fun_headline_variants_meta":{"raw":{"variants":["Neural nets conquer missing-wedge limits in atomic tomography","AI sharpens atomic electron tomography to picometer accuracy","Deep learning recovers surface atoms that standard tomography misses","Neural networks fix missing-wedge artifacts in atomic tomography"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000484,"raw_usage":{"total_tokens":2383,"prompt_tokens":932,"completion_tokens":1451,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":1386}},"tokens_in":548,"tokens_out":1451,"duration_ms":12673,"temperature":1.0,"reasoning_tokens":1386,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:43:43.944759+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a tomogram of a known amorphous or highly disordered structure, such as a simulated metallic glass, run the trained U-Net augmentation, and count whether refined volumes show new well-resolved atomic peaks where the ground-truth density is continuous; a nonzero false-peak rate, or an R-factor that improves while the coordinates diverge from a simultaneously obtained ptychographic reconstruction, would indicate the network is imposing atomicity rather than recovering structure.","supporting_citations":[{"cited_title":"& Yang, Y","cited_arxiv_id":null,"evidence_quote":"Supplies the core demonstration: 3D U-Net augmentation of AET tomograms of a Pt nanoparticle, lifting atom detection from 96.5% to 98.8% and coordinate precision from 26.1 to 15.1 pm, plus the R-factor check."},{"cited_title":"& Xin, H","cited_arxiv_id":null,"evidence_quote":"Introduces the two-step GAN + U-Net pipeline that inpaints missing-wedge sinogram data and suppresses artifacts, the foundation for the missing-wedge recovery claim."},{"cited_title":"& Xin, H","cited_arxiv_id":null,"evidence_quote":"Extends the inpainting approach to 0.7 Å resolution electron tomography of nanoporous gold, supporting the sub-angstrom resolution claim."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the EC-UNETR transformer model that reduces reconstruction RMSE by 5.5%, supporting the claim of further artifact reduction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Demonstrates CNN image inpainting to remove support background, enabling 3D atomic reconstruction of a supported Pd nanoparticle."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies GENFIRE, the iterative reconstruction algorithm whose raw tomograms the U-Net augmentation starts from."},{"cited_title":"S., Brox, T","cited_arxiv_id":null,"evidence_quote":"Defines the 3D U-Net architecture used for tomogram augmentation."}],"review_version":1}