{"id":"4176cf1a-9667-427f-8b9f-74b1f8f0b19d","arxiv_id":"2502.09791","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Gomb-Net identifies atom positions, chemical species, and layer identity in twisted bilayer materials from HAADF-STEM images, and its application suggests selenium substitution sites are not affected by moiré stacking.","lead":"Researchers built a deep learning tool, Gomb-Net, that finds and labels individual atoms in each layer of twisted two-layer materials from electron microscope images. It could let scientists map strain, defects, and chemical substitutions in moiré materials, including a new finding that selenium atoms implant without preferring any particular moiré site.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The no-preference Se-site conclusion rests on a descriptive MSE with no null distribution; with only 333 atoms, a small MSE is not yet evidence.","rationale":"The reader's weakest assumption was simulation-to-real transfer, which is indeed important for the method's validity on experimental data. My concern is distinct and more specific: even granting perfect transfer, the paper's physical conclusion about moiré-site preference is not established because Fig. 4e compares two KDEs without a null distribution. The reader did note that 'no synthetic control with a known preference was run' and that the analysis lacked error bars, which is adjacent to my concern, but the missing permutation/null-distribution analysis is the sharper issue. The methodological core of the paper—the multi-branch U-Net with groupwise combinatorial loss and its simulated test accuracy—is not overturned by this concern. The real-image TBG C-C validation supports positional accuracy, and the Janus reconstruction is visually plausible. However, the central physical finding in the abstract is one of the paper's headline results, and it currently hinges on a descriptive comparison. Since the missing analysis is readily runnable from the public code and data, the appropriate status remains CONDITIONAL, matching the reader's verdict. I therefore recommend no change to the verdict, while emphasizing that the conditional should explicitly require the permutation test or an equivalent inferential check before the no-preference claim is accepted.","tokens_in":12180,"tokens_out":9843,"duration_ms":103582,"concrete_test":"Using the provided code and data, recompute Fig. 4e with a permutation test: (1) build a site-based null by extracting moiré-map values at all chalcogen-column positions identified by Gomb-Net in the six crops (not all pixels); (2) randomly assign 333 Se labels to these sites for 10,000 permutations, recomputing the density-normalized KDE and its MSE against the moiré-site KDE each time; (3) report the empirical p-value for the observed MSE of 4.02 × 10⁻⁶. If the observed MSE falls inside the bulk of the null distribution, the no-preference conclusion is supported; if it is an outlier, the conclusion fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's abstract-level physical claim—that Se atoms occupy chalcogen sites without preference for moiré lattice sites—is based entirely on Fig. 4e: the mean squared error (4.02 × 10⁻⁶) between the KDE of 333 Se-S column positions and the KDE of 1.3 million moiré-map values. This is a point statistic without an inferential calibration. The bootstrap procedure in the SI resamples the Se positions to characterize the green KDE's uncertainty, but it does not generate the distribution of the MSE under the null hypothesis that Se atoms are placed randomly on available chalcogen sites. A small observed MSE can arise from smoothing and small-sample noise even when no preference exists, and conversely a meaningful preference could be masked by KDE bandwidth (0.1) and by comparing against a pixel-area baseline rather than the discrete set of actual substitution sites. Because the no-preference claim is a central advertised result, this missing statistical test is load-bearing. The methodological contribution (Gomb-Net architecture and simulated accuracy) is not directly invalidated, but the headline physical finding is currently unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents Gomb-Net, a multi-branch U-Net with a groupwise combinatorial loss, for simultaneous atom localization, species identification, and layer assignment in HAADF-STEM images of twisted bilayer materials. The model is trained on simulated images for twisted bilayer graphene (TBG) and for a WS2-WS2(1-x)Se2x Janus heterostructure. On 800 simulated TBG test images it reports pixel accuracy 0.98 and mean IOU 0.74, outperforming a standard U-Net (0.86/0.39). On experimental TBG it extracts carbon positions with mean C-C distance 1.39 Å (2.1% from the known value). For the TMD case it finds 98% of Se atoms on the exposed layer and concludes from a KDE comparison that Se occupies chalcogen sites without preference for moiré lattice sites (MSE 4.02 × 10⁻⁶).","tokens_in":12387,"tokens_out":4265,"duration_ms":42814,"significance":"If the quantitative claims hold, Gomb-Net is a useful methodological advance: it would enable layer-specific strain, defect, and dopant mapping in moiré systems that standard segmentation models cannot handle. Strengths include the reproducible code and data links, the clear comparison against U-Net baselines, the use of a known physical bond length as a sanity check on real data, and an explicit statement of the simulation-transfer limitation. The main physical conclusion, however, is currently supported only by a descriptive comparison without inferential calibration, and the simulation-to-experiment transfer is not quantitatively validated. The methodological core is defensible, but the headline finding needs additional statistical work before it can be accepted.","major_comments":[{"comment":"The no-preference conclusion is based solely on the point estimate MSE = 4.02 × 10⁻⁶ between two KDEs, with no null distribution. Because the green KDE is computed from only 333 Se-S columns while the purple KDE is computed from 1.3 million moiré-map values, and because KDE smoothing (bandwidth 0.1) will tend to suppress differences, a small MSE is expected even under random placement. The bootstrap in the SI characterizes uncertainty in the KDE estimate but does not simulate the null hypothesis of random Se placement on chalcogen sites. I request a permutation or parametric null test (e.g., randomly assign the 333 Se atoms to detected chalcogen columns, recompute the MSE, and report the observed MSE percentile), together with a comparison against the discrete distribution of available substitution sites rather than a pixel-area baseline. Without this, the advertised physical conclusion is unsupported.","section":"Fig. 4e, 'final step' paragraph"},{"comment":"The transfer of a network trained only on Gaussian-potential/Airy-disk simulations to experimental HAADF-STEM images is load-bearing for both the TBG C-C validation and the Se-site analysis, yet no quantitative similarity measure or sensitivity analysis is provided. The authors correctly state in the discussion that \"the key factor and limitation ... is the degree of similarity between the training dataset and the real experimental data,\" but this limitation is not addressed by experiment. I suggest adding a small-scale validation on experimental data with known labels (e.g., using an independent method such as multi-slice simulation or ptychographic reconstruction for a subregion), or reporting performance degradation as simulation parameters (probe size, noise, contamination) are varied around the chosen values. The current qualitative agreement in Fig. 3b is suggestive but not quantitative.","section":"Methods (Data Generation), Discussion"},{"comment":"The headline metrics (pixel accuracy 0.98, IOU 0.74) are aggregate over all classes and layers; they do not establish that the rare and physically important Se-S columns are detected reliably. Since the central application counts Se-S columns and maps their positions, please report per-class precision, recall, and IOU for the six classes in the TMD test set, along with the false-positive rate for Se atoms in the wrong layer (currently quoted only as \"on the order of 0.5%\" without a confidence interval or the number of test images/atoms it is based on). This is needed to assess whether a few misclassified Se atoms could bias the stacking-site distribution.","section":"Fig. 1e and 'To evaluate network performance' paragraph"}],"minor_comments":[{"comment":"The displayed Gomb-Loss formula is typeset in a way that is difficult to parse; please provide a clean equation with all symbols defined, including the meaning of the numerator term.","section":"SI Eq. (1)"},{"comment":"The caption describes the operation as \"U_W/X × U_X/W\", while the text says \"Euclidean norm\"; please reconcile the notation.","section":"Fig. 4d"},{"comment":"Add error bars or confidence intervals to the lattice-constant-versus-stoichiometry plot; currently the stoichiometry dependence is presented without uncertainty.","section":"Fig. 3f"},{"comment":"Please provide the actual ranges of the varied parameters (e.g., twist angle range, vacancy counts, phonon sigma values) rather than only distribution names, so that the training-data diversity is reproducible.","section":"Methods (Data Generation), Table S1"},{"comment":"The phrase \"mean KDE of each distribution\" is imprecise; a KDE is a density estimator, not a distribution. Consider rephrasing to \"mean of the KDE estimates\".","section":"Fig. 4e caption"}],"recommendation":"major_revision","confidential_remarks":"The central technical contribution appears sound and the code/data links are a real strength; the main risk is overinterpretation of the Se-site result. I would ask the authors to add the permutation null test and per-class metrics before acceptance, and to verify that the Google Drive data link remains stable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: Gomb-Net is a real and useful tool. The Se-site 'no preference' result is not backed by the statistics as written.\n\nWhat's new: the dual-decoder U-Net with the groupwise combinatorial loss does something the standard U-Net doesn't—it separates the two layers of a moiré bilayer and assigns atoms to layers. The simulated test numbers (0.98 pixel accuracy, 0.74 IOU) are decent, and the real TBG validation is credible: predicted C-C distances average 1.39 Å, within 2.1% of the known value. The reconstructed image from the TMD predictions matches the raw data well, which is a good sign for sim-to-real transfer. Code and data are public. That's a solid methodological contribution.\n\nThe soft spots are concentrated in the physical claim. The paper's main result—Se atoms occupy chalcogen sites without preference for moiré lattice sites—rests entirely on an MSE of 4.02e-6 between the KDE of 333 Se-S column positions and the KDE of 1.3 million moiré map values. There is no null distribution. Bootstrapping the Se positions tells you how uncertain the green KDE is, but it doesn't tell you what MSE to expect under random placement. With 333 points and a KDE bandwidth of 0.1, a small MSE can happen trivially. You'd need a permutation test or a synthetic control with a known preference to calibrate that number. As it stands, the no-preference conclusion is underpowered. The paper also has an inconsistency in how the stacking map is defined (text says Euclidean norm, Fig. 4c caption says product), and the 2% bottom-layer Se fraction lacks error bars; the 0.5% false-positive rate is quoted but not propagated into the site analysis. All this comes from six crops of a single image, so the physical generality is limited.\n\nThe stress-test note is right: the load-bearing physical finding is not supported. The methodological contribution survives it. The authors are honest about the sim-to-real limitation and the possibility that some bottom-layer Se atoms are false positives. That helps.\n\nWho is this for? People doing atomic-scale STEM of twisted bilayers who need layer-resolved atom finding. It will probably be useful despite the weak physical conclusion.\n\nMy recommendation: send it to peer review. A serious referee can push for a proper statistical test on the Se-site claim, or the authors can soften the claim to what the data actually support. The tool itself deserves publication; the headline result needs work.","headline":"A genuinely useful layer-separation tool for moiré STEM, with an underpowered physical conclusion about Se site preferences.","tokens_in":12936,"tokens_out":2900,"would_cite":true,"duration_ms":27667,"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":"Gomb-Net is a multi-branch U-Net with a groupwise combinatorial loss that identifies atomic positions, species, and layer identity in twisted bilayer moiré materials from HAADF-STEM images.","keywords":["moiré materials","HAADF-STEM","atom identification","deep learning segmentation","twisted bilayer graphene","transition metal dichalcogenides","Janus heterostructure","dopant site mapping"],"falsifier":"Take a real HAADF-STEM image of a twisted bilayer whose per-layer atomic species have been established independently, for example by multi-slice ptychographic reconstruction, run a simulation-trained Gomb-Net on it, and compare layer labels; if the per-layer species assignments disagree with the independent reconstruction at a rate far above the roughly 0.5% false-positive error the paper reports for simulated data, the core transfer claim collapses.","tokens_in":11937,"feed_emoji":"🔬","tokens_out":6801,"duration_ms":69129,"temperature":0.7,"pith_summary":"The paper introduces Gomb-Net, a deep-learning model that takes atomic-resolution HAADF-STEM images of twisted bilayer materials and returns, for every atom, a position, an atomic species, and a layer label, effectively undoing the moiré interference that hides the two monolayers inside a single image. If correct, this makes it possible to measure strain, defects, and dopant distributions separately for each layer of a twisted heterostructure, which standard segmentation models cannot do because moiré overlap changes the expected Z-contrast of identical atomic species. The authors demonstrate the method on simulated and experimental twisted bilayer graphene, reaching pixel-wise accuracy of 0.98 in simulation and recovering a C-C bond length of 1.39 Å, which is 2.11% from the accepted 1.42 Å value, on a real image. Applied to a twisted WS2-WS2(1-x)Se2x fractional Janus bilayer, the model finds that Se atoms substitute into chalcogen sites of the exposed layer without preferring any particular moiré stacking site.","feed_headline":"Neural network un-mixes moiré images, atom by atom","feed_subtitle":"Layer-resolved atom maps from HAADF-STEM show where dopants sit in twisted bilayers, with no moiré preference for Se.","key_machinery":"Gomb-Net's load-bearing mechanism is a change to the U-Net decoder and loss function. The architecture routes the shared encoder's bottleneck feature maps into two parallel decoder branches, one assigned to each layer, so the network must factor an image of superimposed lattices into two single-layer segmentations. The groupwise combinatorial loss then measures Dice loss between every pairing of predicted and target layer masks, averaging through reciprocals so the ordering that matches reality dominates, and a numerator term penalizes the two branches collapsing onto identical outputs. This combination of branch specialization and an order-agnostic, false-positive-penalizing loss is what separates atoms by layer rather than by brightness alone.","core_discovery":"The central claim is that a multi-branch U-Net trained with a groupwise combinatorial loss can deconvolute the moiré pattern and correctly identify atoms in each layer of twisted bilayer heterostructures from HAADF-STEM images. The paper argues that the physical symmetry of the imaging process, namely that the out-of-plane layer order does not affect the projected image under kinematic scattering, should be built into the training objective, so Gomb-Loss compares outputs to targets under both possible layer orderings and takes a harmonic-mean-like combination, while the two decoder branches specialize on the two layers. On 800 simulated test images of twisted bilayer graphene, Gomb-Net achieves 0.98 pixel-wise accuracy and 0.74 mean intersection-over-union versus at most 0.86 and 0.39 for the standard U-Net variants. On experimental data, the model finds per-layer carbon positions whose C-C distance distribution peaks at 1.39 Å, and in the fractional Janus bilayer it locates Se-S columns that are 98% on the plume-exposed layer and uniformly distributed across moiré stacking order values.","pith_inferences":["An untested but direct consequence is that the architecture scales structurally to trilayer or more complex moiré stacks by adding decoder branches and loss terms, although the paper does not assess that regime.","The random-looking Se distribution admits two physical explanations the authors leave open, namely energy thresholds that are flat compared with the plume energy or post-implantation diffusion enabled by implantation-induced defects, and a discriminating experiment would measure Se concentration versus plume kinetic energy and sample temperature.","A practical extension would be to use the reconstructed-image agreement as an online confidence metric, flagging regions where the network's per-layer output fails to reproduce the experimental contrast and thereby indicating a need for retraining or simulation refinement.","Because the training images are generated with Gaussian scattering potentials and an Airy-disk probe, the method's transferability suggests even a coarse forward model contains enough physics for layer assignment, so one could probe how much accuracy degrades as those approximations are made coarser."],"forward_implications":["Layer-resolved strain and dopant maps can be computed directly from Gomb-Net coordinates, revealing how the moiré environment modifies each monolayer's local structure.","The same workflow of simulation training, real-image prediction, and blob center-of-mass localization transfers to a harder material class, a two-lattice-constant twisted TMD, needing only a retrained six-class model.","Because predictions run in milliseconds and training takes minutes on a personal computer, the method is compatible with real-time, automated, or autonomous STEM analysis.","The fractional Janus measurement licenses statistical claims about doping-site selection: under the tested pulsed-laser-deposition conditions, Se occupies chalcogen sites with no moiré-lattice preference, and about 2% of bottom-layer chalcogen sites also acquire Se."],"supporting_citations":[{"why":"Supplies the U-Net segmentation architecture that Gomb-Net extends with a second decoder branch and a custom loss.","marker":"[13]"},{"why":"Supplies the conventional semantic-segmentation baseline whose per-layer outputs are compared with Gomb-Net in the accuracy tests.","marker":"[6]"},{"why":"Supplies the dual-decoder U-Net idea adapted here for per-layer segmentation of bilayer images.","marker":"[19]"},{"why":"Establishes the low-energy Se implantation route that produces the fractional Janus WS2-WS2(1-x)Se2x sample.","marker":"[22]"},{"why":"Provides the prior monolayer result that Se first enters the topmost chalcogen sublayer, which the bilayer Se distribution is compared against.","marker":"[23]"},{"why":"Supplies the order-parameter vector used to map local moiré stacking order and to associate Se sites with moiré positions.","marker":"[27]"},{"why":"Provides the known 1.42 Å C-C bond distance used to validate the predicted carbon positions in twisted bilayer graphene.","marker":"[20]"},{"why":"Provides the heterostrain magnitude cited to place the measured 1.39 Å C-C distance within experimental uncertainty.","marker":"[21]"}],"fun_headline_variants":["AI sees through moiré to map atoms layer by layer","Deep learning untangles twisted bilayer atom positions","Gomb-Net: neural network resolves atoms in moiré layers","Unmasking moiré: AI identifies atoms in each crystalline layer","Neural net deconvolutes moiré for layer-resolved atomic maps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results depend on Gomb-Net transferring from simulated images to real microscope images: if the way the training images were generated does not match how actual HAADF-STEM images look, the reported per-layer assignments would not be reliable.","fun_headline_variants_meta":{"raw":{"variants":["AI sees through moiré to map atoms layer by layer","Deep learning untangles twisted bilayer atom positions","Gomb-Net: neural network resolves atoms in moiré layers","Unmasking moiré: AI identifies atoms in each crystalline layer","Neural net deconvolutes moiré for layer-resolved atomic maps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000566,"raw_usage":{"total_tokens":2688,"prompt_tokens":955,"completion_tokens":1733,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":571,"completion_tokens_details":{"reasoning_tokens":1643}},"tokens_in":571,"tokens_out":1733,"duration_ms":10507,"temperature":1.0,"reasoning_tokens":1643,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T20:27:56.251031+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a real HAADF-STEM image of a twisted bilayer whose per-layer atomic species have been established independently, for example by multi-slice ptychographic reconstruction, run a simulation-trained Gomb-Net on it, and compare layer labels; if the per-layer species assignments disagree with the independent reconstruction at a rate far above the roughly 0.5% false-positive error the paper reports for simulated data, the core transfer claim collapses.","supporting_citations":[],"review_version":1}