REVIEW 3 major objections 5 minor 10 references
Atom identification in bilayer moire materials with Gomb-Net
T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash
Pith's one-line read 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.
desk verdict A genuinely useful layer-separation tool for moiré STEM, with an underpowered physical conclusion about Se site preferences. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
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.
What would settle it
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.
Extended reading notes
Core claim
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.
Load-bearing premise
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.
Editorial extensions
If this is right
- 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.
Reading between the lines
- 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.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
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).
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 (3)
- [Fig. 4e, 'final step' paragraph] 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.
- [Methods (Data Generation), Discussion] 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.
- [Fig. 1e and 'To evaluate network performance' paragraph] 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.
minor comments (5)
- [SI Eq. (1)] 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.
- [Fig. 4d] The caption describes the operation as "U_W/X × U_X/W", while the text says "Euclidean norm"; please reconcile the notation.
- [Fig. 3f] Add error bars or confidence intervals to the lattice-constant-versus-stoichiometry plot; currently the stoichiometry dependence is presented without uncertainty.
- [Methods (Data Generation), Table S1] 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.
- [Fig. 4e caption] 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".
Circularity Check
No significant circularity: the core Gomb-Net derivation is self-contained, with only minor non-load-bearing self-citations.
full rationale
The central derivation is not circular. Gomb-Net is trained on simulated HAADF-STEM images with known ground-truth atom positions and evaluated on a held-out simulated test set, and its real-image performance is validated against the known C-C bond length in twisted bilayer graphene (mean 1.39 Å versus 1.42 Å), an external physical value. The Se moiré-site conclusion is not forced by the training labels: although the WS2-WS2(1-x)Se2x training data place Se atoms randomly on chalcogen sites, the network performs per-pixel and per-atom classification from image contrast, so a real moiré preference in the experimental image would still be detectable; the training distribution does not by construction produce the observed KDE match. The paper's own stated limitation, that performance depends on the similarity between simulated training data and real experimental data, is an acknowledged transfer concern, not a circularity. References 22 and 23 are self-citations from the same group used to provide synthesis context and to note consistency with prior PLD experiments on monolayer Janus conversion; these are corroborative and not load-bearing for Gomb-Net's methodological claims or the new moiré-site analysis. The main weakness is statistical, not circular: the no-preference claim in Fig. 4e rests on a pointwise MSE (4.02 × 10^-6) between KDEs without a null-distribution calibration, and the bootstrap in the SI (Kernel Density Estimate Bootstrapping) only characterizes the Se KDE's uncertainty rather than the MSE under random placement. This is an inference gap about the strength of evidence, not a reduction of the prediction to its inputs.
Assumptions & free parameters
free parameters (4)
- Gomb-Loss false-positive weighting coefficient (alpha) =
Not stated in main text; reported in Table S2
- Output gating threshold =
0
- KDE bandwidth =
0.1
- Average lattice parameter d used for moiré stacking normalization =
3.15 Å
assumptions (5)
- domain assumption HAADF-STEM contrast for a bilayer is a 2D projection of atomic scattering potentials, so out-of-plane layer ordering does not affect the image.
- ad hoc to paper Gaussian scattering potentials plus an Airy-disk probe with added Poisson and Gaussian noise yield training images sufficiently close to experimental HAADF-STEM images.
- domain assumption Vegard's law holds for the fractional Janus alloy when relating lattice constant to Se stoichiometry.
- standard math Gaussian KDE with bandwidth 0.1 and normalization to unit integral is a valid way to compare the moiré site and Se site distributions.
- domain assumption The order parameter construction from Maity et al. correctly labels moiré stacking configurations.
Cite this review
Pith. "Pith review of Atom identification in bilayer moire materials with Gomb-Net." pith.science (2026). https://pith.science/paper/PBUIRXJK
@misc{pith2026250209791,
author = {Pith},
title = {Pith review of: Atom identification in bilayer moire materials with Gomb-Net},
year = {2026},
howpublished = {\url{https://pith.science/paper/PBUIRXJK}},
note = {Machine review of arXiv:2502.09791}
}
read the original abstract
Moire patterns in van der Waals bilayer materials complicate the analysis of atomic-resolution images, hindering the atomic-scale insight typically attainable with scanning transmission electron microscopy. Here, we report a method to detect the positions and identities of atoms in each of the individual layers that compose twisted bilayer heterostructures. We developed a deep learning model, Gomb-Net, which identifies the coordinates and atomic species in each layer, effectively deconvoluting the moire pattern. This enables layer-specific mapping of quantities like strain and dopant distributions, unlike other commonly used segmentation models which struggle with moire-induced complexity. Using this approach, we explored the Se atom substitutional site distribution in a twisted fractional Janus WS2-WS2(1-x)Se2x heterostructure and found that layer-specific implantation sites are unaffected by the moire pattern's local energetic or electronic modulation. This advancement enables atom identification within material regimes where it was not possible before, opening new insights into previously inaccessible material physics.
Reference graph
Works this paper leans on
-
[1]
Hjorth Larsen, A. et al. The atomic simulation environment —a Python library for working with atoms. J Phys: Condens Matter 29, 273002 (2017). https://dx.doi.org/10.1088/1361- 648X/aa680e
doi:10.1088/1361- 2017
-
[2]
Maas, A. L., Hannun, A. Y . & Ng, A. Y . Rectifier Nonlinearities Improve Neural Network Acoustic Models. (2013)
work page 2013
-
[3]
Kingma, D. P. & Ba, J. L. Adam: A Method for Stochastic Optimization. ArXiv (2014). https://doi.org/10.48550/arXiv.1412.6980
-
[4]
Liaw, R. et al. Tune: A Research Platform for Distributed Model Selection and Training. ArXiv (2018). https://doi.org/10.48550/arXiv.1807.05118
-
[5]
Akiba, T. et al. Optuna: A Next-generation Hyperparameter Optimization Framework. ArXiv (2019) https://doi.org/10.48550/arXiv.1907.10902
-
[6]
Sudre, C. H. et al. Generalized Dice Overlap as a Deep Learning Loss Function for Highly Unbalanced Segmentations. ArXiv (2017). https://doi.org/10.48550/arXiv.1707.03237
-
[7]
Wang, H. et al. Ultraclean Suspended Graphene by Radiolysis of Adsorbed Water. Nano Lett 24, 8866-8871 (2024). https://doi.org/10.1021/acs.nanolett.4c01440
-
[8]
Wang, K. et al. Interlayer Coupling in Twisted WSe2/WS2 Bilayer Heterostructures Revealed by Optical Spectroscopy. ACS Nano 10, 6612– 6622 (2016). https://doi.org/10.1021/acsnano.6b01486
Show all 10 references
-
[9]
Lin, Y . C. et al. Low Energy Implantation into Transition-Metal Dichalcogenide Monolayers to Form Janus Structures. ACS Nano 14, 3896–3906 (2020). https://doi.org/10.1021/acsnano.9b10196
2020 doi
-
[10]
Harris, S. B. et al. Real -Time Diagnostics of 2D Crystal Transformations by Pulsed Laser Deposition: Controlled Synthesis of Janus WSSe Monolayers and Alloys. ACS Nano 17, 2472–2486 (2023). https://doi.org/10.1021/acsnano.2c09952
2023 doi
Reviewed August 7, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.