{"id":"02f3d723-5df0-490c-bf9f-392e8b92bf3a","arxiv_id":"1908.07959","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A regressional GAN generates graphene/BN hybrid structures whose DFT bandgaps are within about 10% fractional mean absolute error of user-specified target bandgaps.","lead":"This paper trains a generative adversarial network to design graphene/boron nitride hybrid structures that match a requested electronic bandgap, then checks the designs with quantum simulation. It offers a step toward replacing trial-and-error materials search with automated inverse design.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper never specifies how gradients pass through the -1/1 binarization before the regressor; absent a straight-through estimator or relaxation, Eq. 3's regularization term has zero gradient and cannot steer the generator toward the target bandgap.","rationale":"I read the paper as claiming that the regressor loss in Eq. 3 is the mechanism that associates generated structures with continuous bandgap labels. The DFT validation and the PCA/MLLE analyses are real, independent evidence that the trained generator produces plausible structures, and I credit that. The weakest point is internal to the method: the thresholding step described in Supplementary Note 1 has no specified gradient path. Without a straight-through estimator or a continuous relaxation, the exact training procedure in the text cannot transmit the bandgap error back to the generator. The reader's weakest_assumption identified the same issue, and I agree with that assessment. The concern does not by itself disprove the reported results, because the authors may have used an undocumented gradient approximation, but the manuscript as written is incomplete on the point that its central claim depends on. For this reason I would keep the conditional verdict: the authors should disclose or demonstrate the gradient mechanism, and ideally release code, before the claim is accepted as fully supported.","tokens_in":12686,"tokens_out":3205,"duration_ms":36768,"concrete_test":"Reconstruct the described TensorFlow 1.14 training graph exactly: generator output via tanh, threshold at zero to -1/1, then feed the thresholded matrix to the regressor. Compute the gradient of the second term in Eq. 3 with respect to the generator parameters on a batch of 105 samples. If the gradient norm is zero, the paper's stated training procedure cannot steer generation by bandgap, and the authors should either disclose the straight-through estimator or retrain with one. If nonzero, identify which operation provides the gradient and add that detail to the manuscript.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the RGAN autonomously generates graphene/BN hybrids with DFT-validated bandgaps near desired labels via the regressor loss in Eq. 3. Supporting Information Note 1 states that after the generator's tanh layer outputs continuous values in [-1, 1], values below 0 are replaced by -1 and values above 0 by 1 before the structure enters the regressor. This replacement is piecewise-constant with zero derivative almost everywhere. The paper does not mention a straight-through estimator, Gumbel relaxation, or any other mechanism for differentiating through this discrete mapping. Consequently, as written, the gradient of the second term in Eq. 3 with respect to the generator parameters is zero almost everywhere, so the regularization term cannot update the generator. If this is true, the apparent conditioning on Y would have to arise entirely from the discriminator path, which is not demonstrated. The absence of code or training details makes this unresolved. This is load-bearing because if the regressor gradient is zero, the proposed 'regressional' mechanism is not what produces the reported accuracy; the central claim is then unsupported rather than merely under-validated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents a 'regressional GAN' (RGAN) that couples a least-squares GAN with a pre-trained CNN regressor to generate graphene/boron-nitride hybrid structures (4x4 and 5x5 supercells) with specified bandgaps. The generator receives noise and a desired bandgap; the regressor predicts the bandgap of the generated structure and provides latent features that, concatenated with the bandgap, are fed to the discriminator. The generator's loss combines the LSGAN adversarial term with the regressor's L2 error (Eq. 3). The authors report DFT-validated bandgaps of generated structures with MAEF of 9.67% (4x4) and 11.6% (5x5), and use PCA/MLLE to suggest the latent space is shared between generated and real structures.","tokens_in":12889,"tokens_out":4307,"duration_ms":39249,"significance":"If the training mechanism is sound, this is a valuable contribution: it demonstrates inverse design for a continuous property in a 2D materials system with external DFT validation, and the use of regressor latent features to condition the discriminator is a plausible architecture that could transfer to other structure-property problems. However, the central mechanism for passing gradients through the discrete structure encoding is not specified, and the validation lacks baselines and uncertainty estimates. The latent-space analysis is partly circular. These concerns are fixable in revision but currently limit confidence in the main claim.","major_comments":[{"comment":"The manuscript does not specify how gradients flow from the regressor's L2 loss in Eq. (3) back to the generator through the binarization step described in Supplementary Note 1. After the tanh layer, values below 0 are set to -1 and above 0 to 1; this piecewise-constant mapping has zero derivative almost everywhere. Unless a straight-through estimator, Gumbel-softmax relaxation, or similar mechanism is used, the regularization term in Eq. (3) provides no gradient to the generator, so the reported conditioning on Y would have to arise from the discriminator path alone, which is not demonstrated. The authors should state the exact gradient mechanism and provide evidence (e.g., a training curve showing the regressor loss on generated structures decreasing, or an ablation with the regressor term removed).","section":"Supporting Information Note 1; Eq. (3)"},{"comment":"The central quantitative claim (MAEF of 9.67% for 4x4 and 11.6% for 5x5) is based on an unreported number of DFT-validated generated structures, with no confidence intervals or error bars. In addition, no baseline or comparison experiments are reported (e.g., random structure generation, unconditional GAN, or direct optimization with the regressor alone), so it is unclear whether the RGAN's accuracy exceeds what a simpler method would achieve. The authors should specify N for each validation, report the distribution across desired bandgaps, and compare against at least one baseline.","section":"Results and Discussion, Fig. 3"},{"comment":"The PCA and MLLE analyses of latent features are partially circular: the regressor used to encode both real and generated structures was trained on real structures, and the discriminator receives exactly these latent features (concatenated with labels). The overlap in Fig. 4a/e therefore partly reflects the objective the generator was optimized against, not necessarily an independent discovery of physical structure-property relations. The claim that the generator 'caught the physical law' (Fig. 4b-c) should be softened or supported by an analysis on a feature representation that was not part of the training loss.","section":"Results and Discussion, Fig. 4"}],"minor_comments":[{"comment":"'Principle component analysis' should be 'principal component analysis' throughout.","section":"Abstract and Fig. 4 caption"},{"comment":"The shorthand 'MAEF' is used in the main text before it is defined in Supplementary Note 2; please define at first use.","section":"Results and Discussion, Fig. 3"},{"comment":"The notation for π, E, and the subscripted expectations is garbled in the submitted PDF; please ensure all symbols are defined and rendered correctly.","section":"Eqs. (3) and (4)"},{"comment":"Reference [32] (Szegedy et al.) is cited as 'Inception V2' in the text, but the reference is to the GoogLeNet/Inception v1 paper; please correct the version name and citation.","section":"References"},{"comment":"Hyperparameters (e.g., λ=25, learning rate 1e-4, L2 regularization 0.5, batch size 105) are scattered in the text; a summary table would improve reproducibility.","section":"Supporting Information Note 1"}],"recommendation":"major_revision","confidential_remarks":"The binarization gradient issue is the key gate. If the authors can clarify that they used a straight-through estimator or an equivalent differentiable approximation and show an ablation, the paper could be acceptable. Without that, the reported accuracy cannot be attributed to the regressor loss. The paper would also benefit from a baseline comparison and larger DFT validation set, but these are secondary."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here is my read. The paper does something useful: it demonstrates a GAN variant that can generate graphene/BN hybrids with DFT-validated bandgaps close to continuous target values, ~9.7% and ~11.6% mean absolute fractional error on 4x4 and 5x5 supercells. That external validation is the best part, and the 5x5 results with a 0.24% training sample suggest the approach scales. The architecture is a genuine twist on conditional GAN: the regressor outputs both a predicted bandgap and a latent feature vector; the latent vector is concatenated with the continuous label and fed to the discriminator. That is distinct from ACGAN and reg-GAN, and the idea is reasonable.\n\nBut the paper has a load-bearing gap that the stress-test correctly identified. Eq. 3 adds a regressor L2 term to the generator loss. Supplementary Note 1 says the generator output is binarized to -1/1 before entering the regressor. That step has zero gradient almost everywhere, and the paper never mentions a straight-through estimator or relaxation. So as written, the regularization term cannot update the generator. The conditioning might still work through the discriminator path, which sees the latent features and the target label, but the paper does not demonstrate that or provide an ablation. The claim that the RGAN's regressive loss is what steers generation is therefore unsupported on paper.\n\nOther issues are smaller. There is no baseline comparison with ACGAN, reg-GAN, or a plain CGAN, so we cannot tell whether this architecture is necessary or just adequate. Only a few target bandgaps (1.0–2.0 eV) are DFT-validated, so 'any bandgap' is an overstatement. There are no uncertainty estimates for the MAEF values. Code and data are not released, so the gradient question cannot be checked. The PCA/MLLE analysis is partly circular: the latent features come from the same regressor the generator was trained against, so overlap between real and generated features is somewhat expected.\n\nMy verdict: this deserves a serious referee. The application is timely and the external DFT validation is real evidence, not just regressor self-consistency. The missing gradient specification is fixable in revision (either explain the straight-through estimator, remove the Eq. 3 term if it is inert, or show the discriminator path alone suffices). A revision that adds a baseline and releases code would make this a solid contribution.","headline":"Useful inverse-design demonstration with genuine external DFT checks, but the paper never explains how the regressor loss backpropagates through binarized structures—so the advertised mechanism is unsupported as written.","tokens_in":13451,"tokens_out":4043,"would_cite":false,"duration_ms":38823,"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":"A regressional GAN can design graphene/boron-nitride hybrid structures with requested bandgaps, with DFT-validated errors near 10%.","keywords":["inverse design","regressional GAN","generative adversarial network","graphene/boron nitride hybrids","bandgap prediction","latent features","two-dimensional materials","density functional theory"],"falsifier":"Take the trained RGAN, request a batch of structures at a chosen bandgap, recompute their bandgaps with DFT, and compare the fractional mean absolute error; if it is not near the reported 10% level or the structures collapse to a few duplicate layouts, the inverse-design claim fails.","tokens_in":12465,"feed_emoji":"⚛️","tokens_out":10863,"duration_ms":100186,"temperature":0.7,"pith_summary":"The paper tries to flip the usual materials-design workflow: rather than proposing a structure and then computing its bandgap, it trains a network that takes a requested bandgap as input and outputs graphene/boron-nitride hybrid structures expected to realize it. To do this it introduces the regressional GAN (RGAN), a GAN augmented with a supervised convolutional regressor that predicts bandgaps and extracts latent features. Those latent features are concatenated with the bandgap and fed to the discriminator, which lets the generator be conditioned on a continuous numerical property instead of a discrete class. DFT calculations on the generated structures show bandgaps within about 10% of the requested values, and the proposed structures are diverse rather than repetitions of the training set.","feed_headline":"GAN generates graphene/BN hybrids at requested bandgaps","feed_subtitle":"DFT checks confirm generated structures land within about 10% of target bandgaps, without brute-force searching.","key_machinery":"The load-bearing mechanism is the regressor's latent-feature vector $\\pi$. After the regressor, a modified Inception V2 convolutional network, processes either a generated or a real structure, $\\pi$ is concatenated with the corresponding bandgap to form the discriminator input, so the discriminator judges authenticity and label consistency together. The generator loss adds a regularizer $\\lambda = 25$ times the L2 error between the predicted bandgap of the generated structure and the requested bandgap to the least-squares GAN loss; this term is what steers the generator toward the target property. The generator's output matrices are binarized to $-1$ (C-C) and $+1$ (BN) before they enter the regressor.","core_discovery":"The central claim is that an RGAN can autonomously perform inverse structural design of graphene/BN hybrids for requested continuous bandgap values. The regressor, trained on DFT-computed real structures, outputs both a predicted bandgap and a latent feature vector; the generator's objective combines a least-squares adversarial term with an L2 regularization term that penalizes the difference between the regressor's predicted bandgap for the generated structure and the desired bandgap. After the generator's continuous outputs are binarized to $-1$/$+1$ atomic labels, DFT validation on the generated $4\\times4$ and $5\\times5$ structures yields fractional mean absolute errors of 9.67% and 11.6% respectively, with $R^2$ values of 0.82 and 0.87. PCA and MLLE of the latent features show that generated structures follow the statistical distribution of the real structures and trace the physical trend that bandgap rises with BN concentration.","pith_inferences":["If the latent axes really encode physical descriptors, then moving along the PCA or MLLE directions could let a user steer correlated properties such as BN concentration or domain morphology, though the paper does not test this.","The same RGAN recipe should transfer to other continuous labels such as formation energy, magnetic moment, or work function whenever a regressor of comparable accuracy can be trained; the paper only demonstrates bandgap.","The regressor's accuracy is reported on real test structures, so a conservative check before using the RGAN for discovery would probe generated structures outside the training bandgap range, where the regressor's predictions are extrapolation."],"forward_implications":["For a requested bandgap inside the trained range, the generator returns multiple distinct graphene/BN topographies, so inverse design can enumerate candidate structures rather than a single guess.","The same architecture can be retrained for other layered or 2D material families whenever a fast differentiable property predictor and a labeled structure dataset are available.","The $5\\times5$ case works with only 0.24% of all possible configurations sampled, so the approach scales to chemical spaces too large for exhaustive DFT screening.","The latent-space analysis indicates the generator has internalized the structure-property relationship, meaning the learned latent axes can be inspected to extract design rules beyond the training labels."],"supporting_citations":[{"why":"It supplies the DFT-computed graphene/BN dataset, the -1/1 structure encoding, and the CNN bandgap predictor that the RGAN's regressor is built on.","marker":"18"},{"why":"It documents the prior GAN inverse design of metasurfaces with large deviations, the baseline the RGAN's continuous-label conditioning is meant to improve.","marker":"22"},{"why":"It introduces conditional GANs, the discrete-label conditioning scheme that the RGAN extends to continuous quantitative labels.","marker":"24"},{"why":"It defines the auxiliary classifier GAN whose discrete-label limitation motivates feeding latent features plus bandgaps to the discriminator.","marker":"26"},{"why":"It presents the earlier reg-GAN for continuous labels that the RGAN distinguishes itself from by keeping the discriminator's automatic authentication.","marker":"27"},{"why":"It establishes the original generative adversarial network framework that provides the generator-discriminator training paradigm.","marker":"29"},{"why":"It supplies the Inception V2 architecture from which the regressor is modified, including the layer that yields the latent features.","marker":"32"},{"why":"It provides the least-squares GAN losses used as the base objective for the RGAN's generator and discriminator.","marker":"34"},{"why":"It supplies the PCA method used to show that generated latent features overlap the real training distribution and correlate with bandgap.","marker":"35"},{"why":"It supplies the MLLE method used to show that generated and real structures lie on the same low-dimensional manifold.","marker":"36"}],"fun_headline_variants":["RGAN designs graphene/BN hybrids with target bandgaps","AI designs custom-bandgap 2D materials without brute force","Regressional GAN crafts graphene/BN hybrids to order","Inverse design of 2D hybrids via regressional GAN","Machine learning generates 2D materials with desired bandgaps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The target-steering loop depends on the bandgap error being able to guide the discrete -1/1 layout of the generated structure through the rounding step that turns continuous outputs into atomic labels, and the paper does not describe how the gradient crosses that step.","fun_headline_variants_meta":{"raw":{"variants":["RGAN designs graphene/BN hybrids with target bandgaps","AI designs custom-bandgap 2D materials without brute force","Regressional GAN crafts graphene/BN hybrids to order","Inverse design of 2D hybrids via regressional GAN","Machine learning generates 2D materials with desired bandgaps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00087,"raw_usage":{"total_tokens":3800,"prompt_tokens":1006,"completion_tokens":2794,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":622,"completion_tokens_details":{"reasoning_tokens":2706}},"tokens_in":622,"tokens_out":2794,"duration_ms":17643,"temperature":1.0,"reasoning_tokens":2706,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:52:54.145636+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the trained RGAN, request a batch of structures at a chosen bandgap, recompute their bandgaps with DFT, and compare the fractional mean absolute error; if it is not near the reported 10% level or the structures collapse to a few duplicate layouts, the inverse-design claim fails.","supporting_citations":[],"review_version":1}