{"id":"415e521a-2c0f-4f60-8315-928993d96b43","arxiv_id":"2607.18403","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A single equivariant graph neural network generates spin-orbit coupled Wannier Hamiltonians for solids across 69 elements, backed by a post-processing package for subspace projection and linear-scaling spectral calculations.","lead":"This paper introduces G(Wa)NN, a graph neural network that predicts spin-orbit coupled electronic Hamiltonians directly in an orthogonal Wannier basis, trained on more than 111,000 Wannier Hamiltonians spanning 69 elements. It also presents Tailwater, a Python package for projecting those Hamiltonians into low-energy subspaces and computing spectra with linear-scaling Kernel Polynomial methods.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Graph cutoff for inter-site Hamiltonian blocks is never stated; without a locality or convergence analysis, the sparse Hamiltonian could omit long-range hoppings and invalidate predicted spectra and transport properties.","rationale":"I read the manuscript in good faith. The central claim is that G(Wa)NN can directly generate accurate, sparse, spin-orbit coupled Wannier Hamiltonians for arbitrary solid-state systems, enabling transport simulations at 10K+ atoms. For this claim to hold, the predicted Hamiltonian must faithfully represent the target Wannier Hamiltonian, and the sparsity must be physically justified so that omitted entries are negligible. The manuscript never states the graph cutoff radius, and more importantly, provides no convergence analysis demonstrating that the chosen neighbor range captures all significant Wannier hoppings. This is the most load-bearing concern because it directly affects the correctness of the predicted Hamiltonian and the scalability claims. If the cutoff is too short, the predicted band structures, gaps, and topological invariants will be systematically wrong, and the linear-scaling sparse matrix operations will be operating on an incorrect Hamiltonian. Alternative concerns—such as validation on the same GNOME database, lack of public model weights, or the 'first' claim vs. Ref. [30]—are important for generalization and reproducibility, but they are secondary to whether the model actually produces correct Hamiltonians. The reader's weakest_assumption identified exactly this issue, and I agree. The proposed concrete test—measuring hopping decay and convergence with cutoff—would settle whether the concern lands. I therefore recommend keeping the verdict CONDITIONAL (UNCHANGED) until the cutoff is specified and a convergence study is provided.","tokens_in":17916,"tokens_out":5063,"duration_ms":44492,"concrete_test":"Select ~20 diverse test materials (metals, small-gap semiconductors, topological insulators such as Bi2Se3). For each, compute the ground-truth Wannier Hamiltonian with Wannier90. Measure the Frobenius norm of inter-site blocks versus interatomic distance, and compute band structures truncated at cutoffs of 4, 5, 6, 8, 10, and 12 Å. Then run the trained G(Wa)NN using identical graph cutoffs and compare predicted band structures to the DFT/truncated references. If the band-structure MAE (and band-gap error, Z2 invariant) does not converge by the model's default cutoff, the locality assumption is violated. The paper or code should report the actual default cutoff.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III introduces the model as predicting inter-site blocks Hij(R) per directed edge between atom i and a periodic image of atom j, but the graph construction—specifically the cutoff radius—is never reported. The Wannier functions are checked to have spread < 1.5×|v_max| (Sec. II), but that bound does not imply hopping amplitudes have decayed to negligible values at any given neighbor distance. If the graph cutoff is too small, hoppings beyond it are omitted, introducing systematic errors not captured by the reported DOS/band-gap metrics, which are averaged over many states and may be insensitive to small gaps or topological features. The linear-scaling and 10K+ atom claims in the abstract and Sec. V rely on the Hamiltonian being sparse, which is only physically justified if the cutoff is converged. No convergence study with respect to cutoff is presented, and no explicit cutoff is given, so the central claim of generating accurate Wannier Hamiltonians for large systems is not yet supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces G(Wa)NN, an E(3)-equivariant graph neural network that predicts spin-orbit coupled electronic Hamiltonians in an orthogonal Wannier basis directly from crystal structures. The model is trained on more than 111,000 Wannier90-generated Hamiltonians spanning 69 elements, drawn from MP-Traj, C2DB, GNOME, and Alexandria. The authors report a Hamiltonian matrix-element MAE of 10 meV, a normalized DOS MAE of 0.04 on a 2000-material GNOME validation set, a band-gap MAE of 0.08 eV, and successful reproduction of the Z2=1 topological phase and surface Dirac cone of Bi2Se3. They also present a local finetuning workflow for WSe2 bilayers and introduce Tailwater, a Python package for subspace projection and kernel-polynomial-method (KPM) spectral computations. The central claim is that the orthogonal, sparse Wannier representation enables linear-scaling inference and transport simulations at scales of 10K+ atoms, which would be unprecedented for machine-learned electronic-structure models.","tokens_in":18221,"tokens_out":4628,"duration_ms":41838,"significance":"If the central claims are substantiated, this is a significant contribution: direct prediction in an orthogonal, localized Wannier basis avoids the overlap-matrix problem of LCAO-based Hamiltonian models and enables the use of linear-scaling KPM methods for spectral and transport observables. The dataset size and diversity, the focus on gauge-invariant observables rather than raw matrix elements, the topological benchmark on Bi2Se3, and the open-source Tailwater package are clear strengths. The finetuning demonstration on a 78-atom twisted WSe2 system is a useful practical result. However, the paper's headline claims about universality, strict linear scaling, and 10K+ atom capability currently outrun the evidence presented; the missing locality/convergence analysis and the unclear train/test separation are load-bearing issues that must be addressed before the results can be fully credited.","major_comments":[{"comment":"The training set is stated to be built from structures of MP-Traj, C2DB, GNOME, and Alexandria (§II), while the validation set used for all headline benchmarks is described as '2000 materials from the GNOME dataset' (§IV, Figs. 4 and 5). The Fig. 5 caption asserts these materials were 'not seen in training,' but the paper never reports the train/validation split protocol, duplicate removal, or any filtering that would prevent GNOME structures from appearing in both training and validation. If GNOME-derived structures are in the training pool, the reported DOS MAE, band-gap MAE, and R2 values are leakage-inflated. Please state the exact split, the number of GNOME structures in the training set, and the procedure used to ensure no overlap or near-duplicate leakage.","section":"§II and §IV"},{"comment":"The model predicts inter-site Hamiltonian blocks H_ij(R) 'per directed edge between atom i and a periodic image of atom j' (§III), but the graph construction—specifically the neighbor cutoff radius or edge criterion—is never reported. The Wannier spread check of Sec. II (<1.5|v_max|) does not control the decay of hopping amplitudes, so it does not justify omission of longer-range blocks. Without a stated cutoff and a convergence study in terms of physical observables (DOS, band gap, small energy splittings, transport coefficients), the sparse Hamiltonian produced for large systems may contain systematic errors that the averaged DOS/band-gap metrics are insensitive to. Please report the cutoff and provide convergence tests demonstrating that the truncated Hamiltonian faithfully reproduces target spectra and topology.","section":"§III and §V"},{"comment":"The abstract claims 'optimized inference and linear-scaling methods for orthogonal Hamiltonians unlock transport simulations at massive scales (10K+ atoms),' and §V.C states a primary goal of '1K-10K+ atoms.' However, no wall-clock scaling plot is shown for either inference or KPM spectral evaluation, and no calculation at 10K atoms is demonstrated. The runtime comparison in Fig. 7(e-f) is a small set of generation times, not a scaling analysis. 'Strict linear scaling' is asserted but not established. Please provide asymptotic scaling measurements for the full pipeline (graph construction, message passing, Hamiltonian assembly, and KPM observables) and an explicit demonstration at or near 10K atoms, or soften the claims accordingly.","section":"Abstract, §V.C"}],"minor_comments":[{"comment":"The phrase 'first deep-learning model capable of generating the electronic Hamiltonian of solid-state systems in an orthogonal Wannier basis' is stronger than the text supports, given that Ref. [30] already generates Wannier Hamiltonians for a bilayer system. Recommend qualifying with 'at scale' or 'as a foundational model.'","section":"Abstract and §I"},{"comment":"The sentence 'we ensure the total number of states is equal to the number of Wannier functions plus ten additional states' should say 'the number of Kohn-Sham bands' rather than 'states,' which is confusing in a spinor context.","section":"§II"},{"comment":"The composite loss weights λΔ and λ× are said to be 'set so that the spacing and crossing terms refine rather than overwhelm the eigenvalue error,' but their numerical values are not reported. Please provide the values or the tuning procedure.","section":"§III.C"},{"comment":"The 10 meV Hamiltonian matrix-element MAE is mentioned without defining the gauge alignment procedure used before computing the error. Since the model is trained with a gauge-dependent target, this metric is meaningful only with a stated gauge convention; please clarify.","section":"§IV"},{"comment":"The statement 'This sparse matrix can be evaluated directly to extract eigenvalues' should mention that sparse eigensolvers are used, and that their scaling depends on condition number and spectrum. The manuscript's linear-scaling claim should not conflate matrix sparsity with eigensolver scaling.","section":"§V.C"},{"comment":"The code snippet includes USER and PSWRD arguments, implying the API requires credentials. Please state more clearly the current access model (open weights vs. hosted API only) and whether the trained G(Wa)NN weights will be released.","section":"§V.B"}],"recommendation":"major_revision","confidential_remarks":"The core idea is promising and the benchmarks, if valid, would be a meaningful advance. The missing cutoff/convergence analysis, the unclear train/test separation for GNOME, and the unsupported 10K-atom/linear-scaling claims are all fixable with additional experiments and text revisions, but they are load-bearing for the paper's central claims. I recommend major revision rather than rejection because the technical approach appears sound in principle and the required evidence is within the scope of a revised manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a substantial, credible empirical paper. The authors built an equivariant GNN (G(Wa)NN) that predicts spin–orbit coupled Wannier Hamiltonians directly in an orthogonal basis, trained on 111,905 Wannier Hamiltonians spanning 69 elements, and paired it with a post-processing package (Tailwater) that does subspace projection, KPM spectral functions, and surface Green's functions. The Bi2Se3 example is convincing: the model reproduces the band structure, returns Z2=1, and resolves the Dirac surface cone. That is real, useful work.\n\nThe dataset is the most valuable contribution. Wannier Hamiltonians are gauge-fixed, and the authors took care with a fixed per-element orbital basis, a frozen window, and spread checks. The loss design—band-edge weighted eigenvalue error plus level-spacing and crossing-protection terms—is a sensible way to target physically relevant accuracy rather than just matrix-element MAE.\n\nNow the soft spots. The stress-test note is right: the graph cutoff for inter-site blocks is never stated, and there is no convergence study. The spread criterion (<1.5×|v_max|) does not imply hoppings are negligible beyond a given radius. If long-range hoppings are cut, small gaps and topological features can be wrong, and the DOS MAE will not catch it. This needs fixing or at least explicit discussion.\n\nSecond, the validation set comes from GNOME, which also appears in training. 'Universal' generalization is not established by an in-distribution holdout. A held-out set from a different source, or newly computed structures, would be more convincing.\n\nThird, the headline claims outrun the paper. Ref [30] already learned Wannier hopping parameters for bilayers, so 'first deep-learning model... in an orthogonal Wannier basis' is too strong unless you add 'at this scale' and 'for arbitrary stoichiometry.' 'Strict linear scaling' is a property of KPM in an orthogonal sparse basis, but the paper does not show that G(Wa)NN inference itself scales linearly. The benchmark in Fig. 7(e)-(f) is a timing comparison with Uni-HamGNN and QE, not an asymptotic scaling plot. '10K+ atoms' is not demonstrated.\n\nAlso, the model weights and dataset are not publicly available; the API requires credentials, so the central artifact cannot be independently reproduced. Tailwater being open source is good, but the model is the core.\n\nNone of this is fatal. The architecture is a standard equivariant GNN with Clebsch–Gordan tensor products, so the novelty is in the dataset, the orthogonal Wannier target, and the integration, not in the math. The paper deserves a serious referee: it is a genuine engineering contribution that the community will want to engage. I would ask for a cutoff convergence study, a proper out-of-distribution test, softened claims, and ideally a public release of at least a subset of the data and model weights. Send it to review.","headline":"A large-scale, credible ML effort for orthogonal Wannier Hamiltonians with a real engineering contribution, but the 'first' and '10K+ atom' claims outrun the evidence, and the missing graph cutoff is a genuine unresolved issue.","tokens_in":18683,"tokens_out":3691,"would_cite":true,"duration_ms":31311,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["71.15.-m","71.70.Ej","72.10.-d"],"model":"deepseek-v4-flash","headline":"G(Wa)NN claims to be the first deep-learning model that generates electronic Hamiltonians directly in an orthogonal Wannier basis, turning a crystal structure into a sparse spin-orbit coupled tight-binding matrix.","keywords":["Wannier Hamiltonian","deep learning","spin-orbit coupling","equivariant graph neural network","orthogonal basis","kernel polynomial method","tight-binding","transport simulation"],"falsifier":"Take a material known to have significant long-range hoppings, generate its Hamiltonian with the model using a series of increasing graph cutoffs, and compare the predicted band structure and transport coefficients against a converged DFT-Wannier reference; if spectra or small energy splittings shift beyond tolerance when the cutoff is enlarged, the locality assumption fails.","tokens_in":17741,"feed_emoji":"⚛️","tokens_out":4873,"duration_ms":38764,"temperature":0.7,"pith_summary":"The paper introduces G(Wa)NN, a graph neural network trained on over 111,000 Wannier Hamiltonians (150M+ hop matrices) spanning 69 elements. Its central claim: given only a crystal structure, it produces a spin-orbit coupled electronic Hamiltonian in an orthogonal Wannier basis, with the sparsity needed for linear-scaling transport calculations. A companion package, Tailwater, handles subspace projection, KPM spectral functions, and local fine-tuning. If the claim holds, it closes a gap between first-principles accuracy and device-scale simulations. Validation targets include band gaps (0.08 eV MAE on a 2,000-material set), density of states, and a topological invariant Z2=1 with a surface Dirac cone for Bi2Se3.","feed_headline":"Neural net writes spin-orbit Hamiltonians from crystal structure","feed_subtitle":"The first deep-learning model in an orthogonal Wannier basis opens linear-scaling transport at 10,000+ atoms.","key_machinery":"The central object is the spinful Wannier Hamiltonian block, reconstructed by inverting Clebsch–Gordan decomposition: the network emits covariant coefficients that are combined with precomputed coupling tensors into orbital-resolved blocks, and spin is attached through spin-scalar and spin-vector channels. The orthogonal, localized basis is what keeps matrices sparse and enables strict linear-scaling spectral methods. A second load-bearing piece is the momentum-space loss, which anchors eigenvalues to band edges, preserves level spacings, and protects crossings.","core_discovery":"G(Wa)NN is an E(3)-equivariant message-passing network that predicts each Hamiltonian block—on-site and inter-site, including periodic images—as a complex matrix in a fixed spinful Wannier basis. Because the basis is orthogonal, the predicted Hamiltonian is strictly sparse. The training objective combines a real-space matrix-element loss with a momentum-space loss that weights states near the band edges, enforces correct level spacings, and protects symmetry-required band crossings. The paper demonstrates that the predicted Hamiltonians reproduce DFT band structures and density of states, identify the topological invariant of a topological insulator, and can be fine-tuned on a small custom d","pith_inferences":["The paper never states the graph cutoff radius that defines neighboring blocks for Hamiltonian prediction; until a locality-convergence test is published, the claim of universal transferability to systems with long-range hybridization is premature.","The fine-tuning result suggests the pre-trained model could serve as a cheap preconditioner or starting point for hybrid-functional or many-body corrections, though only one example is shown.","The same sparse orthogonal Hamiltonian could feed NEGF and Chebyshev transport codes beyond KPM, potentially extending the acceleration to device-level current calculations."],"forward_implications":["If the model generalizes as claimed, electronic-structure and transport simulations on systems with 10,000+ atoms become routine, with spectral observables computed in strictly linear time.","Band-gap screening on material databases could be done at model speed with accuracy close to dedicated ML gap predictors (0.08 eV MAE).","The sparse orthogonal Hamiltonian format drops directly into existing tight-binding and KPM pipelines, so topological invariants, surface spectra, and spin-Hall conductivities can be computed without an intervening DFT step.","Local fine-tuning on user datasets—including with reduced orbital bases—lets the base model be adapted to different functionals or out-of-distribution moiré systems at negligible cost."],"fun_headline_variants":["AI writes spin-orbit Hamiltonians for 10k+ atom simulations","Neural net predicts spin-orbit Hamiltonians for linear-scaling transport","Spin-orbit Hamiltonians from deep learning in orthogonal Wannier basis","Machine learning models spin-orbit Hamiltonians for large-scale simulations"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The model assumes that all important Hamiltonian matrix elements are determined by atoms within some finite graph neighborhood, but the paper never states the cutoff radius or shows that predictions converge as that radius grows.","fun_headline_variants_meta":{"raw":{"variants":["AI writes spin-orbit Hamiltonians for 10k+ atom simulations","Neural net predicts spin-orbit Hamiltonians for linear-scaling transport","Spin-orbit Hamiltonians from deep learning in orthogonal Wannier basis","Machine learning models spin-orbit Hamiltonians for large-scale simulations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000577,"raw_usage":{"total_tokens":2578,"prompt_tokens":782,"completion_tokens":1796,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":526,"completion_tokens_details":{"reasoning_tokens":1719}},"tokens_in":526,"tokens_out":1796,"duration_ms":11824,"temperature":1.0,"reasoning_tokens":1719,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T15:29:52.609393+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a material known to have significant long-range hoppings, generate its Hamiltonian with the model using a series of increasing graph cutoffs, and compare the predicted band structure and transport coefficients against a converged DFT-Wannier reference; if spectra or small energy splittings shift beyond tolerance when the cutoff is enlarged, the locality assumption fails.","supporting_citations":[],"review_version":1}