{"id":"6fb3acda-ad85-4e93-b04f-b128195fec1e","arxiv_id":"2504.19586","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Uni-HamGNN predicts spin-orbit-coupled Hamiltonian matrices from crystal structure alone, with roughly 0.003 eV error on matrix elements, and flags 120 candidate topological insulators in a high-throughput screen.","lead":"This paper introduces Uni-HamGNN, a graph neural network that predicts spin-orbit-coupled quantum mechanical Hamiltonians for materials across the periodic table. If the accuracy holds, it could replace expensive supercomputer calculations for screening topological and valleytronic materials.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Delta-learning decomposition assumes the spin-independent block of self-consistent SOC-DFT equals the non-SOC H0; this is never tested and could bias TI screening.","rationale":"Reader's conditional verdict is appropriate. The empirical evidence is real: low MAEs on a 5,000-material test set, agreement for ZrSiPt, Ca3(SiIr)4, etc., Berry curvature in WSeS, and twisted-bilayer bands all demonstrate the model is not a trivial failure. The load-bearing weakness is the decomposition's implicit assumption that SOC does not feed back into H0. Methods 4.4 explicitly freezes SOC parameters in Stage 2, so nothing in training can correct a systematic offset in H0. The reported test MAE averages over Materials-Project structures and does not cover the small-gap GNoME candidates where Z2 is sensitive to few-meV shifts. The proposed OpenMX comparison directly measures the size of this offset and whether it changes the screening outcome. I agree with the reader; the verdict stays conditional pending this check.","tokens_in":14440,"tokens_out":21664,"duration_ms":255298,"concrete_test":"Run OpenMX with and without SOC on the three TI examples (HfZr3P4IrRh3, SrGa2IrRh, SrAs12Ru3Pt), the four heavy-element test systems, and 20 unvalidated members of the 120 screened candidates. Extract H0_SOC=(H↑↑+H↓↓)/2 from each SOC run and compare elementwise with the non-SOC H0 used to train channel I. Report the mean absolute difference, the shift in band ordering near the gap, and the change in Z2 indices when H0 is replaced by H0_SOC with SOC parameters fixed. If the H0 difference is comparable to or larger than the claimed 3.58 meV real-part MAE, or if any Z2 index flips, the delta-learning target is biased.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Delta-learning rests on Eq. (2)/Methods 4.4: H = H0⊗I2 + (1/2)ξL·σ, with H0 fit to non-SOC Hamiltonians and ξ fit to imaginary parts. In self-consistent SOC-DFT, the density and scalar potential are recomputed with SOC, so the spin-independent block H0_SOC=(H↑↑+H↓↓)/2 of the SOC calculation generally differs from the non-SOC H0. The paper never trains or tests channel I against H0_SOC; Stage 2 only refines H0 eigenvalues and explicitly leaves SOC parameters frozen. For the heavy 5d and near-gap systems in the GNoME screening pool, even a few meV shift in H0 can move band inversions and flip Z2 invariants. The 3.58 meV real-part MAE is not decisive: it is an average over a Materials-Project test set, includes the analytic (1/2)ξ⟨Lz⟩ diagonal term, and is not reported for the screened candidates, of which only 3 of 120 were verified by DFT.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Uni-HamGNN, a graph-neural-network model that predicts spin-orbit-coupled (SOC) Kohn-Sham Hamiltonians for arbitrary compositions. The method decomposes the total Hamiltonian as H = H0 ⊗ I2 + Hsoc with Hsoc = ξL·σ, then uses a delta-learning scheme: a spin-independent channel is trained on non-SOC Hamiltonians from the Materials Project, and an SOC-strength channel is trained on the imaginary parts of SOC Hamiltonians. A two-stage training protocol first fits Hamiltonian matrix elements and then refines eigenvalues against band structures. The authors report a real-part MAE of 3.58 meV and an imaginary-part MAE of 0.0025 meV on a 5,000-material test set, validate band structures on selected heavy-element compounds, demonstrate transferability to 2D valleytronic materials and twisted TMD heterostructures, and use the model to screen 10,170 heavy-element GNoME structures, identifying 120 candidate topological insulators. The central claim is that a single trained model can replace expensive SOC-DFT calculations for Hamiltonian prediction and high-throughput topological screening.","tokens_in":14671,"tokens_out":4450,"duration_ms":44295,"significance":"If the accuracy and transferability claims hold, Uni-HamGNN would be a valuable tool: it would enable SOC band-structure and Berry-curvature calculations at a small fraction of the cost of self-consistent SOC-DFT, and it would make high-throughput topological-insulator screening practical on large structure databases. The paper has clear strengths: the physically motivated H0 + ξL·σ decomposition is elegant and preserves SU(2) symmetry; the delta-learning strategy is a sensible response to the magnitude disparity between spin-independent and SOC terms; the architecture improves tensor-product efficiency; the code is publicly available; and the validation spans bulk solids, 2D materials, and twisted heterostructures. These are substantive contributions. However, the evidence as presented is conditional: the headline imaginary-part MAE is a generalization metric for the fitted target rather than an independent test of SOC physics, the central delta-learning assumption that SOC does not feed back into H0 is never tested, and the high-throughput screening result is verified on only three of 120 candidates.","major_comments":[{"comment":"The delta-learning decomposition assumes that the spin-independent part of a self-consistent SOC-DFT calculation coincides with the non-SOC H0. The model is trained with H0 from non-SOC DFT and ξ from the imaginary parts of SOC Hamiltonians, and Stage 2 explicitly 'focuses exclusively on the eigenvalues of H0 without modifying SOC parameters.' In a self-consistent SOC calculation, however, the charge density and scalar potential are recomputed with SOC, so the spin-independent block H0_SOC = (H↑↑ + H↓↓)/2 generally differs from the non-SOC H0. The manuscript never trains or tests channel I against H0_SOC. This is load-bearing for the screening claim: for heavy 5d and near-gap systems, even a few meV shift in H0 can move band inversions and change Z2 invariants. I recommend computing H0_SOC for a representative subset of test structures and reporting the channel-I MAE against it; if the deviation is comparable to the reported 3.58 meV, the high-throughput screening conclusions should be re-examined.","section":"§2.2, Methods 4.4, Eq. (2)"},{"comment":"The reported imaginary-part MAE of 0.0025 meV is measured on the same channel used to train ξ, so it is a generalization metric for the fitted target rather than independent evidence that the model captures the physical SOC mechanism. The independent evidence consists of the band-structure and Berry-curvature comparisons in Figs. 4, 6, 7, and 8, which are illustrative rather than statistical. In addition, no error bars, per-structure variance, or quantiles are reported for any MAE, and the real-part MAE is not separated into the H0 contribution and the analytic (1/2)ξ⟨Lz⟩ diagonal term. Please report error bars and a decomposition of the real-part MAE, and state explicitly that the imaginary-part accuracy is a fit-quality measure for the ξ channel.","section":"§2.3, Figs. 3c–3d"},{"comment":"The high-throughput screening result of 120 topological insulators from 10,170 GNoME structures is verified on only three structures (HfZr3P4IrRh3, SrGa2IrRh, SrAs12Ru3Pt) via DFT and Wannier charge centers. This sample size is too small to establish the false-positive rate of the screening pipeline, which is essential for the claim that Uni-HamGNN enables reliable large-scale TI discovery. Please verify a random sample of the 120 predicted TIs and a sample of the trivial candidates, and report the distribution of predicted band gaps and Z2 indices. Without this, the 120-TI result is a candidate list rather than a validated screen.","section":"§2.4, Fig. 5"},{"comment":"The zero-point renormalization correction is written as H~' = H~ − (Σ_ij(H~_ij − H_ij)/Σ_ij S_ij) S_ij, which requires the DFT Hamiltonian H_ij. If this correction is applied only during training, the text should say so explicitly; if it is also applied at inference, the method cannot be used for new materials because the ground-truth H is unavailable. Please clarify the inference-time procedure for estimating the zero-point offset without DFT reference data.","section":"Methods 4.3, Eq. (9)"}],"minor_comments":[{"comment":"The Introduction refers to the 'GMoME' dataset, while the abstract and Section 2.4 use 'GNoME'; please harmonize the terminology.","section":"Introduction"},{"comment":"The number of non-SOC training structures is given as 40,000 in the Introduction and 44,000 in Section 2.2; please reconcile these numbers.","section":"Introduction and §2.2"},{"comment":"Eq. (4) uses a site- and bond-dependent ξij, while Eq. (3) and the text use a global ξ; please clarify how off-site ξij values are parameterized and constrained.","section":"Eq. (4)"},{"comment":"The corrected Hamiltonian is denoted H~′ in Eq. (9) and H′ in Eq. (12); please use a single notation throughout.","section":"Methods 4.3 and 4.4"},{"comment":"The parity plots in Figs. 3c and 3d would be more informative with labeled axes, units, and a color bar indicating point density.","section":"Fig. 3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reports impressive accuracy and a useful architecture, but the headline SOC accuracy and the high-throughput screening result are currently supported by weaker evidence than the text implies. The revision should add the H0_SOC check, error bars, and a statistically meaningful verification of the screening output; these are feasible within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version. Uni-HamGNN is a serious piece of applied ML-for-materials work. It builds a universal SOC Hamiltonian model using a physics-informed L·sigma decomposition and a dual-channel delta-learning scheme, and it reports Hamiltonian MAEs that are impressively low. The real question isn't whether the numbers are good — it's whether the decomposition's central assumption holds.\n\nWhat's actually new: a universal SOC Hamiltonian GNN that works across the periodic table without per-system retraining, with a training set of only ~10k SOC + 40k non-SOC structures. The two-center treatment of SOC beyond the on-site approximation is a sensible choice for heavy elements. The model's ability to reproduce SOC band structures, Berry curvatures, and spin-projected bands for 2D and twisted TMD systems is a genuine transferability test, and it appears to pass. That's real credit. The code is on GitHub and the authors promise to release weights; that is the right behavior.\n\nThe main soft spot is indeed the perturbation assumption. The paper's Eq. (2) and Methods 4.4 treat the SOC term as a delta on top of a H0 taken from non-SOC DFT. But a self-consistent SOC-DFT calculation recomputes the density and scalar potential with SOC, so the spin-independent block of that calculation generally differs from the non-SOC H0. The paper never trains or evaluates channel I against H0_SOC. For the heavy, near-gap systems in the GNoME screening pool, a few meV shift in H0 can flip band inversions and Z2 invariants. This is not a fatal flaw, but it is a load-bearing assumption that deserves an explicit test.\n\nThe screening result is thinner than advertised: 120 predicted TIs, but only 3 verified by DFT. To call it 'high-throughput discovery' you need either a much larger verification sample or a clear argument that the 3 are representative. No error bars on the MAEs is a minor issue but worth noting.\n\nThe imaginary-part MAE of 0.0025 meV is a generalization metric on the channel used to fit xi, not an independent test of SOC physics. That's not circular in a damaging way, because the paper also checks external DFT band structures and Wannier charge centers, which are independent. But the headline number should not be read as proof of physical correctness.\n\nWho is this for? People building ML Hamiltonians, and anyone screening for topological or valleytronic materials. It deserves a serious referee. My recommendation: send it to peer review, and require (1) a test of the H0 perturbation assumption, (2) release of trained weights, and (3) DFT verification on a larger TI sample.","headline":"Uni-HamGNN is a serious universal SOC Hamiltonian model with an untested self-consistency assumption and thinner TI screening validation than the bold claims suggest.","tokens_in":15190,"tokens_out":2560,"would_cite":true,"duration_ms":23499,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"One trained graph neural network predicts full spin-orbit-coupled Hamiltonians for arbitrary compositions across the periodic table.","keywords":["spin-orbit coupling","graph neural network","electronic Hamiltonian","universal model","topological insulator","high-throughput screening","delta learning","Z2 invariant"],"falsifier":"Choose a heavy-element compound with strong relativistic effects and compare the H0 produced by a non-SOC DFT run with the H0 extracted from a self-consistent SOC-DFT run. If the two matrices differ by more than roughly the reported 3.58 meV real-part error, the delta-learning premise is violated and the model's errors on such materials should increase correspondingly.","tokens_in":14235,"feed_emoji":"🧲","tokens_out":5717,"duration_ms":52706,"temperature":0.7,"pith_summary":"The paper introduces Uni-HamGNN, a single graph neural network that predicts the full spin-orbit-coupled (SOC) electronic Hamiltonian of arbitrary crystalline compounds, and argues that this makes expensive SOC-DFT calculations unnecessary for many materials tasks. The central idea is to write the Hamiltonian as H0 ⊗ I2 + ξ L·σ and train the spin-independent part H0 and the SOC strength ξ in separate channels, a delta-learning scheme that prevents the large H0 terms from masking the small SOC terms. On a held-out test set the model reaches mean absolute errors of 3.58 meV on real Hamiltonian entries and 0.0025 meV on imaginary entries, and it reproduces SOC band structures and Berry curvatures for 2D and twisted systems. If these results hold, a researcher can obtain SOC band structures, Berry curvatures, and Z2 topological invariants for unseen materials without running SOC-DFT, including in high-throughput screens.","feed_headline":"Spin-orbit Hamiltonians predicted for any crystal in one pass","feed_subtitle":"One trained model reproduces DFT-level SOC bands and flags 120 topological insulators in a high-throughput screen.","key_machinery":"The load-bearing object is the decomposition H = H0 ⊗ I2 + Hsoc with Hsoc = ξ L·σ, where L is the orbital angular momentum operator, σ the Pauli spin vector, and ξ a learnable spin-orbit strength per orbital pair. This form preserves the SU(2) transformation rule of the full Hamiltonian while reducing the SOC fitting problem to the scalar field ξ, and it includes both on-site and off-site SOC terms. The delta-learning strategy then trains H0 and ξ separately, using the imaginary part of the SOC Hamiltonian as the target for ξ, and a two-stage protocol adds band-structure error as a regularization in the second stage.","core_discovery":"The paper's claim is that a single model trained on about 44,000 non-SOC and 10,000 SOC Hamiltonians generalizes across the periodic table: the predicted Hamiltonian matrices agree with DFT to within a few meV, and derived band structures are visually and quantitatively consistent with DFT for heavy-element solids, 2D valleytronic materials, and twisted bilayer heterostructures. The discovery demonstration is the screening of 10,170 heavy-element structures from a machine-learning-discovered materials database, from which the model identifies 1,383 insulating candidates and, after computing Z2 invariants, predicts 120 topological insulators, three of which are verified by Wannier charge center analysis. The paper presents this as evidence that a universal, transferable SOC Hamiltonian model is feasible.","pith_inferences":["The same spin-independent plus correction decomposition could be applied to magnetic exchange interactions, so a magnetic analogue of the delta-learning scheme seems a natural next step, extending the approach to spin Hamiltonians.","The 0.0025 meV imaginary-part error is far below typical SOC splittings, which suggests the model may resolve weak topological gaps and small spin splittings that could be missed in band-structure comparisons alone.","Because the model was trained on bulk but works on 2D and twisted geometries, it may also generalize to surfaces, interfaces, and finite-temperature structures if the zero-point renormalization can be adapted."],"forward_implications":["SOC-DFT calculations can be replaced by one forward pass of Uni-HamGNN for Hamiltonian-derived quantities, with reported speedups of two to three orders of magnitude.","High-throughput topological-insulator screening can be run on real-space Hamiltonians without constructing Wannier tight-binding models, as the 10,170-structure screen demonstrates.","The model transfers from 3D training data to 2D monolayers and twisted bilayers, so interfacial and valleytronic systems become cheap to evaluate.","The parameterization captures two-center SOC terms, which the paper argues are essential for heavy-element systems where the usual on-site-only approximation fails."],"supporting_citations":[{"why":"Supplies the preceding spinless universal Hamiltonian model that the H0 channel extends.","marker":"[20]"},{"why":"Provides the HamGNN architecture from which HamGNN-V2 is built.","marker":"[17]"},{"why":"Supplies the tight-binding parameterization Hsoc = ξ L·σ for the spin-orbit term.","marker":"[23]"},{"why":"Source of the 44,000 non-SOC Hamiltonians used to train the H0 channel.","marker":"[28]"},{"why":"Source of the 10,170 heavy-element structures used in the topological-insulator screen.","marker":"[33]"},{"why":"Provides the Fukui-Hatsugai-Suzuki method used to compute Z2 invariants from the predicted Hamiltonians.","marker":"[30]"},{"why":"Supplies the DFT code that generated the Hamiltonian training labels.","marker":"[41]"},{"why":"Supplies the self-interaction layers incorporated into HamGNN-V2.","marker":"[27]"}],"fun_headline_variants":["One model maps spin-orbit Hamiltonians across periodic table","Universal SOC Hamiltonian model screens 10k materials, flags 120 topological insulators","Spin-orbit Hamiltonians for any crystal, no retraining","AI predicts spin-orbit effects for all elements, finds 120 topological insulators"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The scheme assumes SOC is a pure perturbation, so the spin-independent H0 appearing in an SOC-DFT calculation is exactly the same as the H0 from a non-SOC calculation, with no feedback of SOC into the charge density or scalar potential.","fun_headline_variants_meta":{"raw":{"variants":["One model maps spin-orbit Hamiltonians across periodic table","Universal SOC Hamiltonian model screens 10k materials, flags 120 topological insulators","Spin-orbit Hamiltonians for any crystal, no retraining","AI predicts spin-orbit effects for all elements, finds 120 topological insulators"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000708,"raw_usage":{"total_tokens":3169,"prompt_tokens":907,"completion_tokens":2262,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":523,"completion_tokens_details":{"reasoning_tokens":2184}},"tokens_in":523,"tokens_out":2262,"duration_ms":14403,"temperature":1.0,"reasoning_tokens":2184,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:49:46.108026+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Choose a heavy-element compound with strong relativistic effects and compare the H0 produced by a non-SOC DFT run with the H0 extracted from a self-consistent SOC-DFT run. If the two matrices differ by more than roughly the reported 3.58 meV real-part error, the delta-learning premise is violated and the model's errors on such materials should increase correspondingly.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the preceding spinless universal Hamiltonian model that the H0 channel extends."},{"cited_title":"npj Comput","cited_arxiv_id":null,"evidence_quote":"Provides the HamGNN architecture from which HamGNN-V2 is built."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the tight-binding parameterization Hsoc = ξ L·σ for the spin-orbit term."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Fukui-Hatsugai-Suzuki method used to compute Z2 invariants from the predicted Hamiltonians."},{"cited_title":"Curran Associates Inc","cited_arxiv_id":null,"evidence_quote":"Supplies the self-interaction layers incorporated into HamGNN-V2."}],"review_version":1}