{"id":"43705fe4-533d-4dc9-bf5a-b3d1dccb98d3","arxiv_id":"2602.16897","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A Hamiltonian-predicting neural network, QHFlow2, reaches NequIP-level force accuracy and 20× lower energy errors than MLIPs when energies and forces are computed directly from its predicted Hamiltonians.","lead":"Machine-learning models that predict quantum Hamiltonians are usually judged on how well they reconstruct the matrix. This paper instead computes energies and forces directly from the predicted Hamiltonian, and shows a new model, QHFlow2, reaches force accuracy on par with a top machine-learned force field while cutting energy errors by over an order of magnitude.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Force errors depend on an unspecified non-SCF analytic-gradient implementation; if it assumes SCF stationarity, the headline NequIP comparison is invalid.","rationale":"The reader's weakest assumption pinpoints the non-SCF force-evaluation pipeline as the critical unverified step. I agree: this is the single most load-bearing concern because the paper's headline ('NequIP-level force accuracy') is a direct quantitative claim built on those force numbers. If the gradient implementation is flawed, the comparison collapses, regardless of the other gaps (e.g., QH9 reference mismatch, HELM omission). The concern is concrete and testable because the paper provides a source-code link and the rMD17 benchmark is recomputed with PySCF, making a finite-difference check straightforward. My verdict remains CONDITIONAL (UNCHANGED relative to the reader) because the paper may have implemented the gradient correctly—e.g., by solving response equations or by first converging SCF from the predicted density—but the manuscript does not document this, and the current description is insufficient to trust the central claim. The proposed check would settle the issue without requiring new experiments or external baselines.","tokens_in":32237,"tokens_out":4630,"duration_ms":44839,"concrete_test":"Inspect the released DFTEval code to determine the exact PySCF routine that computes forces from a predicted Hamiltonian. Then, for one test molecule (e.g., aspirin), compute the reported analytic forces and compare them with finite-difference derivatives of E_KS[ρ] with respect to nuclear displacements, using the same predicted D, basis, and DFT settings. If the maximum absolute deviation exceeds ~1e-4 Hartree/Bohr (or relative error >1%), the analytic gradient is not a true gradient of the stated energy, and the force benchmark does not measure physical forces. Also verify whether SCF is re-converged before the gradient call.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of reaching NequIP-level force accuracy rests entirely on the force MAE numbers computed by the direct-evaluation pipeline. Section 3.1 states that after solving the Roothaan–Hall equation with the predicted Hamiltonian Ĥ and the reference overlap matrix S, the density matrix D is used to evaluate the KS energy functional and the forces as analytic gradients F_i = −∇_{r_i} E_KS[ρ]. Appendix A.1 gives only the standard Hellmann–Feynman/Pulay expression and notes that Pulay terms arise from the basis; it never describes how the density-matrix response is handled when D is not SCF-stationary. For a non-self-consistent density, the energy gradient contains a term proportional to ∂D/∂R that does not vanish because D is not variational. Standard PySCF gradient routines (e.g., grad.RKS) assume SCF convergence; applying them to a non-stationary D is either undefined or yields forces that are not the true gradient of E_KS[D]. The paper does not specify whether the implementation (a) re-converges SCF before differentiation, (b) solves coupled-perturbed response equations for the non-SCF D, or (c) uses only Hellmann–Feynman terms. If (c), all reported forces—the basis for the headline comparison—are missing Pulay and response contributions, invalidating the comparison to NequIP. This is the most load-bearing concern because it directly undermines the paper's central quantitative claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes QHFlow2, an SO(2)-equivariant flow-matching model for predicting Kohn–Sham Hamiltonians, and introduces a benchmark that computes total energies and analytic forces directly from predicted Hamiltonians. The main claims are that on MD17/rMD17 QHFlow2 is the first Hamiltonian model to reach NequIP-level force accuracy while achieving up to 20× lower energy MAE, and on QH9 it reduces energy error by up to 20× compared to MACE. Additional contributions include a two-stage pair update, a recomputed rMD17 benchmark for controlled comparison, scaling studies, and SCF-initialization experiments.","tokens_in":32566,"tokens_out":3630,"duration_ms":36628,"significance":"If the direct-evaluation pipeline is implemented correctly and the comparisons are fair, the paper would provide strong evidence that MLH models are practically viable energy–force predictors, not just Hamiltonian reconstruction tools. The paper's controlled MD17/rMD17 setup—retraining all baselines on the same splits and recomputing reference Hamiltonians—is a methodological strength, as are the scaling and SCF-cycle experiments. The architecture improvements are clearly described and appear to give substantial accuracy gains. However, the central force-accuracy claim depends on an unspecified non-SCF analytic-gradient implementation, and the QH9 energy comparison against published QM9 numbers is not apples-to-apples. These issues need to be resolved before the headline claims can be accepted.","major_comments":[{"comment":"The force evaluation is not specified for a non-self-consistent density matrix. After solving the Roothaan–Hall equation with a predicted Hamiltonian, the density matrix D is not SCF-stationary, so the gradient of E_KS[rho] contains terms proportional to dD/dR that do not vanish. Appendix A.1 gives only the standard Hellmann–Feynman/Pulay expression for SCF-converged densities and never describes how the density-matrix response is treated. Standard PySCF gradient routines assume SCF stationarity; applying them to a non-SCF D would give forces that are not the true gradient of E_KS[rho]. The authors must specify whether they (a) rerun SCF before differentiation, (b) solve coupled-perturbed response equations for the non-SCF D, or (c) use only Hellmann–Feynman terms, and they should numerically verify against finite differences. This is the load-bearing concern because the headline 'NequIP","section":"Section 3.1 and Appendix A.1"},{"comment":"The QH9 energy comparisons against MACE, Equiformer, and EquiformerV2 use published QM9 results from different DFT settings, QH9 labels being B3LYP/def2-SVP, while the published QM9 numbers likely use a different functional/basis combination. Reporting energy MAE in meV and calculating ratios such as '20× lower than MACE' from these numbers is not valid without correcting for the reference electronic-structure settings or retraining/evaluating the MLIPs on QH9. The abstract's QH9 claim is therefore not supported as stated. The authors should either recompute MLIP baselines under the same DFT setup or clearly present the QH9 comparison as cross-references that are illustrative, not quantitative.","section":"Section 4.1 and Table 15 / Figure 3"},{"comment":"The text states that prior Hamiltonian predictors 'fail to produce accurate energies when the initial Hamiltonian is removed,' but Table 12 reports no energy or force values for SchNOrb, PhiSNet, QHNet, or HELM—only Hamiltonian and orbital metrics, with dashes for downstream quantities. This claim is therefore unsupported by the presented data. To make this point, the authors need to run those baselines through the same downstream evaluation pipeline without an initial Hamiltonian, or soften the claim accordingly.","section":"Section G.1, Table 12"}],"minor_comments":[{"comment":"Typos and inconsistent naming: 'Traininig', 'comparion', 'Dimenet', 'Nequip', 'W ANet', 'V on Lilienfeld' (the latter appears in a citation). A careful proofread is needed.","section":"Throughout"},{"comment":"The occupation rule uses a simple eigenvalue ordering without considering near-degeneracies or level crossings. The text should clarify how orbital ordering is resolved for the predicted Hamiltonian, particularly for molecules with near-degenerate frontier orbitals.","section":"Section 3.1, Eq. (3)"},{"comment":"The description of rMD17 says 'we recompute reference quantities with PySCF at the PBE/def2-SVP level,' and in the main text the authors claim consistency with the original rMD17 labels. If the original rMD17 labels were generated with ORCA, the two are not necessarily identical; please clarify whether the benchmark uses only the recomputed labels or a mix of original and recomputed labels.","section":"Section E.1"},{"comment":"The comparison between MLIPs and QHFlow2 on energy MAE is striking (e.g., 0.003 meV vs 0.33 meV for NequIP on ethanol). Since energy MAE is dominated by per-atom reference-energy shifts, the authors should state whether both models are evaluated on exactly the same energy target normalization; otherwise the large apparent advantage might partly reflect different target definitions.","section":"Table 14"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a compelling architecture and a useful benchmark, but the force evaluation is the crux. The lack of implementation details for the non-SCF gradient is the single most important issue; if the authors can document and validate that the forces are true gradients of the stated energy functional, the main MD17/rMD17 claim becomes credible. The QH9 comparison to published QM9 numbers is a fairness problem that likely requires redoing or reframing the comparison. I would ask for the code or a detailed description of the DFTEval step in the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this paper makes a useful shift in how ML Hamiltonians are judged: instead of only matrix-reconstruction metrics, they compute energies and forces from the predicted H through a fixed pipeline and compare to MLIP baselines. That benchmark is genuinely new. Second, the central comparison—QHFlow2 reaching NequIP-level forces—depends on an analytic-gradient implementation the manuscript does not specify. If the forces are computed from a non-self-consistent density without the density-matrix response terms, the numbers may not mean what the abstract says.\n\nWhat's good: the architecture work is credible. QHFlow2 uses an SO(2) backbone with a two-stage pair update, and the MD17/rMD17 experiments are controlled—they retrained NequIP, GemNet-T, DimeNet and the Hamiltonian baselines on the same splits and labels. The scaling plots are clean. The 40% Hamiltonian error reduction with fewer parameters is a concrete, reproducible-looking claim, and the code is linked. The paper is clearly written and does not oversell the SCF acceleration results.\n\nWhere it's soft. The force evaluation is the load-bearing part. Section 3.1 says they solve Roothaan–Hall with H_pred and S, form D, then evaluate E_KS and analytic gradients. Appendix A.1 gives the standard Hellmann–Feynman/Pulay expression but never says how the response of D is handled when D is not SCF-stationary. PySCF's RKS gradient assumes stationarity. If they call grad.RKS on a non-converged density, the output is not the gradient of E[D]. They need to state whether they include response terms or whether they re-converge SCF (which would change what 'direct' means). I did not see this in the main text or appendix; the linked code might resolve it, but a referee should check.\n\nSecondary issues: the QH9 comparison to MACE/Equiformer uses published QM9 numbers from different DFT settings; that's fine as a reference point but shouldn't be read as a controlled comparison. HELM is missing from the main Hamiltonian tables. And Table 12 claims prior models 'fail' without the initial Hamiltonian but doesn't report their energy/force errors—just dashes.\n\nNet: the benchmark and architecture are worth taking seriously. The paper deserves a serious referee, but the referee needs to pin down the gradient implementation before the headline claim can be trusted. I'd want a revision with an explicit derivation or a code inspection.","headline":"The direct-evaluation benchmark is a real contribution and the model results look strong, but the force-evaluation pipeline is underspecified in exactly the place the headline claim rests.","tokens_in":33077,"tokens_out":3302,"would_cite":true,"duration_ms":30263,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","81V55"],"pacs":["71.15.Mb","31.15.Ew","07.05.Mh"],"model":"deepseek-v4-flash","headline":"Under direct evaluation, a Hamiltonian-predicting model (QHFlow2) matches leading force-field accuracy on forces and beats it up to 20× on energies — making predicted Hamiltonians viable energy–force predictors.","keywords":["machine learning Hamiltonians","Kohn–Sham Hamiltonian prediction","energy and force prediction","equivariant neural networks","flow matching","MD17/rMD17 benchmark","QH9 benchmark","interatomic potential comparison"],"falsifier":"Take the released code and several MD17 test molecules: for a fixed predicted Hamiltonian, compute the reported analytic forces and compare them with finite differences of the reported energy under the identical non-self-consistent prescription; disagreement beyond numerical tolerance would show the force numbers do not measure what they claim. Second check: retrain the interatomic-potential baselines on the exact QH9-stable-id split used for QHFlow2 — the paper's QH9 comparison uses published numbers from a different split — and see whether the up-to-20× energy margin survives a strict head-t","tokens_in":32134,"feed_emoji":"⚛️","tokens_out":16615,"duration_ms":145440,"temperature":0.7,"pith_summary":"Machine-learning Hamiltonian models predict the quantum-chemical matrix that encodes a molecule's electronic structure, but they have been judged almost entirely by how well they reconstruct that matrix — not by whether the energies and forces derived from it are physically accurate. This paper asks whether such models can work as direct energy–force predictors, and it builds a benchmark that pushes each predicted Hamiltonian through the Roothaan–Hall equation into a density, a Kohn–Sham energy, and analytic forces. Within that benchmark, the paper's new model, QHFlow2 — a flow-matching network with an SO(2)-equivariant backbone and a two-stage edge update — is the first Hamiltonian model to reach the force accuracy of a leading interatomic-potential model on MD17/rMD17, while reporting up to 20× lower energy errors; on QH9 its energy errors are up to 20× smaller than a leading machine-learning force field. It also reports 40% lower Hamiltonian reconstruction error than the previous best Hamiltonian model with roughly half the parameters, and shows that Hamiltonian, energy, and force errors fall together as model size and data grow. If the benchmark is sound, the payoff is a single model that supplies accurate interatomic forces and, from the same forward pass, orbital energies, densities, and reactivity descriptors — quantities force-field models cannot provide. The authors note the current scope is restricted closed-shell molecules under a fixed functional and basis, with evaluation cost still dominated by the density-functional step.","feed_headline":"Hamiltonian predictor matches NequIP forces, cuts energy error 20x","feed_subtitle":"First Hamiltonian model to reach NequIP-level force accuracy while reporting up to 20× lower energy error.","key_machinery":"The load-bearing object is the predicted Kohn–Sham Hamiltonian matrix Ĥ in an atom-centered orbital basis, and the identity that carries the argument is the Roothaan–Hall equation ĤC = SCϵ — the generalized eigenproblem that, with the geometry-fixed overlap matrix S, converts the predicted matrix into orbital coefficients, a density matrix, a Kohn–Sham energy, and analytic forces. The architecture that makes the prediction accurate enough for this shortcut is QHFlow2's combination of an SO(2)-equivariant message-passing backbone — efficient edge updates carried out in local reference frames while keeping an SO(3)-equivariant feature structure — with a two-stage pair update: pairwise features","core_discovery":"The paper's central claim is that an accurate predicted Hamiltonian, evaluated directly and without self-consistent refinement, yields total energies and atomic forces that are competitive with the best machine-learning interatomic potentials. The pipeline is: predict the Kohn–Sham Hamiltonian Ĥ in an atom-centered orbital basis from the molecular geometry; solve the Roothaan–Hall equation ĤC = SCϵ against the fixed reference overlap matrix S with a simple occupation rule; build the density matrix and evaluate the Kohn–Sham energy functional; take forces as analytic nuclear gradients. On the MD17/rMD17 trajectories, QHFlow2 reports force mean absolute errors between 0.34 and 1.35 meV/Å acros","pith_inferences":["The implied displacement is bigger than the headline: if predicted Hamiltonians are this accurate without self-consistency, then SCF iteration — the main cost of DFT — becomes partially optional for energy and force estimation; a natural test is to measure how energy and force error degrade as the predicted Hamiltonian is deliberately perturbed away from self-consistency.","The reactivity-stretching demonstration (frontier-orbital energies under bond distortion) points to a use case force-field models cannot serve: replacing DFT in reaction-coordinate and descriptor screening. A concrete extension is full reaction-path scans comparing predicted barrier heights and HOMO–LUMO trends against DFT references across many bonds.","The scaling curves invite an extrapolation the authors do not make: if Hamiltonian error continues to fall as a power law in parameters and data, the direct-evaluation route could overtake retrained interatomic potentials at scale, because it produces forces, energies, and electronic structure from one model — worth testing on larger molecular diversity than the six MD17 molecules.","The fixed-functional, fixed-basis training and closed-shell scope suggest a cheap transfer check: test the direct-evaluation pipeline across basis sets (e.g., def2-SVP versus def2-TZVP) to see whether Hamiltonian accuracy transfers or whether conditioning on the basis is required — the authors list this as open future work."],"forward_implications":["Hamiltonian models become directly usable as energy–force predictors: a predicted matrix alone, without any SCF refinement, yields MD-grade forces and accurate energies, giving MLH models a concrete accuracy reference to be held to.","A single model now provides both interatomic forces and electronic-structure observables — occupied orbital energies, HOMO–LUMO gaps, densities, and reactivity descriptors — from the same forward pass, which interatomic-potential models cannot do.","Hamiltonian, energy, and force errors decrease together with model size and training data, so the measured scaling curves imply that more parameters and data will continue to improve downstream simulation accuracy.","Accurate predicted Hamiltonians also accelerate iterative DFT: used as initial guesses, they cut SCF iterations by roughly half or more on QH9 relative to the default start, so the same model can speed up both direct evaluation and converged calculations.","The released MD17/rMD17 benchmark with recomputed, internally consistent Hamiltonians gives future work a controlled, split-matched setting for comparing Hamiltonian predictors against interatomic potentials."],"fun_headline_variants":["QHFlow2: first Hamiltonian model to match NequIP forces","Hamiltonian model hits NequIP forces, 20x lower energy error","QHFlow2: 40% lower Hamiltonian error, NequIP forces, 20x energy","Hamiltonian ML: NequIP force accuracy, energy error cut by 20x"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The central comparison assumes that evaluating the Kohn–Sham energy functional — and especially its analytic nuclear gradients — on a density built from a predicted, non-self-consistent Hamiltonian is a well-defined, correctly implemented prescription for physical energies and forces; the paper specifies the pipeline in Section 3.1 but defers the gradient details to a formula in Appendix A.1 that assumes SCF stationarity, leaving the non-stationary case unspecified.","fun_headline_variants_meta":{"raw":{"variants":["QHFlow2: first Hamiltonian model to match NequIP forces","Hamiltonian model hits NequIP forces, 20x lower energy error","QHFlow2: 40% lower Hamiltonian error, NequIP forces, 20x energy","Hamiltonian ML: NequIP force accuracy, energy error cut by 20x"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0003,"raw_usage":{"total_tokens":1571,"prompt_tokens":745,"completion_tokens":826,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":489,"completion_tokens_details":{"reasoning_tokens":736}},"tokens_in":489,"tokens_out":826,"duration_ms":6994,"temperature":1.0,"reasoning_tokens":736,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T22:24:08.703726+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the released code and several MD17 test molecules: for a fixed predicted Hamiltonian, compute the reported analytic forces and compare them with finite differences of the reported energy under the identical non-self-consistent prescription; disagreement beyond numerical tolerance would show the force numbers do not measure what they claim. Second check: retrain the interatomic-potential baselines on the exact QH9-stable-id split used for QHFlow2 — the paper's QH9 comparison uses published numbers from a different split — and see whether the up-to-20× energy margin survives a strict head-t","supporting_citations":[],"review_version":1}