{"id":"2722a0f8-42a6-4aef-add9-67b3ca26bbb3","arxiv_id":"2412.03090","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A deep neural network with an inverse-Hamiltonian loss function solves the radial Dirac equation for ground and excited states, validated on hydrogen and Woods-Saxon potentials.","lead":"Researchers trained a small neural network to solve the Dirac equation, the relativistic quantum equation for electrons and nucleons, without using labeled training data. The method avoids the usual collapse into negative-energy states and reproduces known energy levels for hydrogen and nuclear Woods-Saxon potentials.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The log-mesh inverse-Hamiltonian loss is only well-defined under a weighted inner product that the paper never states; without it, the Rayleigh quotient has no variational guarantee and may explain the uniform ~7e-4 hydrogen errors.","rationale":"The paper is best read as an empirical demonstration that a DNN plus inverse-Hamiltonian loss can reproduce known Dirac bound states; the hydrogen analytic results and Woods-Saxon GPS benchmarks provide genuine independent support for that claim. The weakest point is not the DNN itself but the discrete operator underlying the loss. The reader identified the Hermiticity of the log-mesh discretization as the weakest assumption, and I agree that this is the load-bearing gap: without a stated, consistent inner product, the inverse-Hamiltonian Rayleigh quotient is not guaranteed to select the targeted state, and the observed hydrogen energy errors are suspiciously uniform at the 1e-3 level while the uniform-mesh cases are much more accurate. However, the flaw may be merely undocumented rather than fatal: if the code internally uses the weighted metric, the discretization is Hermitian and the concern dissolves. Because the fix is a clear statement or explicit implementation of the weighted inner product, and because the benchmark results remain strong evidence, the conditional verdict is appropriate rather than rejection. I do not see a separate concern about the excited-state methods themselves; the recursive orthonormalization and the epsilon' tuning are adequately described, and their error accumulation is acknowledged. Lack of released code weakens reproducibility but is not a logical flaw in the central argument.","tokens_in":15551,"tokens_out":9185,"duration_ms":98911,"concrete_test":"Recompute the hydrogen n=2, kappa=-1 case with M=1700 log-mesh points, keeping the DNN architecture and optimizer identical, but using the weighted inner product <phi|psi> = Delta x sum_i e^{x_i} [F_i psi^F_i + G_i psi^G_i] and the correspondingly symmetrized matrix W^{1/2} H'_Dr W^{-1/2} in the inverse-Hamiltonian loss. If the relative energy error drops from ~6e-4 toward ~1e-6, the reported loss was using an inconsistent metric and the variational guarantee is missing; if the error remains ~6e-4, the Hermiticity issue is not the operative cause and the empirical claim stands on its benchmarks.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on the inverse-Hamiltonian loss selecting the targeted eigenstate by minimizing a Rayleigh quotient. This only has the intended meaning if the discretized Hamiltonian is Hermitian with respect to the inner product used in that quotient. On the log mesh of Eq. (28), the correct Hermitian metric is W = diag(e^{x_i}), under which the derivative matrix in Eq. (30) becomes anti-symmetric; in the unweighted Euclidean metric it is not anti-symmetric, and the assembled H'_Dr is not Hermitian as used. The paper never specifies the discrete inner product for Eq. (15), for the normalization Eq. (12), or for the orthonormalization Eq. (16). If the implementation uses unweighted sums, the loss is actually the Rayleigh quotient of the symmetric part of a non-Hermitian inverse Hamiltonian, so the variational selection of the desired state has no rigorous foundation. This is not merely a formal gap: for the hydrogen log-mesh calculations the inverse-Hamiltonian relative energy errors are uniformly around 5-7e-4 for n=2-6, whereas the uniform-mesh Woods-Saxon calculations, where Eq. (32) is anti-symmetric, reach ~1e-5 or better for the lowest states. The missing metric is therefore the most load-bearing unverified assumption in the paper's argument.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript extends a previously proposed unsupervised deep-neural-network variational method to the radial Dirac equation with spherical symmetry. To avoid variational collapse induced by the Dirac sea, it minimizes the Rayleigh quotient of the inverse Hamiltonian (ε′_nκ − H′_Dr)^{-1}; low-lying excited states are obtained either by choosing the shift ε′_nκ between neighboring eigenvalues or by projecting out lower states via an orthonormalization procedure. The method is applied to hydrogen on a logarithmic mesh and to Woods-Saxon potentials for 16O and 208Pb on a uniform mesh, with energies and wave functions compared against the analytic hydrogen solution and an independent general pseudospectral calculation.","tokens_in":15889,"tokens_out":9942,"duration_ms":100378,"significance":"If the validation is accepted, the paper contributes a practical unsupervised-DNN route to relativistic single-particle bound states and low-lying excited states, circumventing variational collapse in a way that does not fit the benchmark energies. The paper has genuine strengths: the loss function is the inverse-Hamiltonian Rayleigh quotient rather than a fit to known energies; the hydrogen comparison uses closed-form results and the Woods-Saxon comparison uses an independent pseudospectral method; two complementary excited-state strategies are compared; and Appendix A documents failed architectures, providing useful negative results. The main limitation is the unspecified discrete inner product, which currently leaves the logarithmic-mesh hydrogen part of the validation on weaker footing than the uniform-mesh Woods-Saxon part.","major_comments":[{"comment":"The manuscript never specifies the discrete inner product used in the Rayleigh quotient of Eq. (15), in the normalization condition of Eq. (12), or in the orthonormalization of Eqs. (16) and (17). This is load-bearing because the log-mesh derivative matrix in Eq. (30) is not anti-symmetric under the Euclidean metric; the discretized H′_Dr is Hermitian only with respect to a weighted metric such as W = diag(e^{x_i}), with boundary terms vanishing under the Dirichlet conditions. Unless the implementation uses that weighted inner product, the minimized object is not the Rayleigh quotient of a Hermitian inverse Hamiltonian, and the variational selection of the target eigenstate has no rigorous foundation. The issue is not purely formal: the hydrogen inverse-Hamiltonian relative energy errors in Table I are uniformly about 5×10^{-4} to 8×10^{-4} for n=2–6, whereas the uniform-mesh Woods-Saxon results in Tables II and III, where Eq. (32) is anti-symmetric, reach about 10^{-5}–10^{-6} for the lowest states. The authors should state the discrete inner product explicitly and, if unweighted sums were used, repeat the hydrogen calculations with the correct metric; without this, the hydrogen part of the central validation is incomplete.","section":"Sec. III A, Eqs. (15) and (30)"}],"minor_comments":[{"comment":"The statement in Sec. V C that all DNN energies are 'identical (within 0.002 %)' to the benchmark conflicts with Table III, where the largest relative error is 3.52×10^{-4} (0.035 %). The summary statement in Sec. VI that the differences are 'up to 0.15 %' also conflicts with Table I, where the n=6 orthonormal result has relative error 2.16×10^{-3} (0.216 %). These numbers should be harmonized.","section":"Sec. V C and Sec. VI"},{"comment":"The sentence that the inverse Hamiltonian method 'reaches at least 1 × 10^{-4} accuracy' is ambiguous and, taken literally, is not supported by Table I, where the n=2–6 inverse-Hamiltonian relative errors are 5.68×10^{-4} to 7.42×10^{-4}. Please state the per-state accuracy more precisely.","section":"Sec. IV C and Table I"},{"comment":"The stopping criterion for training is never given. Figure 6 shows energy errors as functions of epochs and Tables I–III report final energies, but the reader cannot tell how convergence was decided, what tolerance was used, or how many independent runs produced the displayed values. This should be specified for reproducibility.","section":"Sec. III B and Sec. IV C"},{"comment":"The statement that two hidden layers of 16 units are 'confirmed to be sufficient' is not supported by any reported sensitivity study. A brief description of the tested architectures or hyperparameter variations would make the claim reproducible.","section":"Sec. III B"},{"comment":"The complexity comparison O(XM^2) for the DNN omits the one-time cost of constructing or factorizing the inverse Hamiltonian matrix. If the inverse is precomputed, there is an initial O(M^3) cost; if it is not, each epoch requires a linear solve. Since the comparison with the GPS O(M^3) is given as motivation, this should be clarified.","section":"Sec. V C"},{"comment":"The row labeled ε′_{n−1} is confusing: Eq. (14) defines the shift as ε′_{nκ}, and the table appears to list the chosen shifts. Please make the notation consistent with the text.","section":"Table I"},{"comment":"The introduction contains a duplicated sentence: 'A fundamental challenge in variational treatments of the Dirac equation is variational collapse' appears twice in consecutive lines. This should be removed.","section":"Sec. I"},{"comment":"The observation that the fully connected DNN minimizes H′_Dr directly and converges to the hydrogen ground-state energy should be reconciled with the general claim in Sec. II B that naive energy minimization leads to variational collapse. The mechanism of trapping at a local minimum is interesting and should be stated in the main text rather than only in the appendix.","section":"Appendix A 3"}],"recommendation":"major_revision","confidential_remarks":"The central numerical benchmarks are real and the uniform-mesh Woods-Saxon results are convincing, so I do not doubt the overall feasibility of the method in that setting. The load-bearing issue is the missing discrete inner product for the log-mesh hydrogen calculations; asking the authors to state the metric explicitly, and ideally to provide the relevant code or a small reproducibility check, would settle whether the uniform ~7×10^{-4} hydrogen errors are a discretization artifact. If the requested clarification cannot be supplied, the hydrogen part of the validation should be treated as unverified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is a legitimate extension of the Naito-Hashimoto unsupervised DNN method to the Dirac equation, and it earns credit for facing the Dirac-sea problem head-on. The inverse-Hamiltonian trick and the two excited-state protocols are sensible, and the benchmarks are honest: hydrogen against exact solutions, Woods-Saxon against an independent pseudospectral code. The appendix showing failed network architectures and direct energy minimization is a good sign — it tells you the authors are reporting what actually happens, not just the wins.\n\nThe central weakness is precisely the one flagged in the stress-test note. The log-mesh derivative matrix in Eq. (30) is not anti-symmetric in the Euclidean metric, and the paper never states the discrete inner product used in the Rayleigh quotient, the normalization, or the orthonormalization. Without a weighted inner product (W = diag(e^{x_i}) makes Eq. (30) anti-symmetric), the 'inverse Hamiltonian' is not Hermitian as used, and the loss function is not a genuine Rayleigh quotient. The uniform ~7e-4 errors in the hydrogen log-mesh runs, versus ~1e-5 for the uniform-mesh Woods-Saxon cases, are consistent with that gap. I don't think this destroys the empirical claim — the method clearly lands on the right states and reproduces wave functions — but it does mean the paper's variational justification is incomplete as written. A referee should ask for the inner product, or a manifestly Hermitian discretization, before accepting the explanation.\n\nMinor points: no code is released, the excited-state errors grow to ~1e-3, and for these one-dimensional problems standard solvers are faster and more accurate. The authors acknowledge the latter only implicitly, via a complexity discussion that is a bit hand-wavy. None of this is disqualifying for a proof-of-concept.\n\nBottom line: this is a solid paper for people working on neural-network quantum solvers and relativistic bound states. It deserves a serious referee and probably a conditional accept after the Hermiticity question is settled. I'd want the missing metric said out loud before I'd rely on the method.","headline":"A credible proof-of-concept for DNN-based Dirac solvers, with a real but fixable gap: the log-mesh loss function is only variational under an unstated weighted inner product.","tokens_in":16350,"tokens_out":2859,"would_cite":false,"duration_ms":27346,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81Q05","81-08","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"A deep neural network with an inverse-Hamiltonian loss solves the radial Dirac equation for ground and low-lying excited states, bypassing the variational collapse caused by the Dirac sea.","keywords":["deep neural network","Dirac equation","inverse Hamiltonian method","variational collapse","unsupervised machine learning","excited states","radial wave function","Woods-Saxon potential"],"falsifier":"Compute the discrete spectrum of H_inv on the log mesh with the same derivative matrix and boundary conditions, using a dense linear solver; if spurious eigenvalues appear below the target state energy, the loss would converge to the wrong state on some initializations, directly contradicting the paper's consistency claim.","tokens_in":15364,"feed_emoji":"⚛️","tokens_out":2539,"duration_ms":25176,"temperature":0.7,"pith_summary":"The paper extends an unsupervised deep neural network method from nonrelativistic quantum mechanics to the Dirac equation. Because the Dirac Hamiltonian has negative-energy states, minimizing energy directly would plunge the loss into the Dirac sea; the authors avoid this by using the inverse Hamiltonian as the loss function. They show that both the ground state and low-lying excited states can be obtained accurately for spherically symmetric potentials, with two complementary strategies for excited states. The method is validated against analytical hydrogen solutions and against a high-accuracy pseudospectral benchmark for Woods-Saxon potentials in 16O and 208Pb, achieving agreement at the 0.15% level or better for the listed states.","feed_headline":"Neural network solves the Dirac equation without crashing into the Dirac sea","feed_subtitle":"Inverse-Hamiltonian training reproduces ground and excited states to within 0.15 percent of benchmark results.","key_machinery":"The load-bearing object is the inverse Hamiltonian H_inv = (ε' - H'_Dr)^{-1}, whose spectrum is inverted relative to the original Dirac Hamiltonian, placing the desired bound state at the bottom; the DNN outputs the trial function f(r) = F(r)/r, from which the small component G(r) is reconstructed through the radial Dirac equation, and the loss is the expectation value of H_inv. Two excited-state mechanisms hang off this object: shifting ε' between consecutive eigenvalues, or constructing a state orthogonal to all lower states before applying the inverse-Hamiltonian loss.","core_discovery":"The central claim is that the inverse Hamiltonian method, when used as an unsupervised loss function for a fully connected deep neural network, turns the target bound state of the radial Dirac equation into the lowest eigenstate of the modified operator H_inv = (ε' - H'_Dr)^{-1}, so that energy minimization no longer falls into the Dirac sea. For excited states, the paper demonstrates two routes: adjusting the parameter ε' to lie between adjacent eigenvalues, or projecting out all lower states via orthonormalization before minimizing the inverse-Hamiltonian expectation value. Both routes reproduce benchmark energies and wave functions for the Coulomb potential and for Woods-Saxon potentials, with the inverse Hamiltonian method being simpler to use and the orthonormal method converging faster but accumulating error from lower states.","pith_inferences":["If the log-mesh discretization is made Hermitian by introducing a suitable weighted inner product, the variational guarantee of the inverse-Hamiltonian loss would rest on a rigorous foundation; otherwise the observed agreement with benchmarks is empirical rather than guaranteed.","The approach could be extended to non-spherical potentials by letting a single DNN output both radial components, but the paper's Appendix shows that small-component accuracy suffers when it contributes little to the norm, suggesting a balanced loss would be needed.","Counting nodes in the large component, a technique the paper uses to identify principal quantum numbers, could be automated inside the loss to make the method fully self-tuning for excited states."],"forward_implications":["The method provides a parameter-free variational route to relativistic bound states without the need for kinetic balance or special basis sets.","The computational cost scales as O(X M^2) with epochs X and mesh points M, potentially outpacing the O(M^3) eigen-solvers for large meshes.","The same inverse-Hamiltonian loss can be adapted to relativistic two-body Dirac equations, as the paper notes, since those reduce to a Schrödinger-like eigenvalue problem in the center-of-mass frame.","The orthonormal method's accuracy depends on the accuracy of lower states, so improving lower-state precision automatically improves higher excited states."],"supporting_citations":[{"why":"Supplies the underlying fully connected deep neural network method for ground and low-lying excited states in the nonrelativistic case, which this paper extends to the Dirac equation.","marker":"[28]"},{"why":"Introduces the inverse Hamiltonian method for iteratively solving the Dirac equation, the central loss-function mechanism used here.","marker":"[34]"},{"why":"Provides the analytical hydrogen-atom energies and wave functions used as the first benchmark.","marker":"[38]"},{"why":"Provides the generalized pseudospectral method serving as the numerical benchmark for the Woods-Saxon potential calculations.","marker":"[40]"},{"why":"Supplies the Woods-Saxon potential parameters for 16O and 208Pb used in the nuclear test cases.","marker":"[39]"}],"fun_headline_variants":["Neural net solves Dirac equation via inverse Hamiltonian trick","Deep learning avoids Dirac sea to compute Dirac excited states","Inverse-Hamiltonian loss trains DNN for Dirac ground and excited states","DNN solves Dirac equation without variational Dirac sea failure","Inverse Hamiltonian method unlocks Dirac equation for neural networks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The discretized radial Dirac Hamiltonian on the logarithmic mesh is treated as a Hermitian matrix in the inverse-Hamiltonian loss, but the paper never specifies a discrete inner product that would make the finite-difference derivative matrix Hermitian, so the variational guarantee that should protect against collapse is not rigorously established for those calculations.","fun_headline_variants_meta":{"raw":{"variants":["Neural net solves Dirac equation via inverse Hamiltonian trick","Deep learning avoids Dirac sea to compute Dirac excited states","Inverse-Hamiltonian loss trains DNN for Dirac ground and excited states","DNN solves Dirac equation without variational Dirac sea failure","Inverse Hamiltonian method unlocks Dirac equation for neural networks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000733,"raw_usage":{"total_tokens":3199,"prompt_tokens":783,"completion_tokens":2416,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":399,"completion_tokens_details":{"reasoning_tokens":2335}},"tokens_in":399,"tokens_out":2416,"duration_ms":18480,"temperature":1.0,"reasoning_tokens":2335,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:48:12.984756+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the discrete spectrum of H_inv on the log mesh with the same derivative matrix and boundary conditions, using a dense linear solver; if spurious eigenvalues appear below the target state energy, the loss would converge to the wrong state on some initializations, directly contradicting the paper's consistency claim.","supporting_citations":[{"cited_title":"Armstrong, Lloyd, Relativistic Effects in Atomic Fine Structure, J","cited_arxiv_id":null,"evidence_quote":"Introduces the inverse Hamiltonian method for iteratively solving the Dirac equation, the central loss-function mechanism used here."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the analytical hydrogen-atom energies and wave functions used as the first benchmark."}],"review_version":1}