{"id":"50e9c57d-7864-4954-a14f-9c8f2a1eab0f","arxiv_id":"1908.06198","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A neural-network DFT functional trained with automatic differentiation enforces the exact relation between exchange-correlation energy and potential, improving 1D two-electron calculations including dissociation.","lead":"The authors train a neural network to act as the exchange-correlation energy functional in density-functional theory, with the potential derived automatically as its mathematical derivative. The approach is shown on simple one-dimensional two-electron systems, where it outperforms the standard local-density approximation and improves on the known failures of DFT for molecular dissociation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Inverse-KS label instability and the Exc>-0.55 outlier cut may remove exactly the strong-correlation cases the functional claims to learn; no label-quality check is reported.","rationale":"The reader's weakest assumption is exactly the one I find most load-bearing: the only source of exact v_xc labels is the numerically unstable inverse-KS inversion, and the paper's outlier-removal procedure is not shown to be unbiased. The paper's own Data section admits the instability, and the threshold Exc > -0.55 is a heuristic cut that is never justified in terms of v_xc accuracy. Since the whole method is trained to match these labels, any systematic bias in the cleaning step propagates directly into the learned functional and into the self-consistent results. The test set is intentionally uncleaned, which is honest but means that reported test errors cannot separate model error from label error. The H2 dissociation and 4-nucleus tests are good-faith out-of-distribution checks, but they also cannot resolve this label-quality question. I therefore agree with the reader that the paper should be CONDITIONAL: the methodological contribution is real, but the central quantitative claim needs a label-accuracy audit before it can be accepted as reproducing the exact exchange-correlation functional. My proposed check directly tests the bias by comparing independent inversions and retraining with different cleaning rules, which would settle whether the concern actually lands.","tokens_in":11680,"tokens_out":9811,"duration_ms":106979,"concrete_test":"Take about 100 systems spanning weak to strong correlation, including stretched H2 at R=10, 20, and 40 a.u. Compute exact densities by full configuration interaction; invert the KS equations with two independent solvers, e.g., octopus and a direct 1D inversion method such as the Zhao-Morrison-Parr style procedure, and compare v_xc, Exc, and Exc - integral(v_xc n). Report label differences versus Exc and versus R. Then retrain the kernel-size-30 model on (a) the original cleaned training set and (b) a set that retains all systems whose label differences pass a tolerance, or that removes the Exc > -0.55 cut. If the self-consistent total-energy MAE on the uncleaned test set or the H2 dissociation curve changes by more than the reported margin, the cleaning bias is confirmed; if the labels agree and the retrained results are unchanged, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the ML functional reproduces the exact xc energy and potential rests on the inverted KS labels. The Data section states that the inversion is 'known to be numerically unstable [48]' and that outliers from these instabilities were removed, with an additional cut on Exc > -0.55 a.u. The paper never quantifies the inversion error, never shows a reference comparison for v_xc, and never argues that the removed cases are a random subset. If instability is concentrated in strongly correlated, low-density, or dissociated configurations—the systems most relevant to the paper's motivation—then the training distribution is biased away from the target physics. The test set is deliberately uncleaned, so test errors conflate model error with label noise; and the H2 dissociation test (Fig. 4) probes configurations that may have been underrepresented after cleaning. The 'universal functional' claim therefore needs evidence that the labels are accurate and that the cleaning cut does not select on the physics to be learned.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript trains neural-network exchange-correlation (xc) functionals for one-dimensional two-electron systems, using exact ground-state densities and energies obtained by solving the two-particle Schrödinger equation and then solving the inverse Kohn-Sham problem to obtain xc energies and potentials. The functional is a scanning sum of local network outputs over density neighborhoods (kernel size κ), with the xc potential obtained by automatic differentiation, so the energy-potential relation of Eq. (2) is satisfied by construction. The authors train with a weighted loss on energy, potential, potential derivative, and the energy-potential integral, and they test the functionals in self-consistent Kohn-Sham calculations on held-out systems, on four-nucleus systems, on H2 dissociation, and on the homogeneous electron gas. They report that increasing kernel size reduces total-energy errors relative to a 1D LDA, that the dissociation curve improves with non-locality, and that the approach is a feasible route toward non-local xc functionals with local-scaling cost.","tokens_in":11866,"tokens_out":2854,"duration_ms":32498,"significance":"If the reported results are robust, the paper makes a useful conceptual contribution: it shows that a machine-learned xc functional can be trained simultaneously on energies and potentials, that automatic differentiation enforces the exact functional-derivative relation, and that the resulting functional can be used in self-consistent Kohn-Sham calculations while improving on LDA in a model system where LDA is known to fail. The out-of-distribution tests (four nuclei, H2 dissociation, homogeneous electron gas) are a commendable part of the evaluation, and the explicit use of a held-out test set is a strength. The central limitation is that the training labels come from a numerically unstable inverse Kohn-Sham inversion, and the paper does not yet demonstrate that the label-cleaning procedure is unbiased. Because the label-quality issue bears directly on the validity of the learned functional, the manuscript needs additional evidence before the 'universal functional' claim can be accepted.","major_comments":[{"comment":"The training labels for both Exc and vxc are obtained by solving the inverse Kohn-Sham problem in octopus, and the paper states that 'the inversion is known to be numerically unstable [48]' and that outliers from these instabilities were removed, together with an additional cut Exc > −0.55 a.u. No quantitative information is provided about the magnitude or structure of the inversion error, no comparison of the inverted vxc against any reference is shown, and no analysis is given of which systems are removed by the cleaning procedure. If the instabilities or the Exc cut preferentially remove strongly correlated, low-density, or dissociated configurations, the training distribution is biased away from exactly the physics the functional is meant to describe. This is load-bearing because the claim that the ML functional reproduces the exact xc energy and potential rests entirely on the quality of these labels. I ask the authors to quantify inversion errors (e.g., by convergence checks with respect to grid spacing and inversion algorithm parameters), show at least a small set of representative inverted potentials against independently validated references, and characterize the discarded systems to show that the cleaning is not selecting on the target physics.","section":"Data section, p. 2-3"},{"comment":"Table I reports mean absolute errors for the total energy relative to LDA for each kernel size, but no error bars, no seed variance, and no number of independent training runs are given. The text states that 'models with different hyperparameters were evaluated on a validation set of 250 systems' and that training was 'not completely converged at this stage,' with only the best model per kernel size continued to training and test evaluation. This makes the reported monotonic decrease of error with kernel size and the claimed optimal kernel size around 30 difficult to assess: a single favourable initialization or validation choice could drive the trend. I request repeated training runs over several random seeds for each kernel size, reporting the mean and standard deviation of the test MAE, and a statement of how many hyperparameter configurations were explored.","section":"Table I and Evaluation section, p. 4"},{"comment":"The HEG comparison is weakened by two issues that are acknowledged in the text but not adequately controlled. First, the machine-learned curves are shifted so that they are zero at zero density, a post-hoc adjustment that hides any bias in the network output for vanishing density. Second, the training data contains almost no samples with rs < 1 (as shown in the histogram in Fig. 5), so the behavior at high density is an extrapolation. The statement that 'functionals with larger kernel sizes still generalize on average far better' to the HEG is not supported by any error bars or by a quantitative comparison to the exact/LDA curve; moreover, the text notes that some large-kernel functionals show unphysical behavior. Please provide a quantitative error measure versus rs for each kernel size, with error bars over training runs, and discuss how the zero-density shift affects the comparison.","section":"Homogeneous electron gas, Fig. 5"},{"comment":"The abstract calls the trained network 'the universal exchange-correlation functional of density-functional theory,' but the demonstration is limited to one-dimensional two-electron systems with a specific nuclear potential form, and the Conclusions acknowledge that scaling to three dimensions requires new representations and substantial additional data. The overstatement is consequential because the phrase 'universal functional' invites claims of general applicability that the evidence does not support. Please temper the abstract and title-level claim to 'a step toward' or 'a proof of principle for' a universal functional, and state explicitly in the abstract that the results are for 1D model systems.","section":"Abstract and Conclusions"}],"minor_comments":[{"comment":"Typographical errors: 'fullfil' should be 'fulfill' and 'batchsizes' should be 'batch sizes' in the training paragraph on p. 4.","section":"p. 2"},{"comment":"The loss function combines five terms, but the text says 'weights α, β, γ and δ' are optimized; Eq. (5) indeed has exactly four weights. Please clarify the wording to avoid implying a fifth weight.","section":"Eq. (5)"},{"comment":"Table I only gives ratios MAE(ML)/MAE(LDA). Please also report the absolute MAE for each ML functional so the reader can judge the practical magnitude of the errors, and define the LDA reference error precisely in the table caption.","section":"Table I"},{"comment":"The data-generation description does not specify the distributions of nuclear charges Zk or the minimum separation between nuclei. These details matter for judging the diversity of the training set and for reproducing the results; please add them.","section":"Data section"},{"comment":"The caption should state explicitly that the ML curves are shifted to zero at zero density, as mentioned in the text, so that readers do not interpret the plotted offset as raw network output.","section":"Fig. 5 caption"}],"recommendation":"major_revision","confidential_remarks":"This is a borderline major-revision rather than reject: the core feasibility claim is credible and the authors have made a genuine effort to include out-of-distribution tests. The decisive missing piece is quantitative validation of the inverse-KS labels and of the cleaning procedure's selectivity; without that, the central claim rests on potentially biased training data. The abstract's 'universal functional' wording should also be softened. Given the paper's scope as a letter, these issues are addressable in a revision without changing the overall approach."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is enforcing v_xc = δE_xc/δn during training via automatic differentiation. Earlier ML-DFT papers fit energies or potentials separately; this one makes the functional physically consistent. The self-consistent KS results in 1D are credible: kernel-size scaling, the 2,000-system held-out test, and the out-of-distribution checks (4 nuclei, H2 dissociation, HEG) go well beyond the usual demonstration. The dissociation improvement over LDA with kernel size 30 is a real result, even in 1D.\n\nThe soft spot is the one the authors half-acknowledge: the training labels come from an inverse KS inversion they admit is numerically unstable. They remove outliers and cut all systems with Exc > -0.55 a.u. from training, but never quantify the inversion error or compare a single inverted v_xc against a reference. If the instability and the energy cut preferentially hit strongly correlated, low-density, or dissociated configurations, the training distribution is biased against exactly the physics the functional is meant to learn. The test set is deliberately uncleaned, so the reported errors conflate model error with label noise. The H2 dissociation curve looks good, but it may be probing a region that was largely filtered out. That doesn't sink the method, but it does mean the abstract's 'universal functional' is not supported by the 1D two-electron evidence.\n\nMinor issues: Table I has no error bars or seed variance, no code or data are released, and the HEG curves are shifted to zero energy at zero density, which blurs the zero-density behavior. These are small but worth addressing.\n\nOverall, the core methodological idea holds up and the feasibility claim is supported. This paper deserves a serious referee. I'd bring it to the reading group and would cite it if I worked on ML-DFT. The label-quality issue needs a direct answer—show that the cleaned training labels are accurate, or at least that the removed outliers don't select on the target physics.","headline":"A genuinely consistent ML xc functional enforced by automatic differentiation, with solid 1D self-consistent results whose main weakness is the unquantified inverse-KS label noise.","tokens_in":12446,"tokens_out":2738,"would_cite":true,"duration_ms":27945,"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":"This paper claims that a neural network can learn the exchange-correlation functional of DFT, with the energy and potential reproduced consistently, and demonstrates it in one-dimensional two-electron systems.","keywords":["density functional theory","exchange-correlation functional","machine learning","neural network","automatic differentiation","Kohn-Sham equations","strong correlation","molecular dissociation"],"falsifier":"Run the same network and loss on a fresh set of one-dimensional two-electron systems, but with training potentials computed by an independent exact method such as full configuration interaction; if the self-consistent energies are not systematically closer to exact than LDA on the original test set, the claim of a consistent and accurate learned functional fails.","tokens_in":11449,"feed_emoji":"⚛️","tokens_out":9019,"duration_ms":81198,"temperature":0.7,"pith_summary":"This paper claims that a neural network can be trained as the exchange-correlation functional of density-functional theory in a way that reproduces the exact exchange-correlation energy and the exact exchange-correlation potential at the same time. In Kohn-Sham theory the potential must be the functional derivative of the energy, a relation that many previous machine-learned potentials violate; the paper enforces it by automatic differentiation, so the model is internally consistent. The functional scans the density in a finite neighborhood, so it is non-local on the scale of the kernel but still has the computational scaling of a local approximation. Demonstrated on one-dimensional two-electron systems, the learned functionals improve on the local-density approximation for total energies, exchange-correlation potentials, and the dissociation curve of H2. The reader should care because this is a concrete path toward Kohn-Sham calculations that fix strong-correlation and delocalization errors without the cost of hybrid functionals.","feed_headline":"Machine-learned DFT functional beats LDA in strong-correlation tests","feed_subtitle":"It reproduces exact exchange-correlation energies and potentials, fixes H2 dissociation, and keeps local-functional cost.","key_machinery":"The central object is a kernel-based neural-network functional: a fully connected network that receives a block of $\\kappa$ density values centered on a grid point and outputs a local exchange-correlation energy, with the total energy obtained by summing over all grid points; $\\kappa$ controls the degree of non-locality, from 1 (an ML-LDA) to 180. The load-bearing relation is the exact identity $v_{\\rm xc}(r)=\\delta E_{\\rm xc}/\\delta n(r)$, imposed by automatic differentiation of the network energy rather than by fitting the potential separately. The training loss combines mean-squared errors for the energy, the potential, the spatial derivative of the potential, and the consistency of the energy with $\\int v_{\\rm xc}(r)n(r)\\,dr$. This combination is what allows the model to be highly non-local and still enter standard self-consistent Kohn-Sham iterations.","core_discovery":"The central claim is that imposing the exact mathematical link between energy and potential during training, by obtaining the potential as the automatic derivative of the network's energy, makes a machine-learned exchange-correlation functional usable in self-consistent Kohn-Sham calculations. On a test set of 2,000 one-dimensional two-electron systems outside the training set, the error in the total energy drops to as low as 6.5% of the LDA error as the kernel size is increased, and a network trained only on the energy fails badly on the potential while joint training recovers potentials close to the exact ones. The same functionals reproduce the 1D H2 dissociation curve far better than LDA, even beyond the non-locality range of the kernel. These results are offered as evidence that a consistent, non-local ML functional is feasible and can address errors such as static correlation without abandoning Kohn-Sham efficiency.","pith_inferences":["An immediate testable extension is to use the same consistency loss as a filter for any proposed ML functional: compute the derivative of the network's energy and reject models whose implicit potential disagrees with any separately fitted potential before running self-consistent calculations.","The kernel-size study can be read as a diagnostic of how much non-locality each physical effect actually needs; measuring the required $\\kappa$ per system type could guide data generation in three dimensions.","Since the paper finds larger kernels sometimes produce unphysical dissociation curves, a natural next step is to enforce exact constraints, such as the correct constant-density limit, as soft or hard terms in the loss, which the results here suggest may be necessary for large kernels rather than merely helpful."],"forward_implications":["The same training strategy can be taken to three dimensions using coupled-cluster, full configuration-interaction, or quantum Monte Carlo data, the route the paper names for creating a universal functional.","Because the potential is the functional derivative by construction, trained functionals cannot suffer from the energy-potential inconsistency of 'stray' potentials; remaining failures must come from the training data, the architecture, or the kernel range.","Non-locality over a few atomic units is enough to cure the qualitative failure of LDA for dissociation in the tested model, suggesting that many static-correlation errors do not require infinite-range memory.","Functionals trained on inhomogeneous systems can generalize to the homogeneous electron gas when the training set includes the relevant density range, which matters for use in solids."],"supporting_citations":[{"why":"Supplies the Hohenberg-Kohn existence theorem that justifies learning a functional of the ground-state density.","marker":"[1]"},{"why":"Supplies the Kohn-Sham equations and the total-energy expression in which the learned functional is used.","marker":"[2]"},{"why":"Supplies automatic differentiation, the mechanism used to compute the exchange-correlation potential as the exact derivative of the network's energy.","marker":"[43]"},{"why":"Supplies the earlier neural-network exchange-correlation potentials for molecules that the paper contrasts with, since they do not impose the energy-potential relation.","marker":"[45]"},{"why":"Defines the one-dimensional two-electron model with softened Coulomb interaction that serves as the test bed for strong correlation.","marker":"[46]"},{"why":"Provides the real-space density-functional code used to generate exact exchange-correlation potentials via the inverse Kohn-Sham problem.","marker":"[47]"},{"why":"Documents the numerical instability of the inverse Kohn-Sham problem, which motivates the paper's removal of outliers from the training data.","marker":"[48]"},{"why":"Supplies the one-dimensional LDA reference and the homogeneous-electron-gas data used for performance comparisons.","marker":"[54]"}],"fun_headline_variants":["ML functional nails exact energy and potential in DFT","ML functional matches exact exchange-correlation via auto-diff","Non-local ML functional fixes strong correlation at LDA cost","1D strong-correlation test: ML beats LDA with kernel size","Auto-diff links energy and potential in ML DFT functional"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The trained functional inherits whatever errors are in the exchange-correlation potentials used as training labels, which are obtained by numerically inverting the Kohn-Sham equations and then filtering out cases flagged as unstable; if that inversion is biased in ways that survive the filtering, the entire functional is biased.","fun_headline_variants_meta":{"raw":{"variants":["ML functional nails exact energy and potential in DFT","ML functional matches exact exchange-correlation via auto-diff","Non-local ML functional fixes strong correlation at LDA cost","1D strong-correlation test: ML beats LDA with kernel size","Auto-diff links energy and potential in ML DFT functional"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000983,"raw_usage":{"total_tokens":4136,"prompt_tokens":872,"completion_tokens":3264,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":488,"completion_tokens_details":{"reasoning_tokens":3182}},"tokens_in":488,"tokens_out":3264,"duration_ms":24066,"temperature":1.0,"reasoning_tokens":3182,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:53:11.478508+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same network and loss on a fresh set of one-dimensional two-electron systems, but with training potentials computed by an independent exact method such as full configuration interaction; if the self-consistent energies are not systematically closer to exact than LDA on the original test set, the claim of a consistent and accurate learned functional fails.","supporting_citations":[{"cited_title":"Inhomogeneous electron gas,","cited_arxiv_id":null,"evidence_quote":"Supplies the Hohenberg-Kohn existence theorem that justifies learning a functional of the ground-state density."},{"cited_title":"Self-consistent equations in- cluding exchange and correlation effects,","cited_arxiv_id":null,"evidence_quote":"Supplies the Kohn-Sham equations and the total-energy expression in which the learned functional is used."},{"cited_title":"How to tell when a model Kohn–Sham potential is not a functional deriva- tive,","cited_arxiv_id":null,"evidence_quote":"Supplies automatic differentiation, the mechanism used to compute the exchange-correlation potential as the exact derivative of the network's energy."},{"cited_title":"Automatic differentiation in pytorch,","cited_arxiv_id":null,"evidence_quote":"Supplies the earlier neural-network exchange-correlation potentials for molecules that the paper contrasts with, since they do not impose the energy-potential relation."},{"cited_title":"TensorFlow: Large-scale machine learning on heterogeneous systems,","cited_arxiv_id":null,"evidence_quote":"Defines the one-dimensional two-electron model with softened Coulomb interaction that serves as the test bed for strong correlation."},{"cited_title":"Completing density functional theory by machine-learning hidden messages from molecules","cited_arxiv_id":"1903.00238","evidence_quote":"Provides the real-space density-functional code used to generate exact exchange-correlation potentials via the inverse Kohn-Sham problem."},{"cited_title":"Reference electronic structure calculations in one dimension,","cited_arxiv_id":null,"evidence_quote":"Documents the numerical instability of the inverse Kohn-Sham problem, which motivates the paper's removal of outliers from the training data."}],"review_version":1}