{"id":"29fdf5c0-ae94-4eb9-9f21-0ea667b8ca1b","arxiv_id":"2608.08137","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":6,"one_line_summary":"A fully convolutional neural functional learns the one-body direct correlation of 3D lattice hard cubes from Monte Carlo data and beats lattice fundamental measure theory on bulk and wall observables.","lead":"The paper trains a neural network to act as a density functional for 3D hard cubes on a lattice, using complete 3D density maps instead of cutting them into small windows. It improves predictions of pressure and wall structure over an analytical theory, while pair-structure results are mixed.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified: the central EOS and wall-profile claims are supported by independent GCMC data, and the finite 7x7x7 receptive field is a disclosed limitation that the reported accuracy suggests does not undermine the tested observables.","rationale":"I read the paper in good faith and find the central argument sound. The neural functional is trained on GCMC data in randomized potentials and tested on independent GCMC data for the bulk equation of state and a planar wall; the reported accuracy of the neural model on these observables is internally consistent and supported by the figures and RMSE numbers. The finite 7x7x7 receptive field is the most plausible structural limitation, but it does not rise to a load-bearing concern because the tested observables are reproduced accurately with that field and because the authors explicitly disclose it. The reader's weakest assumption identifies the same architectural feature, but I do not agree that this assumption threatens the central claim; it is a controlled limitation. The reason for the CONDITIONAL verdict—lack of code and data—is a reproducibility issue, not a defect in the physics or the derivation. Therefore the reader's verdict should remain UNCHANGED, and a concrete test with a larger receptive field is the most valuable check that could elevate or qualify the central claim in a future revision.","tokens_in":12501,"tokens_out":22236,"duration_ms":254461,"concrete_test":"A useful verification step: retrain the same architecture with a larger effective receptive field, e.g., 9x9x9 or 11x11x11 first-layer kernels, or by adding a second finite-range spatial convolution. Compare the resulting homogeneous equation of state and planar-wall profile against the independent GCMC data. If predictions are essentially unchanged, the 7x7x7 field is not limiting the central claims; if they differ materially, the finite receptive field becomes a substantive caveat that should be quantified and reported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"No significant scientific objection to the central claim is identified. The paper's core results—markedly improved homogeneous equation of state and planar hard-wall profile relative to the Lafuente-Cuesta functional—are backed by independent GCMC comparisons, and the neural model is evaluated in genuinely self-consistent 3D profile calculations. The 7x7x7 receptive field is a real architectural limitation and is explicitly acknowledged in the Conclusions, but the paper's own tests show that the network nevertheless reproduces the wall contact peak, the packing oscillations, and the bulk equation of state with high accuracy. There is no evidence in the paper that longer-ranged direct correlations are needed for these particular observables; if anything, the favorable wall and EOS results suggest the local neighborhood captures the correlations relevant to the central claims. The pair-structure comparison is honestly reported as mixed for both functionals, and the authors do not overclaim neural superiority there. The principal weakness is not a flaw in the argument but the absence of released code, trained weights, and training data, which prevents independent reproduction of the numerical results; this is a reproducibility concern rather than a correctness concern, and it is appropriately reflected in the CONDITIONAL verdict.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces a fully convolutional neural approximation to the one-body direct-correlation functional c^(1)[ρ] for parallel hard cubes of side length three on a three-dimensional simple cubic lattice. The network is trained on 393 GCMC density profiles generated in randomized external potentials, using a Bernoulli mask on output sites to make full-profile training practical and a projection of the first-layer kernels onto the 48-element cubic point group to enforce exact equivariance. The learned functional is inserted into the Euler–Lagrange equation and solved self-consistently for the homogeneous equation of state, a planar hard-wall slit, and the anisotropic pair distribution around a fixed particle, and the results are compared with independent GCMC data and with the Lafuente–Cuesta lattice functional. The central reported results are a large improvement in the homogeneous packing-fraction isotherm (RMSE 1.99e-4 versus 1.36e-2) and an accurate wall density profile, together with a more nuanced, direction-dependent pair-structure comparison.","tokens_in":12795,"tokens_out":13837,"duration_ms":149555,"significance":"If the results are taken at face value, this is a significant methodological advance in learned classical density functional theory: it is one of the first demonstrations of full three-dimensional profile training and inference, and it shows that convolutional evaluation combined with stochastic output masking is practical at 50^3 field sizes. The data pipeline is careful: long GCMC production runs, twenty replicas per training field, an explicit validation split, independent GCMC checks for the equation of state, wall, and test-particle problems, and replica-resampling error propagation. The authors also report limitations honestly, including the absence of a guaranteed free-energy functional, the finite 7x7x7 receptive field, the training-density cutoff, and the mixed pair-structure accuracy. The finite receptive field is a legitimate architectural constraint, but the paper's own wall and equation-of-state results suggest that it does not undermine the central claims for the tested observables.","major_comments":[],"minor_comments":[{"comment":"The bulk packing fraction for the test-particle comparison is reported as η_b = 0.345 in the text but as η_b = 0.3 in the Fig. 4 caption; because the neural and Lafuente–Cuesta chemical potentials are matched to this value, please correct the discrepancy and state explicitly which value was used.","section":"§IV.C and Fig. 4"},{"comment":"The RMSE values 1.99e-4 and 1.36e-2 are quoted as being evaluated over the 'common supported range', but that range is not defined; please specify the density interval, the number of comparison points, and how the common range is determined.","section":"§IV.A and §V"},{"comment":"The text states that the GCMC data span βμ ∈ [−8, 1.5] and discusses extrapolation beyond η = 0.60, but the figure panels show βμ only down to 0; please extend the axes or explicitly state the plotted range so that the extrapolation claim is visible and checkable.","section":"§IV.A and Fig. 1"},{"comment":"There is a typo in 'we restrict ourself to the convolutional counterpart'; this should read 'we restrict ourselves'.","section":"§III.B"},{"comment":"The exclusion of target sites with c^(1) ≤ −10 is described as removing hard-wall sites and regions with insufficient statistics, but its possible effect on predicted low-density regions is not discussed; please add a sentence explaining how this threshold may influence profiles where the true c^(1) is below −10.","section":"§III.C"},{"comment":"The manuscript contains no data, code, or trained-model availability statement; for a machine-learning functional paper, please state whether the training profiles, GCMC data, trained weights, and evaluation scripts will be released.","section":"Data and code availability"}],"recommendation":"minor_revision","confidential_remarks":"I agree with the stress-test assessment: the 7x7x7 receptive-field concern does not land as a blocker because the authors disclose it and their wall and equation-of-state tests show high accuracy for the observables they emphasize. The most useful improvement for the community would be a data/code availability statement; the η_b discrepancy between Section IV.C and the Fig. 4 caption should be fixed before publication. No concerns about scope or citation practice."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing you should know: this paper actually delivers a 3D full-profile neural c(1) functional that beats the analytical LC functional on bulk EOS and wall profiles, and the validation against GCMC is clean. The second thing: the lack of released code and training data is the only substantial problem; the physics and the ML pipeline are sound.\n\nWhat's new: the Bernoulli-masked full-profile training plus exact Oh symmetrization of first-layer kernels. The convolutional reformulation was already in Glitsch et al. by the same group, but here they make it practical in 3D without window extraction, and the equivariance projection is exact. The 393-profile training set from randomized potentials is careful, and the independent GCMC checks for EOS, walls, and test-particle g(r) give the claims real teeth. The RMSE numbers (1.99e-4 vs 1.36e-2) are convincing.\n\nSoft spots: the c(1) target is fitted to GCMC data, so agreement with simulation is partly interpolation. That's not a flaw in itself, because the validation observables are not training labels, and the test-particle comparison is genuinely independent. The finite 7x7x7 receptive field is a real limitation, but the authors disclose it and the results suggest it captures the relevant correlations for the tested observables. Pair-structure results are honestly mixed; they don't overclaim. The bigger issue is reproducibility: no code, weights, or data are released, so no one can verify the actual numbers without reimplementation. That warrants a conditional rather than unconditional accept.\n\nBottom line: for someone working in classical DFT or machine-learned functionals, this is worth reading. It deserves a serious referee, and the main request should be to release the trained model and the training pipeline.\n\nRecommendation: send it to peer review, with a request for code and data as a condition of acceptance.","headline":"A solid, honest demonstration that full-profile convolutional neural DFT works in 3D for lattice fluids; the main claims are supported by independent GCMC data, and the only real weakness is the absence of released code and data.","tokens_in":13297,"tokens_out":1879,"would_cite":true,"duration_ms":17859,"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":"A network learns the one-body direct-correlation functional of three-dimensional lattice hard cubes from full profiles and improves bulk and hard-wall predictions over the analytic lattice functional.","keywords":["neural density functional theory","classical density functional theory","lattice fluids","hard cubes","convolutional neural network","cubic equivariance","direct correlation functional","grand canonical Monte Carlo"],"falsifier":"Train the same pipeline with a $9\\times 9\\times 9$ (or larger) first-layer kernel on the same training corpus and compare wall contact peaks, directional pair distributions, and the equation of state against GCMC at matched bulk densities. A material and systematic improvement in accuracy with the larger field would show that the $7\\times 7\\times 7$ truncation, not training data or optimization, is the limiting assumption; alternatively, directly estimating $c^{(1)}$ sensitivity to density perturbations beyond three lattice spacings from GCMC would settle the same point.","tokens_in":12270,"feed_emoji":"🧊","tokens_out":8100,"duration_ms":75369,"temperature":0.7,"pith_summary":"This paper constructs a neural classical density functional for parallel hard cubes on a simple cubic lattice, acting directly on complete three-dimensional density profiles instead of pre-extracted local windows. The network represents the one-body direct-correlation functional $c^{(1)}[\\rho]$ as a fully convolutional map with a $7\\times 7\\times 7$ receptive field, projected onto all 48 rotations and reflections of the cubic point group so the density-to-$c^{(1)}$ map is exactly equivariant. A Bernoulli mask on the output sites keeps full-profile training practical. Against independent simulations, the learned functional reproduces the homogeneous equation of state and planar hard-wall density profiles substantially better than the analytic Lafuente-Cuesta lattice functional, while its accuracy for the anisotropic pair distribution is direction-dependent. The authors conclude that full-profile neural density-functional calculations in three dimensions are feasible, with accurate training data, thermodynamic consistency, and pair-structure fidelity as the main challenges for continuum extensions.","feed_headline":"Neural net models 3D hard-cube fluids better than theory","feed_subtitle":"On full 3D profiles, it cuts packing-fraction error by a factor of about 70 versus the standard lattice functional.","key_machinery":"The load-bearing construction is the fully convolutional network with a finite first layer of 32 cubic-symmetrized $7\\times 7\\times 7$ kernels followed by pointwise residual channel mixing. Averaging each kernel over the 48-element cubic point group collapses its 343 entries to linear combinations of 20 cubic-orbit density sums, guaranteeing equivariance without data augmentation. The second key piece is the Bernoulli-masked mean-squared-error loss over profile sites: the forward pass sees entire unexpanded profiles, while only a randomly retained subset of output sites contributes to each gradient update, restoring stochastic variation without storing overlapping windows. The network is trained on profiles from grand-canonical Monte Carlo in randomized external potentials, using the pointwise inversion $c^{(1)}_{\\mathrm{GCMC}} = \\ln\\rho + \\beta V_{\\mathrm{ext}} - \\beta\\mu$ as the supervised target.","core_discovery":"On its own terms, the paper claims that a purely data-driven $c^{(1)}[\\rho]$ can be learned directly on full $50\\times 50\\times 50$ lattice fields and, once solved self-consistently in the Euler equation, yields accurate equilibrium densities for a three-dimensional lattice fluid. Concretely, over the trained packing-fraction range the neural equation of state matches GCMC with RMSE $1.99\\times 10^{-4}$, compared with $1.36\\times 10^{-2}$ for the analytic Lafuente-Cuesta functional, and the neural wall profile tracks the contact peak and oscillations more closely than the analytic one. The test-particle pair distribution is reproduced in its principal shells by both functionals, with neither uniformly better across crystallographic directions. The paper interprets the bulk and wall gains as the practical demonstration, and the mixed structural results as evidence that accurate one-body profiles do not automatically imply accurate pair correlations.","pith_inferences":["Beyond the paper: replacing the cubic projection with O(3)- or E(3)-equivariant convolutions would be the natural continuum counterpart, and the same full-profile training scheme could make three-dimensional hard-sphere functionals practical.","Beyond the paper: learning an equivariant scalar excess free energy instead of $c^{(1)}$ would enforce a symmetric functional Jacobian and path-independent thermodynamics, addressing a limitation the paper names.","Beyond the paper: because the receptive field is the only controlled locality restriction, a systematic sweep of kernel sizes on identical training data would directly quantify how much longer-ranged direct correlations matter for bulk and wall observables."],"forward_implications":["Solving the learned Euler equation for a new external potential yields an equilibrium density directly from full profiles, so wall, cavity, and confinement predictions require no window extraction or reassembly.","The homogeneous equation of state from the network reproduces GCMC packing fractions to RMSE $1.99\\times 10^{-4}$ over the trained range, versus $1.36\\times 10^{-2}$ for the analytic functional.","The Bernoulli-masked full-profile training scheme should carry over to other lattice and, after gridding, continuum models, provided training profiles are sufficiently accurate and diverse.","The test-particle comparison shows that accurate bulk thermodynamics and one-body profiles do not guarantee accurate pair structure, motivating explicit pair-correlation regularization or sum-rule constraints in future functionals.","The method demonstrates some extrapolation beyond the training density cutoff, but the mean packing-fraction cutoff of 0.60 remains a stated controlled limitation."],"supporting_citations":[{"why":"Introduces the neural-functional method and the supervised target $c^{(1)}$ from the Euler-Lagrange equation, which this paper extends to full 3D profiles.","marker":"[20]"},{"why":"Supplies the conceptual basis for learning $c^{(1)}$ rather than a free energy, motivating the architecture choice.","marker":"[21]"},{"why":"Shows how to rewrite the windowed MLP as a fully convolutional network, enabling full-profile evaluation without window expansion.","marker":"[36]"},{"why":"Provides the analytic lattice fundamental-measure functional (LC) used as the reference in every comparison.","marker":"[9]"},{"why":"Gives the group-equivariant convolution construction used to enforce exact cubic equivariance of the first layer.","marker":"[41]"},{"why":"Establishes the test-particle method used to compute the anisotropic pair distribution function.","marker":"[43]"}],"fun_headline_variants":["Neural DFT bests analytic theory for 3D hard-cube fluids","Deep learning improves density functional for 3D lattice fluids","Neural functional cuts packing error 70x for 3D hard cubes","Machine-learned functional nails 3D hard-cube equilibrium densities","Neural DFT improves EOS and wall profiles in 3D lattice fluids"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The assumption that a $7\\times 7\\times 7$ density neighborhood around each site is enough to determine its direct correlation, so any correlation longer than three lattice spacings lies outside what the network can represent.","fun_headline_variants_meta":{"raw":{"variants":["Neural DFT bests analytic theory for 3D hard-cube fluids","Deep learning improves density functional for 3D lattice fluids","Neural functional cuts packing error 70x for 3D hard cubes","Machine-learned functional nails 3D hard-cube equilibrium densities","Neural DFT improves EOS and wall profiles in 3D lattice fluids"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000481,"raw_usage":{"total_tokens":2387,"prompt_tokens":963,"completion_tokens":1424,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":579,"completion_tokens_details":{"reasoning_tokens":1329}},"tokens_in":579,"tokens_out":1424,"duration_ms":11470,"temperature":1.0,"reasoning_tokens":1329,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T00:21:59.389190+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same pipeline with a $9\\times 9\\times 9$ (or larger) first-layer kernel on the same training corpus and compare wall contact peaks, directional pair distributions, and the equation of state against GCMC at matched bulk densities. A material and systematic improvement in accuracy with the larger field would show that the $7\\times 7\\times 7$ truncation, not training data or optimization, is the limiting assumption; alternatively, directly estimating $c^{(1)}$ sensitivity to density perturbations beyond three lattice spacings from GCMC would settle the same point.","supporting_citations":[{"cited_title":"The Journal of Chemical Physics , volume =","cited_arxiv_id":null,"evidence_quote":"Shows how to rewrite the windowed MLP as a fully convolutional network, enabling full-profile evaluation without window expansion."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the group-equivariant convolution construction used to enforce exact cubic equivariance of the first layer."},{"cited_title":"Marques and Levin, Yan and Arenzon, Jeferson J","cited_arxiv_id":null,"evidence_quote":"Establishes the test-particle method used to compute the anisotropic pair distribution function."}],"review_version":1}