{"id":"ccda957a-dac0-4ed6-8778-f7020fd9da75","arxiv_id":"2506.06623","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"In 1D hard-rod cDFT, Fourier neural operators learn the density-to-direct-correlation-function map more accurately than DeepONet variants and dense networks, with squared ReLU giving the best extrapolation.","lead":"This paper benchmarks several neural operators, including FNO and DeepONet variants, for learning the density-to-correlation-function mapping in 1D hard-rod classical density functional theory, and finds FNO the most accurate, especially on an extrapolation test set. It also shows that directly learning potential-to-density maps generalizes poorly, while using the learned functional inside the Euler-Lagrange equation with active-learning surrogates performs best.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'extrapolation' test uses Group V, a linear combination of the Group II and III training potentials (Eq. B.5), so it does not establish out-of-distribution generalization for FNO or operator learning.","rationale":"The paper's main contribution is a controlled comparison of operator-learning architectures on analytically solvable 1D hard-rod cDFT. The FNO result is internally consistent on the reported interpolation test; Table I and Fig. 8 show FNO dominating the other neural operators and most discrete baselines by a large margin. The exact hard-rod data, the explicit parameter settings, and the use of analytical reference solutions are real strengths, and I see no reason to suspect the ranking is a numerical artefact. The single most load-bearing weakness is the design of the 'new data' set. Eq. B.5 defines Group V as the sum of the Group II and Group III potentials with the same parameter ranges used in training. The t-SNE overlap in Fig. 2 confirms that the test inputs are not a new family. Therefore the leave-one-group experiment measures interpolation across known potential shapes, not extrapolation. All conclusions phrased in terms of 'extrapolation performance'—including the abstract, the scaling-law L1 analysis (Table II), and the superiority claim on c1-MSEND—inherit this limitation. This is the same assumption the reader identified; I agree with the reader's CONDITIONAL verdict. A genuinely novel test family, with normalization fit only on training data, would settle whether FNO's generalization advantage is real or an artifact of the test-set construction. Secondary issues (no repeated seeds, the quasi-local-DNN new-data contradiction between Section III.B and Table I, and normalization across all five groups) strengthen the case for conditional acceptance but do not by themselves overturn the benchmark ranking.","tokens_in":31185,"tokens_out":10354,"duration_ms":112646,"concrete_test":"Add a held-out potential family not expressible as a linear combination of Groups II and III, e.g., βVext(x)=A sin(2π x/L) or several Gaussian barriers, with Z-score normalization fit only on Groups I–IV; retrain the same eight models under the same hyperparameters and compare c1-MSEND. If FNO's margin over full-scale DNN and DeepONet stays comparable to Table I, the extrapolation claim survives; if its margin collapses or reverses, 'best extrapolation' must be downgraded to 'best interpolation within the training family.'","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that FNO generalizes best rests on the 'new data' (leave-one-group) test, but that test does not measure a genuine distribution shift. Group V (Eq. B.5) is βVext(x) = mg x − ε[(dhr/x)^3 + (dhr/(L−x))^3], i.e., exactly the sum of the Group III linear potential and the Group II attractive-wall potential, with the same parameter ranges used in training (mg ∈ (0.1,3), ε ∈ (0.1,2.2)). It is generated from the same basis functions as the training families, and Fig. 2 shows its t-SNE embeddings overlapping Groups II and III rather than occupying a new region. A model trained on Groups I–IV has therefore already seen the functional forms needed to represent every Group V input; the task is closer to interpolation between known families than extrapolation to a new potential class. This inflates FNO's apparent advantage on c1-MSEND (Table I: 7.76×10^−6) and the fitted L1 values in the Chinchilla analysis (Table II), so the paper's conclusion that operator learning generalizes better than discrete ML is overstrong. The issue is compounded by the normalization step (Section II D), which computes Z-score mean/standard deviation across all five groups, including the Group V test set; the test distribution thus contributes statistics to the training pipeline. A clean test with a genuinely novel potential family and training-only normalization is needed before 'extrapolation' claims can be credited.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper evaluates several neural operator architectures (DeepONet variants and FNO) for learning the mapping from one-dimensional density profiles to the one-body direct correlation function in classical density functional theory, using analytical hard-rod data as reference. It also examines the direct Vext-to-rho operator, operator-DFT solvers built from the learned rho-to-c1 map, and ALEC-GPR active learning. The models are compared through MSE losses on an in-group test set and a 'new data' test set, excess free-energy predictions, a Chinchilla scaling-law analysis, and transfer experiments across box sizes and to a Lennard-Jones fluid. The main claims are that FNO achieves the best overall accuracy, that squared ReLU is the best activation function, and that operator learning generalizes better than discrete ML approaches in extrapolation.","tokens_in":31498,"tokens_out":6274,"duration_ms":64808,"significance":"If the central claims were fully established, the paper would provide a useful benchmark for neural operators in cDFT: the rho-to-c1 task is validated against the exact Percus/Vanderlick analytical solution, excess free energies are compared with exact hard-rod results, and FNO is tested on multiple system sizes and on an LJ fluid including a phase-separation scenario. The negative result for the direct Vext-to-rho operator is honest and instructive, and the data/code availability statement is a strength. However, the extrapolation claim that distinguishes operator learning from discrete ML is not yet supported by the test design, and the single-run results provide no uncertainty quantification. These issues are fixable but require additional experiments and careful re-analysis.","major_comments":[{"comment":"The 'new data' extrapolation set is Group V, defined as beta Vext(x) = mg x - epsilon[(dhr/x)^3 + (dhr/(L-x))^3] with the same parameter ranges used in Groups II and III. This is a linear combination of the training families, so it lies in the convex hull of the training inputs rather than representing a genuine distribution shift. Figure 2(a-c) indeed shows Group V t-SNE embeddings overlapping Groups II and III. Consequently, the c1-MSEND values in Table I and the fitted 'New Data' L1 values in Table II do not measure out-of-distribution generalization; they measure interpolation among known functional forms. Since the abstract and Section IV base the 'extrapolation' conclusions on this test, the claim that FNO generalizes better than discrete ML is overstrong. Please add a held-out potential family that is not expressible as a sum of the training families (e.g., an oscillatory or localized barrier) and report the new-data results for that family.","section":"II D, Appendix B (Eq. B.5)"},{"comment":"The Z-score normalization computes the mean and standard deviation across all five groups, including the Group V test set. This allows test-set statistics to enter the preprocessing of the training pipeline, which can only improve the apparent new-data performance and weakens the extrapolation analysis. Normalization statistics should be computed from the training groups only (Groups I-IV) and then applied to any held-out set. This issue affects every reported 'new data' MSE and the scaling-law fits in Table II.","section":"II D"},{"comment":"All reported MSE values are single-run point estimates with no indication of seed variation. Because the paper's headline is a ranking of architectures (FNO best; GK-RMSCNN-DeepONet best among DeepONet variants), the results should be accompanied by means and standard deviations over several random initializations, or at least by a statement that the observed gaps are stable across seeds. Without this, Tables I and IV do not establish that the performance differences are reliable, especially for the smaller gaps among the DeepONet variants.","section":"Tables I-V"},{"comment":"The Chinchilla scaling-law analysis concludes that FNO has the lowest irreducible loss L1 and that operator learning generalizes better than discrete ML. This conclusion inherits the Group V test-set flaw: the 'New Data' column is fitted to the same hybrid test set, and the normalization leakage also affects these fits. The claim should be restricted to in-group generalization, or the analysis should be re-run on a properly held-out potential family before being used to support the generalization argument.","section":"III C, Table II"}],"minor_comments":[{"comment":"The text says the list of abbreviations and inputs is 'listed in the beginning of the paper,' but the list actually appears at the end after the references; please update the reference to its actual location.","section":"II B"},{"comment":"The statement that there is no 'new data' test set result for the quasi-local DNN method is contradicted by Table I, which lists c1-MSEND = 7.8e-3 for that model, and by Figure 7(d), which shows New Data curves for the quasi-local DNN; please reconcile the text with the table and figure.","section":"III B"},{"comment":"Several typos and formatting issues remain throughout, including 'convience', 'fir', 'preditions', 'L-GBFS' (should be L-BFGS), and missing spaces in table captions such as 'Variation ofMSE'; a careful proofreading pass is needed.","section":"Various"}],"recommendation":"major_revision","confidential_remarks":"This is a solid engineering benchmark that should be of interest to the cDFT/ML community once the extrapolation claim is fixed. The Group V test set and the normalization leakage are fixable within the manuscript's scope by adding a genuinely novel potential family and by computing normalization statistics only on the training groups. The single-run reporting also needs attention. I do not think rejection is warranted, but the current abstract and conclusion overstate the generalization result, so a major revision is appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this is a solid, well-executed benchmark for learning the rho-to-c1 map in 1D hard-rod cDFT, and the FNO-with-sRelu result is credible. But the 'extrapolation' test is weaker than advertised: Group V is a linear combination of Group II and III potentials, so it lives in the convex hull of the training families.\n\nWhat's new: the systematic comparison across FNO, four DeepONet variants, two DNN baselines, four activations, and the integration with ALEC-GPR. The ranking (FNO best, GK-RMSCNN-DeepONet best DeepONet) is new, and the negative result on direct Vext-to-rho learning is useful. The paper also validates learned c1 against exact Percus/Vanderlick solutions, so the central claim is externally grounded rather than self-referential.\n\nWhat it does well: the MSE tables are internally consistent and the Fex check via thermodynamic integration is a good sanity test. The additional experiments on different box sizes and on a 1D Lennard-Jones fluid provide genuinely out-of-distribution evidence that FNO transfers to a new interaction potential. That partially offsets the weak Group V design.\n\nSoft spots, in order of how much they matter. First, the leave-one-group test is not a genuine distribution shift. Group V (Eq. B.5) is just the sum of Group III's linear term and Group II's attractive-wall term, with the same parameter ranges as training, and the t-SNE shows overlap. So the c1-MSEND numbers and the fitted Chinchilla irreducible losses partly reflect interpolation, and the conclusion that operator learning generalizes better than discrete ML is overstrong. Second, Z-score normalization is computed across all five groups, including the test set, so the test distribution leaks into preprocessing. Training-only normalization is standard and should be used. Third, all results are single-run with no seed variation or error bars; for a benchmark claiming architectural superiority, that's a real gap. Fourth, there's an internal contradiction: the text says no new-data test result exists for the quasi-local DNN, but Table I lists one (7.8e-3). Fifth, code and data are 'will be available' rather than shipped; for a reproducibility-focused benchmark, that should be resolved at submission. The Chinchilla scaling-law parameters are fitted to the same loss curves they then explain; that's mildly circular, but the raw MSE ordering already supports the main claim, so it's a minor point.\n\nWho it's for: anyone working on ML surrogates for cDFT, or on operator-learning benchmarks for physics. The central architectural comparison is likely to be cited. My verdict: send it to peer review, but require the authors to add seed variability, fix the normalization, add a genuinely novel test potential, and release code/data.\n\nRecommendation: engage with it; it deserves a serious referee.","headline":"A credible FNO benchmark for cDFT, but the headline extrapolation test doesn't actually test extrapolation.","tokens_in":32135,"tokens_out":2672,"would_cite":true,"duration_ms":25993,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A Fourier neural operator learns the density-to-correlation map of a hard-rod fluid more accurately than DeepONet or dense networks, and generalizes better to unseen potentials.","keywords":["neural operators","classical density functional theory","Fourier neural operator","DeepONet","direct correlation function","hard-rod fluids","extrapolation","neural scaling law"],"falsifier":"Withhold an entire potential family during training—for example, exclude the constant-potential Group I or the power-law Group IV and test on it—or generate a truly novel family such as sinusoidal wells, then compare FNO against GK-RMSCNN-DeepONet and the quasi-local DNN on that set. If FNO's error advantage narrows to within one order of magnitude, the claim that operator learning generalizes far better than discrete functional learning loses its support.","tokens_in":30933,"feed_emoji":"⚛️","tokens_out":11940,"duration_ms":106168,"temperature":0.7,"pith_summary":"This paper asks whether neural operator architectures, which learn maps between whole functions rather than between fixed-length vectors, can supply the functional relationships that classical density functional theory (cDFT) needs, specifically the map from a fluid's density profile ρ(x) to its one-body direct correlation function c₁(x). Using exact analytical solutions for one-dimensional hard-rod fluids as training data, the authors compare eight machine-learning models—several DeepONet variants, a Fourier neural operator (FNO), and dense-network baselines—and report that FNO is the most accurate on every metric, including extrapolation to a held-out potential family. They further establish that a learned ρ ↦ c₁ operator, when inserted into the Euler–Lagrange equation, reproduces density profiles, whereas a directly learned potential-to-density mapping fails out of distribution. A scaling-law analysis is used to argue that operator learning has a lower irreducible error than discrete functional learning and therefore gains more from added data and compute. If right, this makes FNO a practical route to data-driven free-energy functionals for systems where no accurate analytic functional exists.","feed_headline":"Fourier operator beats 7 rivals at learning fluid density maps","feed_subtitle":"On exact hard-rod data, Fourier neural operators generalize best to unseen potentials and free-energy predictions.","key_machinery":"The load-bearing object is the operator $G_c : \\rho(x) \\mapsto c_1(x)$, an infinite-dimensional map approximated by a Fourier neural operator. FNO works by lifting the discretized input into multiple channels, passing them through Fourier layers in which a truncated fast Fourier transform multiplies the spectral coefficients by a trainable weight matrix before an inverse transform and nonlinear activation, then projecting back to the output function; the spectral multiplication acts as a global convolution, which is how the network captures the nonlocal functional dependence of $c_1$ on $\\rho$. The squared-ReLU activation $\\sigma(x) = x(2-x)$ for $0 < x < 2$, compactly supported and with non-negative derivatives, is identified as the best nonlinearity for extrapolation. Excess free energy is recovered by thermodynamic integration of $c_1$ along a density-scaling path, and density profiles are obtained by solving the Euler–Lagrange equation with the learned $c_1$, a pipeline the paper calls operator-DFT.","core_discovery":"The central claim is that a Fourier neural operator learns the density-to-correlation functional $G_c : \\rho(x) \\mapsto c_1(x)$ for a one-dimensional hard-rod fluid accurately enough to serve as the free-energy engine of a density functional calculation, and that it does so clearly better than DeepONet variants and dense neural networks. On the analytic reference data, FNO reaches a mean-squared error of about $5 \\times 10^{-7}$ on in-group tests and $7.8 \\times 10^{-6}$ on the held-out 'new data' set, roughly three orders of magnitude below the best DeepONet variant on extrapolation. When the learned map is inserted into the Euler–Lagrange equation, the predicted density profiles track the analytical solution, while a neural operator trained directly on the map $V_{\\rm ext} \\mapsto \\rho$ shows much larger extrapolation error, so the paper concludes that the $\\rho \\mapsto c_1$ route is the sound way to build operator-DFT. Free energies computed by thermodynamic integration of the learned $c_1$ reach $R^2 \\approx 0.999$ on all test sets, and the Chinchilla scaling law fits every method, with FNO showing the lowest irreducible loss among them.","pith_inferences":["The full-range versus quasi-local distinction suggests a natural stress test: near a critical point, where the correlation length diverges, quasi-local functionals break down; whether FNO's global convolutions keep their accuracy there is an open question this hard-rod benchmark cannot answer.","Because FNO's trainable parameters live in channels and Fourier modes rather than on the spatial grid, its near-flat error across box sizes $L = 4$ to $12$ hints that one trained operator could be reused at different system sizes with minimal retraining.","Training the operator directly on the integrated free energy, rather than on pointwise $c_1$ values, might sharpen free-energy predictions further; the paper's own results show $F_{\\rm ex}$ is predicted more accurately than the underlying $c_1$ field, indicating that integration averages out errors."],"forward_implications":["FNO is the best of the eight architectures for the $\\rho \\mapsto c_1$ map on both interpolation and extrapolation metrics, with held-out MSE near $8 \\times 10^{-6}$ compared with $10^{-3}$–$10^{-2}$ for the dense-network baselines.","Squared ReLU consistently outperforms ReLU, SeLU, and logistic activations for out-of-distribution prediction in nearly every model, making it the default activation choice for operator-DFT.","The $\\rho \\mapsto c_1$ operator, combined with solving the Euler–Lagrange equation, yields density profiles competitive with the numerical solver, and it serves as a fast sample generator for active-learning Gaussian-process surrogates.","Chinchilla scaling-law fits predict an irreducible new-data MSE of about $2.5 \\times 10^{-6}$ for FNO at infinite data and model size, more than two orders of magnitude below the fitted limits of the full-scale DNN and DeepONet baselines."],"supporting_citations":[{"why":"Analytical density profiles for one-dimensional hard rods; generates the reference data used to train and test all models.","marker":"[28]"},{"why":"Exact hard-rod c₁ functional and exact excess free energy; supplies the target operator and the ground truth for F_ex predictions.","marker":"[33]"},{"why":"Introduces the Fourier neural operator architecture whose performance is the paper's central result.","marker":"[35]"},{"why":"Defines the DeepONet operator-learning framework that the comparison variants are built on.","marker":"[34]"},{"why":"The Neural-DFT quasi-local DNN implementation used as the functional-learning baseline throughout the study.","marker":"[7]"},{"why":"The neural-functional approach for learning c₁ as a functional of ρ, whose philosophy the quasi-local DNN follows.","marker":"[19]"},{"why":"The Gaussian-process and ALEC surrogate for 1D hard-rod fluids that FNO-DFT is benchmarked against and used as a sample generator.","marker":"[24]"},{"why":"Source of the Chinchilla scaling law used to compare irreducible generalization limits across model families.","marker":"[30]"}],"fun_headline_variants":["Fourier neural operator wins hard-rod density functional race","FNO beats DeepONet in classical DFT free-energy accuracy","Operator learning: FNO leads in extrapolation for hard-rod fluids","Fourier nets ace density-to-correlation map in cDFT","Hard-rod DFT: FNO outperforms seven neural architectures"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The extrapolation verdict rests on a single held-out potential family, Group V, which is a linear combination of the Group II and Group III potentials and therefore lies partly inside the range the models have already seen; if a genuinely different potential class is encountered, FNO's measured extrapolation advantage could shrink.","fun_headline_variants_meta":{"raw":{"variants":["Fourier neural operator wins hard-rod density functional race","FNO beats DeepONet in classical DFT free-energy accuracy","Operator learning: FNO leads in extrapolation for hard-rod fluids","Fourier nets ace density-to-correlation map in cDFT","Hard-rod DFT: FNO outperforms seven neural architectures"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000271,"raw_usage":{"total_tokens":1716,"prompt_tokens":1119,"completion_tokens":597,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":735,"completion_tokens_details":{"reasoning_tokens":507}},"tokens_in":735,"tokens_out":597,"duration_ms":6757,"temperature":1.0,"reasoning_tokens":507,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:53:59.757321+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Withhold an entire potential family during training—for example, exclude the constant-potential Group I or the power-law Group IV and test on it—or generate a truly novel family such as sinusoidal wells, then compare FNO against GK-RMSCNN-DeepONet and the quasi-local DNN on that set. If FNO's error advantage narrows to within one order of magnitude, the claim that operator learning generalizes far better than discrete functional learning loses its support.","supporting_citations":[{"cited_title":"Santos-Silva , author P","cited_arxiv_id":null,"evidence_quote":"Analytical density profiles for one-dimensional hard rods; generates the reference data used to train and test all models."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Exact hard-rod c₁ functional and exact excess free energy; supplies the target operator and the ground truth for F_ex predictions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the DeepONet operator-learning framework that the comparison variants are built on."},{"cited_title":"Malpica-Morales , author P","cited_arxiv_id":null,"evidence_quote":"The neural-functional approach for learning c₁ as a functional of ρ, whose philosophy the quasi-local DNN follows."}],"review_version":1}