{"id":"8e72186a-e6a9-45a7-acf9-6bc433a76549","arxiv_id":"2501.18544","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"The authors introduce new equation-of-state-insensitive fits and neural-network models for the radius, logarithmic derivative, and effective gravity on rotating neutron star surfaces, with claimed sub-percent test accuracy.","lead":"This paper uses machine learning to build new universal formulas that describe the shape and surface gravity of rotating neutron stars, trained on 70 different equations of state. The goal is to let astronomers infer neutron star surfaces from X-ray data taken by NICER and future missions.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The EoS-universality claim is not tested against unseen EoS families: the 80/20 split is within EoSs, so leave-one-EoS-out validation is needed to support the claimed universal accuracy.","rationale":"The paper is a careful and substantial empirical study: 70 CompOSE EoSs, 42,694 equilibrium models, enthalpy-based surface localization, new polynomial and ANN fits, and comparisons against prior universal relations. I found no reason to doubt the numerical construction or the reported within-sample accuracy. However, the central claim of EoS insensitivity depends on validation against EoSs not used in training. The current test split is within EoSs, and LOOCV for the polynomial fits is pointwise rather than grouped by EoS, so the validation set shares EoS families with the training set. This is precisely the weakest assumption identified by the reader. A leave-one-EoS-out test is the natural, concrete check: it directly asks whether the relations generalize to a completely unseen equation of state. Until such a test is reported, the appropriate verdict is conditional acceptance of the results as accurate for the 70-EoS ensemble, with universality beyond that ensemble still unverified. I therefore agree with the reader's CONDITIONAL verdict and recommend no change.","tokens_in":59395,"tokens_out":3748,"duration_ms":45528,"concrete_test":"Perform leave-one-EoS-out cross-validation: for each of the 70 EoSs, retrain the polynomial fits and all three ANNs on configurations from the other 69 EoSs only, then evaluate on every configuration of the held-out EoS. Report per-held-out-EoS maximum relative/residual errors against the claimed thresholds (R(µ): 0.25%, d log R/dθ: 8.36e-3, g(µ): 0.91%, and the Eq. (25)-(29) maxima). If one or more held-out EoSs exceed these thresholds substantially, the universality claim must be weakened to accuracy within the 70-EoS ensemble. If all held-out EoSs stay within thresholds, the claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that Eqs. (25)-(29) and the ANN fits (31), (34), (36) are EoS-insensitive universal relations. The reported validation does not establish this. Appendix A states that for each EoS a random 20% of configurations is used as the test set, and the polynomial fits use LOOCV over individual data points. In both cases, every EoS family appears in the training set. The test errors therefore measure interpolation among configurations of already-seen EoSs, not prediction for a new EoS whose microphysics was never seen during training. Because the defining feature of a universal relation is that it holds for equations of state outside the training sample, the headline accuracies—0.25% for R(µ), 8.36e-3 residual for d log R/dθ, 0.91% for g(µ), and the ≤2.79–4.57% maxima for the polynomial fits—are conditional on the 70-EoS ensemble being representative of all physically plausible NS matter. A new EoS family outside this ensemble, for example one with a different phase transition or a more extreme stiffness, could leave the learned manifold, and nothing in the current validation would detect this. This is a load-bearing gap in the universality argument, not an internal inconsistency in the numerics.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper constructs a large dataset of 42,694 relativistic neutron-star equilibrium models (40,015 rotating) from 70 tabulated CompOSE equations of state, extracts the surface radius R(µ), its logarithmic derivative d log R/dθ, and the effective surface gravity g(µ), and proposes a set of EoS-insensitive relations. These include polynomial fits for the polar-to-equatorial radius ratio R(C,σ) (Eq. 25), eccentricity e(C,σ) (Eq. 26), maximum logarithmic derivative (Eq. 27), and polar and equatorial effective gravity (Eqs. 28 and 29), as well as feed-forward neural-network fits for R(µ) (Eq. 31), d log R(µ)/dθ (Eq. 34), and g(µ) (Eq. 36). The authors report test-set accuracies of 0.25% for R(µ), a maximum residual of 8.36e-3 for the logarithmic derivative, and 0.91% for g(µ), and they compare these fits against existing formulas from Morsink et al., AlGendy and Morsink, and Silva et al.","tokens_in":41,"tokens_out":8466,"duration_ms":153854,"significance":"If the proposed relations hold for equations of state outside the training ensemble, they would be practically valuable for pulse-profile modeling, atmospheric modeling, and future X-ray missions, and the paper would offer a substantial improvement in accuracy over existing surface fits for rapidly rotating stars. The analysis has several genuine strengths: a large and diverse numerical dataset, careful leave-one-out cross-validation for polynomial model selection, explicit consistency checks of the enthalpy-based surface localization, construction of synthetic surface data with verified accuracy, and detailed comparison with prior fits in the literature. The central weakness is that the universality claim is not validated against unseen equations of state: every EoS in the ensemble appears in the training set, so the reported accuracies measure interpolation within known EoS families rather than extrapolation to new microphysics.","major_comments":[{"comment":"The validation protocol does not test the central EoS-universality claim. In Appendix A, the test set is formed by randomly selecting 20% of configurations for each EoS (footnote 3), so every EoS family appears in the training set; the polynomial LOOCV in Sec. IV is leave-one-observation-out over individual stellar models, not leave-one-EoS-out. The reported accuracies (0.25% for R(µ), 8.36e-3 residual for d log R/dθ, 0.91% for g(µ), and the 2.79%–4.57% maxima for Eqs. (25)–(29)) therefore measure interpolation among configurations of already-seen EoSs. Because the defining feature of a universal relation is that it holds for equations of state not used in the fit, I request leave-one-EoS-out cross-validation (or at least a holdout of entire EoS families) with per-EoS error reporting, or a clear statement that the relations are conditional on the 70 selected EoSs being representative of all physically plausible neutron-star matter.","section":"Sec. V and Appendix A"},{"comment":"The ANN targets are normalized by endpoint quantities (Rpole, Req, (d log R/dθ)max, gpole, geq) that are not directly observable and must be supplied by the polynomial fits (25)–(29). The reported test errors are for the normalized quantities with exact endpoints. Because the endpoint fits carry maximum errors of 2.79% for R, 3.21% for the logarithmic-derivative maximum, 3.07% for gpole, and 4.26% for geq, the end-to-end surface and gravity errors will be larger than the reported 0.25% and 0.91%. This is a pipeline dependency rather than a circularity, but it should be quantified: either report errors when the endpoints are estimated from Eqs. (25)–(29), or state explicitly that the quoted accuracies assume exact knowledge of the endpoints.","section":"Sec. V A-C, Eqs. (30)-(36)"}],"minor_comments":[{"comment":"Equation (A5) contains a typo: the denominator reads \"max(xi) - mix(xi)\" and should read \"max(xi) - min(xi)\".","section":"Appendix A, Eq. (A5)"},{"comment":"In the two paragraphs following Eq. (29), the text refers to \"our regression model (28)\" when discussing the equatorial-gravity fit; the intended equation is (29) in both places.","section":"Sec. IV B"},{"comment":"The caption of Fig. 15 states that the bottom panel uses \"regression formula (27)\", but the panel shows the equatorial effective gravity geq(C, σ, e); it should cite Eq. (29).","section":"Fig. 15 caption"},{"comment":"The description of the train/test split is ambiguous about whether it is performed separately per EoS. The text in Sec. V says \"For the NS models associated with each EoS, we partition 80% of the data for training\" and the footnote mentions a random selection for each EoS, but this should be stated directly in the main text so the reader immediately knows that every EoS contributes to both training and test sets.","section":"Appendix A and Sec. V"},{"comment":"The dmax columns in Tables XIII and XV are reported as percentages (the text quotes 0.25% and 0.91%), while Eq. (A7) defines dmax as a dimensionless fraction. Please state the units in the table headers or convert the values consistently.","section":"Tables XIII and XV"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern in the reader's report is valid and is the main obstacle to acceptance. The manuscript would be a solid contribution if the authors add leave-one-EoS-out cross-validation or otherwise demonstrate that the fits retain their accuracy for EoS families not represented in the training sample. I would also ask the authors to address the endpoint-dependence of the ANN normalizations, since the quoted headline accuracies assume exact endpoint information. The GitHub repository is mentioned as a future resource; for reproducibility, the code and trained weights should be made available at the time of revision. No concerns about citation practice or scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: careful, genuinely useful fitting work, but the headline 'universal' claim is under-tested. The new fits—Rpole/Req, e(C,σ), (d log R/dθ)max, gpole, geq, plus the ANN surfaces for R(µ), the log-derivative, and g(µ)—are a real improvement on Morsink/AlGendy/Silva for the ensemble studied. The 70-EoS set is the largest used in this line of work, and the enthalpy-based surface localization and LOOCV selection for the polynomials is solid plumbing. If these relations hold up, they will be handy in NICER-style pulse-profile and atmosphere modeling.\n\nThe soft spot is exactly where the stress-test points: the validation does not justify the word 'universal'. The ANN test split is a random 20% of configurations from each EoS; the polynomial LOOCV leaves out individual points, not whole EoS families. So the quoted errors (0.25% on R(µ), 0.91% on g(µ), etc.) measure interpolation among already-seen EoSs. A new EoS with a different phase transition or stiffness could sit outside the learned manifold, and the current validation would not detect it. This is a load-bearing gap in the universality claim, not an arithmetic mistake. Leave-one-EoS-out cross-validation is the natural fix; it would show either that the fits generalize or how far the errors actually grow.\n\nA second, smaller caveat: the ANN fits take Rpole, Req, gpole, geq as inputs, which in practice come from the paper's own polynomial fits with max errors around 2.8–4.3%. The quoted ANN accuracies use exact values during training and testing, so end-to-end error for an observer will be a convolution of fit errors, not just the 0.25%. The paper gestures at this but doesn't quantify it.\n\nMinor issues: a few cross-reference typos, and the code/data repo is promised but not yet up. None of this changes the assessment.\n\nWho it's for: anyone modeling X-ray pulse profiles or cooling tails who wants an accurate oblate-surface recipe across a broad ensemble. It deserves peer review. I'd send it out, and I'd ask the referee to require either leave-one-EoS-out validation or an explicit limitation statement replacing 'universal' with 'accurate for the 70-EoS ensemble'. That is the difference between a solid fitting paper and a claim about unseen physics.","headline":"Careful and useful new fits, but the 'universal' claim is under-tested because the validation never leaves the 70-EoS ensemble.","tokens_in":60400,"tokens_out":3311,"would_cite":true,"duration_ms":37012,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["97.60.Jd","04.40.Dg","07.05.Mh"],"model":"deepseek-v4-flash","headline":"One set of formulas and neural networks describes neutron-star surfaces across 70 equations of state.","keywords":["neutron stars","equation of state","universal relations","machine learning","neural networks","rotating stars","surface gravity","radius"],"falsifier":"Hold out one complete equation of state, retrain the ANN and polynomial fits on the remaining 69, and evaluate on the held-out equation of state; if the maximum error on that equation exceeds the reported $0.25\\%$ for $R(\\mu)$, $0.91\\%$ for $g(\\mu)$, or $2.79\\%$/$4.57\\%$ for the global relations, the universality claim fails for truly unseen equations of state. A second check is observational: a sub-percent X-ray measurement of the surface radius or effective gravity that departs from Eqs. (31) or (36) by more than the quoted maxima would contradict the fits.","tokens_in":59194,"feed_emoji":"⭐","tokens_out":13359,"duration_ms":131741,"temperature":0.7,"pith_summary":"This paper sets out to show that the surface of a rotating neutron star, including its shape and the strength of gravity on it, can be described by formulas that barely depend on the unknown equation of state. Using an ensemble of 70 hadronic, hyperonic, and hybrid equations of state and rotating models up to near the mass-shedding limit, the authors derive polynomial universal relations for the polar-to-equatorial radius ratio, eccentricity, maximum logarithmic surface slope, and polar and equatorial effective gravity, with maximum errors below $4.6\\%$. They then train a feed-forward neural network to reproduce the full surface radius $R(\\mu)$, its logarithmic derivative, and the effective gravity $g(\\mu)$ as functions of compactness, spin, and eccentricity, reaching test-set errors below $0.25\\%$, a residual below $8.36\\times10^{-3}$, and $0.91\\%$, respectively. If these relations hold for real neutron stars, X-ray pulse-profile observations could model hot-spot emission with an accurate surface and gravity prescription without first choosing a particular equation of state.","feed_headline":"Neural net maps rotating neutron-star surfaces to 0.25 percent","feed_subtitle":"New EoS-insensitive radius, shape, and gravity fits could sharpen X-ray mass-radius estimates.","key_machinery":"The load-bearing object is the normalized surface description. For each star, the circumferential radius $R(\\mu)$ is mapped to $(R(\\mu)-R_{\\rm pole})/(R_{\\rm eq}-R_{\\rm pole})$, the logarithmic derivative is divided by its maximum, and the effective gravity is mapped to $(g(\\mu)-g_{\\rm pole})/(g_{\\rm eq}-g_{\\rm pole})$, so every configuration lives in the unit interval and the ensemble collapses onto a near-universal surface parameterized by compactness $C$, reduced spin $\\sigma$, eccentricity $e$ (or ratio $R$), and the angular coordinate $\\mu=\\cos\\theta$. The fits come from least-squares polynomial regression with leave-one-out cross-validation for the global relations, and from a five-hidden-layer feed-forward ANN with LeakyReLU activations and a sigmoid output, trained on Hermite-interpolated surface data. This machinery converts an apparently equation-of-state-dependent stellar shape into a single universal hyperstructure that can be evaluated without solving the field equations.","core_discovery":"On the paper's own terms, the central discovery is that the oblate surface of a neutron star is a nearly equation-of-state-independent hyperstructure once it is described in the right variables. The paper proposes that the polar-to-equatorial radius ratio $R(C,\\sigma)$, the eccentricity $e(C,\\sigma)$, the maximum logarithmic derivative $(d\\log R(\\mu)/d\\theta)_{\\max}(C,\\sigma,R)$, and the effective gravity at the pole and equator follow low-order polynomial fits in compactness $C=M/R_{\\rm eq}$ and reduced spin $\\sigma=\\Omega^2R_{\\rm eq}^3/GM$, with maximum relative errors of $2.79\\%$, $4.57\\%$, $3.21\\%$, $3.07\\%$, and $4.26\\%$. The full surface is then captured by ANN fits that normalize $R(\\mu)$ between $R_{\\rm pole}$ and $R_{\\rm eq}$, the logarithmic derivative by its maximum, and $g(\\mu)$ between $g_{\\rm pole}$ and $g_{\\rm eq}$; this normalization maps every star onto a common universal plane and lets the network reach test-set errors below $0.25\\%$ for $R(\\mu)$, a residual below $8.36\\times10^{-3}$ for the derivative, and $0.91\\%$ for $g(\\mu)$. The paper also shows these fits outperform previous surface and gravity formulas from the literature, especially for rapid rotation, and that the synthetic surface data satisfy the enthalpy condition $H(P)=0$ to high precision.","pith_inferences":["The reported test errors come from a random 20% split within the same 70 equations of state, so the fits have not been tested on an equation of state completely absent from training; a leave-one-equation-of-state-out evaluation would be the sharper test of true EoS insensitivity.","The same min-max normalization strategy could be carried over to other global parameters, such as moment of inertia or tidal deformability, or to differentially rotating stars, whenever a similar low-dimensional hyperstructure exists.","If future X-ray observations measure $R(\\mu)$ or $g(\\mu)$ at sub-percent precision, comparing them to Eqs. (31) and (36) is a direct consistency check: a deviation larger than the quoted maxima would signal either a new class of dense-matter equation of state or the breakdown of universality beyond the sampled ensemble.","The paper leaves open why such a universal hyperstructure exists; a theoretical derivation from the homologous-enthalpy structure of rotating stars would turn these empirical fits into a physical law."],"forward_implications":["Given a star's mass, equatorial radius, and spin, Eq. (31) yields the full surface radius $R(\\mu)$ to better than $0.25\\%$ without any equation-of-state input, which is enough to fix the oblate geometry used in pulse-profile ray tracing.","Eq. (27) supplies the maximum of the logarithmic derivative from $(C,\\sigma,R)$, so Eq. (34) can produce the surface slope needed for beaming-angle and light-curve calculations at arbitrary rotation.","Eqs. (28) and (29) give the polar and equatorial effective gravity, and Eq. (36) interpolates $g(\\mu)$ to $0.91\\%$, a quantity that hydrogen-atmosphere models of X-ray hot spots depend on.","The new fits outperform previous surface and gravity formulas in the test-set comparisons shown, with the largest advantage at high spin, and remain accurate at low spin, so they can replace those formulas in existing analysis pipelines.","When measurements of these surface observables reach the reported precision, deviations from the fits would translate directly into constraints on the equation of state of dense matter."],"supporting_citations":[{"why":"Supplies the earlier oblate-surface approximation and surface-radius fit against which the new R(mu) relation is compared.","marker":"[134]"},{"why":"Introduces the previous universal relations for surface effective gravity that the new pole/equator and g(mu) fits are benchmarked against.","marker":"[139]"},{"why":"Provides updated surface-fit coefficients and the slow/fast elliptical comparison baselines used for R(mu) and eccentricity.","marker":"[140]"},{"why":"The numerical solver that generates the rotating equilibrium models in the dataset.","marker":"[146, 147]"},{"why":"Reference for the stationary axisymmetric spacetime and hydrostationary equilibrium equations underlying the surface condition.","marker":"[148]"},{"why":"The tabulated equation-of-state database supplying the 70 hadronic, hyperonic, and hybrid models.","marker":"[7, 155]"},{"why":"Earlier supervised machine-learning universal relations whose approach the polynomial and ANN fits extend.","marker":"[110]"},{"why":"Universal approximation theorem that justifies using a feed-forward ANN for the surface fits.","marker":"[142]"}],"fun_headline_variants":["Neural nets map rotating neutron stars to 0.25% fits","Universal surface laws for neutron stars from ML","EoS-independent neutron star surface fits to under 1%","Spin and compactness alone predict neutron star surfaces"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim that these fits are universal depends on the 70 tabulated equations of state being representative of all physically plausible neutron-star matter and on the random 20% test split, which shares equation-of-state families with the training set, measuring true generalization to a new equation of state.","fun_headline_variants_meta":{"raw":{"variants":["Neural nets map rotating neutron stars to 0.25% fits","Universal surface laws for neutron stars from ML","EoS-independent neutron star surface fits to under 1%","Spin and compactness alone predict neutron star surfaces"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000291,"raw_usage":{"total_tokens":1823,"prompt_tokens":1188,"completion_tokens":635,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":804,"completion_tokens_details":{"reasoning_tokens":569}},"tokens_in":804,"tokens_out":635,"duration_ms":8332,"temperature":1.0,"reasoning_tokens":569,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T23:02:28.154703+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold out one complete equation of state, retrain the ANN and polynomial fits on the remaining 69, and evaluate on the held-out equation of state; if the maximum error on that equation exceeds the reported $0.25\\%$ for $R(\\mu)$, $0.91\\%$ for $g(\\mu)$, or $2.79\\%$/$4.57\\%$ for the global relations, the universality claim fails for truly unseen equations of state. A second check is observational: a sub-percent X-ray measurement of the surface radius or effective gravity that departs from Eqs. (31) or (36) by more than the quoted maxima would contradict the fits.","supporting_citations":[{"cited_title":"Measuring Nuclear Matter Parameters with NICER and LIGO/Virgo","cited_arxiv_id":"2002.03210","evidence_quote":"Universal approximation theorem that justifies using a feed-forward ANN for the surface fits."}],"review_version":1}