{"id":"cebfc335-d922-4c37-a040-5112a110c910","arxiv_id":"2507.03742","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Using normalizing flows and a neural network on Gaia DR3 data, the authors recover a local pattern speed of 28.2 km/s/kpc and a total matter density of 0.086 solar masses per cubic parsec within 1 kpc of the Sun.","lead":"A data-driven model uses 5.6 million Gaia stars to recover the Milky Way's gravitational potential within 1,000 light-years of the Sun. It reports a local pattern speed of 28.2 km/s per kpc and matter densities, including a non-detection of dark matter.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own diagnostics (Sec. 4.6) show the stationarity assumption fails at a ~20 Myr timescale and is imprinted by the phase spiral; because the quoted uncertainties exclude this systematic, the pattern speed and density errors are lower bounds.","rationale":"The paper is a careful and transparent application of a previously developed method to Gaia DR3, with a thoughtful ensemble-based uncertainty pipeline, released code and trained models, and honest internal diagnostics. The reader's weakest-assumption identification is exactly right: the single rigid-rotation stationarity condition in Eq. 5 is both structurally necessary and demonstrably violated. The paper's own Fig. 9 and Fig. 11 show residual non-stationarity at physically relevant timescales and localized on the phase spiral, and the uncertainty discussion repeatedly warns that algorithmic and stationarity errors are not included. This is not an external disagreement with consensus; it is an internal inconsistency between the method's central premise and the evidence presented. The most load-bearing consequence is that the headline pattern speed and density values, with their small statistical error bars, should be read as conditional on a stronger assumption than the data satisfy. A controlled mock test with injected non-stationarity is the right way to quantify the resulting bias, and until that is done the appropriate verdict remains conditional acceptance. No further objection rises to the same level: the selection-function modeling, open-cluster removal, and bootstrap/error-resampling pipeline are handled carefully, and the authors are explicit about the limitations of their errors.","tokens_in":26271,"tokens_out":2754,"duration_ms":37897,"concrete_test":"Run the public Deep Potential pipeline on a controlled mock dataset with a known potential, an equilibrium background DF, and an injected non-stationary perturbation matching the observed amplitude and phase-space structure of the Gaia phase spiral (e.g., a z-v_z spiral perturbation with ~20 Myr timescale). Vary the perturbation amplitude from zero up to the observed level and compare the recovered Omega_p and rho to the known truth. If, at the observed amplitude, Omega_p shifts by more than 0.1 km/s/kpc or rho by more than 0.01 M_sun/pc^3, the quoted central values are systematics-dominated and the stated uncertainties are confirmed lower bounds.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central inference — recovering the gravitational potential and pattern speed from a snapshot of positions and velocities — is built on the assumption that the tracer distribution function is exactly stationary in a single uniformly rotating frame (Eq. 5, Sec. 2.1). This is the structural condition that converts a six-dimensional phase-space snapshot into an overconstrained system determining Phi(x) and Omega_p. The paper demonstrates that this condition does not hold: Sec. 4.6 and Fig. 9 report residual non-stationarity with a characteristic timescale of ~20 Myr in the rotating frame, and Fig. 11 shows that the residual is concentrated in the z-v_z phase spiral. A ~20 Myr timescale is comparable to vertical dynamical times in the disk, so the missing time-derivative term in Eq. 5 is not a small perturbation. The loss function (Eqs. 8-9) minimizes the magnitude of (df/dt)_Omega, but when no exact solution exists, the optimizer will absorb the unavoidable mismatch into the free neural-network potential and Omega_p. This produces exactly the kind of unphysical density fluctuations seen in Figs. 7-8 and can bias the recovered accelerations and pattern speed. The uncertainty pipeline (Sec. 2.5, Appendix C) explicitly does not include this stationarity-violation systematic, so the stated Omega_p = 28.20 +/- 0.10 km/s/kpc and rho = 0.086 +/- 0.010 M_sun/pc^3 are internal consistency errors, not total errors. The central claim is therefore conditionally supported, not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper applies the 'Deep Potential' method (Green & Ting 2020; Green et al. 2023; Kalda et al. 2024) to 5.6 million upper-main-sequence stars from Gaia DR3 within 1 kpc of the Sun. The method represents the stellar phase-space distribution with a normalizing flow, models the gravitational potential with a neural network, and fits a single pattern speed Ωp for a rotating frame in which the distribution is assumed stationary. The authors report Ωp = 28.20 ± 0.10 km/s/kpc, a local total matter density ρ = 0.086 ± 0.010 M_sun/pc^3, and a dark matter density ρ_DM = 0.007 ± 0.011 M_sun/pc^3, along with acceleration and density maps. The paper explicitly acknowledges that the stationarity assumption is violated: Section 4.6 reports residual non-stationarity timescales of order 20 Myr and a phase-spiral imprint in the residuals, and Section 2.5 states that the quoted uncertainties are lower bounds that do not include the stationarity condition not holding.","tokens_in":26670,"tokens_out":10332,"duration_ms":110284,"significance":"This is a serious, data-rich application of a flexible non-parametric method. Strengths include the public release of code and trained models (Zenodo/GitHub), a smooth neural-network model of the Gaia selection function, careful treatment of open-cluster contamination, and an ensemble-based uncertainty pipeline that propagates several internal sources of error. If the systematic from non-stationarity can be quantified or controlled, the method could become a valuable tool for mapping the local potential and identifying disequilibrium features. At present, however, the headline numbers are conditional on an assumption that the paper itself demonstrates to be violated, so the central claim is not yet established.","major_comments":[{"comment":"The paper's own diagnostic in Section 4.6 shows that the stationarity assumption underlying the method is not satisfied: the residual (∂ ln f/∂t)_Ω has a characteristic timescale of ~20 Myr in the best-fit rotating frame, comparable to vertical dynamical times in the disk, and the residual is imprinted by the phase spiral (Fig. 11). Because the potential Φ(x) and pattern speed Ωp are determined by minimizing the magnitude of (∂f/∂t)_Ω (Eqs. 8-9), a non-stationary DF will bias the recovered potential and pattern speed: the flexible neural network can partially cancel the non-stationarity through spurious force terms. The uncertainty pipeline (Section 2.5, Appendix C) explicitly excludes this 'stationarity condition not holding' contribution, so the quoted Ωp = 28.20 ± 0.10 km/s/kpc and ρ = 0.086 ± 0.010 M_sun/pc^3 are internal consistency errors rather than total errors. To establish the central claim, the authors should quantify the bias using controlled mock data with known non-stationarity (e.g., bar, spiral arms, phase spiral) or add a calibrated systematic floor to all headline values; otherwise the abstract and Table 3 should be rephrased as conditional on exact stationarity.","section":"Section 4.6, Eq. (5)"},{"comment":"The volume-averaged density quoted in Table 3 is ρ(R0) = 0.086 ± 0.010 M_sun/pc^3, but the model density in the midplane shows fluctuations on the order of 50% (Fig. 8), and the phase-spiral residual (Fig. 11) is coherent in the z-v_z plane, which is directly relevant for the vertical density profile. The uncertainty in Table 3 includes only the internal resampling/model-seed variations from Section 2.5; it does not include the coherent non-stationarity signature. The paper therefore provides no evidence that the fluctuations average out unbiasedly, and the density result is not robust until this systematic is quantified.","section":"Section 4.4, Table 3"},{"comment":"The corotation inference R_c = v_circ/Ωp = 8.4 kpc = 1.02 R0 is drawn from Ωp = 28.20 ± 0.10 km/s/kpc without propagating the stationarity-violation systematic. Given that Ωp is defined as the rotation speed that minimizes non-stationarity in a 1 kpc snapshot and that the residual is dominated by the phase spiral, the identification of this fitted quantity with a physical pattern speed, and the corotation conclusion, are not yet supported. Please either include a systematic error term in Ωp or remove/condition the corotation statement.","section":"Section 4.3"}],"minor_comments":[{"comment":"The text states that the characteristic timescales are 5 Myr and 20 Myr for the rotating and non-rotating cases, respectively, while the Figure 9 caption gives 20 Myr and 5 Myr for the same two cases; please reconcile this discrepancy.","section":"Section 4.6 and Figure 9 caption"},{"comment":"The text 'm = 222 (∼ 4 million)' should read 'm = 2^22' (the superscript appears to be missing in the typography).","section":"Section 2.4"},{"comment":"The sentence 'We can compute the matter density ρ = ∇²ρ/(4πG)' contains a typo; it should be ρ = ∇²Φ/(4πG).","section":"Section 4.4"},{"comment":"The expression for binary cross-entropy in Eq. (12) appears to have the wrong sign: the sum as written is maximized when Sθ matches the labels, so the argmin should be applied to the negative of the sum, or the expression should be written with a minus sign.","section":"Section 3.3, Eq. (12)"},{"comment":"The abstract and Table 3 present Ωp and ρ without the caveat given in Section 2.5 that the quoted errors are lower bounds and exclude the stationarity-condition systematic; consider adding a qualifying phrase to avoid overstatement.","section":"Abstract and Table 3"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for an astrophysics journal and the method is promising. The central issue is the unquantified systematic from the violated stationarity assumption, which the authors themselves document. I would support publication after the authors either provide a simulation-based calibration of the induced bias or add a systematic floor to the headline uncertainties. The transparency of Sections 2.5 and 4.6 is commendable, but the abstract and conclusion go beyond what the evidence currently supports."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: this is a careful, honest application of the Deep Potential method to real Gaia DR3 data. The core method is not new—Green & Ting and Green et al. developed it, and Kalda et al. tested it on N-body simulations. What's new here is the real-data implementation: a smooth neural-network selection function, an ensemble uncertainty pipeline, and the first claimed recovery of a local pattern speed and 3D potential from 5.6 million stars. On its own terms, the paper is well put together: the math is clear, the code and trained models are public, and the authors are unusually explicit about what can go wrong.\n\nThe main result is a local pattern speed of 28.2±0.1 km/s/kpc, a total density of 0.086±0.010 M_sun/pc^3, and a dark matter density of 0.007±0.011—a non-detection, which they say outright. The circular velocity at ~237 km/s agrees with the literature, and the comparison with binary pulsar accelerations is a nice external check, even if not yet constraining.\n\nNow the soft spots, in proportion. The central assumption is that the tracer population is statistically stationary in a single rotating frame. The paper itself shows this fails: Section 4.6 reports residual non-stationarity on a ~20 Myr timescale, and the residual is concentrated in the phase spiral. That timescale is comparable to the vertical dynamical time, so the missing time-derivative term in Eq. 5 is not a small perturbation. When no exact stationary solution exists, the optimizer will absorb the mismatch into the neural-network potential and the fitted Omega_p. The paper acknowledges this and explicitly states that the quoted errors are lower bounds, but the abstract and tables present these values without that caveat front and center. The density maps show fluctuations of order 50%, which they attribute to the model and to non-stationarity. A skeptical reader can't fully separate the signal from the systematics.\n\nIs this fatal? No. The method has been tested on simulations with non-stationarities, and the paper's own diagnostics are transparent. But the central numbers should be presented with a systematic error term that reflects the stationarity violation, and the authors should show that the recovered potential and pattern speed are stable when the non-stationary part of the DF is masked or downweighted. As is, the quoted uncertainties are internal consistency errors, not total errors.\n\nWho is this for? Anyone working on Milky Way dynamics, dark matter in the solar neighborhood, or normalizing-flow applications to galactic data. It deserves peer review, not desk rejection. I'd recommend conditional acceptance: ask for systematic uncertainties on the headline values and a discussion of how the phase-spiral non-stationarity biases the recovered potential. A serious referee will find things to push on, but the paper is honest and technically solid enough to engage with.","headline":"Solid first application of Deep Potential to real Gaia data, with honest caveats—but the stationarity violation means the headline uncertainties are lower bounds.","tokens_in":27189,"tokens_out":2475,"would_cite":true,"duration_ms":27536,"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 data-driven \"Deep Potential\" analysis of 5.6 million Gaia stars claims to recover the Milky Way's gravitational potential in a 1 kpc sphere around the Sun, finding a local pattern speed of 28.20 ± 0.10 km/s/kpc and a total matter…","keywords":["Galactic dynamics","Gravitational potential","Normalizing flows","Neural networks","Milky Way pattern speed","Local dark matter density","Gaia DR3","Solar neighborhood"],"falsifier":"The strongest direct test would be to run the identical pipeline on a high-resolution N-body simulation of a Milky-Way-like galaxy with a known potential and known non-axisymmetric features, and check whether it recovers the input pattern speed and density within the quoted uncertainties; disagreement at the level of the residuals seen in Figure 8 would indicate the method's systematics dominate over the statistical error bars.","tokens_in":26093,"feed_emoji":"🌌","tokens_out":4021,"duration_ms":33482,"temperature":0.7,"pith_summary":"The gravitational potential of the Milky Way is usually inferred under strong assumptions like axisymmetry or a fixed form for the distribution function. This paper claims that a nearly assumption-free recovery is possible from a single snapshot of six-dimensional Gaia phase-space data, provided the tracer population is statistically stationary in some rotating frame. Using 5.6 million upper-main-sequence stars within 1 kpc of the Sun, the authors fit a flexible distribution function with a normalizing flow and a flexible gravitational potential with a neural network, simultaneously solving for the rotation speed that makes the system appear most stationary. They recover a local pattern speed of $\\Omega_p = 28.20 \\pm 0.10\\,\\mathrm{km\\,s^{-1}\\,kpc^{-1}}$, a total matter density of $\\rho = 0.086 \\pm 0.010\\,M_\\odot/\\mathrm{pc}^3$, and a dark matter density of $\\rho_\\mathrm{DM} = 0.007 \\pm 0.011\\,M_\\odot/\\mathrm{pc}^3$ that is consistent with zero. The significance is that this is a path toward mapping the full three-dimensional matter distribution, including dark matter, from kinematic snapshots with minimal structural priors.","feed_headline":"Deep learning maps the Milky Way's local gravity field","feed_subtitle":"5.6 million Gaia stars yield a pattern speed of 28.2 km/s/kpc and a total density of 0.086 solar masses per cubic parsec.","key_machinery":"The central object is the stationarity condition combined with the collisionless Boltzmann equation: $$\\left(\\frac{\\partial f}{\\partial t}\\right)_\\$\\Omega$ = \\sum_i \\left[(u_i - v_i)\\frac{\\partial f}{\\partial x_i} + \\left(\\frac{\\partial \\Phi}{\\partial x_i} + w_i\\right)\\frac{\\partial f}{\\partial v_i}\\right] = 0.$$ The distribution function $f(\\mathbf x, \\mathbf v)$ is modeled nonparametrically with a normalizing flow (a chain of invertible transformations mapping a Gaussian to the observed phase-space density), and the gravitational potential $\\Phi(\\mathbf x)$ is modeled with a feed-forward neural network whose second derivatives give the density via the Poisson equation $\\rho = \\nabla^2\\Phi/(4\\pi G)$. The machine presses the potential and pattern speed against the sampled gradients of the flow, penalizing any residual non-stationarity and any negative density, which is how a six-dimensional snapshot becomes enough to solve for a three-dimensional field.","core_discovery":"The paper's central claim is that the three-dimensional gravitational potential, and hence the matter density, in a 1 kpc volume around the Sun can be recovered directly from a snapshot of stellar positions and velocities, without assuming axisymmetry or a parametric form for either the distribution function or the potential. The method treats the Milky Way as a collisionless system that is statistically stationary in a frame rotating with a single angular speed $\\Omega$ about the Galactic Center, and finds the potential and $\\Omega$ that best satisfy the resulting collisionless Boltzmann equation. Applied to 5.6 million upper-main-sequence stars from Gaia DR3, it yields a best-fit pattern speed $\\Omega_p = (28.20 \\pm 0.10)\\,\\mathrm{km/s/kpc}$, which makes the solar neighborhood appear most stationary and places the corotation radius near the Sun at about $1.02\\,R_0$. The same model gives a local total matter density of $\\rho = 0.086 \\pm 0.010\\,M_\\odot/\\mathrm{pc}^3$, and after subtracting a baryonic model, a dark matter density of $0.007 \\pm 0.011\\,M_\\odot/\\mathrm{pc}^3$, which is not a statistically significant detection of dark matter. The recovered acceleration field deviates from the best-fitting axisymmetric model by only about 2% in most of the volume, but shows a systematic azimuthal acceleration opposite to Galactic rotation, indicating non-axisymmetric structure such as spiral arms locally.","pith_inferences":["The stationarity assumption is the load-bearing premise, and the paper itself demonstrates it is violated: Section 4.6 shows residual non-stationarity imprinted by the phase spiral on a ~20 Myr timescale, so the quoted potential should be read as the best-fit single rotating frame, not a unique ground-truth potential.","The method could be stress-tested by applying it to a high-resolution simulation of a Milky-Way-like galaxy with a known potential, and checking whether the recovered $\\Omega_p$ and $\\rho$ reproduce the input values and whether the residuals trace the same phase-space features seen here.","A testable prediction is that a different tracer population (e.g., red giants rather than upper-main-sequence stars) should yield a different best-fit pattern speed, because different populations experience different resonances and non-stationarities; comparing them would quantify how far the single-rotating-frame picture is from reality."],"forward_implications":["If the method works as claimed, the local matter density and pattern speed can be recovered independently of parametric models of the Milky Way's disk, bar, spiral arms, and halo shape.","The recovered pattern speed of about $28\\,\\mathrm{km/s/kpc}$ implies corotation near the Sun, which would affect models of how the spiral arms and bar torque the local disk and how stars migrate radially.","The inability to detect dark matter locally ($\\rho_\\mathrm{DM} = 0.007 \\pm 0.011\\,M_\\odot/\\mathrm{pc}^3$) sets an upper bound on the local dark matter density that is consistent with but much less constraining than standard estimates.","The residual non-stationarity, shown to be imprinted by the Gaia phase spiral with a characteristic timescale of about 20 Myr, provides a new observable diagnostic for disequilibrium features in the solar neighborhood.","The same framework can map the potential beyond 1 kpc as more luminous tracers and future data releases are added, potentially revealing the three-dimensional dark matter distribution in the disk."],"supporting_citations":[{"why":"Introduces the Deep Potential concept of fitting a distribution function and potential together under a stationarity assumption.","marker":"Green & Ting (2020)"},{"why":"Lays out the theoretical motivation and demonstrates the method on toy models with observational errors and non-stationarities.","marker":"Green et al. 2023"},{"why":"Extended Deep Potential to concurrently fit the rotating frame and demonstrated recovery of acceleration and density in an N-body barred galaxy simulation.","marker":"Kalda et al. 2024"},{"why":"A parallel normalizing-flow-based approach applied to RGB stars within 4 kpc to estimate local dark matter density; this paper's main comparison and methodological contrast.","marker":"Lim et al. 2023"},{"why":"The review on Milky Way dynamics from which the pattern-speed comparison literature values are adapted.","marker":"Hunt & Vasiliev 2025"},{"why":"Provides the Gaia DR3 catalog that is the observational foundation of this work.","marker":"Gaia Collaboration et al. 2023a"},{"why":"Supplies the fixed values of the distance to the Galactic Center and its velocity relative to the Solar System used to set the rotating frame.","marker":"GRAVITY Collaboration et al. 2022"}],"fun_headline_variants":["AI maps local gravity from 5.6M Gaia stars","Deep learning recovers Sun's 3D gravity field","Neural net finds Milky Way's local pattern speed","AI infers galaxy's potential from star snapshot"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The stars in the sample are statistically stationary in a single frame rotating at one angular speed around the Galactic Center, but the paper itself shows the Milky Way is not perfectly stationary in such a frame (the phase spiral's ~20 Myr signature remains).","fun_headline_variants_meta":{"raw":{"variants":["AI maps local gravity from 5.6M Gaia stars","Deep learning recovers Sun's 3D gravity field","Neural net finds Milky Way's local pattern speed","AI infers galaxy's potential from star snapshot"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000757,"raw_usage":{"total_tokens":3447,"prompt_tokens":1107,"completion_tokens":2340,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":723,"completion_tokens_details":{"reasoning_tokens":2275}},"tokens_in":723,"tokens_out":2340,"duration_ms":21536,"temperature":1.0,"reasoning_tokens":2275,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:02:54.297751+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"The strongest direct test would be to run the identical pipeline on a high-resolution N-body simulation of a Milky-Way-like galaxy with a known potential and known non-axisymmetric features, and check whether it recovers the input pattern speed and density within the quoted uncertainties; disagreement at the level of the residuals seen in Figure 8 would indicate the method's systematics dominate over the statistical error bars.","supporting_citations":[{"cited_title":"Deep Potential: Recovering the gravitational potential from a snapshot of phase space","cited_arxiv_id":"2011.04673","evidence_quote":"Introduces the Deep Potential concept of fitting a distribution function and potential together under a stationarity assumption."},{"cited_title":"M., & Ghosh , S","cited_arxiv_id":null,"evidence_quote":"Extended Deep Potential to concurrently fit the rotating frame and demonstrated recovery of acceleration and density in an N-body barred galaxy simulation."},{"cited_title":"H., Putney , E., Buckley , M","cited_arxiv_id":null,"evidence_quote":"A parallel normalizing-flow-based approach applied to RGB stars within 4 kpc to estimate local dark matter density; this paper's main comparison and methodological contrast."}],"review_version":1}