{"id":"d52ecd1f-e4d6-4508-86d3-9cbb648071fd","arxiv_id":"2411.14317","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A model-free neural network estimates local entropy production from trajectories, revealing that time-reversal breaking in a flocking model concentrates at the interface between order and disorder.","lead":"This paper trains a neural network to infer the probability current of a stochastic system directly from trajectory data, without knowing the underlying equations. It uses that current to map where entropy is produced and consumed in a flocking model, showing that time-reversal breaking concentrates at the flock's boundary.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No quantitative validation of the learned current: the N=64 EPR maps rest on an unverified ĝ, with finite-Δt bias and GNN expressivity errors unquantified.","rationale":"The analytical core of the paper is coherent: the variational objective (12) is a valid regression for g, the identities (7) and (11) follow from the path-measure and transport arguments, and the symmetries in the GNN are physically sensible. The weakness lies entirely in the step from 'the unique minimizer of (12) is g' to 'the trained network's output is close to g in the N=64 system.' The reader's weakest assumption captured this correctly. My stress-test sharpens it by naming the specific unquantified sources: the finite-Δt bias in (23), the restricted expressivity of the single-sum-pooling GNN (24), and the absence of any quantitative low-dimensional benchmark. The N=2 d=1 study is presented as intuition, but it could have been a quantitative validation by solving the 2D Fokker-Planck equation on a grid; its absence means the method has never been checked against a known answer. The high-dimensional results, including the central physical claim about spatial localization of EPR, therefore rest on an unverified approximation. I do not see an internal inconsistency that would force rejection; the method may well work. The appropriate disposition is conditional: keep the current verdict, require the proposed validation before accepting the physical conclusions. This does not change the reader's verdict, hence 'UNCHANGED' with agreement on the weakest assumption.","tokens_in":16075,"tokens_out":19318,"duration_ms":177837,"concrete_test":"Solve the stationary Fokker-Planck equation for the N=2, d=1 displacement system (60) on a fine grid to obtain exact ρ, g, and the exact EPR fields via (7) and (11); train the same network (or the GNN (24) with N=2) on data generated from (1) with the same Δt; compute pointwise relative L2 error and Pearson correlation between learned and exact fields, and repeat for Δt = 0.1, 0.01, 0.001 in units of 1/γ. If the error is large or the spatial pattern of ˙ssys differs from the exact solution, the N=64 conclusions are not established; if the fields converge as Δt→0, the method is validated in the tractable case.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that EPR can be estimated from trajectories alone via the current velocity g, and the headline N=64 result (TRS violations localized on the flock interface) is read off from the learned fields ˙ssys and ˙stot computed from ĝ through (7) and (11). These identities are exact only for the true g. The learned ĝ is obtained by minimizing the discrete loss (23) with the graph neural network (24), and three uncontrolled approximations intervene: (i) the discretization in (23) replaces the Stratonovich integral in (12) by a single-step midpoint rule, introducing a bias of order Δt in the minimizer; the value of Δt is not reported; (ii) the GNN (24) parameterizes ĝ as a single sum over pairwise terms followed by a readout, which may not represent the many-body score term −γv_∗^2 ∇_v log ρ in (4) in the 256-dimensional phase space; no capacity or convergence study is given; (iii) no quantitative validation is provided even for the N=2 d=1 system—the phase portraits in Fig. 2 are qualitative and no comparison is made against a numerical solution of the Fokker-Planck equation (60). Consequently, the spatial EPR patterns in Fig. 3 are not yet supported by evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This Letter proposes a machine-learning method to estimate probability current velocities g from stochastic trajectories of inertial active-matter systems, and uses them to compute local total (s_tot) and system (s_sys) entropy production rates. The authors derive exact relations: s_tot = |g^R|^2/(γv_*^2) (Eq. 7) and s_sys = ∇_v·g (Eq. 11), and show that g is the unique minimizer of the variational objective (12). They demonstrate the method on a two-particle Vicsek-like system, where a two-dimensional phase-space visualization is possible, and on a 64-particle system in 256-dimensional phase space. The headline physical claim is that entropy is produced and consumed on the spatial interface of a flock, with intermittent dynamics and 1/f noise in the EPR time series. The analytical derivations appear coherent, but the numerical evidence for the high-dimensional results is not validated quantitatively, and several implementation details needed to assess the accuracy of the learned g are omitted.","tokens_in":16425,"tokens_out":2641,"duration_ms":29118,"significance":"If the numerical estimates of g are reliable, the paper offers a genuinely useful tool: it gives a trajectory-only route to local entropy production in high-dimensional inertial systems, with exact formulas (7) and (11) and a rigorously characterized variational objective whose unique minimizer is g (SI Section D). The low-dimensional phase-space pictures provide intuitive validation of the physics. However, the strength of the physical conclusions rests on unverified approximations: the discrete loss (23), the neural-network ansatz (24), and the absence of any quantitative comparison against an exact or reference solution. The paper's main contribution would be much stronger with a convergence or validation study. In its current form, the central claim that the method accurately reveals the spatial structure of entropy production in the N=64 flocking system is not yet supported by the evidence presented.","major_comments":[{"comment":"The N=64 results are not validated against any exact or reference solution, nor are error bars or convergence studies provided. The EPR fields, time series, and power spectra are all computed from the learned graph neural network ĝ of Eq. (24), and the reliability of these quantities is simply assumed. Since the identities (7) and (11) are exact only for the true g, the central physical claim about spatial localization on the flock interface requires either validation of ĝ in this setting or at least a capacity/convergence study showing that the estimator is not introducing artifacts.","section":"§3 (High-dimensional system), Figs. 3–5"},{"comment":"The discrete loss (23) replaces the Stratonovich integral in (12) by a one-step central-difference approximation, which introduces a bias of order Δt in the minimizer. The value of Δt used for the reported simulations is never given, and there is no demonstration that the estimated EPR quantities are insensitive to Δt or to the number of trajectories n. This is load-bearing because the quantitative statements in Figs. 4 and 5, such as the distribution asymmetries and the power-law exponents, depend on the accuracy of the learned g.","section":"End Matter, Eq. (23)"},{"comment":"For the N=2, d=1 system, where a full validation is feasible, no comparison is made against a numerical solution of the Fokker–Planck equation (60) or against an independently computed g, e.g. from conditional averages of the SDE. The phase portraits and EPR maps in Fig. 2 and the statistics in Fig. 6 are qualitative; reporting a quantitative error, such as a relative L2 error in the learned g or in ⟨s_tot⟩ and ⟨s_sys⟩, would establish that the neural-network minimization actually recovers the true probability current.","section":"§3 (Low-dimensional system), Eq. (60)"},{"comment":"The power spectral density analysis quotes exponents p=1.16 and p=1.27 without error bars, fit ranges, or details of the fitting procedure. The claim that the EPR exhibits 1/f noise is a quantitative statement, and the current evidence does not establish its precision or its robustness with respect to the time-series length, binning, or detrending choices.","section":"§3, Fig. 5"}],"minor_comments":[{"comment":"The smoothing parameter β for the kernel is not reported; since β controls the sharpness of the interaction and affects the learning problem, a value should be specified for both the two-particle and the 64-particle simulations.","section":"SI §F.1, Eq. (58)"},{"comment":"There is a typo in the sentence 'we learn usingwa simple four-layer fully-connected network'; it should read 'using a simple'.","section":"SI §F.2.1"},{"comment":"The notation for the velocity scale is inconsistent: the main text uses v_* while the SI uses v_0 (e.g., Eq. (25) and Eq. (47)). This should be harmonized.","section":"Throughout, SI §C–E"},{"comment":"The text says T is arbitrary, yet the discrete loss (23) uses T=Δt only. It should be clarified how stationarity is used to justify this reduction and how the initial samples (x_0,v_0)∼ρ are generated in practice.","section":"End Matter, §Estimating the loss"},{"comment":"The references to 'the accompanying movie here' and 'the longer movie here' are not self-contained; URLs or permanent DOIs should be provided in the caption or supplementary material.","section":"Fig. 3 caption"}],"recommendation":"major_revision","confidential_remarks":"The analytical parts of the paper are solid and the proposed method is potentially significant, but the numerical demonstration currently lacks the validation needed to support the headline physical conclusions. I would be willing to reconsider after a revision that adds quantitative checks, including a comparison with an exact or reference solution in the two-particle case, a convergence study in the discrete timestep and network capacity for the 64-particle case, and error estimates for the reported power-law exponents."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper is worth your time, but read it with the numerical validation in mind. The analytic part is genuinely clean: they define the current velocity g as the conditional mean of ˙v, show it is the unique minimizer of a trajectory-based loss, and derive two local EPR formulas: ˙stot = |gR|^2/(γv*^2) and ˙ssys = ∇·g. Those derivations are sound, and the relations themselves are useful. The model-free twist—estimating g from trajectories only, without knowing the drift or noise amplitude—is a real step beyond their earlier PNAS paper.\n\nThe application to a Vicsek-like model is also interesting. The two-particle example gives an intuitive picture of alternating entropy production and consumption, and the N=64 results suggest that TRS breaking concentrates on the flock interface. If that is right, it's a nice observation. But the evidence for the N=64 claims is thin. There is no quantitative check that the learned g matches the true current in any test case. For N=2, they could solve the Fokker-Planck equation on a grid and compare g, ˙ssys, and ˙stot pointwise; they only show phase portraits. For N=64, they report no error bars, no convergence study, and no checks that the learned g satisfies known identities within statistical error. The one-point loss in (23) uses a midpoint discretization that introduces an O(Δt) bias, and Δt is not reported. The GNN parameterization is plausible but its expressivity for the many-body score is unexamined. The PSD exponents come with no uncertainty estimates. These are all fixable, but they are central to the paper's claim, not cosmetic.\n\nI also noticed that the SI has a few typos ('usingwa'), but that is minor.\n\nThe citation pattern is fine; they cite their own prior work where it is directly relevant, and the derivation of the loss stands on its own.\n\nBottom line: the paper deserves a serious referee. I would send it out, but with clear instructions to demand quantitative validation in the low-dimensional case, full training details, and error bars or alternative consistency checks for the high-dimensional results. As it stands, the physical conclusions are conditional on an unverified approximation. If the authors can close that gap, the paper would be a solid contribution to the active-matter and stochastic-dynamics communities.\n\nI'd happily bring it to the reading group to discuss the method; the conversation is worth having.","headline":"Solid analytic core and a plausible model-free estimator, but the high-dimensional results are under-validated; the paper needs a round of careful numerical checks before the conclusions can be trusted.","tokens_in":16867,"tokens_out":3312,"would_cite":true,"duration_ms":32794,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper establishes that entropy production rates can be learned from trajectory data alone and localizes the breakdown of time-reversal symmetry to the interface of a flock.","keywords":["entropy production rate","active matter","flocking","time-reversal symmetry","probability current","machine learning","graph neural network","nonequilibrium steady state"],"falsifier":"Run the same estimator on a system whose entropy production is exactly computable, such as the two-particle Vicsek pair reduced to one dimension by (60), solving the stationary Fokker-Planck equation for the true $g$ and comparing the learned $\\dot s_{\\rm sys}$ and $\\dot s_{\\rm tot}$ fields pointwise; if the learned fields deviate in low-probability regions or fail to reproduce the known macroscopic EPR, the 64-particle spatial-localization conclusions would be unsupported.","tokens_in":15892,"feed_emoji":"🐦","tokens_out":14752,"duration_ms":123383,"temperature":0.7,"pith_summary":"This paper claims that for inertial active systems, the entropy production rate (EPR) can be estimated directly from stochastic trajectories, with no knowledge of the forces or equations of motion. The route is to learn the probability current velocity $g(x,v)=f(x,v)-\\gamma v-\\gamma v_*^2\\nabla_v\\log\\rho(x,v)$, the conditional mean acceleration of particles on the nonequilibrium steady state; two identities tie $g$ to the local total EPR and the local system EPR. Applied to a Vicsek-like flocking model, the learned fields show that time-reversal symmetry is broken mostly at the interface between ordered flocks and the disordered gas, and that the system EPR is negative when particles align (order creation) and positive when they anti-align (order destruction). The paper's value is that it removes the need for a known dynamical model, opening entropy-production diagnostics to experimental trajectory data and giving spatially resolved maps of where a system is out of equilibrium.","feed_headline":"Trajectories alone reveal where flocking breaks time-reversal symmetry","feed_subtitle":"A neural network reads raw trajectories and shows where flocking creates or destroys order.","key_machinery":"The load-bearing object is the current velocity $g(x,v)=f(x,v)-\\gamma v-\\gamma v_*^2\\nabla_v\\log\\rho(x,v)$, which equals the conditional mean acceleration $\\langle\\dot v_t\\,|\\,(x_t,v_t)=(x,v)\\rangle$ and whose flow lines $(v,g)\\rho$ form the probability current on the steady state. The argument rides on two identities: $g$ squared gives the local total EPR and the velocity divergence of $g$ gives the local system EPR. The computational machinery is the variational objective $L[\\hat g]=\\frac{1}{T}\\mathbb{E}\\left[\\int_0^T|\\hat g(x_t,v_t)|^2\\,dt-2\\hat g(x_t,v_t)\\circ dv_t\\right]$, which has $g$ as its unique minimizer and is discretized over a single timestep, then minimized with a permutation-equivariant and translation-invariant graph neural network.","core_discovery":"On its own terms, the paper's central discovery is that two local measures of irreversibility in a nonequilibrium steady state are both determined by one vector field, the current velocity $g$. The total entropy production rate satisfies $\\dot s_{\\rm tot}(x,v)=|g^R(x,v)|^2/(\\gamma v_*^2)$ with $g^R(x,v)=-g(x,-v)$, and the system entropy production rate satisfies $\\dot s_{\\rm sys}(x,v)=\\nabla_v\\cdot g(x,v)$. Because $g$ is the unique minimizer of a variational objective built from trajectory increments, it can be learned without specifying the active force $f$, and the authors do so for a 256-dimensional Vicsek-like system with a graph neural network. The learned fields indicate that entropy is produced and consumed on the spatial interface of a flock as alignment and fluctuation create and destroy order, with total EPR spikes marking flock breakup and mergers.","pith_inferences":["If the learned current velocity is faithful at $N=64$, the paper's localization picture implies that the thermodynamic cost of maintaining a flock is paid at its boundary, so entropy production in larger flocks should scale with interface length rather than system volume.","The variational objective itself contains no noise-strength parameter, so $g$ can be learned even when friction or noise amplitude is unknown; converting $g$ into an EPR with physical units through (7) and (11) would then require estimating $\\gamma$ and $v_*$ separately.","A finite-size test of the two-particle mechanism is natural: with open boundaries, where collisions are not forced by periodic conditions, the heavy negative tail of $\\dot s_{\\rm sys}$ should become less prominent if it is collision-driven.","Since the probability flow $\\dot x=v,\\ \\dot v=g$ preserves the steady state, the learned $g$ is also a generative model of the nonequilibrium process, not just an EPR estimator."],"forward_implications":["Entropy-production diagnostics become available for systems whose dynamics are unknown, such as experimental active matter or animal groups, provided trajectories can be tracked.","The two identities give per-particle decompositions of both EPRs, so the spatial location of time-reversal symmetry breaking can be visualized even in a 256-dimensional phase space.","In the flocking model, particles on the boundary of a flock drive almost all entropy production; deep inside an aligned flock the system EPR vanishes.","The sign of the system EPR distinguishes order creation from order destruction, while the total EPR does not carry that signed information.","The EPR time series show heavy-tailed statistics and power spectra compatible with $1/f$ noise, indicating intermittency tied to flock formation and breakup."],"supporting_citations":[{"why":"Defines the stochastic entropy of a trajectory and its integral fluctuation theorem, the basis of the system entropy production rate in (9)-(11).","marker":"Seifert, 2005"},{"why":"Establishes the Gallavotti-Cohen-type symmetry and large-deviation functional through which time-reversal symmetry breaking defines the total EPR.","marker":"Lebowitz and Spohn, 1999"},{"why":"Supplies the path-measure entropy-production fluctuation relation used to write the total EPR as a Kullback-Leibler divergence between forward and time-reversed path measures.","marker":"Crooks, 1999"},{"why":"Provides the path-integral Lagrangian representation of diffusion used to derive the total EPR identity (7) and recover the macroscopic total EPR (8).","marker":"Onsager and Machlup, 1953"},{"why":"Companion path-integral treatment for systems with kinetic energy, used in the same derivation.","marker":"Machlup and Onsager, 1953"},{"why":"Gives the fluctuation relations and reverse-time dynamics for diffusion processes behind the time-reversed equation (3).","marker":"Chetrite and Gawedzki, 2008"},{"why":"Introduces the canonical self-propelled particle model whose flocking dynamics are the subject of the numerical study.","marker":"Vicsek et al., 1995"},{"why":"Provides the variations on the Vicsek model, including the smooth alignment kernel used in the simulations.","marker":"Chaté et al., 2008"},{"why":"Prior deep-learning computation of EPRs in active matter that required knowing the dynamics; the model-free claim here is the new step.","marker":"Boffi and Vanden-Eijnden, 2024"},{"why":"Defines the graph network architecture used to estimate the current velocity with permutation equivariance and translation invariance.","marker":"Battaglia et al., 2018"}],"fun_headline_variants":["Neural net maps where flocks create and destroy order","Learning probability flows exposes flocking's entropy hotspots","From raw trajectories: where flocking defies time reversal","AI finds where flocking leaks entropy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the trained neural-network estimate of the current velocity is accurate enough in the 256-dimensional phase space that the computed entropy-production maps and their statistics reflect the true dynamics; the paper reports no validation against an exact solution, no error bars, and no convergence study.","fun_headline_variants_meta":{"raw":{"variants":["Neural net maps where flocks create and destroy order","Learning probability flows exposes flocking's entropy hotspots","From raw trajectories: where flocking defies time reversal","AI finds where flocking leaks entropy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000156,"raw_usage":{"total_tokens":1204,"prompt_tokens":916,"completion_tokens":288,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":532,"completion_tokens_details":{"reasoning_tokens":228}},"tokens_in":532,"tokens_out":288,"duration_ms":3544,"temperature":1.0,"reasoning_tokens":228,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:19:03.306381+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same estimator on a system whose entropy production is exactly computable, such as the two-particle Vicsek pair reduced to one dimension by (60), solving the stationary Fokker-Planck equation for the true $g$ and comparing the learned $\\dot s_{\\rm sys}$ and $\\dot s_{\\rm tot}$ fields pointwise; if the learned fields deviate in low-probability regions or fail to reproduce the known macroscopic EPR, the 64-particle spatial-localization conclusions would be unsupported.","supporting_citations":[{"cited_title":"Entropy production along a stochastic trajectory and an integral fluctuation theorem","cited_arxiv_id":null,"evidence_quote":"Defines the stochastic entropy of a trajectory and its integral fluctuation theorem, the basis of the system entropy production rate in (9)-(11)."},{"cited_title":"Fluctuation Relations for Diffusion Processes","cited_arxiv_id":null,"evidence_quote":"Gives the fluctuation relations and reverse-time dynamics for diffusion processes behind the time-reversed equation (3)."},{"cited_title":"Novel Type of Phase Transition in a System of Self - Driven Particles","cited_arxiv_id":null,"evidence_quote":"Introduces the canonical self-propelled particle model whose flocking dynamics are the subject of the numerical study."}],"review_version":1}