{"id":"99156c7c-004a-4e3b-a12d-32d6c7604529","arxiv_id":"2607.29364","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An online-trained Fourier Neural Operator with a memory window reproduces linear and nonlinear Landau damping in a 1D electrostatic fluid model and interpolates across initial amplitudes.","lead":"This paper trains a neural network to replace the missing heat-flux physics in a one-dimensional plasma fluid model, and shows the model then reproduces both the early linear damping and the later nonlinear trapping of plasma waves. It is a step toward fast fluid simulations that retain memory of kinetic effects, which could lower the cost of large-scale plasma modeling.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Solver-coupled effective flux undercuts the 'broadly applicable closure' claim; all tests use the training solver/grid/k, and upward amplitude extrapolation fails.","rationale":"I read the paper in good faith. The online-training strategy is clearly motivated, the numerical scheme is described in detail, and the paper honestly reports the limitations that are most damaging to the broad reading of the claim. The reader's conditional verdict is appropriate. My stress-test focuses on the same load-bearing spot: the closure is only ever tested inside the exact differentiable solver, grid, and wavenumber used to train it. The paper explicitly concedes scheme dependence in §3.2 and §4.4, and the abstract's 'outside the training set' generalization is not uniform—above-range amplitudes fail. A cross-solver/cross-grid deployment test would settle whether the learned object is a physical closure or a solver-specific correction. The nonzero-mean ∂xq design choice (§3.2) reinforces this: the output can act as a global source/sink that no periodic heat flux can represent, so the 'closure' is partly an engineered energy-exchange term. None of this invalidates the paper's narrower contribution—a demonstration of stable online-trained closures for one-dimensional single-mode Landau damping—but it does mean the results should not be read as a broadly applicable plasma closure without further evidence. I also note that no code or data are released, which makes the proposed independent-solver test the clearest path to resolving the concern.","tokens_in":27324,"tokens_out":14582,"duration_ms":165853,"concrete_test":"Take the trained multi-amplitude FNO and deploy it in an otherwise identical fluid solver with a different spatial discretization (e.g., spectral derivative or 4th-order WENO) or a different grid (Nx=256) at the same k and A values; compare WE(t) and per-moment errors to kinetic ground truth for A=1e-3, 1e-2, 1e-1. Also compute the spatial average of the FNO's ∂xq output over a nonlinear run: if |mean(∂xq)| contributes materially to d<p>/dt while the kinetic ⟨∂xq⟩ is near zero, the closure is operating as a global source rather than a heat-flux closure. If cross-solver deployment fails or the mean-source term is large, the abstract and conclusions should be narrowed to 'solver-coupled online-trained closure'.","verdict_should_be":"UNCHANGED","load_bearing_attack":"To support the central claim, the learned ∂xq must function as a physical moment-to-flux closure, not as a correction to one numerical method. The paper's own limitations say otherwise: §3.2 states the trained FNO 'is tightly coupled to this particular numerical method, which may limit its portability to other fluid codes', and §4.4 states the specific effective flux is 'not a unique or scheme-independent object'. Every online test deploys the closure in the same differentiable SSPRK3 + 2nd-order-central-difference solver, on the same 128-point grid and the same k=0.35 single-mode cos(kx) family. The only out-of-training amplitude successes are below the training minimum; A=0.2 and 0.3 fail (§4.3, Fig. 5). Thus the evidence establishes an online-trained, solver-embedded correction for one mode family, not a broadly applicable plasma closure. Additionally, because the network outputs ∂xq directly with no zero-mean constraint, it can inject a spatially uniform source/sink into the pressure equation (§3.2), a term no periodic heat flux can produce; this further separates the learned object from a standard closure and makes scheme-dependence expected.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a data-driven closure for the one-dimensional electrostatic fluid equations, replacing the heat-flux-gradient term in the pressure equation with a Fourier Neural Operator (FNO) trained online: the loss is computed on trajectories of a differentiable SSPRK3 solver closed by the FNO, rather than on pointwise fits to kinetic heat-flux data. The closure is non-Markovian, acting on a trailing K=50-step memory window of the resolved moments, with an amplitude-conditioning channel logA_scale. The authors train single-amplitude closures for linear (A=1e-3) and nonlinear (A=1e-1) Landau damping, and a multi-amplitude closure trained on A=1e-3 to 1e-1 at four points per decade. They report that the online-trained closures reproduce the kinetic electric-field energy evolution in both regimes, that the multi-amplitude closure interpolates across held-out amplitudes and extrapolates to amplitudes below the training range, and that the learned flux should be interpreted as an effective, scheme-dependent closure rather than the true kinetic heat flux. Comparison is made against an offline-trained FNO (which fails in the matched configuration) and against a calibrated Hammett-Perkins closure (which fails in the nonlinear regime). A sensitivity/ablation analysis is used to argue that the closure computes a genuine moment-to-flux relation with physically structured memory use.","tokens_in":27648,"tokens_out":3498,"duration_ms":39626,"significance":"If the central claims hold, the paper would be a useful demonstration that an online-trained, non-Markovian neural closure can reproduce both linear and nonlinear Landau damping with a single model, and that a posteriori training yields stable deployment in the same solver. The memory-window design, the amplitude-conditioning channel, and the explicit ablation/HSIC analysis are valuable contributions, and the authors are appropriately candid about the scheme dependence of the learned object. However, the paper's headline claims currently outrun the evidence: the generalization claim is only true below the training envelope, all reported errors come from single runs without seed variability, and the closure is only tested in the exact solver/grid/wavenumber used for training. The Hammett-Perkins benchmark is calibrated to the kinetic damping rate, so its linear-rate agreement is by construction. These issues are load-bearing for the abstract's wording about generalization and broad applicability, though the underlying method remains plausible and worthy of revision.","major_comments":[{"comment":"The abstract states the closure 'generalises to initial perturbation amplitudes outside the training set,' but the evidence shows this is only true for extrapolation below the training minimum. At A=0.2 and A=0.3, which lie above A_max=0.1, the damping rate departs sharply from the kinetic value and the WE(t) envelope is over-predicted (Fig. 5a and the last two panels of Fig. 6). The paper itself describes this as a 'failure to extrapolate above the training envelope.' The claim in the abstract and Section 1 should be qualified to specify interpolation and downward extrapolation, and the failure mode should be acknowledged in the abstract.","section":"§4.3, Fig. 5(a), Fig. 6"},{"comment":"The closure outputs ∂xq directly with no zero-mean constraint, so it can inject a spatially uniform source/sink into the pressure equation. On a periodic domain, no true heat-flux gradient can have a nonzero spatial mean, since ∫∂xq dx=0. The authors justify this as allowing net heating from Landau damping, but it means the learned object is not a physical closure in the standard sense; it is a solver-embedded effective term. This is consistent with the paper's own admission that the flux is 'not a unique or scheme-independent object' (§4.4) and that the FNO is 'tightly coupled to this particular numerical method' (§3.2). This combination undercuts the 'broadly applicable learned plasma closure' framing. The authors should either impose the zero-mean constraint and assess the consequences, or explicitly reframe the contribution as a solver-specific correction and provide a portability te","section":"§3.2 and §4.4"},{"comment":"The linear-regime claim of reproducing Landau damping is stated without the important test-window caveat. The fitted damping rate in the test window (t>75) has relative error 1.82e-1 (18%), and the electric-field error crosses 10% only 1.4 ω_p^{-1} after leaving the training window, reaching 50% at 45 ω_p^{-1} and 100% at 55 ω_p^{-1}. The paper does acknowledge this drift as 'expected,' but the abstract's unqualified 'reproduces both linear and nonlinear Landau damping' masks a substantial generalization gap in the linear case. The authors should either report the test-window degradation in the abstract/claims or define the reproduction claim only over the trained window and the nonlinear long-time behavior, where the performance is genuinely better.","section":"§4.1, Table 4, Table 6"},{"comment":"The Hammett-Perkins benchmark coefficient is calibrated by matching the kinetic linear damping rate (χ=√(4.8/π)≈1.24, 22% below the canonical √(8/π)). Therefore, the statement that 'the calibrated closure reproduces the damping rate' (Table 4 caption) is true by construction and cannot serve as independent validation. The meaningful comparison is the oscillation frequency, where HP is shown to have a ~5% error. The paper should clearly label the HP linear-rate agreement as a calibration check, not a predictive success, and should avoid implying that HP validation strengthens the case for the FNO.","section":"§3.6, Table 4"},{"comment":"All quantitative results are based on a single trained model for each configuration, with no repeated training seeds, no error bars, and no measure of training variability. Since the FNO training is stochastic (AdamW, random initialization, data sampling), the reported accuracy could be seed-dependent. The authors should either train multiple seeds and report mean/standard deviation for the key metrics (damping rates, errors, t_div), or at least state that all results are single-run and cannot be used to infer statistical robustness. This is particularly important for the claimed visual agreement in Figs. 1 and 3, which may not be representative.","section":"§4.1, §4.2, §4.3; Tables 4-6"},{"comment":"The portability limitation is acknowledged in §3.2, but the paper still frames the contribution as a step toward closures 'suitable for large-scale fluid simulation.' All online tests use the same SSPRK3 integrator, second-order central differences, 128-point grid, and fixed k=0.35; no test is performed with a different spatial resolution, timestep, integrator, or wavenumber. As the reader's report notes, the learned effective flux is only defined through its action on this specific discretized dynamics. A minimal transfer test (e.g., varying Nx or Δt) would materially strengthen the portability claim. Without it, the conclusions should be restricted to the exact numerical configuration used.","section":"§3.2, §5"}],"minor_comments":[{"comment":"The panel title reads 'Linear (A=10^-1) (Offline Training)' but the case is the nonlinear amplitude A=10^-1. This is a typographical inconsistency with the surrounding text and should be corrected to 'Nonlinear'.","section":"Figure 3(b) caption"},{"comment":"The legend entry 'Landau' is ambiguous; it likely denotes the theoretical linear Landau rate. Please label it explicitly, e.g., 'linear Landau rate γ_Landau = ...'.","section":"Figure 5(a) legend"},{"comment":"The notation M_t is used for the window of K slots, and later M^(1), M^(2) for substage windows. The distinction between the accepted-state buffer and the provisional overwritten slot is described in prose but would be clearer if the buffer update rule were written as an explicit algorithm or pseudocode.","section":"§3.2, Eq. (3.7)"},{"comment":"The HSIC lag analysis is carefully caveated regarding autocorrelation, but the same caveat applies to Figure 12's correlation analysis where inputs are asinh-normalized; please state the normalization explicitly in the text or caption to avoid confusion.","section":"§4.5, Figure 10"},{"comment":"The statement that 'the batch size is selected dynamically to make use of available GPU memory' and that 'the value of 50 corresponds to the 40GB A100' is helpful, but it would be useful to report the total number of training iterations per epoch (or total snapshot count) for reproducibility.","section":"§3.4, Table 3"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of J. Plasma Phys. and addresses a timely topic. The core idea—online training of a neural closure within a differentiable fluid solver for Landau damping—is promising, and the authors are transparent about several limitations. However, the abstract and Section 1 overstate the generalization and broad-applicability results relative to the evidence. The most important fixes are (i) qualifying the amplitude-extrapolation claim, (ii) addressing the zero-mean/∂xq issue, (iii) adding seed variability or explicitly reporting single-run results, and (iv) softening or supporting the portability claim with a transfer experiment. With those changes, the paper could be a solid contribution. I would not reject it, but I would not accept it in its current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the online/a-posteriori training of a plasma closure inside a differentiable fluid solver, with a trailing memory window. That combination is the plasma analogue of what Frezat and MacArt did for neutral fluids, and it is executed carefully. A single FNO trained on multiple amplitudes captures both linear and nonlinear Landau damping, interpolates between training amplitudes, and stays stable in long rollouts. The effective-vs-kinetic heat flux distinction is the right way to think about what such a closure learns, and the ablation/HSIC analysis is a real step beyond just reporting loss curves. The authors also earn credit for openly stating the main limitations themselves: the closure is tightly coupled to this particular SSPRK3/finite-difference solver, the effective flux is not a unique or scheme-independent object, and upward amplitude extrapolation fails at A=0.2 and 0.3.\n\nSoft spots, in proportion. First, every claim about generalization is interpolation below the training envelope; the two out-of-training amplitudes above it fail, so the abstract's \"generalises to initial perturbation amplitudes outside the training set\" should be qualified. Second, the Hammett-Perkins benchmark coefficient is calibrated to the kinetic damping rate, so its linear-rate agreement is by construction; the frequency mismatch is the meaningful part and the paper does report that. Third, all results come from single runs; no repeated seeds, no error bars. Fourth, no code or data released, which makes the solver-coupling claim harder to assess. The nonzero-mean ∂xq output is worth a moment: it means the learned object is not a physical periodic heat flux but an effective term that also fixes the mean-pressure evolution. The paper says this explicitly and frames it as capturing net heating, so I would not call it a flaw, but it reinforces the scheme-dependence point.\n\nDoes the central argument hold? For one-dimensional, single-wavenumber, electrostatic Landau damping, yes. The method is plausible, the diagnostics are informative, and the paper does not hide its limitations. The reader's conditional verdict and the stress-test note are both fair; the stress-test concern is largely a restatement of the paper's own Section 3.2 caveat, so it is a scope restriction, not a fatal flaw.\n\nWho is this for: people working on learned closures for plasma fluids, and anyone applying operator learning to kinetic-fluid reduction. It deserves a serious referee. My advice to the editor: send it out, but ask for tempered claims in the abstract, error bars or repeated training runs, and either code/data or a much clearer portability statement. This is a solid contribution in a small domain, not a breakthrough for all of plasma physics.","headline":"Honest, well-scoped demonstration of online-trained non-Markovian FNO closure for 1D Landau damping; the core result holds for the single-mode setup, but the abstract oversells generalization and the closure is explicitly solver-coupled.","tokens_in":28080,"tokens_out":1616,"would_cite":true,"duration_ms":22277,"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 single Fourier Neural Operator closure, trained online inside a differentiable fluid solver, reproduces both linear and nonlinear electrostatic Landau damping and generalises across initial perturbation amplitudes.","keywords":["Landau damping","Fourier Neural Operator","plasma fluid closure","moment hierarchy","non-Markovian closure","Mori-Zwanzig","effective heat flux","a posteriori training"],"falsifier":"Deploy the trained FNO in the same physical setup but with a different time integrator or grid (e.g., classical RK4 or 256 spatial points) and compare the closed-fluid trajectory to the Vlasov ground truth; if it becomes unstable or drifts beyond the error level reported in the paper, the portability claim is falsified. A sharper test: train at k=0.35 and run at k=0.5 with the same amplitudes; failure there would show the closure has learned the specific mode, not the damping mechanism.","tokens_in":27270,"feed_emoji":"⚡","tokens_out":5741,"duration_ms":59350,"temperature":0.7,"pith_summary":"This paper tries to show that a machine-learned closure can replace the missing heat-flux term in a fluid plasma model and thereby recover the full kinetic physics of one-dimensional electrostatic Landau damping, in both its linear and nonlinear (particle-trapping) regimes. The approach trains a Fourier Neural Operator online—inside a differentiable fluid solver, with the loss evaluated on the closed-fluid trajectories against kinetic moment data—rather than fitting it to kinetic snapshots. The central claim is that one such FNO, acting on a trailing memory window of resolved moments, reproduces both damping regimes, interpolates and slightly extrapolates across initial perturbation amplitudes, and remains numerically stable in independent fluid simulations. This matters because analytic closures such as Hammett-Perkins capture only the linear regime, while offline-trained neural closures can be inaccurate or unstable once coupled to a solver. The paper also establishes that the learned heat flux is an effective flux, not a pointwise match to the kinetic heat flux, and that the model's use of memory is physically structured.","feed_headline":"One neural closure reproduces both Landau damping regimes","feed_subtitle":"Trained inside a fluid solver, it generalizes to unseen wave amplitudes and stays stable in deployment.","key_machinery":"The load-bearing object is the Fourier Neural Operator closure for the heat-flux gradient, ∂xq, entering the pressure equation of a 1D electrostatic fluid model. The FNO has eight Fourier layers with 32 retained modes and 64 hidden channels; its 251 input channels consist of five resolved fields {n, u, p, E, ∂xq} over a trailing window of K=50 timesteps (a non-Markovian memory horizon of 5 omega_p^-1), plus a spatially constant log-amplitude conditioning channel that lets one model modulate between linear and nonlinear regimes. Inputs are normalised by the RMS density perturbation amplitude and passed through an inverse-hyperbolic-sine transform to compress dynamic range. Online training is","core_discovery":"The paper's central claim is that a single Fourier Neural Operator, trained a posteriori by backpropagating through a differentiable fluid solver, supplies the heat-flux-gradient closure that the one-dimensional electrostatic moment hierarchy needs to reproduce Landau damping. Deployed in the same fluid solver, the FNO recovers the linear damping rate to within about 0.2 percent in the training window and reproduces the nonlinear saturation and bounce oscillations of the trapped state over hundreds of plasma periods. One FNO, trained on amplitudes from 10^-3 to 10^-1 and conditioned on a log-amplitude channel, tracks the damping rate through the linear-to-nonlinear transition and interpolate","pith_inferences":["Editorial extension: the fixed wavenumber k=0.35 is the true test boundary; the same recipe should be tried across wavenumbers, since the closure's memory of phase mixing is likely k-dependent, and the paper does not claim transfer across k.","Editorial extension: the sensitivity diagnostics suggest a testable refinement—if the memory window were shortened below the bounce period in the nonlinear regime, accuracy should degrade; scanning K would map the memory requirement quantitatively.","Editorial extension: because the effective flux is scheme-dependent, portability between fluid codes may require either retraining per discretisation or a parametrisation that conditions on discretisation metadata—an open problem the paper flags as future work.","Editorial extension: the same online-training-with-memory recipe may apply to other unresolved kinetic effects (e.g., pressure anisotropy in 2D/3D or electromagnetic fluctuations), where the Mori-Zwanzig memory integral is similarly non-trivial; the 1D electrostatic case is a proof-of-concept."],"forward_implications":["A learned fluid closure can replace analytic linear-theory closures like Hammett-Perkins for problems where both linear and nonlinear regimes matter, using a single model rather than a regime switch.","Online-trained closures are stable in deployment by construction; the paper's matched comparison suggests that a-priori-optimal closures can be stable or accurate in pointwise flux but still fail to reproduce resolved-moment trajectories.","The heat flux in a truncated fluid model should be judged by its effect on the resolved moments, not by pointwise agreement with the kinetic flux; the learned effective flux is a legitimate closure even where it diverges from the kinetic one.","The amplitude-conditioning channel and input normalisation allow one model to span several decades of perturbation amplitude, so training sets need not cover every amplitude, only the regimes.","Memory-window dependence is physically structured (peak sensitivity at lags separated by the oscillation period), suggesting that finite-memory neural closures can encode genuine kinetic memory rather than being arbitrary fits."],"fun_headline_variants":["FNO fluid closure nails linear and nonlinear Landau damping","Single neural closure reproduces both Landau damping regimes","Neural operator closure learns Landau damping from solver trajectories","Memory-based neural closure captures linear and nonlinear damping","Data-driven fluid closure generalizes Landau damping to new amplitudes"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The learned closure is defined only through its effect on one specific discretised fluid solver—SSPRK3 time stepping, second-order central differences, a 128-point grid, and fixed wavenumber k=0.35—so if the effective heat flux does not transfer to other solvers, grids, or wavenumbers, the claim of a broadly applicable learned plasma closure collapses.","fun_headline_variants_meta":{"raw":{"variants":["FNO fluid closure nails linear and nonlinear Landau damping","Single neural closure reproduces both Landau damping regimes","Neural operator closure learns Landau damping from solver trajectories","Memory-based neural closure captures linear and nonlinear damping","Data-driven fluid closure generalizes Landau damping to new amplitudes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000247,"raw_usage":{"total_tokens":1360,"prompt_tokens":707,"completion_tokens":653,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":451,"completion_tokens_details":{"reasoning_tokens":588}},"tokens_in":451,"tokens_out":653,"duration_ms":6191,"temperature":1.0,"reasoning_tokens":588,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T08:24:42.329591+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Deploy the trained FNO in the same physical setup but with a different time integrator or grid (e.g., classical RK4 or 256 spatial points) and compare the closed-fluid trajectory to the Vlasov ground truth; if it becomes unstable or drifts beyond the error level reported in the paper, the portability claim is falsified. A sharper test: train at k=0.35 and run at k=0.5 with the same amplitudes; failure there would show the closure has learned the specific mode, not the damping mechanism.","supporting_citations":[],"review_version":1}