{"id":"c5088c43-97c3-4320-ba65-23a5379a7b96","arxiv_id":"2506.16619","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"The authors fit a neural-network kernel, parameterized by LMV-style free-path and density-weighted distance plus time, to particle-in-cell heat flux data and show qualitative agreement on three held-out 1D cases.","lead":"This paper trains a neural network to learn time-dependent, nonlocal electron heat-flux kernels from 1D particle-in-cell plasma simulations and tests them against the standard LMV and Spitzer-Harm models. A generalist should read it to see whether machine-learned closures can replace analytic transport kernels in fusion and astrophysical plasma modeling, though the paper's own text limits the approach in strongly nonlocal cases.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Eq. 6 kernel ansatz fixes Q_SH to the initial temperature profile, and the paper itself concedes this is wrong in the strongly nonlocal regime, so the 'across regimes' claim is unsupported.","rationale":"The reader's weakest-assumption analysis identifies the same load-bearing issue: the kernel convolution uses a local SH flux computed from the initial temperature profile, and the paper itself says this is not correct in strongly nonlocal conditions. My reading of the full text confirms this. The central claim of 'strong agreement with kinetic benchmarks across regimes' is contradicted by the authors' own caveat that the kernel-based approach breaks down in strongly nonlocal regimes, which are explicitly included in the test set and in the abstract's scope. The absence of quantitative error metrics and the saturation-time normalization additionally weaken the demonstration, but they are secondary to the modeling-ansatz failure. A narrowed claim restricted to moderately nonlocal and local regimes might be defensible, but the manuscript as written overstates its regime coverage. Since the reader already rejected on this basis, the verdict is unchanged. I am not raising any additional independent objection beyond this conceded limitation, because that limitation alone is sufficient to block the central claim.","tokens_in":8545,"tokens_out":4135,"duration_ms":45363,"concrete_test":"For the Fig. 4(a) strongly nonlocal test case (LT/lambda_free ~ 33), recompute the Eq. 6 prediction twice: once with Q_SH(x') computed from the initial temperature profile, as in the manuscript, and once with Q_SH(x') computed from the time-evolved temperature profile at the same output time, keeping the learned kernel fixed. If the evolved-profile prediction deviates from the PIC ground truth by more than about 10% in peak heat flux, or if it differs substantially from the initial-profile prediction, the kernel cannot absorb profile evolution and the cross-regime claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that learned kernels produce 'strong agreement with kinetic benchmarks across regimes.' That claim rests on Eq. 6: Q_e(x,t) = sum_{x'} W(lambda(x'), X(x,x'), t) * Q_SH(x'). The manuscript's own discussion of Fig. 4(a) states that in the strongly nonlocal regime the initial temperature profile undergoes non-negligible evolution, and 'convolving the local SH flux, which is calculated from the initial temperature profile, across the whole space is not correct,' concluding that 'the kernel-based approach breaks down in strongly nonlocal regimes.' Because Fig. 4(a) is one of only three test cases offered as evidence of cross-regime performance, and because the abstract explicitly claims agreement across regimes, the fixed-profile convolution ansatz is load-bearing and is conceded to fail in a regime the paper claims to cover. The dynamic-kernel formulation is also coupled to a time coordinate normalized by each case's known saturation time t_sa, which makes the kernel non-predictive without an oracle for t_sa; however, the fixed-SH-flux ansatz is the more fundamental problem.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces LINN, a physics-informed neural network that learns a spatiotemporal heat flux kernel W(λ, X, t) from 1D OSIRIS PIC simulations that span local to strongly nonlocal electron heat transport. The kernel is used in Eq. (6) to convolve the local Spitzer-Härm flux QSH, computed from the initial temperature profile, into a nonlocal heat flux. The authors train on 829 × 2 samples, then compare kernel shapes and heat flux profiles against PIC ground truth for three held-out cases with LT/λfree ≈ 33, 333, and 5000. They report that the learned kernels are time-dependent and better confined than LMV kernels, and the abstract claims 'strong agreement with kinetic benchmarks across regimes.'","tokens_in":8770,"tokens_out":4561,"duration_ms":43901,"significance":"If the predictive capability were established, this would be a useful step toward data-driven closures for nonlocal heat transport, combining a large PIC dataset with an interpretable, time-evolving kernel. The dataset is substantial and the use of physical inputs (λ, X, t) is a strength. However, the central 'across regimes' claim is not supported by the evidence: the paper itself concludes that the kernel-based approach breaks down in the strongly nonlocal regime, and the time normalization by the target's saturation time tsa compromises predictive utility and evaluation validity. As presented, the work is a proof-of-concept with overclaimed scope rather than a validated model.","major_comments":[{"comment":"The ansatz in Eq. (6) fixes QSH(x') to the initial temperature profile. The manuscript's own discussion of Fig. 4(a) states that in the strongly nonlocal regime the initial temperature profile undergoes non-negligible evolution and 'convolving the local SH flux, which is calculated from the initial temperature profile, across the whole space is not correct,' concluding that 'the kernel-based approach breaks down in strongly nonlocal regimes.' Because Fig. 4(a) is one of only three test cases offered as evidence for the abstract's 'strong agreement ... across regimes' claim, the core modeling assumption is conceded to fail in a regime the paper claims to cover.","section":"Section 4, Eq. (6), Fig. 4(a)"},{"comment":"The time coordinate input is normalized by the saturation time t_sa of each sample. At inference, t_sa would not be known a priori for an arbitrary plasma configuration, so the model as presented is not a closure for hydrodynamics codes. Furthermore, since t_sa is extracted from the same simulation whose heat flux is being predicted, the held-out test evaluation leaks information from the target. The manuscript does not specify how t_sa would be estimated in practical applications.","section":"Section 4, Fig. 4, Eq. (6)"},{"comment":"The claimed agreement with PIC is supported only by qualitative visual comparison of heat flux profiles. No quantitative error metric (e.g., L2 norm, relative error, or coefficient of determination) is reported for the three test cases, making the strength of the claimed agreement difficult to assess. This is particularly important given the acknowledged breakdown in Fig. 4(a).","section":"Section 4, Fig. 4"},{"comment":"The training and test samples are drawn from the same parametric family (same temperature profile functional form, Eq. (1), and isobaric density profile). Because the network is a supervised fit to PIC samples, agreement on held-out samples from the same family is partly a test of interpolation within the training distribution, not of generalization to genuinely new physical conditions. The paper should state this limitation explicitly and temper the generalization claim.","section":"Sections 2 and 3"}],"minor_comments":[{"comment":"The names 'Schurtz Nicolaï Busquet' and 'Luciani Mora Virmont' should be typeset with en-dashes/hyphens (e.g., 'Schurtz–Nicolaï–Busquet') for consistency with standard literature usage.","section":"Abstract and Section 1"},{"comment":"There is a typo in the title: 'Elecron' should be 'Electron.'","section":"Reference [10]"},{"comment":"The caption and text indicate that L is adjusted for each R using LT = 2L(R+1)/(R-1), but the specific L values for each color are not given. Please list the L values or provide a table so that the parameter sampling is reproducible.","section":"Fig. 1 and Section 2"},{"comment":"The typeset kernel arguments are garbled (W((x′) X(x,x′) t)). Please define W(λ(x′), X(x,x′), t) unambiguously, including the normalization of each input (e.g., λ/dx, dx removed from X) and the summation convention.","section":"Eq. (6)"},{"comment":"The phrase 'All 829 × 2 samples (including their mirrored counterparts...)' should clarify whether the factor of 2 refers solely to mirror symmetry and how the mirrored data are constructed and partitioned into train/validation/test sets.","section":"Section 3"}],"recommendation":"reject","confidential_remarks":"The stress-test concern is valid and the paper's own text concedes the central limitation. The manuscript could potentially be revised into a more modest methodological study focusing on the moderate nonlocal regime, with quantitative error metrics and a strategy for estimating t_sa in applications. As it stands, the abstract's across-regimes claim is not defensible, and the t_sa normalization undermines the predictive claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a legitimate proof-of-concept for learning time-dependent heat-flux kernels from PIC data, but the paper's headline claim — strong agreement across regimes — is not supported by the evidence, including the authors' own concession that the kernel ansatz breaks down in strongly nonlocal conditions.\n\nWhat is genuinely new: the idea of parameterizing a spatiotemporal kernel with LMV-style features (X, λ) plus a time coordinate, trained on 1D PIC heat flux profiles, does not appear in the prior ML-closure literature. The dataset itself, 829 samples spanning LT/λ from 10 to 5000, is real work. The NN architecture is standard but reasonable. I also credit the authors for explicitly stating that Eq. (6), which convolves SH flux from the initial temperature profile across all space, is incorrect in the strongly nonlocal regime, and that the kernel approach breaks down there. That is an honest limitation.\n\nWhere it gets soft: that limitation lands directly on the central claim. The abstract says predictions show strong agreement with kinetic benchmarks across regimes. Fig. 4(a) is one of three test cases, and in that case the model's own input assumption fails. The time coordinate is normalized by the sample's saturation time t_sa, which is a leak if the model is meant to be predictive — you need to know the answer to normalize the input. Validation is visual only; no error metrics, no SNB baseline even though SNB is cited in the motivation, and no code, data, or weights are provided. Finally, agreement on the same simulation family used for training is partly circular.\n\nNone of this means the approach is worthless. For moderate and local regimes, where the fixed-profile convolution is reasonable, the learned kernels appear to capture the dynamics and even fix the LMV kernel's nonphysical concave feature. That is a useful result within a restricted scope.\n\nWho this is for: anyone building ML closures for nonlocal transport will want to know this exists, and the authors' own limitation statement is a useful warning. It deserves a serious referee — the construction is new and the data is substantial — but the current manuscript overclaims. I would recommend sending it to peer review with a clear expectation that the claims be scaled back to the regimes where the kernel ansatz actually holds, and that quantitative benchmarks and release of code/data be added.","headline":"A real proof-of-concept for learned time-dependent transport kernels that overclaims its own reach; the authors concede the kernel ansatz fails in a regime the abstract claims to cover.","tokens_in":9311,"tokens_out":1909,"would_cite":false,"duration_ms":18278,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["52.25.Fi","52.65.Rr"],"model":"deepseek-v4-flash","headline":"This paper argues that a neural network trained on kinetic particle-in-cell simulations can learn spatiotemporal heat flux kernels that reproduce nonlocal plasma heat transport, including time-dependent behavior that static LMV and SNB…","keywords":["nonlocal heat transport","plasma","neural network","transport kernel","Spitzer-Harm","Luciani-Mora-Virmont","particle-in-cell","kinetic simulation"],"falsifier":"Run the trained network on an unseen strongly nonlocal test case with L_T/lambda_free near 10 and compare the heat flux from Eq. (6) with the PIC ground truth at saturation; a deviation larger than the simulation noise would confirm that the kernel-over-initial-profile ansatz does not hold in that regime.","tokens_in":8314,"feed_emoji":"🔥","tokens_out":7392,"duration_ms":70508,"temperature":0.7,"pith_summary":"The paper tries to establish that the kernel connecting local Spitzer-Harm flux to the true nonlocal heat flux can be learned directly from kinetic simulation data, and that the learned kernel should depend on time as well as space. This matters because existing nonlocal transport models use static, semi-empirical kernels and fail in moderately or strongly nonlocal conditions where the electron mean free path approaches the temperature gradient scale. The proposed LMV-Informed Neural Network (LINN) outputs a kernel as a function of density-weighted distance, local mean free path, and time, and the heat flux is reconstructed by convolving this kernel with the local Spitzer-Harm flux. If the claim holds, hydrodynamic codes could adopt a flexible, data-driven closure for nonlocal heat transport and gain an interpretable picture of how far heat travels as a plasma evolves.","feed_headline":"Net learns time-evolving plasma heat kernels from kinetic data","feed_subtitle":"Unlike fixed LMV and SNB kernels, learned kernels adapt in time and match PIC benchmarks across local to nonlocal regimes.","key_machinery":"The central object is the heat-flux kernel W(X,lambda,t), a learned, time-dependent generalization of the LMV kernel used in Eq. (2). The LMV-Informed Neural Network maps the density-weighted distance X, the effective propagation range lambda, and a normalized time coordinate to a positive scalar kernel weight through residual MLP blocks with a Softplus output. Heat flux is then assembled via Eq. (6), Q_e(x,t)=sum_x' W(x,x',t) Q_SH(x'), where Q_SH is the local Spitzer-Harm flux. The time coordinate is the key addition that makes the kernel dynamic, while the X and lambda features keep the operator interpretable in the language of nonlocal transport theory.","core_discovery":"On its own terms, the paper's discovery is that a neural network can learn the nonlocal heat flux kernel W(X,lambda,t) and that convolving it with the Spitzer-Harm flux through Eq. (6) reproduces kinetic heat flux from strongly nonlocal to local regimes. The learned kernels are time-dependent, narrowing or broadening as the plasma evolves, whereas LMV and SNB kernels are static. The network also suppresses the unphysical long-range concave structure that appears in LMV kernels. The paper is explicit that in strongly nonlocal cases convolving the local Spitzer-Harm flux, computed from the initial temperature profile, across the whole space is not correct, and the kernel ansatz therefore breaks down in that regime.","pith_inferences":["The paper's own limitation statement suggests a testable extension: replacing the fixed initial-profile Q_SH in Eq. (6) with a time-updated local flux could extend the same learned-kernel architecture to strongly nonlocal cases without abandoning the convolution representation.","The reported collapse of the data onto L_T/lambda_free for fixed temperature ratio suggests the network may mostly be learning a function of the regime parameter rather than of detailed profile shapes; a test with identical L_T/lambda_free but different temperature profiles would separate regime sensitivity from profile sensitivity.","Normalizing the time coordinate by saturation time may limit transfer to non-isobaric or continuously driven plasmas where no single saturation scale exists; an absolute-time or local-collision-time input would be a meaningful extension.","The current geometry is one-dimensional, so a multidimensional extension would need angular information in the kernel and new training data before the approach can serve as a general hydrodynamic closure."],"forward_implications":["Hydrodynamic codes could compute nonlocal heat flux with a convolution of the same cost structure as LMV or SNB, but with kernels that evolve in time before saturation.","The learned kernels provide a direct diagnostic of energy transport range: broad on the hot side, narrow on the cold side, without the unphysical concave tails seen in LMV.","Held-out test cases suggest the same trained kernel function can be evaluated for unseen combinations of temperature ratio, density, and gradient steepness within the covered parameter range.","Embedding such kernels into radiation-hydrodynamics simulations would improve predictions in inertial confinement fusion and astrophysical plasmas where L_T/lambda_free drops below about 500."],"supporting_citations":[{"why":"Supplies the local Spitzer-Harm heat flux whose convolution with the kernel defines the target in Eq. (6).","marker":"[6]"},{"why":"Supplies the nonlocal LMV kernel-convolution structure that the network inherits, along with the static baseline it must improve upon.","marker":"[8]"},{"why":"Gives the SNB static-kernel model, the other standard baseline for nonlocal heat transport in hydrodynamic codes.","marker":"[9]"},{"why":"Provides the binary collision model needed for PIC simulations to capture collisional heat transport.","marker":"[12]"},{"why":"Extends the binary collision treatment to relativistic regimes and non-uniform particle weights, underpinning the PIC transport data.","marker":"[13]"},{"why":"Provides the PIC code used to generate the 829 training and test samples spanning the local-to-nonlocal parameter space.","marker":"[17]"},{"why":"Supplies the prior connection between machine learning and transport processes that motivates the data-driven kernel-learning route.","marker":"[19]"}],"fun_headline_variants":["NN learns time-evolving plasma heat kernels, beats static models","Learned plasma heat kernels adapt over time, matching kinetic simulations","Time-dependent heat kernels from neural networks match plasma simulations","Neural network learns evolving heat flux kernels for nonlocal plasma transport","Kinetic-trained NN yields time-adaptive heat kernels for plasma"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that true heat transport can be written as a convolution of a learned kernel with the local Spitzer-Harm flux computed from the initial temperature profile; the paper itself acknowledges this premise fails in strongly nonlocal cases.","fun_headline_variants_meta":{"raw":{"variants":["NN learns time-evolving plasma heat kernels, beats static models","Learned plasma heat kernels adapt over time, matching kinetic simulations","Time-dependent heat kernels from neural networks match plasma simulations","Neural network learns evolving heat flux kernels for nonlocal plasma transport","Kinetic-trained NN yields time-adaptive heat kernels for plasma"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000649,"raw_usage":{"total_tokens":2908,"prompt_tokens":806,"completion_tokens":2102,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":422,"completion_tokens_details":{"reasoning_tokens":2018}},"tokens_in":422,"tokens_out":2102,"duration_ms":14707,"temperature":1.0,"reasoning_tokens":2018,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:22:45.075629+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained network on an unseen strongly nonlocal test case with L_T/lambda_free near 10 and compare the heat flux from Eq. (6) with the PIC ground truth at saturation; a deviation larger than the simulation noise would confirm that the kernel-over-initial-profile ansatz does not hold in that regime.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the nonlocal LMV kernel-convolution structure that the network inherits, along with the static baseline it must improve upon."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the SNB static-kernel model, the other standard baseline for nonlocal heat transport in hydrodynamic codes."},{"cited_title":"Sentoku, K","cited_arxiv_id":null,"evidence_quote":"Provides the binary collision model needed for PIC simulations to capture collisional heat transport."},{"cited_title":"Sentoku and A","cited_arxiv_id":null,"evidence_quote":"Extends the binary collision treatment to relativistic regimes and non-uniform particle weights, underpinning the PIC transport data."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the PIC code used to generate the 829 training and test samples spanning the local-to-nonlocal parameter space."},{"cited_title":"Miniati and G","cited_arxiv_id":null,"evidence_quote":"Supplies the prior connection between machine learning and transport processes that motivates the data-driven kernel-learning route."}],"review_version":1}