{"id":"f6b1e05f-9671-401c-805f-04389e0cdb97","arxiv_id":"2501.16709","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A neural network emulator trained on CAMELS simulations predicts FRB dispersion measure distributions and reveals non-monotonic dependence of the scatter parameter F on galaxy feedback parameters.","lead":"This paper builds a neural network that predicts the distribution of radio burst dispersion measures from CAMELS simulation parameters, then uses it to show that the scatter-based feedback parameter F behaves non-monotonically with feedback strength. The results suggest the small CAMELS boxes underestimate the scatter seen in FRB observations, a caution for future constraints.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The unvalidated independence assumption in Eq. (3) for assembling long DM sightlines directly sets the variance that defines F, so the non-monotonic F result may be an artifact of the pipeline.","rationale":"The paper makes a useful contribution by training a neural network emulator on the full CAMELS Latin Hypercube and by being transparent about inherited box-size limitations, with public code available. The NN itself shows acceptable test-set correlation for σ_DM, so the weakest link is not the regression but the target it learns. Equation 3 in §2.2 builds long DM sightlines by drawing independent samples from the one-point CDF of a single 25 h^-1 Mpc snapshot at each 100 h^-1 Mpc step. This assumes zero covariance between adjacent path segments. Since F is defined as the scatter of the summed DM, the covariance structure directly determines F. Feedback changes not only the marginal electron-density distribution but also the spatial clustering of gas, so a one-point CDF construction can distort the dependence of F on feedback parameters. The authors acknowledge differences with Medlock et al. (2024) and defer a comparison project to future work, but the non-monotonic F extrema are central to the paper's message and rest entirely on this unvalidated assumption. A direct test in a larger simulation with identical subgrid physics, as described in concrete_test, would settle whether the assumption holds. Until such a test is performed, the F-based conclusions should remain conditional rather than definitive. The reader's weakest_assumption identifies the same issue, and my assessment does not move the verdict away from CONDITIONAL.","tokens_in":14165,"tokens_out":6932,"duration_ms":67110,"concrete_test":"In IllustrisTNG-300 (or a 100 h^-1 Mpc SIMBA-Large box if available), ray-trace straight sightlines of length 100 h^-1 Mpc at z ≈ 0.5 directly through the gas and record the DM. Separately, take a 25 h^-1 Mpc sub-box of the same simulation snapshot, build the single-box CDF, and generate synthetic 100 h^-1 Mpc DMs by summing four independent CDF draws as in Eq. (3). Compare the variance of the direct ray-traced DMs with the variance of the synthetic sums over the same number of realizations. If the ratio of variances deviates from unity by more than the jackknife error, the independence assumption fails and the F values in Table 3.3 need to be recomputed with a method that preserves correlations.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.2 constructs a sightline to z_frb by drawing an independent uniform deviate at each 100 h^-1 Mpc segment and mapping it through the single-box CDF of the nearest snapshot (Eq. 3). This is equivalent to assuming that DM fluctuations in different path segments are statistically independent and that the one-point distribution fully describes the density field. Real sightlines cross the same filaments and voids over several segments, so the covariance between segments is generally non-zero. F ≡ σ_DM z^{1/2} is the standard deviation of the summed DM, so any error in the covariance structure translates directly into F. Positive correlations increase σ_DM; negative correlations decrease it. The size and sign of these correlations will depend on feedback, which changes the spatial clustering of gas, not just the marginal distribution. Consequently the non-monotonic trends in Table 3.3 and Fig. 3 could be produced by the independence assumption rather than by the simulations. The authors note discrepancies with Medlock et al. (2024) and attribute them to different DM computation methods, but they do not validate their method against direct ray-traced long sightlines or against a larger-volume simulation with the same subgrid physics.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a neural-network emulator for the cosmic dispersion measure (DM) distribution in the CAMELS simulation suite, trained on the Latin Hypercube (LH) set of SIMBA and TNG simulations. The authors compress each snapshot's p(DM) using the Macquart fitting function into four parameters (⟨DM⟩, σ_DM, α, β), then train a NN to map the six CAMELS parameters plus redshift to these summary statistics. Using the trained NN, they explore the F ≡ σ_DM(Δ) z^{1/2} parameter across the feedback parameter space at fixed cosmology, reporting non-monotonic behavior and identifying F extrema at parameter combinations not present in the LH set. They compare their F values with the observational constraint from Baptista et al. (2024) and argue that CAMELS box sizes limit the achievable scatter.","tokens_in":14408,"tokens_out":7851,"duration_ms":74572,"significance":"If validated, the NN emulator would provide a fast and flexible tool for exploring how multidimensional feedback parameters shape the FRB DM distribution, going beyond the one-at-a-time 1P analyses. The paper is honest about several limitations, particularly the small CAMELS box size and the resulting insufficiency for direct observational comparison. A key strength is the use of the full LH suite, which allows the exploration of joint parameter dependencies. However, the central claims—especially the non-monotonic F behavior and the F extrema—rest on an unvalidated line-of-sight construction and on NN extrapolation at sparsely sampled parameter boundaries, so the results are not yet established at the level claimed.","major_comments":[{"comment":"The sightline construction assumes that the DM contribution at each path segment is an independent random draw from the redshift-interpolated CDF, and the paper does not test this independence assumption against directly ray-traced long sightlines or against a larger-volume simulation. Since F is defined as the standard deviation of the summed DM, any covariance between adjacent segments would directly change σ_DM and hence F. The text also leaves ambiguous whether the CDF used in Eq. (3) corresponds to a single 25 h^-1 Mpc CAMELS box or to the full 100 h^-1 Mpc segment, and how the periodic tiling of the small box is handled. This is load-bearing because the non-monotonic F trends in Figure 3 and the extrema in Table 1 are derived from this assumed covariance structure.","section":"Section 2.2, Eq. (3)"},{"comment":"The F extrema in Table 1 are NN predictions at parameter combinations that are not present in the LH training set; for example, the SIMBA maximum-F point (A_SN1, A_AGN1, A_SN2, A_AGN2) = (0.26, 0.26, 1.71, 1.01) sits at the boundary of the A_SN1 and A_AGN1 ranges. With only 1000 LH points in a six-dimensional space, such corner regions are sparsely sampled, and the paper provides no validation of the NN predictions at these locations against direct simulations or against the CAMELS EX/CV sets. The central claim that these extrema are properties of the simulations rather than NN artifacts therefore needs a dedicated test.","section":"Section 3.3, Table 1"},{"comment":"The authors concede that the NN predictions for α and β exhibit high scatter. Because the stated goal is to emulate p(DM), and α and β jointly control the shape of the Macquart fitting function, poor α/β performance means the NN does not reliably reproduce the full p(DM), even if ⟨DM⟩ and σ_DM are well predicted. The paper should quantify the impact of α/β scatter on the reconstructed p(DM) (e.g., by comparing predicted and true CDFs or by reporting the R^2 of the predicted p(DM)), and should state whether the F results depend only on σ_DM, which would mitigate this issue.","section":"Section 3.2, Figure 2"}],"minor_comments":[{"comment":"The definition y = e^{log(DM)}/⟨DM⟩ is just DM/⟨DM⟩, and the notation f(x) introduces an undefined variable x; the transformation to the logarithmic distribution should be written more clearly.","section":"Eq. (5)"},{"comment":"The notation σ_DM(Δ) is used for the standard deviation of the normalized quantity Δ = DM/⟨DM⟩, while σ_DM in Eq. (4) is used for the same quantity in the Macquart fit; the paper should explicitly distinguish the normalized and unnormalized standard deviations to avoid confusion with the observational F parameter.","section":"Section 3.3, Eq. (10)"},{"comment":"Typographical issues include 'Meahwhile' in Section 2.1 and 'balck dotted line' in the Figure 4 caption; these should be corrected.","section":"Section 3.1, Figure 1"},{"comment":"The comparison with Medlock et al. (2024) is presented only qualitatively; showing the actual Medlock et al. curves or a quantitative measure of the difference would strengthen the discussion.","section":"Section 3.3, Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The paper makes a useful methodological contribution and is appropriately cautious about the CAMELS box-size limitation. However, the two load-bearing gaps—the untested independence assumption in the sightline assembly and the unvalidated NN predictions at the F-extremum corners—prevent the main physical claims from being accepted as they stand. Both issues are addressable within the manuscript's scope: the authors could test the independence assumption by directly ray-tracing long sightlines through stacked CAMELS boxes, and they could validate the F extrema against simulations at the nearest LH points or in the EX/CV sets. I therefore recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does one genuinely new thing: it trains a neural network to emulate p(DM) across the full CAMELS Latin Hypercube, and it uses that emulator to argue that the F parameter is not a monotonic function of individual feedback parameters. The tool itself is practical and clearly described, with public code and data, and the authors are appropriately cautious about the small-box limitation, even explicitly saying that CAMELS is not yet suitable for interpreting observed FRB DMs. That honesty is welcome.\n\nThe soft spot is real and load-bearing. The sightline assembly in Eq. (3) treats each 100 h^-1 Mpc segment as an independent draw from the redshift-interpolated CDF. That assumption directly sets the variance of the summed DM, which is exactly what F measures. If adjacent segments are correlated along real sightlines — and they will be, because the same filaments and voids span multiple segments — the covariance contributes to sigma_DM, and the size and sign of that contribution can depend on feedback physics. The paper never validates this against direct ray-traced long sightlines or a larger-volume simulation with the same subgrid model. The authors note discrepancies with Medlock et al. (2024) and attribute them to different DM computation methods, but that is precisely why an external validation is needed. As it stands, the non-monotonic F trends in Fig. 3 and Table 1 could be artifacts of the stitching procedure rather than properties of the simulations themselves.\n\nA second, related issue: the F extrema in Table 1 come from NN predictions at parameter corners that are not present in the LH training set. The NN performance on <DM> and sigma_DM is good, but alpha and beta show high scatter, and the authors concede that. Without direct simulation checks at those corners, the specific extremal values should be treated as tentative. The comparison to the 1P set in Fig. 3 is a reasonable sanity check, but it inherits the same sightline-construction method, so it cannot validate the independence assumption.\n\nThe paper is still worth engaging. The emulator is a useful community resource, and the central caution — that F is not a simple monotonic proxy for feedback strength — is worth stating even if the quantitative trends are provisional. A serious referee should ask for two things: a validation of the sightline assembly against full ray-traced sightlines or a larger box, and direct simulation checks of the F extrema. With those, the paper could become a solid reference for FRB-DM work. I would not desk-reject it.","headline":"A useful but incremental emulator paper whose headline F result rests on an unvalidated independence assumption in the sightline construction.","tokens_in":14976,"tokens_out":2969,"would_cite":false,"duration_ms":32402,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A neural network trained on the full CAMELS simulation suite emulates FRB dispersion-measure distributions and reveals that the scatter parameter F depends non-monotonically on supernova and AGN feedback strength.","keywords":["fast radio bursts","dispersion measure","CAMELS simulations","galaxy feedback","neural network emulator","Macquart relation","F parameter","cosmological simulations"],"falsifier":"Ray-trace a continuous sightline through a sequence of concatenated CAMELS boxes (or through a larger-volume counterpart such as TNG-300) out to z = 1, measure the resulting p(DM), and compare its σ_DM with the CDF-sampling method: if the independently-sampled variance differs from the directly-simulated variance at the level that would change F by more than the network's test-set error, the paper's core assumption is falsified.","tokens_in":13972,"feed_emoji":"📡","tokens_out":9796,"duration_ms":76993,"temperature":0.7,"pith_summary":"The paper argues that a neural network trained on the full Latin Hypercube of the CAMELS simulation suite can emulate the probability distribution of cosmic dispersion measures (DM) at any redshift up to z = 1, as a function of the cosmological and feedback parameters that control the simulations. Using this emulator, the authors study the parameter F = σ_DM z^(1/2), the standard summary of the scatter in p(DM), and find that F does not respond monotonically to any single feedback knob: the joint state of supernova and AGN feedback sets the effective scatter. They also find that the largest F values in the CAMELS parameter space fall short of current observational constraints on F from fast radio bursts, which they attribute to the small simulation boxes missing the most extreme overdensities and voids. The intended payoff is a tool that turns the thousands of CAMELS simulations into a continuous map from feedback parameters to FRB dispersion statistics, usable for future inference with larger simulation suites.","feed_headline":"Galaxy feedback bends FRB dispersion scatter in non-monotonic ways","feed_subtitle":"Because the relation is non-monotonic, seeing a large spread in FRB dispersion does not by itself mean strong feedback.","key_machinery":"The machinery is a feedforward neural network with five to fifteen hidden layers and fifteen to thirty-five neurons per layer, rectified-linear activations, mean-square-error loss, and hyperparameters chosen by an automatic search; its input is the parameter vector (Ωm, σ8, A_SN1, A_AGN1, A_SN2, A_AGN2) together with redshift z, and its output is the set (⟨DM⟩, σ_DM, α, β) that describes p(DM) through a quasi-universal fitting function. Each snapshot's p(DM) is first compressed into those four numbers by a Markov-chain fit, so the network never sees the raw histograms. To reach redshifts beyond a single box, the paper assembles sightlines by summing 100 h^(-1) Mpc segments whose DM values are drawn independently from CDFs linearly interpolated in redshift between snapshots; this is the assumption that ultimately sets σ_DM. The scalar that carries the astrophysical claim is F = σ_DM z^(1/2), the redshift-scaled scatter, which is also the quantity compared with observed FRB samples.","core_discovery":"The central claim is that the effect of galaxy feedback on the FRB dispersion-measure distribution p(DM) can be captured by a feedforward neural network whose inputs are the six CAMELS parameters plus redshift, and whose outputs are four summary parameters of a standard fitting form for p(DM): the mean DM, the scatter σ_DM, and two shape parameters. With the network trained on the thousand Latin Hypercube realizations for both the SIMBA and TNG variants of CAMELS, the paper reports good predictive performance on held-out boxes. From the network's predicted σ_DM it computes F = σ_DM z^(1/2) over a dense grid of feedback parameters, and finds that F has interior extrema: in the SIMBA variant the maximal F occurs at low supernova and quasar-mode AGN strength with relatively strong jet-mode AGN, while in the TNG variant the maximum occurs at maximal supernova wind energy; setting every feedback parameter to its maximum does not maximize F, and some one-dimensional trends reverse direction relative to earlier one-at-a-time analyses. The paper interprets these non-monotonicities as the joint action of feedback channels (for instance, supernova feedback suppressing black-hole growth and thus weakening the effective AGN feedback), and it cautions that the low overall F values compared with the observed value reflect the limited dynamic range of the 25 h^(-1) Mpc boxes.","pith_inferences":["If the independence of adjacent 100 h^(-1) Mpc CDF draws is violated by large-scale structure coherence, the true p(DM) scatter in CAMELS volumes would differ from the paper's construction; a direct ray-trace through concatenated CAMELS boxes would settle this, and the paper's own caveat about cosmic variance suggests it is a real risk.","The non-monotonic F(θ) surface suggests that the shape parameters α and β, which the paper's emulator also predicts, may carry complementary information that breaks some of the degeneracies among feedback parameters; the paper does not explore this, but its own network makes it straightforward.","The discrepancy in mean DM at z = 1 between this work and earlier 1P analyses (by roughly 10–20% depending on the DM computation code) implies the emulator's calibration inherits a systematic from the chosen ray-tracing or gridding method; a comparison project across codes would be needed before applying the network to real FRB data."],"forward_implications":["With the trained emulator, p(DM) at any redshift up to z = 1 and any combination of the six simulation parameters is available without running a new hydrodynamical simulation.","The non-monotonicity of F in the feedback parameters implies that a single measured F value cannot be mapped uniquely back to a feedback strength; inference must treat the four feedback parameters jointly.","The maximal F values from the emulator are still smaller than the current observed central value, so the paper concludes that CAMELS' small boxes cannot yet be used to fit FRB DM observations directly.","The same emulation strategy is transferable to future CAMELS-like suites with larger boxes, which the paper identifies as the route to constraining feedback from growing FRB samples."],"supporting_citations":[{"why":"Supplies the CAMELS simulation suite, including the Latin Hypercube and 1P sets whose DM distributions are the training and testing data.","marker":"Villaescusa-Navarro et al. 2021"},{"why":"Provides the quasi-universal fitting function for p(DM) that the paper uses to compress each distribution into four parameters.","marker":"Macquart et al. 2020"},{"why":"The earlier CAMELS p(DM) analysis on the 1P set whose one-dimensional F trends the paper compares and partly contradicts.","marker":"Medlock et al. 2024"},{"why":"The current observational measurement of F from 76 FRBs that the paper uses to judge the realism of CAMELS scatter.","marker":"Baptista et al. 2024"},{"why":"The grid-based method for computing DM maps from hydro simulations that the paper adapts for CAMELS-SIMBA.","marker":"Batten et al. 2021"},{"why":"The neural-network emulation template and hyperparameter search strategy that the paper follows.","marker":"Nicola et al. 2022"},{"why":"Defines the SIMBA subgrid feedback model whose parameters are varied in the CAMELS-SIMBA suite.","marker":"Davé et al. 2019"},{"why":"Defines the IllustrisTNG subgrid model underlying CAMELS-TNG.","marker":"Weinberger et al. 2017"}],"fun_headline_variants":["Neural network maps feedback's non-monotonic imprint on FRB dispersion","AI predicts feedback's surprising control of FRB dispersion scatter","Neural net untangles feedback's nonlinear grip on FRB dispersion","FRB dispersion scatter's feedback twist decoded by neural network","CAMELS neural net uncovers feedback's odd twist on FRB dispersion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The construction of long sightlines assumes that the DM contribution of each successive 100 h^(-1) Mpc segment is an independent random draw from the redshift-interpolated CDF, so any correlation in gas density between adjacent segments is ignored; if that independence fails, the scatter σ_DM and therefore the F parameter would be systematically wrong.","fun_headline_variants_meta":{"raw":{"variants":["Neural network maps feedback's non-monotonic imprint on FRB dispersion","AI predicts feedback's surprising control of FRB dispersion scatter","Neural net untangles feedback's nonlinear grip on FRB dispersion","FRB dispersion scatter's feedback twist decoded by neural network","CAMELS neural net uncovers feedback's odd twist on FRB dispersion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00156,"raw_usage":{"total_tokens":6303,"prompt_tokens":1086,"completion_tokens":5217,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":702,"completion_tokens_details":{"reasoning_tokens":5127}},"tokens_in":702,"tokens_out":5217,"duration_ms":33971,"temperature":1.0,"reasoning_tokens":5127,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T11:14:40.278252+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Ray-trace a continuous sightline through a sequence of concatenated CAMELS boxes (or through a larger-volume counterpart such as TNG-300) out to z = 1, measure the resulting p(DM), and compare its σ_DM with the CDF-sampling method: if the independently-sampled variance differs from the directly-simulated variance at the level that would change F by more than the network's test-set error, the paper's core assumption is falsified.","supporting_citations":[],"review_version":1}