{"id":"0d0b2fb5-432b-4c83-89e8-656d7ffa1c26","arxiv_id":"2506.23885","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":12,"one_line_summary":"A CNN surrogate for proton-synchrotron and hybrid lepto-hadronic blazar models reproduces SOPRANO spectra and fits the multimessenger SEDs of TXS 0506+059 and PKS 0735+178.","lead":"This paper trains a neural network to imitate a slow numerical code that models blazar jets with protons, making it fast enough to fit gamma-ray and neutrino data. The authors then use the tool to model two blazars linked to IceCube neutrinos, TXS 0506+059 and PKS 0735+178.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Surrogate error in the posterior region is unquantified: validation is internal to SOPRANO and no direct SOPRANO check of the best-fit/posterior SEDs is performed, so the P-syn/hybrid conclusions could be CNN artifacts.","rationale":"The reader's weakest assumption correctly identifies the absence of surrogate-error propagation as the key vulnerability. My read agrees and sharpens it: the validation metrics in §3.3 are computed on spectra drawn from the same simulation database used for training, and they are aggregate numbers that do not characterize the specific high-B, gamma_p,max~10^8 corner occupied by the P-syn fits. The paper never verifies the fitted CNN spectra against direct SOPRANO runs, so the reported P-syn/hybrid preference for PKS 0735+178 and the difference between the Gaussian and Poisson TXS fits could be driven by surrogate bias rather than by the physics of SOPRANO. This is not an internal inconsistency in the CNN construction; it is a missing validation step that is directly testable. The paper otherwise gives credit where due: it uses a large training set, transparent parameter ranges, and makes the tool available on MMDC, though without releasing weights or code for independent reproduction. Because the reader already set a CONDITIONAL verdict reflecting this concern, my stress-test does not change the verdict; it only reinforces the need for the direct SOPRANO check before the astrophysical conclusions are treated as robust.","tokens_in":16826,"tokens_out":6144,"duration_ms":73170,"concrete_test":"Take 200-1000 posterior samples from each of the three fits in Table 2 (plus the best-fit points) and run SOPRANO at those parameters to compute electromagnetic and neutrino spectra; compare directly with the CNN outputs and with the plotted SEDs. Then refit PKS 0735+178 with the Poisson likelihood using SOPRANO evaluations for a subset (e.g., every 10th MultiNest sample or 500 random posterior draws) and compare the P-syn/hybrid posterior weights and log-evidence. If the CNN/SOPRANO spectra agree within the reported MAE and the bimodality persists, the concern is resolved; if not, the surrogate bias is load-bearing and the fits need correction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the CNN 'effectively reproduces' SOPRANO's electromagnetic and neutrino outputs well enough for Bayesian inference. The only quantitative support is §3.3: average R2=0.66, MSE=2.25e-3, MAE=5.05e-3 on a random validation split drawn from the same 7e6-spectrum database used for training, with no region-resolved metrics and only 'similar results hold for neutrinos.' Section 4 then performs all fits with the CNN alone; the plotted best-fit SEDs, gray posterior envelopes, and neutrino spectra are CNN predictions, not SOPRANO outputs. In particular, the P-syn solutions for TXS 0506+059 (Poisson) and PKS 0735+178 sit at B~10^2.7-2.9 G and gamma_p,max~10^8, inside the P-syn-focused training region whose boundary behavior is not separately reported. Because validation error is not propagated into the MultiNest posteriors, a localized surrogate bias could shift posterior modes or create the PKS 0735+178 P-syn/hybrid bimodality reported in §4.3. The paper itself acknowledges this general risk only implicitly through the bimodality discussion; it never checks a single fitted spectrum against SOPRANO. This is the load-bearing gap: without such a check, the astrophysical conclusions are conditional on both SOPRANO and on unverified CNN fidelity at the fitted parameters.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper extends the authors' prior CNN-based surrogate modeling program to hadronic and hybrid lepto-hadronic blazar emission models. A convolutional neural network is trained on 7e6 synthetic spectra generated with the SOPRANO kinetic code over a wide parameter space, augmented by focused sampling in proton-synchrotron and hybrid regions. The trained surrogate, comprising separate CNNs for photon and neutrino outputs, is coupled to MultiNest and applied to fit the multimessenger SEDs of TXS 0506+059 and PKS 0735+178 under two neutrino likelihood treatments: a Gaussian likelihood using assumed neutrino flux points and a Poisson likelihood based on one IceCube event in one year. The paper reports best-fit parameters for both sources, finds hybrid or proton-synchrotron solutions depending on the likelihood and source, and makes the trained model available through the MMDC web platform.","tokens_in":17205,"tokens_out":4783,"duration_ms":55024,"significance":"If the emulator's fidelity is confirmed at the fitted parameter values, this is a useful and timely contribution: it would make Bayesian fitting of computationally expensive hadronic blazar models practical and would provide the community with a public tool for multimessenger SED interpretation. The scale of the training database (7e6 SOPRANO spectra), the integration into MMDC, and the explicit treatment of neutrino likelihoods are concrete strengths. The central claim, however, is that the CNN 'effectively reproduces' SOPRANO's electromagnetic and neutrino outputs; the support offered for that claim is internal validation only, and the reported average R2=0.66 is modest. The significance of the astrophysical applications is therefore conditional on an additional validation step that is currently missing.","major_comments":[{"comment":"The validation metrics reported in §3.3 (average R2=0.66, MSE=2.25e-3, MAE=5.05e-3 across the four photon models) are not sufficient to support the claim of 'great accuracy' in the Figure 1 caption, and no region-resolved or spectral-band-specific metrics are provided. The statement 'Similar results hold for neutrinos' is given without any quantitative support, even though the neutrino output is the decisive input for the Gaussian-versus-Poisson comparison in §4.2. The authors should report per-region accuracy, worst-case residuals, and neutrino-specific metrics (e.g., R2 or MAE on the relevant energy range) before claiming that the surrogate reliably reproduces multimessenger spectra.","section":"§3.3"},{"comment":"All fits in Section 4, including the plotted best-fit SEDs, gray posterior envelopes, and neutrino spectra, are CNN predictions; no spectrum is checked against SOPRANO after fitting. Because the validation in §3.3 is drawn from the same simulation database used for training, the reported accuracy measures emulator self-consistency rather than agreement with an independent calculation. Since the proton-synchrotron solutions for TXS 0506+059 (Poisson) and PKS 0735+178 lie at log10(B/[G])~2.7-2.9 and log10(γp,max)~8, inside the focused P-syn sampling region, a localized surrogate bias could shift the posterior modes or create the PKS 0735+178 bimodality discussed in §4.3. The authors should run SOPRANO on the best-fit parameters and on a sample of posterior draws, compare the electromagnetic and neutrino spectra, and either propagate surrogate uncertainty into the posteriors or explicitly quantify its effect on the reported parameter constraints.","section":"§4"},{"comment":"The astrophysical conclusions rest on several ad hoc choices that are not stress-tested: the assumed two-point neutrino flux at 183 TeV and 4.3 PeV with 10% uncertainty in the Gaussian likelihood, the one-year/one-event IceCube assumption in the Poisson likelihood, and the fixed log10(γe,min)=2.0 in the Poisson runs. The TXS 0506+059 result changes from a hybrid to a proton-synchrotron interpretation when the neutrino likelihood is changed (Table 2), so the model-selection statement is sensitive to exactly these choices. The authors should add sensitivity tests over exposure time, event count, and γe,min priors, and state which of the reported conclusions are robust to these choices.","section":"§4.2-4.3"}],"minor_comments":[{"comment":"The caption states that the results have 'great accuracy,' but the reported average R2=0.66 is modest; please either provide stronger region-specific metrics or soften the claim to match the quantitative results.","section":"Figure 1 caption"},{"comment":"The text states that Lp=1.5e51 erg/s is at the 'upper boundary of the parameter range used to train' the network, but Table 1 lists the maximum as log10(Lp)=52, which is a factor of several above the fitted value; please clarify which boundary is meant.","section":"§4.2"},{"comment":"The Poisson likelihood expression as written, log10 L = 2Σ(t mi) - Si log10(t mi) + log10(Si!), does not match the standard Cash statistic convention; please clarify the definition of L, the role of the factor 2, and the units of t mi.","section":"§4.1, Eq. (5)"},{"comment":"The text alternates between 'P-sync' and 'P-syn'; please use a single abbreviation consistently.","section":"Throughout"},{"comment":"The Data Availability statement says the network 'can be shared on a reasonable request' while §5 says it is publicly available through MMDC; please make the accessibility statement consistent.","section":"§5"}],"recommendation":"major_revision","confidential_remarks":"The paper is well within the scope of the journal and the public MMDC tool is a genuine asset. My main concern is the load-bearing validation gap: the emulator is validated only internally, and no SOPRANO cross-check is performed at the fitted parameters that drive the astrophysical conclusions. I would request that the authors add such a check and temper the accuracy claims, but I do not see this as requiring rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this is a real extension of the CNN surrogate program to proton-synchrotron and hybrid lepto-hadronic blazar models, trained on 7e6 SOPRANO spectra, and the authors are transparent about most of their assumptions. Second: the central accuracy claim is only weakly supported. The validation in Section 3.3 is entirely internal — CNN spectra are compared to SOPRANO spectra drawn from the same database used for training, and the reported R2=0.66 and MAE=5e-3 are modest for a claim of \"great accuracy\". No region-resolved metrics are given, and there is no direct SOPRANO computation at any of the fitted parameter values in Section 4. That gap matters because all posterior SEDs, neutrino spectra, and gray envelopes are CNN predictions, so surrogate bias could masquerade as a preferred model — for example, the P-syn/hybrid bimodality for PKS 0735+178.\n\nWhat is actually new: the previous papers built surrogates for SSC and EIC; here the hadronic case includes fast spectral transitions and neutrino outputs, which is a non-trivial extension. The application to TXS 0506+059 and PKS 0735+178 with both a Gaussian and a Poisson neutrino likelihood is a useful demonstration, and the discussion of the PKS bimodality is honest: the data are not sufficient to distinguish the two models. Making the trained CNN available through MMDC is a practical community resource.\n\nThe main soft spot is the missing external check. I agree with the stress-test: without a single fitted SED validated against SOPRANO, the astrophysical conclusions are conditional on both the simulator and unverified CNN fidelity at the fitted parameters. The Gaussian likelihood assumes an unmeasured neutrino flux level, the Poisson runs fix gamma_e,min=100 and assume one event over one year, and the TXS proton luminosity sits at the training boundary. These are stated in the text, but their effect on the inferred parameters is not quantified.\n\nThat said, the paper is a plausible methodological advance and the authors are clear about limitations. I'd send it to a serious referee, but the referee should ask for direct SOPRANO validation of the best-fit spectra and a propagation of surrogate error into the posterior distributions. Without that, the physics is interesting but not yet convincing.","headline":"Genuine extension of the CNN surrogate to hadronic scenarios, but the accuracy claim rests on internal validation alone and the posteriors are never checked against direct SOPRANO runs.","tokens_in":17805,"tokens_out":2573,"would_cite":false,"duration_ms":31125,"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 convolutional neural network can reproduce proton-synchrotron and hybrid blazar emission fast enough for Bayesian multimessenger fits.","keywords":["blazars","hadronic models","proton synchrotron","hybrid lepto-hadronic models","convolutional neural networks","multimessenger astronomy","neutrino emission","spectral energy distributions"],"falsifier":"Run SOPRANO at the best-fit parameter sets in Table 2 and at random parameter draws from the fitted posteriors, then compare the resulting electromagnetic and neutrino spectra with the CNN predictions; a mismatch larger than the quoted mean absolute error in the regions that drive the fits would show that the surrogate's posterior is biased.","tokens_in":16641,"feed_emoji":"🔭","tokens_out":10182,"duration_ms":101727,"temperature":0.7,"pith_summary":"This paper tries to establish that hadronic and hybrid models of blazar jets, which normally require minutes of computation per spectrum, can be replaced by a fast convolutional neural network without losing the essential physics. The network is trained on seven million spectra from the SOPRANO kinetic code, covering proton-synchrotron and hybrid lepto-hadronic scenarios, and it reproduces both the electromagnetic and the neutrino output. If the surrogate is as accurate as claimed, multimessenger fitting that combines radio-to-gamma-ray data with IceCube neutrino constraints becomes practical enough for full statistical exploration of the parameter space. The paper demonstrates the approach on TXS 0506+059 and PKS 0735+178, two blazars associated with IceCube neutrinos, showing that the preferred scenario can shift depending on how the neutrino information is encoded.","feed_headline":"CNN surrogate makes neutrino-blazar fitting tractable","feed_subtitle":"A fast CNN reproduces the code's radio-to-neutrino spectra, making full Bayesian fits of hadronic blazar models practical.","key_machinery":"The load-bearing object is a composite convolutional neural network trained on spectral outputs of SOPRANO, an implicit kinetic code that evolves the coupled Fokker–Planck equations for electrons, protons, photons, neutrons, and secondary particles (Equations 3 and 4). The network maps ten physical parameters — Doppler factor, blob radius, magnetic field, electron and proton injection indices, minimum and maximum electron Lorentz factors, maximum proton Lorentz factor, and electron and proton luminosities — to electromagnetic and neutrino spectra. The photon parameter space is split into four overlapping luminosity regions, each with its own network whose outputs are averaged in overlaps, and spectral derivatives are included in the training target to suppress oscillations; separate networks handle the neutrino spectra. This surrogate is what makes full posterior sampling of hadronic models computationally affordable.","core_discovery":"The central claim is that a convolutional neural network trained on $7\\times10^6$ spectra generated by the SOPRANO code can act as a surrogate for the full time-dependent hadronic model, reproducing the broadband electromagnetic and neutrino spectra of both proton-synchrotron and hybrid lepto-hadronic scenarios. The authors report an average validation $R^2=0.66$ and mean absolute error $5.05\\times10^{-3}$ across the four photon networks, with similar behaviour for the neutrino networks. Coupled to a Bayesian sampler, the surrogate recovers best-fit parameter sets for TXS 0506+059 and PKS 0735+178: a Gaussian likelihood on an assumed neutrino spectrum favours a hybrid model for TXS 0506+059, while a Poisson event-count likelihood favours a proton-synchrotron model for the same source, and the PKS 0735+178 posterior is bimodal between the two scenarios.","pith_inferences":["The validation metrics are measured only on spectra drawn from the same simulation database used for training; the paper does not propagate surrogate error into the posterior, so a locally biased region of the CNN could masquerade as a preferred model, for instance in the PKS 0735+178 bimodality.","A direct test of reliability would be to run SOPRANO at the best-fit parameters and at random posterior draws and compare the predicted spectra; the reported $R^2$ alone does not guarantee uniform accuracy across the parameter space.","The Poisson-likelihood results assume one detected neutrino in a one-year IceCube exposure and use a specific effective area; altering those assumptions would shift the inferred proton luminosity and could change the hybrid-versus-P-syn balance.","Because the paper explicitly excludes external photon fields, the current surrogate is most directly applicable to BL Lac-type jets; adding external radiation could reduce the very high proton luminosity found for TXS 0506+059."],"forward_implications":["For TXS 0506+059, the preferred model depends on the neutrino likelihood: a Gaussian likelihood on an assumed neutrino spectrum favours a hybrid scenario, while a Poisson likelihood based on one IceCube event per year favours proton synchrotron.","For PKS 0735+178, the Poisson-likelihood fit favours a proton-synchrotron solution, but the posterior is bimodal with a secondary hybrid mode, so the current multimessenger data cannot cleanly distinguish the two scenarios.","Because the CNN is fast, full posterior sampling becomes feasible, allowing model selection between P-syn and hybrid scenarios to be driven by the data rather than fixed before the fit.","The trained network is released as an online fitting service, letting other researchers fit their own multimessenger SEDs without recomputing the seven-million-spectrum simulation database.","A stated next step is to include external photon fields, which would extend the surrogate to sources with accretion disks or broad-line regions."],"supporting_citations":[{"why":"Supplies the SOPRANO kinetic code and simulation framework that generated the training spectra.","marker":"Gasparyan et al. 2022"},{"why":"Introduces the CNN architecture and spectral-derivative training method on which the hadronic surrogate is built.","marker":"Bégué et al. 2024"},{"why":"Extends the same CNN method to external inverse Compton models, providing the template for the neutrino-sector networks.","marker":"Sahakyan et al. 2024b"},{"why":"Provides the IceCube-170922A detection and the TXS 0506+059 SED and neutrino constraints used in the first application.","marker":"IceCube Collaboration et al. 2018a"},{"why":"Provides the PKS 0735+178 multimessenger SED and the earlier P-syn/hybrid modeling to which the CNN fit is compared.","marker":"Sahakyan et al. 2023"},{"why":"MultiNest is the Bayesian sampler used to explore the posterior distributions in Section 4.","marker":"Feroz et al. 2009"},{"why":"Supplies the EBL absorption model applied to the CNN spectra before comparison with observations.","marker":"Domínguez et al. 2011"},{"why":"Supplies the Poisson likelihood used for the neutrino event-count fits.","marker":"Cash 1979"},{"why":"Supplies the IceCube effective area used to convert predicted neutrino flux into expected counts.","marker":"Blaufuss et al. 2019"}],"fun_headline_variants":["CNN surrogate speeds hadronic blazar fits","Neural net models proton and hybrid blazars","Fast CNN fits neutrino-blazar spectra","AI surrogate for hadronic blazar emission","CNN emulator enables Bayesian blazar fits"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire fitting exercise rests on the trained network being a faithful stand-in for the numerical model, even though it was only validated against spectra drawn from the same simulation database used for training.","fun_headline_variants_meta":{"raw":{"variants":["CNN surrogate speeds hadronic blazar fits","Neural net models proton and hybrid blazars","Fast CNN fits neutrino-blazar spectra","AI surrogate for hadronic blazar emission","CNN emulator enables Bayesian blazar fits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000257,"raw_usage":{"total_tokens":1600,"prompt_tokens":988,"completion_tokens":612,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":604,"completion_tokens_details":{"reasoning_tokens":545}},"tokens_in":604,"tokens_out":612,"duration_ms":6876,"temperature":1.0,"reasoning_tokens":545,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:29:32.266248+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SOPRANO at the best-fit parameter sets in Table 2 and at random parameter draws from the fitted posteriors, then compare the resulting electromagnetic and neutrino spectra with the CNN predictions; a mismatch larger than the quoted mean absolute error in the regions that drive the fits would show that the surrogate's posterior is biased.","supporting_citations":[{"cited_title":"The Next Generation of IceCube Realtime Neutrino Alerts","cited_arxiv_id":"1908.04884","evidence_quote":"Supplies the IceCube effective area used to convert predicted neutrino flux into expected counts."}],"review_version":1}