{"id":"13c0d405-37f5-402f-adb2-c14e30a4087f","arxiv_id":"2507.23297","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Neural density estimators paired with nested sampling recover JUNO's energy response parameters with bias below 0.55% and uncertainties consistent with statistics, in simulation-based closure tests.","lead":"The paper trains two neural networks to mimic how the JUNO neutrino detector converts particle energy into light signals, then uses these fast surrogates to infer detector calibration parameters from simulated calibration data. If it works on real data, this approach could make detector tuning thousands of times faster for JUNO and future neutrino experiments.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The near-zero bias claim is validated only against the JUNO simulator, not the physical detector; simulator fidelity is the untested load-bearing premise.","rationale":"I read the paper in good faith and find no internal flaw in the closure-test argument. The training/validation/testing grid design is careful: test points lie midway between training points, testing dataset 2 provides 1,000 independent realizations per exposure, and the comparison of estimated uncertainties with the actual spread of best-fit values is a valid calibration check. The missing factor of Delta-x in the TEDE Poisson likelihood (Eq. 14) is formally a typo but is absorbed by the free normalization constants N_s and does not affect the inference on the physics parameters. The reader's weakest assumption correctly identifies the load-bearing premise: every dataset in the paper comes from the same JUNO simulation, so the closure tests validate the pipeline against the simulator, not the simulator against the detector. The paper's own framing as a 'template' and its data-availability statement are consistent with this limitation, but the abstract's unqualified 'near-zero systematic bias' claim extends beyond what is demonstrated. Therefore the CONDITIONAL verdict is appropriate, and no change to the reader's verdict is needed.","tokens_in":31813,"tokens_out":10940,"duration_ms":139354,"concrete_test":"Generate a pseudo-observed dataset with the official JUNO software after applying a controlled perturbation that is not representable by (kB, fC, Y) alone, for example a 5% change in PMT quantum efficiency or a 2% change in scintillator attenuation length, while fixing (kB, fC, Y) at a known test point. Run the trained TEDE and NFDE inference pipelines on this dataset and compare the inferred parameters and their uncertainties to the known values. If the inferred parameters shift by more than the statistical uncertainties reported in Figure 8, the 'near-zero systematic bias' claim is conditional on exact simulator fidelity; if the shifts are negligible, this class of simulator mismatch would not threaten the claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that TEDE and NFDE recover JUNO's energy response parameters with 'near-zero systematic bias and uncertainties limited only by statistics' is established entirely through closure tests in which the 'observed' data are generated by the same official JUNO software used to produce the training data (Section IV A). The tests in Section II B demonstrate that the neural surrogates accurately emulate the simulator and that the Bayesian inference is statistically calibrated, but they cannot detect any mismatch between the simulator and the real detector. If the simulator's PMT quantum efficiency, optical attenuation, trigger model, or other response ingredients differ from the physical detector, the inferred (kB, fC, Y) will be biased in a real application by an amount that no test in this paper constrains. This is not an internal inconsistency; it is a boundary on the scope of the central claim as stated in the abstract.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two neural likelihood estimators, TEDE (a transformer-based histogram density estimator) and NFDE (a conditional normalizing flow), as surrogates for the expensive JUNO Monte Carlo simulation in the tuning of the energy-response parameters (kB, fC, Y). The estimators are trained on roughly 600 million simulated calibration events from five sources and are subsequently used in Bayesian nested sampling, Metropolis-Hastings MCMC, and MIGRAD+MINOS fits. The central result is a set of closure tests on parameter grids deliberately offset from the training points, with 1,000 independent fits per statistical exposure and 100 fits per parameter-space point, showing median biases below about 0.25% in kB, 0.55% in fC, and 0.02% in Y, and estimated 1-sigma uncertainties that match the spread of best-fit values.","tokens_in":31978,"tokens_out":13071,"duration_ms":145964,"significance":"If the validation holds, the paper provides a valuable methodological template for simulation-based inference in large liquid-scintillator neutrino detectors, where analytic likelihoods are unavailable. The internal validation is unusually thorough: validation and testing grids are offset from training points to test interpolation, 1,000 fits per exposure are used to assess bias and calibration, and three independent inference methods give consistent results. The code is publicly available. The main caveat is that all training, validation, and 'observed' data are produced by the same official JUNO software, so the demonstrated unbiasedness is conditional on simulator fidelity; this is a boundary on the physical claim rather than an internal inconsistency.","major_comments":[{"comment":"The abstract's claim of 'near-zero systematic biases' and 'uncertainties limited only by statistics' is established entirely through closure tests in which the 'observed' datasets are generated by the same official JUNO software [37] used for the training data (Section IV A). Any mismatch between the simulator and the physical detector would bias the inferred parameters in a real application, and no test in Sections II B or II B 2 can detect that mismatch. The abstract and conclusions should explicitly qualify these statements as simulation-based validation and identify simulator fidelity as the key assumption that must be tested on real calibration data.","section":"Abstract and Section IV.A"},{"comment":"Equation (14) defines the expected bin count as mu_{s,i} = N_s \\hat{p}_i^{TEDE}, but \\hat{p}_i^{TEDE} is a probability density per Np.e. (with the 1/Delta x factor in Eq. (1)), so sum_i \\hat{p}_i = 1/Delta x rather than 1. The formula should either use the bin probability (softmax output without the 1/Delta x factor) or read mu_{s,i} = N_s \\hat{p}_i^{TEDE} \\Delta x. Because N_s is a free parameter, the profile likelihood for (kB, fC, Y) is invariant under this rescaling and the central parameter estimates are likely unaffected; nevertheless, the equation as written contradicts the stated interpretation of N_s as the expected total number of events and must be corrected.","section":"Section IV.E.1.a, Eq. (14)"}],"minor_comments":[{"comment":"The early-stopping patience for NFDE is stated as 50 in the first paragraph of Section IV.D but as 100 in the later training description and in the caption of Figure 13; these values should be reconciled.","section":"Section IV.D and Figure 13"},{"comment":"The prior bounds for the flat priors used in the Bayesian fits are not stated; specifying them (presumably the extent of the training grid) would improve reproducibility.","section":"Section IV.E.3"},{"comment":"The y-axis ranges in Figure 8 are dominated by the uncertainty bands, which makes the claimed sub-percent-level bias curves visually indistinguishable; a zoomed inset or a separate axis for the median bias would improve readability.","section":"Figure 8"},{"comment":"There are minor language issues: 'both methods provides' should be 'both methods provide', and the Table I caption says 'three statistical distance' instead of 'three statistical distances'.","section":"Section I and Table I caption"}],"recommendation":"major_revision","confidential_remarks":"The paper is internally sound and the validation is strong; the main issue is that the abstract and conclusions overstate the scope of the claims relative to the simulator-only validation. The correction of Eq. (14) is straightforward and does not, as far as I can tell, affect the reported parameter estimates, but it should be fixed. The authors' membership in JUNO makes the use of the official software appropriate, and the data-sharing restrictions are clearly explained."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read: this is a solid SBI methods paper with unusually careful closure testing, applied to JUNO MC tuning. The central claim — that both a transformer histogram estimator and a planar-flow normalizing estimator recover the three energy-response parameters with well-calibrated uncertainties — holds for simulated data. Just be aware that the \"near-zero systematic bias\" claim is about the surrogates matching the simulator, not about the simulator matching the physical detector.\n\nWhat's genuinely good: the validation design. Training on a 21^3 grid and testing exactly midway between grid points is the right way to probe interpolation, and the aggregated 1,000-fit and 100-fit bias and coverage checks go beyond what most SBI papers do. The uncertainty scaling with 1/sqrt(N) and consistency across three inference methods (ultranest, MCMC, MINOS) give real evidence that the calibration is working. The code is on GitHub.\n\nSoft spots: the simulator-fidelity issue is real. All \"observed\" data come from the same JUNO software, so any mismatch with the real detector (PMT quantum efficiency, optical attenuation, trigger model) would shift the inferred parameters in a way this study cannot see. The abstract overreaches slightly when it says \"near-zero systematic biases\" without qualification; \"near-zero bias with respect to the simulation model\" would be accurate. Also, the data are not public per JUNO policy, so independent replication is limited to code-level checks. Minor: the fC bias at 0.55% is a bit larger than the other two; not a problem, but a line of comment would have been nice. The paper cites its own earlier conference paper, which is fine given this is a substantial extension.\n\nWho it's for: people doing MC tuning or SBI for large detectors; it's a useful template. The internal logic is clean — no circularity in the parameter-recovery test. I'd send it to review. It deserves referee time, with the main comments being about the abstract's scope and a request to discuss what would be needed to validate on real calibration data.","headline":"Solid SBI closure study for JUNO MC tuning; the near-zero bias claim is simulator-scope only, but the validation design is genuinely careful.","tokens_in":32866,"tokens_out":1638,"would_cite":true,"duration_ms":19991,"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":"The paper claims that two neural likelihood estimators, a transformer-based binned model and a normalizing-flow model, can substitute for JUNO's slow simulation and recover the three energy-response parameters with near-zero bias and…","keywords":["simulation-based inference","neural likelihood estimation","normalizing flows","transformer density estimator","Monte Carlo tuning","JUNO detector","energy response calibration","Bayesian nested sampling"],"falsifier":"Fit genuine JUNO calibration-source data with the trained surrogates and compare the recovered ($k_B$, $f_C$, $Y$) against independent determinations, or check the residuals of the predicted photoelectron spectra against observed spectra; a systematic shift or poor goodness-of-fit would show that the uncertainties are not purely statistical.","tokens_in":31614,"feed_emoji":"⚛️","tokens_out":6756,"duration_ms":69672,"temperature":0.7,"pith_summary":"Monte Carlo tuning of a large liquid-scintillator neutrino detector normally means rerunning the full simulation at every candidate parameter value, which becomes impractical when the likelihood has no closed form. This paper claims that two neural density estimators—a transformer that predicts binned spectra and a normalizing flow that gives continuous densities—can stand in for JUNO's simulation and, combined with Bayesian nested sampling, recover the three energy-response parameters ($k_B$, $f_C$, $Y$) from five calibration sources. In closure tests built from roughly one billion simulated events, both estimators show median biases below 0.25% for $k_B$, 0.55% for $f_C$, and 0.02% for $Y$, and the fitted uncertainties match the actual spread of best-fit values across 1,000 fits. Because the surrogate evaluates a spectrum in milliseconds rather than hours, systematic uncertainty studies become feasible. The claim matters because JUNO's physics goals require sub-percent control of the energy scale and resolution.","feed_headline":"Neural net surrogates tune JUNO's energy response with near-zero bias","feed_subtitle":"Two density estimators replace hours-long Monte Carlo runs, recovering kB, fC, and Y at the statistical limit.","key_machinery":"The load-bearing object is the learned conditional density $p(x|\\psi)$, where $x$ is the total number of photo-electrons and $\\psi = (k_B, f_C, Y, S)$ combines the three physics parameters with the calibration-source type. TEDE maps $\\psi$ to a temperature-scaled softmax over 800 bins, so the inference uses a binned Poisson likelihood; NFDE composes 120 planar flows conditioned on $\\psi$ to map $x$ into a standard normal, so the inference uses an extended unbinned likelihood with an exact Jacobian. Nested sampling over the three physics parameters plus five source normalizations produces the posterior. These surrogates compress roughly $10^6$-fold the cost of evaluating the full simulation at a parameter point.","core_discovery":"The paper establishes that neural likelihood estimation can replace the expensive simulation chain for detector energy-response tuning without losing statistical accuracy. The TEDE model conditions on the parameters and source type to output an 800-bin histogram of the photoelectron spectrum, while the NFDE model conditions on the same inputs to output an exact continuous density through a change of variables; both are trained on the official JUNO simulation output. Integrated with Bayesian nested sampling, they produce posteriors for ($k_B$, $f_C$, $Y$) whose medians sit on the true values within the quoted uncertainties and whose credible intervals track the true scatter of fits. The authors present this as a general template: if a simulation is too slow for iterative tuning, pre-train a likelihood surrogate on a parameter grid and then use standard Bayesian or frequentist inference on real data.","pith_inferences":["A real test the closure tests cannot provide: fit actual JUNO calibration data with the trained surrogates and check whether the pull distributions remain centered; any simulator-to-detector mismatch would show up there, not in the simulated-data tests.","The same training scheme could be extended to time-dependent calibration or to include the 3-inch PMT system; the paper names these directions but does not demonstrate them.","The unbinned normalizing-flow estimator may show a clearer advantage at lower statistics than the exposure levels used here, where its precision already matches the binned transformer."],"forward_implications":["JUNO's energy-response parameters can be tuned with near-zero systematic bias and uncertainties that shrink as $1/\\sqrt{N}$, so repeated fits across many parameter settings become practical.","The same trained surrogates work under frequentist minimization and two Bayesian samplers, giving consistent best-fit values and interval estimates.","A spectrum evaluation that took hours per parameter point drops to milliseconds, making experiment-wide systematic uncertainty quantification feasible.","The modeled conditional densities capture the non-linear and strongly correlated effects of $k_B$, $f_C$, and $Y$, so combined multi-source fits can break parameter degeneracies.","The approach carries over to other large neutrino and particle-physics experiments whose likelihoods are intractable."],"supporting_citations":[{"why":"supplies the official JUNO simulation whose output defines the ground-truth spectra for training and evaluation","marker":"[37]"},{"why":"defines the five calibration sources and the strategy of combining them to constrain the correlated parameters","marker":"[45]"},{"why":"establishes the non-linear energy-response model and the roles of kB, fC, and Y that the inference targets","marker":"[36]"},{"why":"provides the transformer-encoder architecture at the core of the TEDE density estimator","marker":"[46]"},{"why":"introduces planar and radial normalizing flows used to build the NFDE transformations","marker":"[48]"},{"why":"supplies the conditional density-estimation formulation that lets NFDE produce exact likelihoods","marker":"[49]"},{"why":"provides the nested sampling engine used for Bayesian posterior inference","marker":"[50]"},{"why":"grounds the extended unbinned likelihood that NFDE uses for parameter estimation","marker":"[58]"}],"fun_headline_variants":["Neural surrogates tune JUNO energy response to statistical limit","Neural density estimators make JUNO calibration bias-free","Simulation-free inference for JUNO tuning via neural likelihoods","Neural likelihood surrogates achieve statistical limit in JUNO"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the official JUNO simulation faithfully reproduces the real detector's energy response, because every training spectrum and every 'observed' dataset in the tests come from that same simulation; if the simulation is wrong, the near-zero biases measured here would not hold on real data.","fun_headline_variants_meta":{"raw":{"variants":["Neural surrogates tune JUNO energy response to statistical limit","Neural density estimators make JUNO calibration bias-free","Simulation-free inference for JUNO tuning via neural likelihoods","Neural likelihood surrogates achieve statistical limit in JUNO"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000685,"raw_usage":{"total_tokens":3078,"prompt_tokens":890,"completion_tokens":2188,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":506,"completion_tokens_details":{"reasoning_tokens":2114}},"tokens_in":506,"tokens_out":2188,"duration_ms":17678,"temperature":1.0,"reasoning_tokens":2114,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T10:51:51.333684+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fit genuine JUNO calibration-source data with the trained surrogates and compare the recovered ($k_B$, $f_C$, $Y$) against independent determinations, or check the residuals of the predicted photoelectron spectra against observed spectra; a systematic shift or poor goodness-of-fit would show that the uncertainties are not purely statistical.","supporting_citations":[{"cited_title":"An et al","cited_arxiv_id":null,"evidence_quote":"supplies the official JUNO simulation whose output defines the ground-truth spectra for training and evaluation"},{"cited_title":"Abusleme et al","cited_arxiv_id":null,"evidence_quote":"defines the five calibration sources and the strategy of combining them to constrain the correlated parameters"},{"cited_title":"Brehmer et al","cited_arxiv_id":null,"evidence_quote":"establishes the non-linear energy-response model and the roles of kB, fC, and Y that the inference targets"},{"cited_title":"Abusleme et al., JUNO sensitivity on proton decay p → νK + searches, CPC 47, 113002 (2023)","cited_arxiv_id":null,"evidence_quote":"provides the transformer-encoder architecture at the core of the TEDE density estimator"},{"cited_title":"Abusleme et al","cited_arxiv_id":null,"evidence_quote":"introduces planar and radial normalizing flows used to build the NFDE transformations"},{"cited_title":"Abusleme et al","cited_arxiv_id":null,"evidence_quote":"supplies the conditional density-estimation formulation that lets NFDE produce exact likelihoods"},{"cited_title":"Lin et al","cited_arxiv_id":null,"evidence_quote":"provides the nested sampling engine used for Bayesian posterior inference"},{"cited_title":"Abusleme et al., Calibration Strategy of the JUNO Experi- ment, JHEP 03, 004","cited_arxiv_id":null,"evidence_quote":"grounds the extended unbinned likelihood that NFDE uses for parameter estimation"}],"review_version":1}