{"id":"815312f2-cda1-4a81-8b17-860e893e6939","arxiv_id":"2607.22192","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A validated procedure to produce continuous-energy adjoint nuclear data from ENDF-B/VIII.0, covering all reaction laws including TSL and URR, for adjoint Monte Carlo transport.","lead":"The authors build a complete library of adjoint nuclear data for continuous-energy Monte Carlo neutron transport, covering elastic/inelastic scattering, fission, thermal scattering laws, and unresolved resonance range from ENDF-B/VIII.0. The library enables adjoint ('reversed') Monte Carlo simulations that sample particles backward from a detector, aiming to accelerate shielding calculations where the source is much larger than the detector.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Energy-only benchmark cannot validate adjoint flight kernel; spatial heterogeneity and flight weight correction are untested.","rationale":"The reader's weakest_assumption focuses on the variance of the single-sample thermal estimator. That is a real concern, but it is an efficiency concern: unbiasedness is preserved by the product-of-independent-unbiased-estimators argument, and the benchmark actually exercises the estimator across the 238U resonance region (the simulation uses TSL/SVT below 400kBT and sources/detectors force adjoint particles to traverse those energies). The observed 20x slowdown may partly reflect this estimator, but it does not invalidate the correctness claim. The more load-bearing gap is that the infinite-medium benchmark reduces the adjoint flight operator to a trivial case: the weight ratio is 1, and no spatial dependence, boundaries, or material interfaces are present. The central claim of the paper is 'adjoint Monte Carlo simulations' for shielding configurations, which inherently involve spatial transport. The conclusion explicitly defers spatial-heterogeneity studies to future work, so the validation as presented cannot support the general claim. A heterogeneous reciprocity test would directly settle whether the flight kernel is correctly implemented. If it passes, the CONDITIONAL verdict can be upgraded; if it fails, the method as described has a real error. Thus UNCHANGED is appropriate: the reader's CONDITIONAL is well-founded, and our concern sharpens the condition.","tokens_in":42329,"tokens_out":10960,"duration_ms":118477,"concrete_test":"Run a spatially heterogeneous reciprocity benchmark, e.g., a 238U slab adjacent to an H2O slab, with localized source/detector regions and the same 100-pair t-test as Section V.C. If the adjoint results deviate from forward beyond statistics, the flight weight correction Σt(r′,E′)/Σt(r,E) or boundary handling is at fault; agreement would validate the flight operator.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the proposed method produces a full adjoint nuclear data library and that the strategy is validated by reproducing forward Monte Carlo responses on a continuous-energy benchmark. However, the only validation configuration (Section V.B/V.C) is an infinite homogeneous medium; the text states 'the problem only depends on the energy variable.' In this geometry the adjoint flight operator (Section II.B.1) reduces to the trivial case where the weight correction Σt(r′,E′)/Σt(r,E) is identically 1 (E′=E and Σt is spatially constant). Thus any error in the implementation of the adjoint flight kernel, in the treatment of boundary crossings, or in the spatially varying cross-section ratio would go undetected by the reciprocity test. The 100 source–detector pairs all probe the same zero-dimensional geometry; they exercise collision kinematics over a wide energy range but not the spatial part of transport. The paper itself acknowledges in the Conclusions that 'a detailed study of the behavior of adjoint Monte Carlo games in the case involving flights and spatially heterogeneous media is needed.' Since real shielding applications are precisely the spatial problems the adjoint method is intended to accelerate, the evidence presented does not yet support the general validity of the method for those problems. The single-sample thermal estimator variance (Eq. (77)/(83)), by contrast, affects efficiency rather than correctness and is at least exercised in the benchmark through the 238U resonance region.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper develops a general methodology for producing adjoint nuclear data libraries for continuous-energy Monte Carlo transport. Building on the authors' prior adjoint-sampling framework (Ref. [12]), it formulates adjoint cross sections and adjoint distribution laws from ENDF data using a flux-shape prior g(E), and details processing schemes for elastic and inelastic scattering, fission, continuum and Kalbach-type reactions, Doppler-broadened elastic scattering, thermal scattering laws, and unresolved resonance range data. It also presents adaptive discretization algorithms for unidimensional and bidimensional tabulated data. The authors produce an adjoint library for ENDF-B/VIII.0 and validate the collision sampling by a reciprocity test on an infinite homogeneous 238U/1H mixture, comparing 100 direct/adjoint source-detector responses; 3 of 100 t-values exceed 2σ and none exceed 3σ. The paper concludes that spatial-heterogeneity and flight-kernel tests remain future work.","tokens_in":42720,"tokens_out":6442,"duration_ms":72369,"significance":"If the claims hold, the paper removes a major obstacle to practical continuous-energy adjoint Monte Carlo: the preparation of adjoint data covering the full diversity of modern nuclear data formats. The systematic treatment of TSL, URR, thermal broadening, and all ENDF reaction laws is a substantial contribution, and the production of a full ENDF-B/VIII.0 adjoint library is an impressive engineering result. The reciprocity test is a meaningful consistency check that goes beyond simple unit tests. The paper is also honest about the limits of its validation: the benchmark is energy-only, the adjoint runs are reported to be 20× slower than direct runs, and several approximations are acknowledged but not quantified. These limitations do not invalidate the method, but they materially affect the strength of the paper's broader claims.","major_comments":[{"comment":"The validation is performed in an infinite homogeneous medium, so the problem 'only depends on the energy variable.' In this geometry the adjoint flight operator (Eq. (5)) reduces to a trivial case: for energy-conserving flights in a spatially constant medium, the weight correction Σt(r′,E′)/Σt(r,E) is identically 1, and boundary crossings are never exercised. The 100 reciprocity pairs therefore test the collision kernel and its energy dependence, but not the spatial transport components of the adjoint game. The Conclusions explicitly state that 'a detailed study of the behavior of adjoint Monte Carlo games in the case involving flights and spatially heterogeneous media is needed.' Since the stated goal is a full adjoint Monte Carlo capability for shielding applications, the present evidence does not yet support the general validity of the method for those problems. A spatially heterogen","section":"Section V.B/V.C and Section II.B.1"},{"comment":"The on-the-fly evaluation of the Doppler-broadened scattering kernel relies on a single-sample Monte Carlo estimate of the integral in Eq. (77) and on a rejection-count estimator for C−1(v) in Eq. (83). The authors correctly argue that these estimators are unbiased in expectation, but no variance or convergence analysis is provided. For strongly resonant 0 K cross sections, the single-sample estimate can have very large fluctuations, which would not bias the adjoint game but could destroy the efficiency that motivates the method. This concern is concrete: Section V.C reports that the adjoint simulations took, on average, 20 times longer than the direct simulations to reach comparable statistical uncertainty. The manuscript should at least report empirical variances of these thermal estimators on the benchmark, or a theoretical bound, and discuss the implications for practical use.","section":"Section III.H.2, Eqs. (72)-(83)"},{"comment":"The discretization tolerances (10% for unidimensional, 20% for bidimensional data) and the 'leveler' mechanism are introduced to control storage size and variance, but no quantitative connection is made between these tolerances and the observed benchmark results. The weight-correction factor scales as 1/rd/o, where rd/o = fd/fo; in the 'leveled' regions, rd/o can be arbitrarily large, potentially producing very large particle weights. The paper states that the chosen settings 'have been found to work well on several test cases,' but no sensitivity study is provided. Since the central practical claim is that the discretized adjoint data preserve the variance-reduction properties of the idealized definitions in Eq. (12), the authors should either quantify the actual deviations of the produced data from the target tolerances or show that benchmark results are insensitive to the tolerance ch","section":"Section IV.E and Section V.A"}],"minor_comments":[{"comment":"The sentence 'we use the sampling' appears to contain a typo; it should probably read 'we use the sampling procedure' or 'we use the DBRC strategy.'","section":"Section III.H.2"},{"comment":"The expression for P(E→E′) has a somewhat unbalanced use of parentheses around the error functions; this makes the formula harder to read. A cleaner bracketing would help readers verify the signs for E<E′ and E>E′.","section":"Appendix A, Eq. (84)"},{"comment":"The 100 t-values are not independent, because for a given source the 10 detector responses are tallied in the same simulation and hence correlated. The statement that the t-values 'should follow a standard normal distribution' is therefore only an approximation. This does not affect the main conclusion, but it should be acknowledged.","section":"Section V.C, Table II"},{"comment":"The note that the two-stage Douglas-Peucker application yields a final tolerance 'twice as large as the individual tolerance' is potentially confusing. It should be clarified whether each of the two stages is run with the stated tolerance, so that the total error bound is the sum, or whether the individual tolerances are halved to keep the final bound.","section":"Section IV.D"},{"comment":"The choice to identify the fission spectrum χf with the spectrum at the highest incident energy when U is negative, and the arbitrary −10 MeV threshold for treating the spectrum as energy-dependent, are heuristic. The text explains the variance rationale, but it would be useful to state explicitly that this choice cannot cause bias because the weight correction uses the true direct distribution.","section":"Section III.F"}],"recommendation":"major_revision","confidential_remarks":"The paper is a substantial methods contribution, and the analytical derivations are generally careful. My main concern is that the validation does not exercise the spatial part of the adjoint transport operator, and that the thermal on-the-fly estimator is not analyzed for variance. These are not fatal flaws, but they should be addressed before publication. The reported 20× slowdown of the adjoint runs should also be discussed more prominently; otherwise the reader may be left with the impression that the method already provides the expected computational gains."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is the most complete treatment I've seen of producing continuous-energy adjoint nuclear data from ENDF/B-VIII.0. The authors go reaction by reaction—elastic, discrete inelastic, continuum, fission with all spectrum laws, SVT/DBRC thermal, TSL, URR—and give implementable formulas and discretization schemes. That alone is a substantial advance over Hoogenboom, MCBEND, and Diop, who all treated subsets. The reciprocity test is well designed: 100 source–detector pairs, 3 t-values above 2σ and none above 3σ, which is exactly what you expect by chance. There's no fitting-to-answer circularity; adjoint data are constructed from the ENDF data plus the user-supplied prior g(E). Credit where due: the detailed treatment of the Doppler-broadened kernel and the unbiased single-sample estimator for the integral is clever, and the discussion of the rejection-count estimator for C^{-1}(v) is honest about variance.\n\nThe soft spots are real but not deal-breakers. The validation is energy-only: infinite homogeneous medium, so the adjoint flight kernel's weight correction Σt(r',E')/Σt(r,E) is trivially 1. Errors in the flight implementation or in spatially varying cross-section ratios would go undetected. The authors acknowledge this in the conclusions. The variance of the single-sample thermal estimator is unquantified; it doesn't bias the game, but could kill efficiency in strongly resonant media. Also, adjoint was 20x slower in this benchmark; they call it problem-dependent, which is fair but leaves the efficiency motivation unproven. No code/library released, which makes reproduction harder, but the formulas are detailed enough to reimplement.\n\nWho should read it: anyone trying to implement adjoint Monte Carlo in a production code, or working on zero-variance games. It deserves a serious referee. I'd send it to peer review, but ask for at least one spatially heterogeneous test case and some variance diagnostics for the thermal estimator.","headline":"Complete adjoint nuclear data library from ENDF/B-VIII.0, with a solid collision-physics validation; spatial transport remains untested.","tokens_in":43199,"tokens_out":4727,"would_cite":true,"duration_ms":40340,"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":"Adjoint Monte Carlo neutron transport can be run with a complete library of adjoint nuclear data, covering every reaction in modern nuclear data evaluations, and matches forward simulations on a 100-case reciprocity test.","keywords":["adjoint Monte Carlo","neutron transport","adjoint nuclear data","reciprocity theorem","Doppler broadening","thermal scattering law","unresolved resonance range","nuclear data processing"],"falsifier":"Compute or measure the variance of the single-sample estimator for the Doppler kernel on a strongly resonant nuclide at thermal energy (e.g., U-238 at 600 K or a synthetic narrow resonance) and check whether the adjoint game's figure of merit degrades dramatically compared with the forward game; if the estimator variance diverges with resonance strength, the efficiency claim fails even though bias may be absent.","tokens_in":42237,"feed_emoji":"☢️","tokens_out":6906,"duration_ms":67621,"temperature":0.7,"pith_summary":"This paper aims to make adjoint Monte Carlo simulation—where fictitious particles are born at the detector and travel backward to the source—practical for real shielding problems by producing the missing ingredient: a complete adjoint nuclear data library. The method converts every reaction channel in modern evaluated nuclear data (elastic, discrete and continuum inelastic, fission, Kalbach, thermal scattering laws, unresolved resonance range, and Doppler-broadened elastic scattering) into adjoint cross sections and adjoint outgoing-energy/angle laws. The key maneuver is to define these adjoint laws by folding the direct data against an arbitrary importance-spectrum prior g(E), and to preserve unbiasedness through a weight correction that uses the same tabulated distributions for sampling and for the correction. The authors generate a full adjoint library for 555 nuclides and validate it on an infinite-medium U-238/H-1 benchmark: for 100 source–detector pairs, adjoint and direct Monte Carlo responses agree within statistical uncertainty (three pairs above two standard deviations, none above three). If the approach holds, adjoint games can exploit source–detector reciprocity to speed up calculations where a large source illuminates a small detector.","feed_headline":"Full adjoint nuclear data library passes 100-case reciprocity test","feed_subtitle":"Backward-sampling versions of every reaction type reproduce direct simulation results across 100 source–detector pairs.","key_machinery":"The central object is the non-normalized adjoint distribution f̃†(E′→E) = g(E) σ(E) ν(E) f(E→E′, μ), whose integral over E gives the adjoint cross section and whose normalized version is the adjoint scattering law. It is stored with a two-dimensional mesh of piecewise-linear (unit-based) interpolations; sampling is done by stochastic interpolation between bracketing incident-energy distributions and inverse-transform sampling of the cumulative. The load-bearing consistency requirement is that the density used in the weight correction is exactly the density of the distribution sampled, which is what makes arbitrary discretization, angular simplifications, and Monte Carlo estimates of the Dopp","core_discovery":"The paper establishes that a fully general treatment of adjoint collision sampling is possible: for every reaction present in modern evaluated nuclear data, an adjoint cross section and an adjoint distribution law can be precomputed and tabulated so that the adjoint game remains unbiased and efficient. Adjoint data are defined from direct data by folding them against a user-chosen prior guess g(E) of the direct flux shape; because the weight correction in the sampling equation is evaluated with the same discretized density used for sampling, tabulation errors do not bias results—they only degrade variance. The same principle covers thermal Doppler broadening, where the direct kernel is estim","pith_inferences":["If the method scales past infinite-medium benchmarks, the same reciprocity test could be run in a spatially heterogeneous geometry (e.g., layered shielding) to check whether the weight-correction sampling remains practical when adjoint flights cross material boundaries.","The paper leaves the variance of the single-sample Doppler-kernel estimator unexamined; for a strongly resonant thermal scatterer, that estimator is unbiased but could produce very large weights, so a variance analysis or an adaptive number of samples near resonances is a natural extension.","The arbitrary prior g(E) is implicitly a tuning lever: choosing it from a preliminary forward calculation, or iterating between forward and adjoint runs, could push the game toward the zero-variance limit the authors mention, at the cost of an extra production step.","Because the formalism is stated for general particle transport, the same adjoint-data preparation could extend to photon transport or coupled neutron-photon games, though the paper demonstrates only neutrons."],"forward_implications":["Adjoint Monte Carlo can be implemented with continuous-energy nuclear data covering all reaction types, not just multigroup approximations, including thermal scattering laws and the unresolved resonance range.","Any detector response computable by a forward simulation is also computable by the adjoint game through reciprocity, so shielding configurations with a large source and a small detector become natural targets for variance reduction.","A complete adjoint library can be produced from a modern evaluated data set within about 20 CPU-minutes and about 2 GB of storage, making on-the-fly adjoint sampling feasible in a production code.","Discretization error, isotropic angular approximations, and even stochastic estimation of the thermal kernel do not bias the game, provided the weight correction uses the same sampled density—so the method degrades gracefully under approximation.","The approach is a stepping stone to adjoint transport in a next-generation production Monte Carlo code."],"fun_headline_variants":["Adjoint nuclear data library now covers every reaction type","Complete adjoint nuclear data library for unbiased Monte Carlo","Backward-sampling Monte Carlo: full nuclear data library built","All modern reactions in new adjoint nuclear data library"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The practical convergence of the whole scheme rests on the claim that one Gaussian sample of the Doppler-broadened elastic-scattering kernel, together with a rejection-count estimate of its normalization, is an unbiased estimator; this holds in expectation for any cross section, but if the target cross section is strongly resonant the estimator's variance can explode, and without a variance bound the adjoint game may lose the efficiency it is meant to provide.","fun_headline_variants_meta":{"raw":{"variants":["Adjoint nuclear data library now covers every reaction type","Complete adjoint nuclear data library for unbiased Monte Carlo","Backward-sampling Monte Carlo: full nuclear data library built","All modern reactions in new adjoint nuclear data library"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000206,"raw_usage":{"total_tokens":1209,"prompt_tokens":697,"completion_tokens":512,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":441,"completion_tokens_details":{"reasoning_tokens":447}},"tokens_in":441,"tokens_out":512,"duration_ms":5622,"temperature":1.0,"reasoning_tokens":447,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T05:29:10.510426+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute or measure the variance of the single-sample estimator for the Doppler kernel on a strongly resonant nuclide at thermal energy (e.g., U-238 at 600 K or a synthetic narrow resonance) and check whether the adjoint game's figure of merit degrades dramatically compared with the forward game; if the estimator variance diverges with resonance strength, the efficiency claim fails even though bias may be absent.","supporting_citations":[],"review_version":1}