{"id":"bc7ddc1e-fada-46d2-ab38-1b432e0cd169","arxiv_id":"2504.17275","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A physics-embedded Bayesian neural network with a hand-crafted shell factor predicts fine structure and energy dependence of fission product yields.","lead":"A Bayesian neural network for fission yields got a new input feature: a hand-coded 'shell factor' marking known shell-closure masses. It now reproduces fine structure and energy trends in fission product yields, which could improve nuclear data for reactors and forensics.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The energy dependence of fine structure is written into the shell-factor input via exp(-E/kT), so the claimed 'emergent' behavior is partly imposed by construction; the abstract's 'energy-independent' wording is inconsistent with Eq. (5).","rationale":"The paper has real strengths: the out-of-sample Tonchev 2025 data were not used in training, the WAIC architecture selection is principled, and the qualitative agreement with prompt neutron multiplicities (not trained on) is supportive. Those points keep the paper publishable as a conditional contribution. My concern is not that the model fails predictively, but that a central interpretive claim is overstated. The shell factor in Eq. (5) contains an explicit energy-damping term, so the energy dependence of fine structure is substantially imposed by the input rather than discovered by the network. The abstract's 'energy-independent' wording is internally inconsistent with this. A targeted ablation (remove the Boltzmann factor, or make kT effectively infinite) would settle whether the energy-dependent attenuation of A = 134/138/143 is learned or inherited. This is the same area the reader flagged (hand-set shell-factor parameters), but it is more specific: even if W1, W2, W3, sigma, and kT are perfectly chosen, the claim of 'emergent' behavior needs to be qualified. Because the predictive validations are genuine and the fix is interpretational/experimental rather than fatal, I would keep the reader's CONDITIONAL verdict unchanged.","tokens_in":14068,"tokens_out":7652,"duration_ms":75471,"concrete_test":"Retrain the PE-BNN with the Boltzmann damping removed from SF, e.g., set SF = W1(N1+N4) + W2(N2+N5) + W3(N3+N6) or set kT to a very large value so that exp(-E/kT) is essentially constant, using the same training/validation split and WAIC-based architecture selection as in the paper. Then compare the predicted energy dependence of the A = 134, A = 138, and A = 143 peaks (Figs. 3-4) and the log-likelihood gain. If the fine-structure damping largely disappears or the model must learn it from data, the original model's energy dependence is carried by the hand-set exp(-E/kT) term; if the behavior persists, the network learns it from the FPY training data despite the removal.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equation (5) defines the shell-factor input as SF = exp(-E/kT) [W1(N1+N4)+W2(N2+N5)+W3(N3+N6)] with kT = 1.5 MeV, Gaussian centers at A = 134, 140, 144 (and mirrors), sigma = 1, and weights W1 = 4, W2 = 2, W3 = 1. This input therefore contains not only the fine-structure positions and relative strengths but also an explicit exponential suppression of all shell effects with excitation energy. At E_n = 14 MeV, exp(-E/kT) is about 1e-4, so the model receives essentially no shell signal at high energy by construction. The abstract's phrase 'energy-independent phenomenological shell factor' is inconsistent with Eq. (5) and with the body text's 'excitation-energy damping term.' More importantly, Section III states that the stabilization of A = 134 and the energy-dependent attenuation of A = 138 and A = 143 'emerge naturally from the trained network, without explicitly imposing such behaviors during training.' But the relative persistence of A = 134 (weight 4) and the suppression of the other peaks (weights 2 and 1, near A = 140 and A = 144), together with the common Boltzmann damping, are exactly what Eq. (5) injects. The reported log-likelihood gains show that the network uses the SF, but they do not show that the energy dependence of fine structure is learned from FPY data rather than inherited from the hand-set functional form. The independent Tonchev validation is valuable, but it tests the full model including the hand-set SF and therefore cannot separate learned from imposed behavior.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a physics-embedded Bayesian neural network (PE-BNN) for predicting fission product yields (FPYs) as functions of fragment mass, fissioning nucleus, and excitation energy. The central novelty is a hand-constructed 'shell factor' input, Eq. (5), of the form SF = exp(-E/kT) [W1(N1+N4) + W2(N2+N5) + W3(N3+N6)], with Gaussian peaks at A = 134, 140, 144 and mirror partners on the light-fragment side, weights W1 = 4, W2 = 2, W3 = 1, width sigma = 1, and a Boltzmann damping scale kT = 1.5 MeV. The network is trained with NUTS-based full Bayesian inference on a dataset composed of JENDL-5, EXFOR, and theoretical yields, with architecture selection by WAIC. The authors report that including the shell factor raises log-likelihood by 21.3% on training and 34.6% on test data, and they validate the predictions against energy-dependent experimental cumulative yields for 235U, 238U, 239Pu, and 232Th, including the out-of-sample Tonchev 2025 dataset, as well as against independently measured prompt-neutron multiplicities. The paper claims that the fine-structure energy dependence, in particular the persistence of A = 134 and the attenuation of A = 138 and A = 143, 'emerge naturally from the trained network,' and that the framework provides a physically interpretable, predictive tool for unmeasured incident neutron energies.","tokens_in":14452,"tokens_out":8619,"duration_ms":79485,"significance":"If the central claims hold, the framework is a practically useful addition to nuclear data evaluation: it produces complete mass distributions with Bayesian credible intervals at arbitrary neutron energies, is validated on a genuinely out-of-sample dataset (Tonchev 2025), and shows a clean, quantified log-likelihood gain from the physics-embedded input. The consistency between FPY energy trends and independently measured prompt-neutron multiplicities, with no neutron-emission data in the training set, is a notable emergent check. Strengths of the paper include the use of full NUTS-based Bayesian inference rather than variational approximations, explicit WAIC-based model selection, a transparent definition of the physics input, and validation across several isotopes. The main caveat is that the fine-structure hierarchy (A = 134 persistent relative to A = 138/143) is largely inherited from the hand-set shell-factor weights and damping, so the more ambitious interpretive claim is not yet established; the engineering value as a prediction tool is more solid than the claim of discovering the underlying energy-damping law.","major_comments":[{"comment":"The claim that the stabilization of A = 134 and the energy-dependent attenuation of A = 138 and A = 143 \"emerge naturally from the trained network, without explicitly imposing such behaviors during training\" is not supported as stated, because the dominant part of this behavior is written into the input feature. Equation (5) places Gaussian peaks at A = 134, 140, 144 with weights W1 = 4, W2 = 2, W3 = 1 and a common damping exp(-E/kT) with kT = 1.5 MeV; at E_n = 14 MeV the damping factor is about 10^-4, so the shell factor carries essentially no fine-structure signal at high energy by construction. The attenuated A = 138 and A = 143 structures fall on the tails of the weighted 140 and 144 Gaussians, so their decline relative to A = 134 is inherited from the hand-set weights and damping. The reported log-likelihood gain relative to the no-SF model shows that the network uses the SF feature; it does not show that the energy dependence of the fine structure is learned from FPY data rather than imposed by the functional form. A control experiment (equal weights, permuted centers, or removal of the exponential damping) is needed to support the attribution of this behavior to learning rather than construction.","section":"§III, \"Similarly, the prediction results for 232Th...\"; Eq. (5)"},{"comment":"The five shell-factor constants (W1, W2, W3, kT, sigma) and the peak centers are hand-set and described as adjustable hyperparameters, yet no sensitivity analysis is reported. The 21.3%/34.6% log-likelihood improvements compare the SF model with a no-SF model; they do not establish that the validation agreement in Figs. 2, 6-8 is robust to the specific choices, as opposed to being a product of tuning these constants to the training set. Since the paper contrasts itself with GEF, which it describes as relying on a large set of adjustable parameters, the claim that the present approach is \"transparent and reproducible... without relying on iterative parameter retuning\" is weakened by the five hand-set SF constants. A WAIC or validation scan over, for example, kT in 0.5-3 MeV, sigma in 0.5-2, and alternative weight assignments (including equal weights) is the minimal evidence needed to show that the out-of-sample performance is a property of the PE-BNN method rather than of this particular hand-tuned input.","section":"§II, Eq. (5) and surrounding text; §III log-likelihood paragraph"},{"comment":"The main predictive validation compares experimental cumulative yields with predicted independent mass yields, and the paper twice acknowledges this mismatch. Its justification, that \"the difference between cumulative and independent yields is expected to be small in the present incident-energy range,\" is asserted without a quantitative estimate for the specific mass chains plotted. The magnitude of the cumulative-independent correction is mass- and energy-dependent, and for some of the plotted nuclides (e.g., A = 132, A = 140) beta-decay feeding is not obviously negligible on the relevant timescales. Because the panels in Figs. 6-8 are the primary evidence for predictive capability, the authors should either quantify the correction (for instance from evaluated charge distributions and decay data, or by comparing JENDL-5 independent and cumulative yields for the same chains) or explicitly restrict the claims to trend-level agreement.","section":"§III, Figs. 6-8; §II data paragraph"}],"minor_comments":[{"comment":"The abstract as listed describes the shell factor as an \"energy-independent phenomenological shell factor,\" which contradicts Eq. (5) and the full-text abstract's \"phenomenological shell-related input feature with an excitation-energy damping term.\" The two abstracts should be aligned.","section":"Abstract (arXiv listing)"},{"comment":"References [15] and [16] are identical (Naik et al., Nucl. Phys. A 913, 185 (2013)); one is presumably intended to be a different publication and should be corrected.","section":"References [15] and [16]"},{"comment":"The caption states that results are shown \"in the case of selective learning of nuclides with high practical demand,\" but this selection procedure is not defined anywhere in the text; the caption should either explain the selection or refer to the section that does.","section":"Fig. 5 caption"},{"comment":"The observation noise sigma in the log-likelihood definition is not defined; state explicitly whether it is the training-data uncertainty, a model parameter, or a fixed constant, and how it is obtained.","section":"§III, Eq. (9)"},{"comment":"The paper motivates reproducibility but provides no code or data availability statement; given that the method depends on specific implementations of NUTS (NumPyro), WAIC, and the data-weighting scheme, a statement on availability of code and of the assembled training/test datasets would be valuable.","section":"§II (Computational Methods)"}],"recommendation":"major_revision","confidential_remarks":"For the editor: this is a solid applied ML-for-nuclear-data paper whose main empirical results (log-likelihood gains, out-of-sample validation on Tonchev 2025 data, consistency with prompt-neutron systematics) appear reliable and useful. My principal concern is the gap between the interpretive language in Section III ('emerge naturally from the trained network') and the construction of Eq. (5), which encodes the reported hierarchy and energy damping by hand. This is fixable within the manuscript's scope via a control experiment on the shell-factor parameters and softened attribution language; it does not, in my view, invalidate the predictive claims. I also note a version inconsistency between the arXiv abstract and the full-text abstract that should be resolved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one if you care about energy-dependent fission yields or ML in nuclear data. The novel piece is the shell factor SF in Eq. (5): a single hand-built input with Gaussians pinned to 134, 140, 144 and mirror partners, weighted 4/2/1 and Boltzmann-damped with kT=1.5 MeV. Adding it raises log-likelihood by 21.3% on training and 34.6% on test, and the gains are consistent across architectures. That is genuine evidence the feature is doing work, not just noise. The independent validation on Tonchev's cumulative-yield data and the consistency with prompt neutron multiplicity—despite not training on neutron emission—are real selling points. If you work in reactor physics or data evaluation, this is directly useful.\n\nThe soft spots are real but mostly fixable. First, the abstract says 'energy-independent phenomenological shell factor,' but Eq. (5) has an explicit exp(-E/kT). The body text correctly calls it an excitation-energy damping term, so this is a wording bug, not a fatal one. More substantively, the paper claims that stabilization of A=134 and attenuation of A=138/143 'emerge naturally from the trained network.' That is too strong. The network is given the hierarchy (weight 4 vs 2 vs 1) and the damping scale; the output behavior is in large part inherited from the input. The log-likelihood comparison shows the network uses the feature, not that it learned the energy dependence from FPY data. The Tonchev comparison validates the full model including the hand-set SF, so it cannot separate learned from imposed behavior. Second, the shell-factor hyperparameters (centers, widths, weights, kT) are not varied or uncertainty-quantified; a sensitivity analysis would strengthen the claim considerably. Third, comparing independent mass-yield predictions to cumulative experimental yields is explicitly trend-level; that is appropriately hedged in the text, but it caps how much strong point-by-point validation the paper can claim. No code or data artifacts are provided, which limits reproducibility but is common in this subfield.\n\nNone of this sinks the paper. The central result—that a physics-embedded feature plus WAIC model selection gives a usable energy-dependent FPY model with fine structure—holds up. The paper deserves a serious referee. I would recommend borderline accept after minor-to-moderate revision: fix the 'energy-independent' wording, add a sensitivity study of the SF parameters, and tone down the 'emergent' language. If you are in the field, bring it to reading group and cite it once the revision addresses the damping issue.","headline":"A useful extension of the authors' own BNN program—the shell-factor input is doing real work—but the paper overstates 'emergence' when the energy damping is in the input by construction.","tokens_in":14965,"tokens_out":2857,"would_cite":true,"duration_ms":25072,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["25.85.-w","02.50.Tt","07.05.Mh"],"model":"deepseek-v4-flash","headline":"A single shell factor gives fission-yield predictions their fine structure","keywords":["fission product yields","fine structure","shell effects","Bayesian neural network","energy dependence","WAIC","prompt neutron multiplicity","nuclear data"],"falsifier":"Retrain the PE-BNN with the shell-factor Gaussian centers shifted by two mass units, or with the damping scale $kT$ set to 1.0 or 3.0 MeV, and compare against the out-of-sample energy-dependent yields of reference [43]. If the predictions remain within the same credible intervals, the fine structures are not actually coming from the shell factor; if they degrade sharply, the hand-coded values are load-bearing. Applying the frozen shell factor to a fissioning system far below the actinides, where the 134/140/144 shell closures do not apply, would test the factor's generality beyond its calibration scope.","tokens_in":13872,"feed_emoji":"⚛️","tokens_out":5476,"duration_ms":48243,"temperature":0.7,"pith_summary":"This paper claims that a Bayesian neural network can predict energy-dependent fission product yields, including the sharp fine structures in the heavy and light mass peaks, when given a single hand-crafted input: a phenomenological shell factor encoding the positions and relative strengths of shell-stabilized fragment masses. The shell factor is fixed at mass numbers 134, 140, and 144 with weights 4:2:1 and mirror partners on the light side, damped by a Boltzmann factor $\\exp(-E/kT)$ with $kT=1.5\\,\\mathrm{MeV}$, a scale the authors motivate from dynamical fission simulations. Trained on a mix of evaluated, experimental, and theoretical yields, the network reproduces fine structures that are missed without this input, and its energy-dependent predictions agree with out-of-sample experiments and with prompt neutron multiplicity trends even though neutron emission data were never used in training. If true, this gives a practically useful way to interpolate yields at neutron energies where no measurements exist, for reactor physics and nuclear data applications.","feed_headline":"One shell factor unlocks fine-structure fission yield predictions","feed_subtitle":"Trained only on yield data, it matches energy trends and prompt-neutron systematics across actinides.","key_machinery":"The central object is the shell factor $$SF = \\exp\\left(-\\frac{E}{kT}\\right)\\left(W_1(N_1+N_4)+W_2(N_2+N_5)+W_3(N_3+N_6)\\right),$$ where $N_1,N_2,N_3$ are Gaussian bumps centered at $A=134,140,144$ with width $\\sigma=1$, $N_4,N_5,N_6$ mirror these features on the light fragment side, $W_1=4$, $W_2=2$, $W_3=1$, and $kT=1.5\\,\\mathrm{MeV}$. It encodes double shell closure at $A=134$ and deformed shell structures at $A=140$ and $144$, together with a Boltzmann factor that suppresses shell effects as excitation energy rises. Combined with WAIC-based selection of the 11-11 hidden-layer architecture, this input is what allows the network to reproduce fine structures and their energy dependence instead of interpolating smooth global trends.","core_discovery":"The central discovery is that most of the missing physics needed to capture fission-yield fine structures and their energy dependence can be compressed into one input feature rather than learned or parameter-tuned. The PE-BNN is a two-hidden-layer Bayesian neural network mapping the compound-nucleus charge, mass, excitation energy, fragment mass, and the shell factor to post-neutron independent mass yields, trained with a weighted dataset of evaluated, experimental, and theoretical yields. Adding the shell factor raises the validation log-likelihood by about 35% compared with the same network without it. The predicted yields show shell-effect damping with incident neutron energy: the doubly magic $A=134$ peak persists, while the deformed-shell peaks at $A=138$ and $A=143$ fade, a hierarchy that emerges from training rather than being imposed. The model also predicts that light-fragment yields barely move with energy while heavy-fragment ridge positions shift downward, consistent with the measured increase of prompt neutron multiplicity in the heavy fragment.","pith_inferences":["A natural stress test is to retrain the PE-BNN with the shell-factor Gaussian centers shifted off 134/140/144, or with $kT$ set to 1.0 or 3.0 MeV, and measure out-of-sample log-likelihood: the paper's claim implies a sharp optimum at the physical values.","The same single-physics-input recipe should transfer to other observables with known systematic features, such as charge yields or isomeric yield ratios, where a hand-coded feature could substitute for additional training data.","Because the shell factor is symmetric in fragment mass about half the compound mass, the framework implicitly assumes mirror-shell effects on the light fragment; if measurements for very neutron-rich fissioning systems break that symmetry, an asymmetric feature would be needed.","The emergence of prompt-neutron systematics from yield-only training suggests that FPY data alone constrain the mean number of prompt neutrons per mass chain, which could be exploited to produce correlated FPY and $\\bar\\nu$ predictions for nuclear data libraries."],"forward_implications":["Yields at unmeasured incident neutron energies, such as between 0.5 and 14 MeV, can be predicted with quantified Bayesian credible intervals, replacing linear interpolation between the three standard evaluated energies.","Fine-structure behavior is captured for multiple actinides with the same fixed shell factor, so no per-reaction retuning of phenomenological parameters is needed.","The hierarchy of shell damping, with $A=134$ robust and $A=138$, $A=143$ fading, emerges from training on yields alone, providing a physically interpretable account of shell effects.","The framework is explicitly scoped to low-to-intermediate neutron energies; at higher energies where multi-chance fission dominates, the paper states that an explicit treatment of individual fission channels would be required."],"supporting_citations":[{"why":"Supplies the prior BNN architecture and weighted data augmentation method that the present work extends.","marker":"[7]"},{"why":"JENDL-5 is the main evaluated dataset, split into 80% training and 20% validation.","marker":"[5]"},{"why":"Provides energy-dependent experimental FPY data used for training and as a benchmark for the fine-structure predictions.","marker":"[19]"},{"why":"Dynamical fission simulations motivating the $kT=1.5$ MeV damping scale of the shell factor.","marker":"[30]"},{"why":"Recent out-of-sample cumulative-yield measurements over a broad incident-energy range used to test predictive capability without retraining.","marker":"[43]"},{"why":"Bayesian shell-effect treatment in nuclear mass predictions that inspired the phenomenological shell factor input.","marker":"[23]"}],"fun_headline_variants":["One shell factor captures fission yield fine structure","Bayesian net: single physics feature predicts fission yields","Fission yield energy dependence from one shell feature","Physics-embedded model uses one input for yield predictions","One input feature explains fine structure in fission yields"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The shell factor with fixed Gaussian centers at $A=134,140,144$, weights 4:2:1, width 1, and damping scale $kT=1.5$ MeV correctly encodes the physical shell effects; if these hand-set numbers are wrong for a given fissioning system, the fine-structure predictions will be biased by construction regardless of what the network learns.","fun_headline_variants_meta":{"raw":{"variants":["One shell factor captures fission yield fine structure","Bayesian net: single physics feature predicts fission yields","Fission yield energy dependence from one shell feature","Physics-embedded model uses one input for yield predictions","One input feature explains fine structure in fission yields"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000502,"raw_usage":{"total_tokens":2410,"prompt_tokens":856,"completion_tokens":1554,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":472,"completion_tokens_details":{"reasoning_tokens":1482}},"tokens_in":472,"tokens_out":1554,"duration_ms":13435,"temperature":1.0,"reasoning_tokens":1482,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:43:48.027152+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the PE-BNN with the shell-factor Gaussian centers shifted by two mass units, or with the damping scale $kT$ set to 1.0 or 3.0 MeV, and compare against the out-of-sample energy-dependent yields of reference [43]. If the predictions remain within the same credible intervals, the fine structures are not actually coming from the shell factor; if they degrade sharply, the hand-coded values are load-bearing. Applying the frozen shell factor to a fissioning system far below the actinides, where the 134/140/144 shell closures do not apply, would test the factor's generality beyond its calibration scope.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the prior BNN architecture and weighted data augmentation method that the present work extends."},{"cited_title":"Iwamoto, N","cited_arxiv_id":null,"evidence_quote":"JENDL-5 is the main evaluated dataset, split into 80% training and 20% validation."},{"cited_title":"Gooden, C","cited_arxiv_id":null,"evidence_quote":"Provides energy-dependent experimental FPY data used for training and as a benchmark for the fine-structure predictions."},{"cited_title":"Ishizuka, M","cited_arxiv_id":null,"evidence_quote":"Dynamical fission simulations motivating the $kT=1.5$ MeV damping scale of the shell factor."},{"cited_title":"Tonchev, J","cited_arxiv_id":null,"evidence_quote":"Recent out-of-sample cumulative-yield measurements over a broad incident-energy range used to test predictive capability without retraining."},{"cited_title":"Niu and H","cited_arxiv_id":null,"evidence_quote":"Bayesian shell-effect treatment in nuclear mass predictions that inspired the phenomenological shell factor input."}],"review_version":1}