{"id":"0d967d2a-86d3-45a1-8ec8-c24943a7d83a","arxiv_id":"2507.07109","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A neural network trained on about 15% hand-labeled events and self-generated labels for the rest extracts atomic charge state and atomic number from VAMOS++ data, claiming a 9% resolution gain.","lead":"This paper uses a deep neural network with only a small set of hand-labeled examples to sort heavy ions detected by the VAMOS++ spectrometer by atomic charge state and atomic number. It claims to reduce the analysis time from months to hours while improving peak resolution, which matters for nuclear physics experiments that need fast, reproducible particle identification.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported accuracy on unlabeled events is circular: targets for 82% (q) and 87% (Z) of events are rounded network predictions, and the FWHM/9% gains are computed on those self-labeled data.","rationale":"The reader's weakest assumption is exactly the problem I find most load-bearing: the unlabeled-range targets are generated by rounding the network's own predictions, and the reported FWHM and 9% improvement are evaluated on those same self-generated labels. I agree with the reader that this makes the accuracy claim circular. The hard pseudo-labeling mechanism strengthens the concern: since the loss is minimized by moving the output toward the nearest integer, narrow peaks in the final distribution are partly an artifact of the training objective, not necessarily evidence of correct integer assignment. The paper's existing labeled anchors for q=25-34 and Z=30-40, plus physically motivated inputs such as A/q and positions, make the method plausible, so this is not a reason to reject the work outright. However, without an external check on the unlabeled region, the claimed 1.0-1.2% FWHM and 9% improvement cannot be taken as established. A gamma-ray coincidence validation, which the paper itself notes as standard in Section 3, would settle the question. No other concern appears more load-bearing: the architecture description, training convergence, and comparison plots are consistent with the stated implementation. The appropriate verdict remains CONDITIONAL, so I recommend no change from the reader's verdict.","tokens_in":8136,"tokens_out":4961,"duration_ms":56035,"concrete_test":"Select a subset of events in the unlabeled ranges q_u=[35,50] and Z_u=[41,63] and verify their q_NN and Z_NN assignments against independent gamma-ray coincidence gates using known transitions of the fission fragments, as suggested in Section 3. Quantify the fraction of events for which the network's integer assignment disagrees with the gamma-gated assignment; a systematic disagreement above the few-percent level would invalidate the claim that the self-labeled FWHM values reflect true classification accuracy.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the network autonomously and correctly classifies the large unlabeled region. The evidence for this is not independent. In Section 4.1, for unlabeled q in [35,50] (82% of events) the training target is q_t = floor(q_p + 0.5), i.e. the network's own current prediction rounded to the nearest integer; Section 4.2 does the same for Z in [41,63] (87% of events). The FWHM values and the 9% improvement over conventional analysis quoted in Sections 4.1, 4.2 and 5 are computed on this self-labeled population. Because the objective pushes outputs toward the nearest integer, narrow final distributions can be manufactured even if the absolute q/Z assignments are wrong: hard pseudo-labeling reinforces initial errors, and the random 80/20 validation split is not an external reference. The physical inputs (energies, A/q, positions) and the few labeled low-state events constrain the problem, so the method may well work, but the paper does not demonstrate that the unlabeled-region labels are correct. This is the load-bearing weakness: without external validation, the headline resolution numbers rest on circular targets.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a deep neural network approach to determine the atomic charge state q and atomic number Z of heavy ions detected by the VAMOS++ magnetic spectrometer. The network is trained on an experimental dataset where only 18% of events (q in [25,34]) and 13% of events (Z in [30,40]) are labeled with integer values; for the remaining unlabeled events (q in [35,50] and Z in [41,63]) the training target is the network's own current prediction rounded to the nearest integer. The authors report FWHM/q_NN from 1.0% to 1.2%, FWHM/Z_NN from 1.1% to 1.2%, and a 9% improvement in FWHM over conventional analysis, with analysis time reduced from months to hours.","tokens_in":8509,"tokens_out":5253,"duration_ms":54149,"significance":"If the claimed accuracy on the unlabeled region were independently verified, the method would be a practical and time-saving contribution to heavy-ion spectrometer data analysis, and the use of a minimal labeled fraction is an interesting and potentially valuable idea. The paper also explicitly uses raw detector energies and positions without providing calibration relationships, which is a useful goal for reducing human bias. However, the central quantitative claims rest on self-generated labels, and the current manuscript does not provide any external check on the correctness of the unlabeled-region predictions. The reported FWHM values and the 9% improvement measure self-consistency of the network rather than physical accuracy.","major_comments":[{"comment":"The training targets for the unlabeled majority are the network's own rounded predictions: for q in [35,50] (82% of events) the target is q_t = floor(q_p + 0.5), and for Z in [41,63] (87% of events) the target is Z_t = floor(Z_p + 0.5). The validation split is subject to the same pseudo-labeling, and early stopping uses the validation loss computed on these self-generated targets. Therefore the validation RMSD values (0.177 for q, 0.227 for Z), the FWHM values in Sections 4.1, 4.2 and 5, and the 9% improvement over conventional analysis are not independent measurements of accuracy. A network can produce arbitrarily narrow distributions on the unlabeled region by learning to output values close to the nearest integer, even if the absolute q or Z assignments are systematically wrong. The central claim that the network 'autonomously and accurately classifies the remaining events' (Abstract) is therefore not established. I request an external validation on the labeled region only (e.g., reporting FWHM and peak positions for q in [25,34] and Z in [30,40] where ground truth is known), or a physics-based check such as gamma-ray coincidences as mentioned in Section 3, or a comparison against a fully independently analyzed subset of the unlabeled region.","section":"Sections 4.1 and 4.2"},{"comment":"The comparison between the neural-network results and the 'conventional analysis' results (q_conv, Z_conv) is not quantified in enough detail to support the claimed 9% FWHM improvement. It is not stated how q_conv and Z_conv are obtained (the paper says only that they come from 'a complete process of the conventional analysis method'), what peak-fitting procedure is used, how many events are included, and whether the improvement is dominated by the self-labeled region. If the same unlabeled events are used for both q_NN and q_conv, the comparison is partly circular because the q_NN distribution is generated under the pseudo-label objective. Please provide the exact definition of FWHM used, the uncertainty on the quoted values, and a breakdown of the improvement for the labeled and unlabeled ranges separately.","section":"Sections 5 and Figures 5(d), 6(d)"},{"comment":"The paper provides no evidence that the pseudo-labeled training converges to the true physical q and Z rather than to a self-consistent but inaccurate solution. Because only 18% (q) and 13% (Z) of events carry externally supplied labels, the network's behavior on the unlabeled range is constrained only by the smoothness of the input-output mapping and by the rounding rule, which can reinforce initial systematic errors. I suggest the authors perform a stability analysis by retraining from several random initializations and reporting the variance of the predicted q and Z for the unlabeled events; if the predictions are stable across seeds, this still does not prove correctness, but it would at least demonstrate that the result is not a single-run artifact. Additionally, the paper mentions that coincident characteristic gamma rays can verify isotopic identification, but this verification is not performed or reported; applying it to a subset of the data would provide the missing external reference.","section":"Section 4"}],"minor_comments":[{"comment":"The input list states 'the reciprocal of the mass-over-charge, (A/q), and of the Lorenz term, (gamma-1)' which is confusing: (A/q) and (gamma-1) are not reciprocals. Please clarify whether the inputs are q/A and 1/(gamma-1) or A/q and (gamma-1), and correct the typo 'Lorenz' to 'Lorentz'.","section":"Section 4.1"},{"comment":"Equation (3.4) as written is dimensionally inconsistent with the definition of u as 931.494 MeV/c^2: the denominator 1 u * (gamma-1) * (A/q) has units of MeV/c^2, not MeV. Please either define u as an energy (931.494 MeV) or include the factor c^2 explicitly, and state the assumed units of E_tot.","section":"Equation (3.4)"},{"comment":"The caption of Figure 4 does not indicate which curves correspond to q and which to Z, although the text refers to both. Please add labels or a legend to the figure.","section":"Figure 4"},{"comment":"The sentence beginning 'Figure demonstrates a sharp definition...' is missing a figure number. It should refer to Figure 6(a) and (b).","section":"Section 4.2"},{"comment":"The derivation of q in equation (3.4) uses the total energy E_tot, but the text says E_tot = E_um + Delta E + E_res, while the input variables to the network for the q analysis include raw energies and the reciprocal of A/q; it would be helpful to state explicitly whether E_tot is used in the network inputs or only in the reference q_conv.","section":"Section 3"}],"recommendation":"major_revision","confidential_remarks":"The central issue is the circular pseudo-labeling: the validation and performance metrics are computed on targets that are the network's own rounded predictions. This is a load-bearing flaw that cannot be fixed by rewording; the authors need to provide an independent check, such as performance on the labeled subset, gamma-ray coincidence verification, or a simulated dataset with known ground truth. If such validation is not possible with the existing data, the paper's claims would need to be substantially weakened to a demonstration of self-consistent labeling without quantitative accuracy claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper reports a deep regression network for atomic charge state and atomic number analysis in the VAMOS++ spectrometer, trained on an experimental dataset where only a small fraction of the events carry integer labels. The method is clearly described, the dataset is real and linked to a DOI, and the practical payoff is real: a manual iterative calibration that takes months is reduced to hours. The authors also build sensibly on their earlier trajectory-reconstruction work. I believe the technique will be useful to anyone working with heavy-ion spectrometers.\n\nThe soft spot is the one the stress-test flags, and it is load-bearing. For the unlabeled majority (82% of q events, 87% of Z events), the training target is floor(prediction + 0.5), i.e., the network's own current output rounded to the nearest integer. The reported FWHM values and the 9% improvement over conventional analysis are computed on that self-labeled population. A random 80/20 validation split does not break the circularity, because both splits are labeled by the same rounding rule. So the paper does not actually demonstrate that the network correctly identifies the unlabeled high-q and high-Z events in an absolute sense. The physical inputs (A/q, energy losses, positions) and the labeled low states constrain the problem enough that I suspect the method works, but the evidence presented is not independent.\n\nI also note two smaller issues. The paper does not cite the substantial semi-supervised learning literature on pseudo-labeling and self-training, which this method is an instance of. And the abstract's language about \"optimal\" and \"unprecedented\" results is stronger than what the validation supports.\n\nIf I were refereeing this, I would ask for external validation on the unlabeled region: gamma-ray coincidences, or a comparison on a different experiment where q and Z are known from another method. That would turn a plausible method into a demonstrated one. As it stands, the paper is a solid engineering contribution with a gap between its claims and its evidence.\n\nFor a reading group, it is a good case study in how easy it is to fool yourself with self-generated labels. I would not cite it myself in the next year, but I would send it to peer review because the method is new and the concern is addressable with more work.","headline":"A practical time-saver for VAMOS++ analysis, but the claimed accuracy on unlabeled events is not independently established because the network's own rounded predictions are used as training targets.","tokens_in":8964,"tokens_out":1576,"would_cite":false,"duration_ms":18792,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A deep network can sort the full spectrum after seeing only the lowest, cleanest charge states and atomic numbers labeled by hand.","keywords":["deep neural networks","atomic charge state","atomic number","magnetic spectrometer","fractional labeling","pseudo-label training","ionization chamber","isotopic identification"],"falsifier":"Take a held-out sample from the unlabeled ranges, determine the true charge state and atomic number independently, for instance through coincident characteristic gamma rays or a complete manual analysis of a few thousand events, and compare with the network's predictions. A systematic offset in the unlabeled region would show that the self-labeling loop is not being checked against the ground truth.","tokens_in":7845,"feed_emoji":"⚛️","tokens_out":7486,"duration_ms":69951,"temperature":0.7,"pith_summary":"This paper claims that a deep regression network can take over the two hardest calibration steps in magnetic-spectrometer analysis, finding an ion's atomic charge state $q$ and atomic number $Z$, using only a small slice of precisely labeled events. On a fission-fragment dataset, the network was trained with integer labels for only the lowest, best-resolved charge states and atomic numbers, while all higher values were labeled by rounding the network's own prediction in each training step. The resulting spectra reach FWHM($q_{NN}$)/$q_{NN}$ between 1.0% and 1.2% and FWHM($Z_{NN}$)/$Z_{NN}$ between 1.1% and 1.2%, a 9% tighter resolution than the conventional step-by-step analysis, and the whole analysis runs in hours rather than months. If correct, this would make high-quality isotopic identification routine and reproducible for this class of spectrometers.","feed_headline":"Neural net labels ion spectra 9 percent sharper than manual analysis","feed_subtitle":"Trained on a fraction of hand-labeled events, it resolves charge state and atomic number in hours, not months.","key_machinery":"The working engine is a dense feed-forward regression network, eight layers of 32 units, trained on $5\\times10^7$ experimental events split 80/20 into training and validation sets. In the labeled region the target is a fixed integer; in the unlabeled region the target is the network's own output rounded to the nearest integer at each training step. The input mix is deliberately low-level: the ten raw ionization-chamber row energies, positions on the entrance window and focal-plane detectors, and the reciprocal mass-over-charge and Lorentz-factor terms, so the network has to infer the calibration-related correlations itself rather than receiving them as formulas.","core_discovery":"At its center, the paper reports a training scheme that supplies true integer labels only for the low-lying, well-resolved states and lets the network generate its own labels for the rest: for every unlabeled event the target is updated to the rounded value of the network's current prediction, so the model effectively pulls the entire ridge of charge-state and element bands into integer steps. With raw ionization-chamber energies and position variables as inputs, the trained network removes the window-deformation and position dependencies visible in the conventional spectra and yields charge-state and atomic-number distributions with 9% better FWHM resolution than the complete conventional analysis. The authors present this as a complete replacement for the months-long iterative correction procedure, producing the same or better isotopic identification from the same experimental data.","pith_inferences":["If the self-labeling scheme generalizes, it becomes a template for any detector where a smooth response function distorts an intrinsically integer-valued observable: anchor a few resolved integers and let the network fill in the ladder.","An independent check is still missing: identities for events in the unlabeled ranges, established by gamma-ray coincidences or a manual reanalysis of a small subset, would settle whether the self-generated labels match true charge states and elements.","Run-to-run transfer may require retraining, since the window deformation and electronics response can change between experiments; the paper demonstrates one dataset rather than a universal trained model."],"forward_implications":["Routine VAMOS++ charge-state and $Z$ analysis would take hours instead of months, with no manual iteration.","The 9% FWHM improvement over the conventional method propagates directly into cleaner mass-number identification when $q_{NN}$ is combined with mass-over-charge.","The procedure makes the analysis reproducible: replaying the same trained network on the same raw data removes the human-bias component.","Because accurate absolute energy calibration of the ionization chamber is not needed, the method can be applied when calibration sources are unavailable."],"supporting_citations":[{"why":"Supplies the $5\\times10^7$ fission-fragment events that form the training and validation dataset.","marker":"[6]"},{"why":"Provides the seven-dimensional trajectory reconstruction that yields the mass-over-charge, velocity, and rigidity inputs used by the networks.","marker":"[5]"},{"why":"Describes the ionization-chamber energy analysis and the conventional trajectory-reconstruction method that the neural approach replaces.","marker":"[4]"},{"why":"Establishes the VAMOS performance baseline and the identification requirements that the improved analysis must meet.","marker":"[1]"},{"why":"Documents the VAMOS++ detector and acceptance, the system whose charge-state and Z analysis is under study.","marker":"[2]"},{"why":"Supplies the dual position-sensitive MWPC telescope whose entrance and focal-plane positions are input variables.","marker":"[7]"}],"fun_headline_variants":["Neural net labels ion spectra 9% sharper in hours","DL cuts ion spectra analysis from months to hours","Fraction-labeled DNNs sharpen ion spectra by 9%","Self-labeling network resolves ion spectra 9% sharper"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the network's self-generated rounded labels for the unlabeled events are correct enough to serve as training targets, because the reported resolutions and the 9% improvement are measured on those self-labeled events rather than on independently confirmed identities.","fun_headline_variants_meta":{"raw":{"variants":["Neural net labels ion spectra 9% sharper in hours","DL cuts ion spectra analysis from months to hours","Fraction-labeled DNNs sharpen ion spectra by 9%","Self-labeling network resolves ion spectra 9% sharper"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000617,"raw_usage":{"total_tokens":2857,"prompt_tokens":931,"completion_tokens":1926,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":547,"completion_tokens_details":{"reasoning_tokens":1859}},"tokens_in":547,"tokens_out":1926,"duration_ms":12837,"temperature":1.0,"reasoning_tokens":1859,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:13:14.032762+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a held-out sample from the unlabeled ranges, determine the true charge state and atomic number independently, for instance through coincident characteristic gamma rays or a complete manual analysis of a few thousand events, and compare with the network's predictions. A systematic offset in the unlabeled region would show that the self-labeling loop is not being checked against the ground truth.","supporting_citations":[{"cited_title":"10.26143/GANIL-2022-E826_21","cited_arxiv_id":null,"evidence_quote":"Supplies the $5\\times10^7$ fission-fragment events that form the training and validation dataset."},{"cited_title":"Rejmund and A","cited_arxiv_id":null,"evidence_quote":"Provides the seven-dimensional trajectory reconstruction that yields the mass-over-charge, velocity, and rigidity inputs used by the networks."},{"cited_title":"Lemasson and M","cited_arxiv_id":null,"evidence_quote":"Describes the ionization-chamber energy analysis and the conventional trajectory-reconstruction method that the neural approach replaces."},{"cited_title":"Pullanhiotan, M","cited_arxiv_id":null,"evidence_quote":"Establishes the VAMOS performance baseline and the identification requirements that the improved analysis must meet."},{"cited_title":"Rejmund, B","cited_arxiv_id":null,"evidence_quote":"Documents the VAMOS++ detector and acceptance, the system whose charge-state and Z analysis is under study."},{"cited_title":"Vandebrouck, A","cited_arxiv_id":null,"evidence_quote":"Supplies the dual position-sensitive MWPC telescope whose entrance and focal-plane positions are input variables."}],"review_version":2}