{"id":"60b3f4b4-db74-46a4-b72d-c611ce89a463","arxiv_id":"2502.07359","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Linear and nonlinear frequency-domain parametric models trained on charge-normalized CLYC pulses match or beat existing neutron/gamma discriminators, especially at low sampling rates and under noise, when judged on adversarial hard examples.","lead":"This paper applies simple neural networks to the frequency spectrum of scintillator pulses to separate neutron and gamma signals, and proposes testing classifiers on deliberately hard examples that remove the energy difference between the two particle types.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Charge normalization plus adversarial selection may let the model exploit per-class integral-dependent artifacts rather than energy-independent pulse shape; the fixed FCI cutoff tuning and missing per-fold variance leave the 10 MS/s comparison underdetermined.","rationale":"The reader's weakest assumption points to charge normalization as the key premise. I agree that normalization is the most load-bearing preprocessing step, but I see the deeper problem as a compound of three insecurities: (1) the paper does not demonstrate that charge normalization preserves shape information equally for neutrons and gammas, and the 'Energy' and 'Linear w/o norm.' baselines showing identical constant error counts across all conditions suggest the adversarial example set is static and may be dominated by amplitude ordering, not shape; (2) the FCI baseline is hand-tuned per condition, which is not a fair comparison with a fixed trained FDPM; and (3) no variance or code is provided, so the headline 0.999 vs 0.751 at 10 MS/s could be an artifact of a single seed or the tuning protocol. The reader identified the normalization issue, but I would reframe the central concern as an evaluation-protocol issue as much as a preprocessing issue. The paper is honest about the organic dataset where charge comparison is competitive, which tempers the headline but does not undermine the 10 MS/s inorganic claim. Since the concern is specific and testable, the conditional verdict is appropriate. I do not see a reason to move to reject: the experiments are plausible, the data is open, and the proposed test would settle the matter. The 'UNCHANGED' verdict reflects that my concern does not change the reader's CONDITIONAL recommendation, though it sharpens the conditions.","tokens_in":16499,"tokens_out":1322,"duration_ms":13589,"concrete_test":"Re-run the 10 MS/s experiment with: (1) the same adversarial examples selected once, then re-run FCI with a single cutoff frequency fixed before seeing test data (e.g., chosen on the training set) and report the AUC; (2) train the nonlinear FDPM on charge-normalized pulses, then evaluate on raw pulses (no normalization) to see if the learned shape features transfer; if the AUC collapses, the classifier is not truly energy-insensitive but tuned to the normalization artifact; (3) run 5 training seeds and report mean and standard deviation of adversarial AUC for FCI, Linear, and Nonlinear at 10 MS/s. If the FDPM advantage persists under these conditions, the concern is resolved; if the gap shrinks to within the seed variance or the no-normalization test fails, the central claim needs qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that frequency-domain parametric models (FDPM) are more accurate and robust than other discriminators, especially at 10 MS/s where nonlinear FDPM AUC is 0.999 vs FCI 0.751. The load-bearing premise is that charge normalization (Section 4.1 step 5) removes energy while preserving shape information. The paper shows the linear model without normalization nearly matches the Energy baseline (AUC 0.675 vs 0.685), which supports the premise, but it does not show that normalization is shape-preserving for both classes. If normalization distorts neutron and gamma pulses differently (e.g., through baseline estimation, the random starting-point resampling, or the 10-bit digitizer's discrete amplitudes), the learned frequency weights in Fig. 10 could be encoding these class-dependent artifacts rather than intrinsic pulse shape. The constant error counts for Energy and Linear w/o norm. (53 and ~56-59 errors across all sampling rates and noise levels) are a red flag: those numbers are exactly what would result if the Energy baseline is computed on a fixed set of amplitude-rank-selected adversarial examples; the same adversarial examples are reused for all conditions. The FCI baseline is hand-tuned per condition (cut-off frequencies adjusted for best performance per sampling rate), which biases the comparison, while FDPM is trained once and evaluated across conditions. Without code, seeds, or per-fold variance, the 0.999 vs 0.751 gap at 10 MS/s could partly reflect the FCI tuning disadvantage, not a genuine robustness advantage. The paper's own Section 4.6 shows charge comparison (plus min-max normalization) beats FDPM on the organic dataset (AUC 0.916 vs 0.874), so the headline is not universal. The central unresolved point is whether the adversarial-sampling superiority is a property of the learned model or of the preprocessing plus a favorable evaluation protocol.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes linear and nonlinear frequency-domain parametric models (FDPM) for neutron/gamma discrimination. The models take FFT-derived features from charge-normalized pulses and learn a small set of weights via cross-entropy training. The authors argue that figure-of-merit (FoM) is inadequate for high-precision discriminators and introduce ROC curves on adversarially sampled events as a more stringent evaluation. Using an open-source CLYC dataset, they report that the nonlinear FDPM achieves near-perfect AUC on adversarial examples at sampling rates from 100 MS/s down to 10 MS/s (AUC 0.999–1.000), outperforming the frequency classification index (FCI), which degrades at low rates, as well as conventional methods. Additional experiments address noise robustness, model interpretability, and an organic scintillator dataset.","tokens_in":16803,"tokens_out":7876,"duration_ms":72603,"significance":"If the results are reproducible, the paper offers a simple, interpretable, and hardware-friendly classifier that improves neutron/gamma discrimination on difficult events, and it proposes an evaluation criterion that is more informative than FoM for high-precision applications. Strengths include the use of an open dataset, explicit ablations (w/o FFT, w/o normalization), and a clear statement that charge normalization is essential for energy insensitivity. However, several load-bearing aspects of the evaluation protocol need clarification or strengthening before the central claims can be fully accepted.","major_comments":[{"comment":"The FCI baseline has its cut-off frequencies 'adjusted for best performance according to the sampling rate,' but the tuning procedure (search range, optimization criterion, use of validation or test set) is not described. Because FDPM parameters are learned on a fixed training split, the comparison is not on equal footing unless FCI is tuned on the same training/validation split only. Please specify the tuning protocol and report the chosen cut-off values and their sensitivity.","section":"Sec. 4.2, FCI description"},{"comment":"The Energy baseline reports exactly 53 errors (both on all data and on adversarial examples) and AUC 0.685 for every sampling rate and every noise condition. This is suspicious: if the Energy discriminator is applied to subsampled or noisy waveforms, some variation in error counts would be expected. Please clarify whether the Energy baseline is computed on the original full-rate, noiseless waveforms for all conditions. Also state the number of adversarial examples (the error counts imply 200) and whether the same adversarial set is reused across all conditions; report variability by repeating the adversarial selection and by training multiple models with different seeds.","section":"Table 1 and Sec. 4.2/4.3"},{"comment":"It is unclear whether a separate model is trained for each sampling rate and each noise condition, or whether a single trained model is evaluated under all conditions. The input dimension N changes with subsampling, so separate models appear necessary, but the text does not state this. Please specify the training protocol for every row of Table 1, including whether noise is added before the train/test split or only to the test set. This is central to the claim of 'adaptability to variations of data integrity.'","section":"Sec. 4.1, preprocessing steps 3–4, and Sec. 4.2/4.3"},{"comment":"Conventional methods (charge comparison, zero crossing, PCNN, ladder gradient, FGA, FEPS) are evaluated with min-max normalization, while FDPM and CNN/LSTM are evaluated with charge normalization. This difference in preprocessing could disadvantage the conventional methods. Please justify the choice of min-max normalization for each conventional method, or evaluate them under the same charge normalization when their definitions permit, and report the sensitivity of the results to the normalization choice.","section":"Sec. 4.5, Table 3"},{"comment":"The adversarial sampling algorithm is underspecified. The text says examples are 'shared by several discrimination methods' and that for CLYC, neutron events with smallest amplitudes and gamma events with largest amplitudes are selected, but the exact number of events, the threshold criteria, and which methods contribute to the 'shared' condition are not stated. It is also not explicit that adversarial examples are selected only from the test set. Please provide the full algorithm and confirm that no information from the test set leaks into training or validation.","section":"Sec. 2.2 and Sec. 4.1"}],"minor_comments":[{"comment":"The term 'city-block approach' for |Re(f^c)| + |Im(f^c)| is nonstandard; please define it clearly and justify why this combination is used instead of the complex magnitude.","section":"Eq. (5)"},{"comment":"Please specify the target integral value used for charge normalization and how zero-integral or near-zero-integral signals are handled.","section":"Sec. 4.1, step 5"},{"comment":"At 50 MS/s the nonlinear FDPM has AUC 0.965 versus FCI 0.996, so the statement that FDPM models 'significantly improve' performance is too strong; consider wording such as 'improve robustness across sampling rates.'","section":"Sec. 4.2, discussion of Fig. 4"},{"comment":"The legend of Fig. 5 is crowded with many curves; consider splitting into separate panels or presenting the AUC values in a table for readability.","section":"Sec. 4.3, Fig. 5"},{"comment":"Please define how 'noise level' is computed in Fig. 6a and 6b (e.g., standard deviation divided by maximum amplitude) and explicitly relate the dB SNR values to the plotted noise levels.","section":"Sec. 4.4, Fig. 6"},{"comment":"The paper notes that LSTM converged in only 2 of 5 folds on the organic dataset; please state this limitation in the table caption and consider excluding the non-converged folds from the reported metrics.","section":"Sec. 4.6, LSTM results"},{"comment":"Typo: 'base on FDPM' should be 'based on FDPM.' Also, 'the larger value of the two targets will be the predicted class' requires a comma or semicolon for readability.","section":"Sec. 1, contributions list"},{"comment":"For the dataset reference [23], please include the version or access date, as Zenodo datasets may be updated.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of Nuclear Engineering and Technology and addresses a topic of practical interest. The main concern is the rigor of the evaluation protocol: the FCI tuning procedure, the constant Energy baseline, the unclear train/test protocol across conditions, and the lack of repeated-run statistics all make the quantitative claims fragile. If the authors can clarify these points and provide variance estimates, the paper could be suitable for publication. I would also encourage the authors to release code and exact hyperparameters for reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Straight to it: the paper is worth a serious look, and the main result is probably real. A linear/nonlinear model on FFT magnitudes, trained with charge normalization, holds near-perfect ROC/AUC on adversarially selected CLYC events down to 10 MS/s where the frequency classification index collapses to 0.751 AUC. That is a practically useful result for portable detectors, and the adversarial-sampling ROC recipe is a sensible way to stress-test discriminators that would otherwise look perfect.\n\nWhat is genuinely new: applying a minimal parametric model to the full discrete spectrum instead of fixed FCI bands, plus the evaluation protocol itself. The paper is also honest in places—Section 4.6 admits charge comparison beats FDPM on the organic dataset, and the ablation showing the linear model without normalization mimics the energy baseline (AUC 0.675 vs 0.685) is the right kind of check. The interpretability plot (Fig. 10) is a nice touch.\n\nThe soft spots are the usual ones, and they are real but not disqualifying. No error bars or repeated runs; all ROC curves are single draws. The FCI baseline has its cutoffs hand-tuned per sampling rate, which puts the comparison at a disadvantage; a stated validation procedure would fix that. The constant error counts for Energy and Linear w/o norm. across all conditions (53 and ~56–59) are suspicious at first glance, but the likely explanation is that the same fixed adversarial subset is reused everywhere, so those numbers are not evidence of cheating—just something the paper should state explicitly. The bigger conceptual issue is that energy-insensitivity is largely enforced by charge normalization: if every pulse has the same integral, the classifier cannot use energy. Calling the model energy-insensitive is then true by construction, not a discovery. The open question—whether the learned weights encode intrinsic shape differences or normalization artifacts—is worth a careful look. The paper shows the linear w/o norm model matches the energy baseline, which supports the premise, but it does not test whether normalization is shape-preserving for both classes.\n\nWho is this for: anyone working on neutron/gamma PSD, especially with limited sampling rates or embedded constraints. It deserves a proper referee. The requests should be code, seeds, per-fold error bars, and an explicit description of how the adversarial subset is fixed across conditions.","headline":"A simple FFT-based learned classifier with charge normalization that plausibly beats FCI at low sampling rates, wrapped in a useful adversarial-ROC evaluation—not a result that changes physics, but one worth referee time.","tokens_in":17427,"tokens_out":2005,"would_cite":true,"duration_ms":18971,"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 minimal feedforward network on the FFT spectrum, trained on charge-normalized pulses, outperforms fixed discriminators for neutron/gamma separation, especially at low sampling rates and under added noise.","keywords":["neutron/gamma discrimination","pulse shape discrimination","frequency-domain parametric models","charge normalization","adversarial sampling","CLYC scintillator","ROC curve","figure-of-merit"],"falsifier":"Train the nonlinear FDPM on charge-normalized events split into narrow energy bins, then measure AUC within each bin; if AUC varies strongly across energy or drops well below 0.999 in some bins, the energy-insensitivity claim fails. Alternatively, generate synthetic pulses with identical shapes but different integrals, normalize them, and check whether the network still separates them—if it does, charge normalization leaks energy information.","tokens_in":16255,"feed_emoji":"⚛️","tokens_out":6621,"duration_ms":56416,"temperature":0.7,"pith_summary":"The paper tries to establish that simple frequency-domain parametric models (FDPMs)—basic feedforward networks fed with the FFT spectrum of a pulse—are better neutron/gamma discriminators than fixed indexes like the frequency classification index (FCI), once the data are charge-normalized. It argues that figure-of-merit (FoM) is too coarse for high-precision discrimination and proposes evaluating classifiers with ROC curves on adversarially selected events that sit near the classification boundary. On an open-source CLYC dataset, the nonlinear FDPM reaches near-perfect adversarial AUC at every tested sampling rate, including 10 MS/s where FCI drops to 0.751, and it remains the best method under added Gaussian noise. If this holds, cheap, low-rate digitizers with simple neural-network inference could replace high-rate fixed-index discrimination.","feed_headline":"FFT-based neural classifier keeps neutron/gamma accuracy at low rates","feed_subtitle":"Charge-normalized FFT spectra push adversarial ROC AUC to 0.999 at 10 MS/s, where the fixed index falls to 0.751.","key_machinery":"The machinery is the FDPM: a full N-point FFT of each baseline-subtracted pulse, transformed to real features by $f^r = |\\mathrm{Re}(f^c)| + |\\mathrm{Im}(f^c)|$, then either a linear map $p_l = W^l f^r + b^l$ or a one-hidden-layer nonlinear map $p_n = W^n_2 \\sigma(W^n_1 f^r + b^n_1) + b^n_2$, trained with softmax cross-entropy. Charge normalization (scaling every pulse to the same integral) is the step that forces the model to learn energy-independent features. The evaluation machinery is adversarial sampling: selecting neutron events with the smallest amplitudes and gamma events with the largest amplitudes so that the two classes overlap in energy, making ROC curves informative where FoM cannot distinguish methods.","core_discovery":"The central discovery is that a learned combination of the full discrete FFT spectrum separates neutrons from gammas using energy-independent pulse-shape information, not just the energy gap. A linear model (one matrix) already beats FCI at 20 and 10 MS/s, and a nonlinear model with one hidden layer reaches adversarial AUC 1.000 at 100 and 20 MS/s and 0.999 at 10 MS/s, while FCI falls to 0.751. Charge normalization is essential: without it, the linear model's ROC collapses to the energy-only baseline (AUC 0.675–0.684), showing that the network would otherwise exploit the neutron/gamma energy gap. The same approach transfers to a TOF-labelled organic scintillator dataset, where FDPM and CNN give similar ROC curves and the best conventional method is slightly better on the AUC metric.","pith_inferences":["A natural next test is to apply FDPM to pile-up events or lower-bit ADC data; the frequency-domain features may degrade differently than time-domain charge comparison, and the current paper does not test these regimes.","The charge-normalization trick could be replaced by other energy-invariant transformations, such as normalizing by peak amplitude or by a low-frequency bin, and comparing those would isolate whether the energy insensitivity comes from normalization or from the FFT features themselves.","If the method generalizes, learned frequency weights could serve as a physically interpretable replacement for manually tuned discriminator indexes, and the same adversarial-sampling ROC protocol could be adopted as a standard benchmark for high-precision neutron/gamma classifiers."],"forward_implications":["At 10 MS/s, the nonlinear FDPM keeps adversarial AUC 0.999 and 98.5% accuracy, while FCI falls to 0.751 and 61.5%, so low-rate digitizers should no longer be ruled out for high-precision neutron/gamma discrimination.","Under 30 dB Gaussian noise, the nonlinear FDPM still leads across sampling rates, and it degrades much less with correlated noise and dark pulses than with white noise.","The linear FDPM's learned frequency weights are interpretable and concentrate in the low-frequency region, matching the physical intuition behind fixed frequency indexes while being optimized for the data.","Because the linear FDPM at 10 MS/s needs only 256-point FFTs and matrix operations, the method is suitable for FPGA or microcontroller implementation and online inference."],"supporting_citations":[{"why":"Supplies the frequency classification index (FCI) that FDPM extends and beats, and the frequency-domain analysis inspiration.","marker":"[22]"},{"why":"Supplies the open-source CLYC SiPM dataset used for training and testing all models.","marker":"[23]"},{"why":"Provides the CNN architecture used as a machine-learning baseline comparison.","marker":"[13]"},{"why":"Provides the LSTM baseline used for comparison at 10 MS/s.","marker":"[12]"},{"why":"Supplies the conventional neutron/gamma discriminators (charge comparison, zero crossing, FGA, and others) used as baselines.","marker":"[31]"},{"why":"Provides the TOF-labelled organic scintillator dataset used to test transfer of the method.","marker":"[32]"},{"why":"Motivates the city-block conversion of complex FFT coefficients to real features.","marker":"[25]"}],"fun_headline_variants":["FFT-based linear model beats fixed index at low sample rates","Charge-normalized FFT pushes neutron/gamma AUC to 0.999 at 10 MS/s","Learned FFT spectra yield robust neutron/gamma discrimination at low rates","Adversarial sampling reveals energy-independent FFT features for neutron/gamma","Parametric FFT models hit AUC 1.000 at high sample rates"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim that the classifier is energy-insensitive rests on charge normalization removing energy information without distorting neutron/gamma pulse-shape differences; if normalization instead encodes class-dependent artifacts, the reported near-perfect separation is a preprocessing effect rather than true pulse-shape discrimination.","fun_headline_variants_meta":{"raw":{"variants":["FFT-based linear model beats fixed index at low sample rates","Charge-normalized FFT pushes neutron/gamma AUC to 0.999 at 10 MS/s","Learned FFT spectra yield robust neutron/gamma discrimination at low rates","Adversarial sampling reveals energy-independent FFT features for neutron/gamma","Parametric FFT models hit AUC 1.000 at high sample rates"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001684,"raw_usage":{"total_tokens":6673,"prompt_tokens":940,"completion_tokens":5733,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":5632}},"tokens_in":556,"tokens_out":5733,"duration_ms":38100,"temperature":1.0,"reasoning_tokens":5632,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T13:00:58.150731+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the nonlinear FDPM on charge-normalized events split into narrow energy bins, then measure AUC within each bin; if AUC varies strongly across energy or drops well below 0.999 in some bins, the energy-insensitivity claim fails. Alternatively, generate synthetic pulses with identical shapes but different integrals, normalize them, and check whether the network still separates them—if it does, charge normalization leaks energy information.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the frequency classification index (FCI) that FDPM extends and beats, and the frequency-domain analysis inspiration."},{"cited_title":"Argueta, I","cited_arxiv_id":null,"evidence_quote":"Supplies the open-source CLYC SiPM dataset used for training and testing all models."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the CNN architecture used as a machine-learning baseline comparison."},{"cited_title":"Fabian, G","cited_arxiv_id":null,"evidence_quote":"Provides the LSTM baseline used for comparison at 10 MS/s."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the conventional neutron/gamma discriminators (charge comparison, zero crossing, FGA, and others) used as baselines."},{"cited_title":"Maedgen, B","cited_arxiv_id":null,"evidence_quote":"Provides the TOF-labelled organic scintillator dataset used to test transfer of the method."},{"cited_title":"Alharbi, Distance metrics for digital pulse-shape discrimination of scintillator detectors, Radiation Physics and Chem- istry 156 (2019) 205–209","cited_arxiv_id":null,"evidence_quote":"Motivates the city-block conversion of complex FFT coefficients to real features."}],"review_version":1}