{"id":"5f694960-52b3-4909-aee7-b8dc05415278","arxiv_id":"2504.17272","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An upsampled convolutional autoencoder reconstructs simulated scintillator pulse times with 0.19 ns MAE and amplitudes with 23.2 mV MAE, about four to five times more accurate than the baseline model.","lead":"Neural networks can reconstruct pulse arrival time and amplitude in calorimeter waveforms. An upsampled autoencoder, informed by an explainability analysis, cuts arrival-time error by about four times on simulated data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The UMAC-versus-MAC comparison is not apples-to-apples: Eq. (4)'s weighted arrival-time correction appears to be applied only to UMAC, and Figure 8's uncorrected 0.75 ns MAC offset would shrink the claimed 4x timing gain if Eq. (4) were also applied to MAC.","rationale":"The reader's weakest assumption identifies the most load-bearing issue: the central claim rests on a comparison that may not be controlled. The paper's own Figure 8 and Table 1 both show the 0.75 ns rounding offset for MAC, while Section 4 states that UMAC results are post-processed with Eq. (4). This asymmetry directly affects the headline 4x timing improvement. The concern is concrete and testable by re-running the MAC evaluation with identical post-processing. The reader's CONDITIONAL verdict is appropriate: the issue is addressable, and the paper does have independent strengths, including the occlusion-sensitivity demonstration of sub-bin sensitivity and the internally plausible UMAC statistics. No change to the verdict is needed beyond the conditions already stated.","tokens_in":9789,"tokens_out":5633,"duration_ms":57822,"concrete_test":"Recompute Table 1 from the MAC model outputs by applying Eq. (4) to MAC exactly as stated for UMAC: for each reconstructed pulse, form t_weighted = sum(A_i t_i) / sum(A_i) over the nonzero output cluster within the merging window T, then recompute arrival-time MAE, MSE, and the Figure 8 histogram. If the MAC arrival-time MAE drops from 0.77 ns to below roughly 0.35 ns, the claimed 4x gain mostly reflects withheld post-processing rather than the UMAC architecture. Also check whether the MAC amplitude is read at the weighted or the unweighted position, and use the same definition for both models before comparing amplitude errors.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central Table 1 comparison requires identical post-processing, but the paper never states that the baseline MAC received the weighted arrival-time formula (Eq. 4). Section 4 introduces Eq. 4 as a change in post-processing prompted by the xAI study, then says the UMAC results are post-processed using the weighted arrival-time determination; it is not stated that MAC is re-processed the same way. Figure 8 is the main evidence: the MAC Gaussian fit is centered at 0.7521 ns, and the text attributes this 0.7 ns mean exactly to the difference between the true arrival time and the rounded label position—precisely the offset Eq. 4 is designed to remove. Table 1's MAC arrival-time MAE of 0.77 ns and error-distribution mean of 0.75 ns are consistent with an unweighted argmax or integer-bin readout, not with Eq. 4. If Eq. 4 were applied to the MAC output, the dominant 0.75 ns bias would be removed (or at least substantially reduced), so the MAC MAE could drop well below 0.77 ns. Even a partial correction would shrink the reported 4x UMAC timing advantage. The amplitude rows may be affected as well if the amplitude is read at a position determined by different post-processing for each model. Until the same post-processing is demonstrably applied to both models, the headline MAC-versus-UMAC numbers do not establish a controlled comparison.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents two convolutional autoencoder models, MAC and UMAC, for reconstructing pulse arrival time and amplitude from simulated 1024-sample scintillator waveforms. The MAC is first analyzed with Occlusion Sensitivity; the xAI results are used to motivate a weighted arrival-time post-processing formula (Eq. 4). The UMAC modifies the decoder with two upsampling layers and uses 4096-element labels with 0.25 ns precision. On a common held-out simulated test set, the paper reports arrival-time MAE/MSE of 0.19 ns/0.16 ns^2 for UMAC versus 0.77 ns/0.86 ns^2 for MAC, and amplitude MAE/MSE of 23.2 mV/988.6 mV^2 versus 111.6 mV/17,045.8 mV^2, concluding that the upsampling model reconstructs both quantities more accurately.","tokens_in":9981,"tokens_out":8337,"duration_ms":77045,"significance":"If the reported improvement is real, the UMAC is a practically valuable step for calorimeter pulse reconstruction, and the use of xAI to motivate a concrete post-processing change is a nice illustration of explainability-driven development. The paper has real strengths: the comparison uses a large common held-out simulated dataset; the metrics and error distributions are presented transparently; and the central numbers are internally consistent on that dataset. Its main weakness is not the statistical basis but the lack of a controlled comparison: the UMAC and MAC differ in post-processing, label precision, and architecture, so the headline gain cannot yet be attributed to the upsampling design as claimed.","major_comments":[{"comment":"The paper never states that the MAC baseline was re-processed with the weighted arrival-time formula, and the evidence suggests it was not. After introducing Eq. (4), the text says only that the UMAC results are 'post-processed using the weighted arrival time determination'; no analogous statement is made for MAC. Figure 8 shows the MAC arrival-time error distribution centered at 0.7521 ns, which the text attributes to the difference between the true arrival time and the rounded 1 ns label position—exactly the bias Eq. (4) is designed to remove—and Table 1 gives a MAC error-distribution mean of 0.75 ns. If MAC predictions were read out by unweighted argmax while UMAC used Eq. (4), the reported 4x timing improvement conflates the post-processing change with the model change. The authors should apply Eq. (4) to MAC outputs and report the resulting Table 1 and Figure 8, or explicitly document and justify any difference in post-processing.","section":"Section 4, Eq. (4) and Table 1/Figure 8"},{"comment":"The UMAC design changes two factors simultaneously: the label precision is increased from 1 ns to 0.25 ns (4096 labels per event) and the decoder is augmented with two upsampling layers. The conclusion 'upsampling models do perform better for reconstructing both the arrival time and for the signal amplitude' is therefore ambiguous as a statement about architecture: the improvement could come entirely from the finer labels. A controlled comparison (for example, MAC trained on 0.25 ns labels, or UMAC with labels rounded to 1 ns) is needed to attribute the gain to the upsampling layers; otherwise the conclusions should be rephrased to refer to the combined UMAC design rather than to 'upsampling models.'","section":"Section 4, UMAC description and Section 5 conclusions"}],"minor_comments":[{"comment":"The text first says the output layer has 'no activation function' and then says 'The activation of the output layer is also set to ReLU'; please reconcile these statements.","section":"Section 2"},{"comment":"Please specify the summation range in Eq. (4); as written it appears to sum over all positions, whereas the earlier merging-window description suggests only positions within the window T around the maximum should contribute.","section":"Section 4, Eq. (4)"},{"comment":"Please state the number of test events used for Table 1 and report uncertainties on the MAE/MSE values; Figures 8 and 9 contain 97,493 entries.","section":"Table 1 and Figures 8-9"},{"comment":"The UMAC histogram mean (0.02121 ns) differs from the Gaussian fit mean (0.000608 ns); please clarify which quantity corresponds to the 'error distribution mean' in Table 1.","section":"Figure 8"},{"comment":"Adding a short description of the simulation (pulse multiplicity distribution, amplitude distribution, noise level, and the generation of t0 with 10^-6 ns precision before rounding) would improve reproducibility.","section":"Section 2"},{"comment":"The statement that the model might remove the need for energy calibration is based only on the simulated test-set mean offset; please qualify it with the caveat that the real-data application described in Section 2 required additional calibration.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The post-processing issue is the key technical risk. If the authors can demonstrate that applying Eq. (4) to MAC still leaves UMAC substantially better, the paper would be publishable after further revision. The compound design change (label precision plus upsampling layers) should be acknowledged as an attribution limit. The citation pattern is reasonable; the authors' earlier work is appropriately cited for the MAC baseline and the real-data application."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuine news here is that occlusion sensitivity can expose sub-bin timing information in a pulse waveform, and the authors turn that into a concrete architectural change: upsampled labels at 0.25 ns precision. That part is good, and it is the reason this paper deserves a real referee rather than a desk reject.\n\nThe xAI section is careful and the connection between saliency patterns and the label-rounding artifact is clearly drawn. The UMAC is a natural extension of the authors' own MAC, and the simulated error distributions are reported cleanly. The paper is short and readable, and the authors are honest about the resource cost of upsampling.\n\nThe main soft spot is the comparison in Table 1. Eq. (4), the weighted arrival-time formula, is introduced only in the section describing UMAC, and Figure 8 still shows MAC with a 0.75 ns mean offset—exactly the rounding offset that Eq. (4) is designed to remove. The MAC's 0.77 ns timing MAE is essentially that bias plus scatter. If the same weighted readout were applied to MAC, its timing MAE could drop to roughly 0.3–0.4 ns, and its amplitude MAE from 111.6 mV to perhaps 50–60 mV once the -111.8 mV bias is corrected. UMAC may still win, but the win is likely closer to 2x, not the claimed 4–5x.\n\nSecond, UMAC changes both architecture and label precision at once. The finer labels alone could account for a large part of the improvement, so attributing all of it to upsampling is not yet justified. A cleaner test would retrain MAC with 0.25 ns labels or evaluate a non-upsampled model with weighted post-processing. This is an addressable confound, but the paper does not address it.\n\nThird, no code or training configuration is released. The data availability statement says “on request,” which does not count as reproducible. Fourth, the suggestion that UMAC could remove the need for energy calibration on real data is speculative, because no real-data UMAC test is presented.\n\nThese issues are real but fixable. The core mechanism—xAI-guided refinement of the label grid and readout—is sound, and the simulated numbers are internally consistent. This is a paper for people working on calorimeter pulse reconstruction or ML-based timing extraction. I would send it to peer review, but with a clear request for major revision: apply identical post-processing to both models, disentangle architecture from label precision, and make the code and data available.","headline":"Occlusion sensitivity points to sub-bin timing information, but the UMAC-vs-MAC gain is overstated because the two models are compared under different post-processing.","tokens_in":10683,"tokens_out":3002,"would_cite":false,"duration_ms":28885,"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 convolutional autoencoder with upsampled output reconstructs calorimeter pulse timing to 0.19 ns and amplitude to 23 mV on simulated scintillator signals.","keywords":["machine learning","signal reconstruction","explainable AI","calorimeters","autoencoder","upsampling","Occlusion Sensitivity","pulse arrival time"],"falsifier":"Re-run the MAC model on the same independent test set after applying the weighted arrival-time post-processing (Eq. 4) that UMAC uses, then recompute the arrival-time MAE and MSE in Table 1. If the MAC values move close to UMAC's $0.19$ ns / $0.16~\\mathrm{ns}^2$, the fourfold timing improvement would be attributable to post-processing rather than to the upsampling architecture.","tokens_in":9447,"feed_emoji":"⚛️","tokens_out":8214,"duration_ms":67974,"temperature":0.7,"pith_summary":"This paper claims that a convolutional autoencoder trained on raw digitized waveforms can reconstruct the arrival time and amplitude of pulses in a scintillating-crystal calorimeter, and that an upsampled version of the model (UMAC) is substantially more accurate than the earlier modified autoencoder (MAC). On a simulated test set, UMAC lowers arrival-time mean absolute error from 0.77 ns to 0.19 ns and amplitude mean absolute error from 111.6 mV to 23.2 mV. The authors also use occlusion-sensitivity analysis to show that the signal front and maximum are the parts of the pulse the network relies on, and they use that observation to motivate finer arrival-time labels and a weighted arrival-time formula. If the comparison is correct, this is a practical route to better time and energy resolution in pile-up-prone calorimeters without changing the detector hardware.","feed_headline":"Upsampled autoencoder reaches 0.19 ns pulse timing in calorimeters","feed_subtitle":"Four-times-finer time labels and weighted post-processing also cut amplitude error about fivefold.","key_machinery":"The load-bearing mechanism is the upsampled modified autoencoder (UMAC): a convolutional autoencoder whose decoder inserts two factor-2 upsampling layers after the first two transposed convolutions, expanding the 1024-sample input to a 4096-sample output so the pulse arrival-time label can be rounded at 0.25 ns instead of 1 ns. Around it sits a weighted arrival-time post-processing formula $t_{\\mathrm{arrival}} = \\frac{\\sum_i A_i t_i}{\\sum_i A_i}$, which uses the model's amplitude values on neighboring bins to estimate a sub-bin arrival time. The same architecture also serves as the inference target of the Occlusion Sensitivity analysis, which masks stretches of the waveform and watches the prediction loss to identify the signal rise and maximum as the features the network uses.","core_discovery":"On the paper's own terms, the central discovery is that finer time labels alone, realized by adding two factor-2 upsampling layers to the decoder, turn a moderate autoencoder reconstruction into one that reads pulse timing at a fraction of the digitizer's 1 ns bin. The reported numbers for the same independent test set are arrival-time MAE/MSE of $0.19$ ns / $0.16~\\mathrm{ns}^2$ and amplitude MAE/MSE of $23.2$ mV / $988.6~\\mathrm{mV}^2$ for UMAC, against $0.77$ ns / $0.86~\\mathrm{ns}^2$ and $111.6$ mV / $17{,}045.8~\\mathrm{mV}^2$ for MAC. The arrival-time error distribution for UMAC is centered at 0.02 ns with standard deviation 0.30 ns, while MAC sits at about 0.75 ns, and the amplitude error for UMAC is centered near $-2.7$ mV instead of $-111.8$ mV. From this the authors conclude that upsampling models reconstruct both arrival time and amplitude better, and that the amplitude result is centered enough to suggest less energy calibration will be needed when the model is applied to real detector data.","pith_inferences":["A test the paper does not report: apply the weighted arrival-time formula of Eq. (4) to the MAC baseline on the same test set. If the MAC arrival-time MAE drops close to 0.19 ns, part of the claimed fourfold timing gain would be attributable to post-processing rather than to the upsampling architecture.","UMAC's added upsampling layers increase the parameter count and runtime, so a natural extension is to check whether the same sub-bin accuracy survives quantization or distillation for FPGA-based triggers.","The same upsampling-plus-weighted-arrival-time recipe could transfer to other binned waveform-extraction problems, such as drift-chamber timing or photodetector pulse-height analysis, wherever the label rounding is coarser than the signal's intrinsic arrival-time precision."],"forward_implications":["On the reported test set, UMAC improves arrival-time MAE by roughly a factor of 4 (0.77 to 0.19 ns) and amplitude MAE by roughly a factor of 5 (111.6 to 23.2 mV), so the reconstruction can be noticeably more accurate before any detector changes.","Because the arrival-time error for UMAC is centered near zero rather than offset by roughly 0.75 ns, the need for an empirical timing calibration of the model may be reduced compared with MAC.","The 0.25 ns label precision and the weighted post-processing give the model sub-bin timing information, which the authors identify as important for separating close-in-time pulses.","The xAI analysis suggests that the signal front and maximum carry most of the information used for recognition, which can guide where to spend model capacity or where to mask inputs in future architectures."],"supporting_citations":[{"why":"Supplies the modified autoencoder baseline and the simulated pulse-reconstruction setting that the paper extends.","marker":"[14]"},{"why":"Documents an earlier application of this model family to real calorimeter data, the real-data context that motivates the accuracy improvements.","marker":"[13]"},{"why":"Supports the use of convolutional autoencoders for signal reconstruction, the architecture family the paper modifies.","marker":"[15]"},{"why":"Defines the CAEN V17XX digitizer output format that fixes the 1024-sample, 1 ns binning of the generated training data.","marker":"[16]"},{"why":"Introduces the Occlusion Sensitivity method whose single-pulse and modified variants drive the paper's explainability analysis.","marker":"[19]"},{"why":"Compares the ML reconstruction against conventional peak finding on real annihilation events, supporting the claim of better time resolution at low energies.","marker":"[17]"}],"fun_headline_variants":["Upsampled autoencoder reads pulses at 0.19 ns in calorimeters","Four-times finer labels slash pulse timing error to 0.19 ns","Sub-nanosecond timing from upsampled autoencoder in calorimeters","Upsampling layers push autoencoder timing to 0.19 ns","Autoencoder with upsampling achieves 0.19 ns pulse timing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that MAC and UMAC were compared under identical post-processing; the paper never explicitly states that the weighted arrival-time formula (Eq. 4) was applied to MAC, and the 0.75 ns offset left in the MAC timing distribution suggests it may not have been.","fun_headline_variants_meta":{"raw":{"variants":["Upsampled autoencoder reads pulses at 0.19 ns in calorimeters","Four-times finer labels slash pulse timing error to 0.19 ns","Sub-nanosecond timing from upsampled autoencoder in calorimeters","Upsampling layers push autoencoder timing to 0.19 ns","Autoencoder with upsampling achieves 0.19 ns pulse timing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000766,"raw_usage":{"total_tokens":3362,"prompt_tokens":874,"completion_tokens":2488,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":490,"completion_tokens_details":{"reasoning_tokens":2389}},"tokens_in":490,"tokens_out":2488,"duration_ms":15882,"temperature":1.0,"reasoning_tokens":2389,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:44:45.478775+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the MAC model on the same independent test set after applying the weighted arrival-time post-processing (Eq. 4) that UMAC uses, then recompute the arrival-time MAE and MSE in Table 1. If the MAC values move close to UMAC's $0.19$ ns / $0.16~\\mathrm{ns}^2$, the fourfold timing improvement would be attributable to post-processing rather than to the upsampling architecture.","supporting_citations":[{"cited_title":"Pattern recognition and signal parameters extraction using machine learning methods","cited_arxiv_id":null,"evidence_quote":"Supplies the modified autoencoder baseline and the simulated pulse-reconstruction setting that the paper extends."},{"cited_title":"Using Artificial Intelligence in the Reconstruction of Signals from the PADME Electromag- netic Calorimeter","cited_arxiv_id":null,"evidence_quote":"Documents an earlier application of this model family to real calorimeter data, the real-data context that motivates the accuracy improvements."},{"cited_title":"Convolutional Autoencoders for Signal Reconstruction and their Application to Damage Signature Extraction","cited_arxiv_id":null,"evidence_quote":"Supports the use of convolutional autoencoders for signal reconstruction, the architecture family the paper modifies."},{"cited_title":"6; V1751 Technical Information Manual Rev","cited_arxiv_id":null,"evidence_quote":"Defines the CAEN V17XX digitizer output format that fixes the 1024-sample, 1 ns binning of the generated training data."},{"cited_title":"Visualizing and Understanding Convolutional Networks","cited_arxiv_id":null,"evidence_quote":"Introduces the Occlusion Sensitivity method whose single-pulse and modified variants drive the paper's explainability analysis."},{"cited_title":"Machine learning assisted reconstruction of positron-on-target annihilation events in the PADME experiment","cited_arxiv_id":null,"evidence_quote":"Compares the ML reconstruction against conventional peak finding on real annihilation events, supporting the claim of better time resolution at low energies."}],"review_version":1}