{"id":"6f1468e9-eb1f-4e0d-a1cb-420f5641a7be","arxiv_id":"2608.01943","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Long-delayed afterpulses peaking near 260 μs (dynode) and 90/550 μs (MCP), at yields of 10^-3 per primary photoelectron, were measured in JUNO 20-inch PMTs out to 20 ms.","lead":"JUNO's 20-inch PMTs emit small afterpulses hundreds of microseconds to milliseconds after a large flash, with distinctive timing patterns for dynode and microchannel-plate tubes. The measured yields, around one afterpulse per thousand primary photoelectrons, are large enough to affect neutrino detector background modeling after cosmic muons.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Yield normalizations rest on an unquantified linearity assumption; the concern is real and should be explicitly bounded before the per-PE numbers are relied upon.","rationale":"A conditional accept is the appropriate disposition because the paper is an empirical detector-characterization study with two separable layers: (1) the timing morphology (260 us dynode peak, 90/550 us MCP peaks, small millisecond-scale tail, consistency between the 1.8 ms and sliding-window methods) and (2) the absolute per-PE yields in Table 2 and Sec. 4.3. Layer (1) is well supported: the sliding-window method reproduces the several-hundred-microsecond features observed in the direct window, the LED-on/LED-off comparison demonstrates a real correlated excess, and the time profiles are robust to a global normalization error. Layer (2) is the weakest link. The authors state the large-signal linearity limitation in Sec. 5.1 and cite Ref. [34] showing that measurable nonlinearity exists in exactly these PMTs, yet they provide no bound on its magnitude or on the resulting yield bias. The 3 mV threshold (0.6-0.67 of mean SPE amplitude) also risks losing small afterpulse avalanches, especially under MCP gain suppression. Since the central quantitative claim is stated as a number with only statistical uncertainties, the missing systematic is load-bearing. The stress-test pass therefore agrees with the reader that the verdict should remain CONDITIONAL: publication is reasonable provided the systematics are bounded in revision or the quantitative claims are relaxed to qualitative statements. I see no reason to move to ACCEPT or REJECT: the concern is not about internal inconsistency or fraud, and the empirical content is strong enough to warrant publication after revision. Verification of the exact nonlinearity magnitude would be the single decisive check, and it is feasible through the already-cited JUNO linearity study rather than requiring new data-taking.","tokens_in":9875,"tokens_out":1817,"duration_ms":14776,"concrete_test":"Recompute the primary-PE normalization using the linearity data already cited as Ref. [34]: apply the measured high-intensity charge nonlinearity curves for the HPK R12860 and NNVT GDB-6201 at ~2-3x10^4 PE to the integrated primary-pulse charges, and propagate the resulting shift in N_PE into the Table 2 slopes and the 2.0-20.0 ms probabilities. If the corrected N_PE moves any headline yield by more than its statistical uncertainty band, the per-PE yields should be re-reported with this systematic included or downgraded to qualitative statements.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central quantitative claim—afterpulse yield per primary PE at the 10^-3 level and approximately linear growth with primary intensity—depends on converting the measured primary-pulse charge into a photoelectron count via the single-photoelectron charge (Sec. 3). This conversion treats the dynode and MCP PMTs, plus the CAEN DT5751 readout, as linear up to ~2.8x10^4 PE (dynode) and ~1.8x10^4 PE (MCP), precisely the regime where the authors themselves cite dedicated JUNO measurements [34] showing measurable charge nonlinearity in high-intensity pulse mode (Sec. 5.1). A nonlinear compression of the primary-pulse charge would bias N_PE low and hence bias all quoted yields high; for the MCP PMT, ALD-related gain suppression after large signals could instead bias yields low by pushing afterpulses below the fixed 3 mV threshold. Both effects are pointed out in Sec. 5.1, but neither is quantified, so the Table 2 slopes and the 2.0-20.0 ms integrated probabilities in Sec. 4.3 are not yet anchored to a stated systematic scale. The timing features (peak positions, distinct shapes, extension to 20 ms) are not affected by this concern; the paper itself largely separates qualitative timing conclusions from quantitative yields. The gap is that the quantitative values are presented as the paper's headline and are the input for JUNO background modeling, so an unquantified common-mode normalization error matters directly.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports a dedicated measurement of long-delayed afterpulses in two types of JUNO 20-inch PMTs: an HPK R12860 dynode PMT and an NNVT GDB-6201 MCP-PMT. Using high-intensity LED illumination (about 10^4 PE) and a 1.8 ms digitizer window, the authors reconstruct afterpulse time profiles after template subtraction and dark-noise subtraction. They observe a broad component peaking near 260 us in the dynode PMT and components near 90 us and 550 us plus a millisecond-scale tail in the MCP-PMT. A sliding-window readout extends the effective profile to 20 ms. Afterpulse yields in selected windows from 10 us to 1.8 ms are fitted to linear functions of the primary PE number, giving slopes at the 10^-3 level per primary PE, and integrated probabilities in the 2.0-20.0 ms window are reported. The authors discuss possible physical origins and explicitly acknowledge limitations from large-signal nonlinearity and gain suppression.","tokens_in":10234,"tokens_out":6851,"duration_ms":60971,"significance":"If the quantitative results hold, this is a valuable contribution to JUNO background modeling and to PMT characterization more generally. The timing features are robust: they appear in both the direct 1.8 ms window and the sliding-window data, and the LED-off subtraction and template subtraction are carefully implemented. The observed difference between dynode and MCP structures is a clean, falsifiable result. However, the absolute yields are not yet anchored to a stated systematic scale, because the primary-PE normalization relies on large-signal linearity that is acknowledged but not quantified, and the fixed 3 mV threshold imposes an amplitude-dependent detection efficiency. The paper would be strengthened by bounding these systematics or by scaling back the quantitative claims accordingly. The authors are to be credited for clearly separating the robust timing conclusions from the more uncertain yield normalization.","major_comments":[{"comment":"The primary-PE normalization is obtained by dividing the integrated primary-pulse charge by the mean SPE charge, assuming linear response up to about 2.8e4 PE for the dynode PMT and 1.8e4 PE for the MCP-PMT. The authors cite dedicated JUNO linearity measurements in Ref. [34] showing measurable charge nonlinearity in this regime and state that this could bias the yield normalization, but they do not quantify the effect. Since Table 2 and the integrated probabilities in Sec. 4.3 are central quantitative results, the paper should either propagate a systematic uncertainty from the measured nonlinearity curve in Ref. [34] or perform an explicit linearity check. Without this, the per-PE yields are not anchored to a stated systematic scale.","section":"Sec. 3 and Sec. 5.1"},{"comment":"The fixed 3 mV reconstruction threshold (0.6-0.67 times the mean SPE amplitude) is a hand-chosen value that directly affects the measured afterpulse yields. For the MCP-PMT, the additional ALD-related gain suppression after large signals, acknowledged in Sec. 5.1, can push small afterpulses below threshold and bias the yield low. The paper should provide a threshold scan or an efficiency correction based on the SPE amplitude distribution, and it should quantify the gain-suppression effect using the references cited. This is load-bearing for the absolute yield values and the linear dependence shown in Fig. 8.","section":"Sec. 3 and Table 2"},{"comment":"The sliding-window phase reconstruction is not fully specified. The text states that the delay axis was reconstructed according to the trigger sequence and that time stamps were used to identify missing triggers, but it does not explain how the absolute phase offset is anchored (for example, using events where the primary pulse appears in the window) or how missing triggers are handled in the phase assignment. If a trigger is missed, the phase offset for subsequent events would shift by 10 us, and without correction this would smear the reconstructed delay profile and bias the integrated probabilities in the 2.0-20.0 ms window. A detailed description of the phase reconstruction algorithm and the treatment of missing triggers, along with an estimate of the resulting timing uncertainty, is needed to support the 20 ms results.","section":"Sec. 2.2 and Sec. 4.3"}],"minor_comments":[{"comment":"The phrase 'pre-peak region' is used to estimate the residual background level, but the exact time range of this region is not defined; please specify the delay interval used for the background subtraction.","section":"Sec. 3"},{"comment":"The lower panels show relative deviations from the linear fits, but the error bars on the deviations are not displayed; please include them or state that the deviations are shown without uncertainties.","section":"Fig. 8"},{"comment":"The caption states no systematic uncertainties are included, but it would be clearer to explicitly repeat in the caption that the listed uncertainties are statistical only, as the text already notes.","section":"Table 2"},{"comment":"The mean delay of 6.9 ms and standard deviation of 2.6 ms for the MCP-PMT millisecond component are reported without uncertainties; please indicate how these were computed and provide their statistical precision.","section":"Sec. 4.3"},{"comment":"The sentence 'residual nonlinearity in the PMT or readout chain could underestimate the primary PE' mixes two distinct effects; the CAEN DT5751 digitizer linearity should be discussed separately from the PMT charge nonlinearity, ideally with a reference to the digitizer specifications.","section":"Sec. 5.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is honest about its limitations in Sec. 5.1, but those limitations bear directly on the headline quantitative claims in the abstract and Table 2. The authors should be encouraged to either quantify the nonlinearity and threshold systematics (even approximately) or soften the quantitative statements to 'at the level of 10^-3 under the stated assumptions.' The sliding-window phase reconstruction also needs more detail for reproducibility. These are fixable within the scope of a revision, so I do not recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a competent, honestly-written measurement paper. The timing structures are new and robust; the per-photoelectron yields are provisional until the authors quantify the large-signal linearity systematics they themselves flag in Sec. 5.1.\n\nWhat's actually new: they extend afterpulse measurements for the two JUNO 20-inch PMT models out to 20 ms using a sliding-window readout, and they resolve distinct long-delayed components—a broad ~260 us component in the dynode PMT and ~90 us plus ~550 us components in the MCP-PMT. The 2.0-20.0 ms integrated probabilities, 7e-5 and 2e-4 per primary PE, are tiny but non-negligible after high-intensity muon events. The several-hundred-microsecond structures reproduce across the 1.8 ms direct window and the sliding-window method, so the timing features are solid.\n\nThe paper does several things well. The sliding-window strategy is clever. They do proper SPE calibration, dark-noise subtraction, and template subtraction. They also resist speculating about mechanisms and explicitly list the normalization caveats.\n\nThe soft spot is the absolute yield normalization. The primary light intensity is obtained by dividing the integrated primary-pulse charge by the SPE charge, which assumes PMT and readout linearity up to ~2x10^4 PE. Ref. [34] shows measurable charge nonlinearity in exactly these PMTs in high-intensity pulse mode. If the primary charge is compressed, N_PE is underestimated and every yield is overestimated by a common factor. For the MCP-PMT, ALD-related gain suppression could push afterpulses below the fixed 3 mV threshold, biasing yields in the opposite direction. Both effects are acknowledged in Sec. 5.1 but not bounded. The fixed threshold (0.6-0.67 SPE) and the chosen integration windows are additional hand choices, and Table 2's slopes summarize data rather than predict anything. So the timing conclusions are safe, but the absolute per-PE numbers should not be treated as final until the systematics are quantified or the claims are relaxed to qualitative.\n\nThis is a measurement paper with no circular reasoning, and the citation pattern is appropriate. The intended readers are JUNO collaborators working on muon vetoes and delayed-energy reconstruction, plus anyone modeling PMT afterpulse backgrounds. It deserves a serious referee. I'd send it to peer review with the request that the authors either provide a linearity correction or a quantitative bound on its size, or revise the headlines to present yields as estimates subject to an unquantified common-mode systematic. With that, it's publishable as a detector characterization note.","headline":"A solid, honest measurement of long-delayed afterpulses: timing features are robust, but per-PE yields need a quantified linearity systematic before they are used in JUNO background modeling.","tokens_in":10754,"tokens_out":2474,"would_cite":true,"duration_ms":21477,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["85.60.Ha"],"model":"deepseek-v4-flash","headline":"JUNO's 20-inch photomultiplier tubes emit long-delayed afterpulses out to 20 ms, and the amount of delayed charge scales linearly with the size of the primary flash.","keywords":["photomultiplier tubes","afterpulse","long-delayed afterpulses","JUNO","MCP-PMT","dynode PMT","detector characterization","sliding-window readout"],"falsifier":"Record the same LED-triggered PMT signal with a deep-memory oscilloscope capturing one contiguous 20 ms window and compare the reconstructed afterpulse profile and the 2–20 ms integrated probability with the sliding-window result; if the ~260 μs dynode peak, the ~90 μs MCP peak, and the ~$10^{-4}$ late probabilities do not reproduce, the extended profiles are a readout artefact.","tokens_in":1948,"feed_emoji":"⚡","tokens_out":8121,"duration_ms":154381,"temperature":0.7,"pith_summary":"Large photomultiplier tubes (PMTs), the vacuum light sensors of the JUNO liquid-scintillator detector, continue to emit small secondary pulses, or afterpulses, for tens of milliseconds after a bright flash, not just in the few-microsecond range usually studied. This paper measures those long-delayed afterpulses in the two 20-inch PMT types used by JUNO—a dynode tube and a microchannel-plate (MCP) tube—and shows that their delay structure is different for the two designs: a broad component peaking near 260 $\\mu$s for the dynode tube, and components near 90 $\\mu$s, 550 $\\mu$s, and a broad millisecond-scale tail for the MCP tube. The per-photoelectron yields in the microsecond-to-millisecond windows are at the $10^{-3}$ level and rise approximately linearly with primary light intensity, so after a high-energy event the accumulated delayed charge is sizeable. Understanding this instrumental response matters because delayed physics signals, such as neutron-capture gamma rays at about $220\\,\\mu$s after a cosmic muon, sit in exactly the time range where these afterpulses appear.","feed_headline":"Afterpulses linger up to 20 ms in JUNO's 20-inch PMTs","feed_subtitle":"Small per flash but linear in light intensity, they can contaminate post-muon physics windows in JUNO.","key_machinery":"The key machinery is the long-window waveform readout combined with a per-photoelectron normalization. A direct 1.8 ms digitized window captures the first part of the delay spectrum, and a sliding-window strategy—driving the LED at a 20 ms period and the digitizer trigger at 20.01 ms, so the relative phase drifts—lets short 10 $\\mu$s snapshots tile the full 20 ms delay axis while keeping 1 GS/s sampling. Afterpulse candidates are found by subtracting an averaged primary-pulse template (which removes the large-signal baseline undershoot and recovery), applying a fixed 3 mV threshold, and subtracting a uniform pre-peak dark-noise level. The primary photoelectron number is obtained from the integrated primary-pulse charge divided by the single-photoelectron charge from SPE calibration. This combination is what turns raw waveforms into time profiles and per-primary-PE yields.","core_discovery":"The paper's central claim is that long-delayed afterpulse components, previously seen only as short few-microsecond features in these tubes, extend to tens of milliseconds and are quantitatively non-negligible. In the 1.8 ms direct window, the authors identify a broad dynode afterpulse component peaking at about 260 $\\mu$s, a pronounced MCP component near 90 $\\mu$s, a weaker MCP structure near 550 $\\mu$s, and, from the sliding-window extension to 20 ms, a small broadly distributed millisecond-scale component—centered near 6.9 ms for the MCP tube—whose integrated probability in the 2–20 ms window is $(7.01\\pm0.36)\\times10^{-5}$ per primary photoelectron for the dynode tube and $(2.23\\pm0.15)\\times10^{-4}$ for the MCP tube. The normalized yields in the selected 10 $\\mu$s–1.8 ms windows are at the $10^{-3}$ level per primary photoelectron and increase approximately linearly with primary light intensity, demonstrating that these delayed pulses are correlated with the primary signal and not with dark noise.","pith_inferences":["If the hundreds-of-microsecond peaks are ion time-of-flight features, their centroids should move with the applied high voltage; a voltage scan that leaves the 90 μs and 260 μs peaks fixed would instead point to delayed electron emission or optical feedback.","The approximately linear intensity dependence suggests the ion-production probability per avalanche is roughly constant over the probed range; extending the measurement to lower intensities would check whether a threshold or saturation effect appears near the single-photon limit.","For future large detectors, these results imply that afterpulse contamination can be reduced at the source by choosing PMTs whose long-delayed components are smallest in the physics-critical windows—an optimization that the difference between the dynode and MCP patterns makes possible.","If the same sliding-window analysis were applied to LED-off or low-rate data from the running JUNO detector itself, it could turn the detector's own muon events into a large-statistics check of these laboratory yields."],"forward_implications":["JUNO's reconstruction of neutron-capture signals after cosmic muons must include a correlated charge component from these afterpulses; the dynode peak near 260 μs falls close to the about 220 μs hydrogen-capture time.","Muon veto windows must be long enough or supplemented with afterpulse models: the MCP tube still shows afterpulses with a mean delay of 6.9 ms, so a veto of a few milliseconds will not clean the detector.","Because yields grow linearly with primary light intensity, the delayed background scales with event size; the largest signals produce proportionally more long-delayed activity.","PMT response simulations for JUNO should assign different long-delay spectra to dynode and MCP PMTs rather than one generic afterpulse shape.","The sliding-window readout offers a way to characterize long-delayed PMT afterpulses in other detectors whose digitizers cannot store millisecond-long contiguous waveforms."],"supporting_citations":[{"why":"Previous measurement of afterpulse components in JUNO 20-inch PMTs; the short-delay components reported here match it and the long-delay study extends past its range.","marker":"[28]"},{"why":"Earlier reports of hundred-microsecond to millisecond afterpulse components in other PMTs and detectors, motivating and framing the long-delay measurement.","marker":"[23–25]"},{"why":"Earlier afterpulse study of a 20-inch HQE-MCP-PMT for JUNO, including flight-time interpretation of short-delay groups that this work extends to longer delays.","marker":"[29]"},{"why":"Proposes delayed electron emission as a source of long-delayed PMT pulses, one candidate mechanism for the observed components.","marker":"[18]"},{"why":"Establishes the single-photoelectron counting calibration used to convert primary-pulse charge into photoelectron number.","marker":"[33]"},{"why":"Shows measurable charge nonlinearity of 20-inch dynode and MCP PMTs at high intensity, the main systematic limitation on per-PE yield normalization.","marker":"[34]"},{"why":"Documents temporary gain suppression and recovery in ALD-coated MCP-PMTs after large signals, a bias affecting reconstructed afterpulse yield for the MCP tube.","marker":"[35–37]"}],"fun_headline_variants":["Afterpulses in JUNO PMTs extend to 20 ms, worst for MCP","JUNO PMTs show long afterpulses up to 20 ms after big hits","New long-delayed afterpulse components in JUNO phototubes","Afterpulse yields rise linearly with light: JUNO PMT study","Dynode vs MCP: distinct afterpulse timing in JUNO tubes"],"cache_read_input_tokens":12800,"weakest_assumption_plain":"The primary pulse's photoelectron count is obtained by dividing its integrated charge by the single-photoelectron charge, which assumes the PMT and readout chain stay linear up to roughly $2\\times10^4$ photoelectrons; if large-signal nonlinearity is significant there, every per-photoelectron yield is scaled wrong.","fun_headline_variants_meta":{"raw":{"variants":["Afterpulses in JUNO PMTs extend to 20 ms, worst for MCP","JUNO PMTs show long afterpulses up to 20 ms after big hits","New long-delayed afterpulse components in JUNO phototubes","Afterpulse yields rise linearly with light: JUNO PMT study","Dynode vs MCP: distinct afterpulse timing in JUNO tubes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000513,"raw_usage":{"total_tokens":2567,"prompt_tokens":1092,"completion_tokens":1475,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":708,"completion_tokens_details":{"reasoning_tokens":1363}},"tokens_in":708,"tokens_out":1475,"duration_ms":12029,"temperature":1.0,"reasoning_tokens":1363,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:02:52.310352+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record the same LED-triggered PMT signal with a deep-memory oscilloscope capturing one contiguous 20 ms window and compare the reconstructed afterpulse profile and the 2–20 ms integrated probability with the sliding-window result; if the ~260 μs dynode peak, the ~90 μs MCP peak, and the ~$10^{-4}$ late probabilities do not reproduce, the extended profiles are a readout artefact.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Earlier afterpulse study of a 20-inch HQE-MCP-PMT for JUNO, including flight-time interpretation of short-delay groups that this work extends to longer delays."},{"cited_title":"Dossi, A","cited_arxiv_id":null,"evidence_quote":"Establishes the single-photoelectron counting calibration used to convert primary-pulse charge into photoelectron number."},{"cited_title":"Study on the linearity of 20\" dynode and MCP PMTs","cited_arxiv_id":"2212.11514","evidence_quote":"Shows measurable charge nonlinearity of 20-inch dynode and MCP PMTs at high intensity, the main systematic limitation on per-PE yield normalization."}],"review_version":1}