{"id":"107474bc-3083-4b82-a39b-82dc34964ca3","arxiv_id":"2411.15360","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Machine-learning classifiers let transition-edge sensors resolve photon numbers at 800 kHz, roughly four times their thermal recovery limit, with accurate assignment up to five photons.","lead":"Researchers used machine learning to classify overlapping pulses from transition-edge sensors, boosting the usable detection rate from about 200 kHz to 800 kHz while still resolving photon numbers up to five. The method, tested on coherent and squeezed light, could speed up photonic quantum experiments that require fast photon-number measurements.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 800 kHz supervised-classification claim depends on linear superposition of 100 kHz traces, but Supplement SII documents trace flattening at high rates; the margin at 800 kHz for n≤5 should be checked directly.","rationale":"The paper is a solid engineering demonstration. The POVM reconstruction at 800 kHz with KNN is a meaningful test, and the heralded-Fock results for HDBSCAN at 800 kHz provide independent evidence for the unsupervised route. The strongest claim, however, is specifically about the supervised method's ability to classify at 800 kHz up to five photons. That method's only source of high-rate training examples is the synthetic overlap of low-rate traces, so the validity of this step is the load-bearing assumption and every subsequent benchmark inherits it. The supplement itself documents that the assumption fails at 900 kHz/1 MHz, with trace flattening attributed to the low-pass filter and/or thermal saturation. The scaling of this effect with rate and photon number is not quantified, so the exact position of the 800 kHz operating point relative to the onset of nonlinearity is unknown. This does not refute the paper; it defines the check needed to turn the conditional verdict into an accept. The reader's weakest_assumption identifies the same point, and the recommended verdict remains CONDITIONAL, hence UNCHANGED.","tokens_in":14602,"tokens_out":3779,"duration_ms":37070,"concrete_test":"Re-run the supervised classifier on 800 kHz coherent data with a training set built from actual 800 kHz traces whose photon-number labels are obtained independently, for example by collecting the same pulses at 100 kHz with equal per-pulse power and applying the IP method, or by using heralded TMSV events with a known idler photon number. Compare the resulting diagonal POVM terms and TVD against the linear-overlap-trained KNN results in Fig. 5(c) and Fig. 4. Additionally, compute the mean residual between actual 800 kHz traces grouped by IP-assigned photon number and the corresponding linearly overlapped synthetic traces for n=0..5; if the normalized residual exceeds the noise floor for any n≤5, the linear-superposition assumption is violated in the claimed regime.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the supervised KNN classifier maintains accurate photon-number assignment up to at least five photons at 800 kHz rests on the training-set construction in Section 3.1: 100 kHz voltage traces are overlapped 'to emulate the detector's response to light pulses at the target repetition rate.' This is valid only if the TES response is linear and additive. Supplement SII shows that at 900 kHz and 1 MHz the voltage traces flatten, that the training step does not account for this shape change, and that accumulated heat may drive the tungsten film into the normal-conducting phase, temporarily diminishing PNR capability. If this nonlinearity already begins at 800 kHz, especially for the higher photon numbers within the n≤5 claim, then the synthetic training distribution is biased relative to the real 800 kHz signal distribution, and the reconstructed POVM diagonals could be optimistic. The POVM reconstruction and heralded-Fock benchmarks provide independent support, but they do not isolate the training-assumption mismatch: a classifier trained on actual 800 kHz calibration data could in principle perform differently. The concern is not that the method cannot work; it is that the linear-overlap training procedure may conceal a rate-dependent systematic error that the current benchmarks are not designed to expose.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript reports two machine-learning post-processing methods for photon-number assignment from transition-edge sensor (TES) voltage traces at repetition rates above the detector's thermal recovery limit. The supervised method trains a K-nearest-neighbour classifier on synthetic high-rate traces formed by overlapping 100 kHz calibration traces labelled by the inner-product method; the unsupervised method applies HDBSCAN clustering to the first two PCA factor scores. The authors benchmark against coherent states using total variation distance and POVM reconstruction, and against two-mode squeezed vacuum light using heralded Fock-state preparation, claiming a four-fold increase in operation rate to 800 kHz while preserving accurate photon-number assignment up to at least five photons.","tokens_in":14821,"tokens_out":4064,"duration_ms":41460,"significance":"If the 800 kHz claim holds, the result is practically valuable: it offers a hardware-agnostic way to extend the rate of TES-based photon-number-resolving detection without modifying the detector, which could benefit photonic quantum information experiments that require fast feed-forward. The paper has clear strengths: the code is publicly available, the POVM reconstruction includes bootstrap error bars in Fig. 5, and the squeezed-light benchmark provides an independent test on a different photon-number distribution. The methods themselves are standard ML tools, so the novelty is mainly in the application and the careful benchmarking rather than in new algorithmic ideas, but the demonstrated four-fold rate improvement is a useful experimental contribution.","major_comments":[{"comment":"The central 800 kHz claim rests on the training-set construction in Section 3.1, where 100 kHz voltage traces are overlapped to emulate the detector response at the target repetition rate. This assumes linear, additive detector response. Supplement SII explicitly states that at 900 kHz and 1 MHz the voltage traces flatten and that the training step does not account for this change, and it identifies accumulated heat as a possible cause. The authors should provide direct evidence that this effect is absent or negligible at 800 kHz for the photon-number range n<=5. A concrete test would be to compare the PCA distribution of synthetic training traces with the PCA distribution of real 800 kHz traces, or to report a quantitative mismatch metric. Without such a check, the reconstructed POVM diagonals at 800 kHz could be optimistic if nonlinear thermal saturation already begins at that rate.","section":"Section 3.1 and Supplement SII"},{"comment":"The TVD curves in Fig. 4 are the primary quantitative evidence that the supervised methods maintain accuracy up to 800 kHz, yet they are plotted without error bars, confidence intervals, or bootstrap estimates. The differences between methods at a given rate are therefore difficult to judge: a TVD value of 0.05 may be within statistical uncertainty of a value of 0.10 if the sample sizes are not reported. The authors should add error bars or state the number of independent repetitions used to generate each curve, especially for the KNN and IP curves where the improvement is claimed to be significant.","section":"Section 4.1.1, Fig. 4"},{"comment":"The fidelity metric in Eq. (3) compares the reconstructed POVM at high repetition rates against a reference POVM obtained from the same inner-product method at 100 kHz that also supplies the KNN training labels. This means the fidelity is partly a self-consistency check between the classifier and the IP labelling procedure, not an absolute calibration of photon-number assignment accuracy. The paper acknowledges this choice, but the limitation should be stated explicitly in the main text. In addition, the diagonal POVM terms in Fig. 5(a-c) show only the reconstructed conditional probabilities; reporting the associated posterior assignment probabilities or a separate absolute calibration with a calibrated power meter would strengthen the claim.","section":"Section 4.1.2, Eq. (3)"}],"minor_comments":[{"comment":"The caption contains a typo: the third panel is labelled '(a) mu=5.29' instead of '(c) mu=5.29'. Please correct this.","section":"Fig. 4 caption"},{"comment":"The caption states that KNN classifies traces up to 13 photons at 500 kHz and 12 photons at 800 kHz, but the POVM reconstruction in Section 4.1.2 claims accurate assignment only up to five photons at 800 kHz. The distinction between visible cluster separation and calibrated POVM accuracy should be clarified in the caption or the main text.","section":"Fig. 3 caption"},{"comment":"The sentence 'This is verified against the tomographic reconstruction of the TES's POVM at 100 kHz' is slightly ambiguous because the verification is against the 100 kHz POVM used as a reference, not an independent absolute POVM. Rephrasing to 'This is benchmarked against the tomographic reconstruction obtained at 100 kHz' would be clearer.","section":"Section 5, Conclusions"},{"comment":"The abstract states 'at least a four-fold improvement' without specifying the baseline. The introduction clarifies that the intrinsic thermal recovery limit is about 200 kHz, so 800 kHz is four-fold, but stating this baseline in the abstract would make the claim self-contained.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The manuscript fits the scope of a quantum-optics or applied-photonics journal. The main revision requested is a direct experimental validation of the linear-overlap training assumption at 800 kHz, which is the load-bearing step for the headline claim. The POVM reconstruction and squeezed-light benchmarks already provide useful independent support, so I do not see a reason to reject the paper, but the missing direct check at 800 kHz should be addressed before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me give you the quick version. This paper shows a practical way to speed up TES photon-number-resolving detection by a factor of four (to 800 kHz) using machine learning on overlapping pulses. The supervised KNN method, trained on overlapped low-rate traces, works for coherent states up to about five photons; the unsupervised HDBSCAN method works for thermal/squeezed states. The central claim is believable, and the benchmarking is more careful than most: POVM reconstruction with bootstrap error bars, heralded Fock state tests, and honest discussion of where each method fails.\n\nWhat's genuinely new is the specific combination: using overlapped 100 kHz traces as a synthetic training set to emulate higher rates, then applying KNN for coherent and HDBSCAN for thermal states. The paper also compares several classifiers (RF, SVM, GBDT, CNN) and reports timings. That's useful practical information for anyone building fast PNR systems.\n\nThe soft spots are real but not fatal. The training-set construction assumes linear superposition of traces. The supplement shows trace flattening at 900 kHz and 1 MHz, and attributes it to accumulated heat driving the film normal. The question is whether that effect begins to bite already at 800 kHz for the higher photon numbers in the n≤5 claim. The TVD benchmark in Fig. 4 has no error bars, so we don't know the uncertainty on the 800 kHz point. The POVM diagonals have bootstrap errors, but those errors reflect the fit, not the training-assumption mismatch. There's also a mild circularity: the KNN is trained on IP labels at 100 kHz and compared against the IP at 100 kHz as reference, so any IP bias is inherited. That's a calibration choice, not an invalid loop, but it means the reported accuracy is relative to the IP method, not absolute.\n\nThe abstract oversells the unsupervised method slightly: HDBSCAN fails for coherent light at 800 kHz, and the general statement 'extend the TES operation rate to 800 kHz, maintaining accurate assignment up to at least five photons' is really only true for the supervised method on coherent states and for the unsupervised on thermal states. The body is clear about this; the abstract is a bit too compressed.\n\nData are not public, only the code on GitHub without a commit hash. That limits reproducibility but doesn't undermine the claims.\n\nBottom line: this is a solid, useful paper for anyone working with TES detectors or photonic experiments that need fast PNR. The core result is credible, the limitations are mostly disclosed, and the presentation is honest. It deserves a proper peer review; I'd send it out with a request to add error bars to the TVD plots and to comment explicitly on the linearity assumption at 800 kHz. I'd also suggest slightly softening the abstract about the unsupervised method. But these are revisions, not rejection reasons.","headline":"A practical four-fold TES speed-up with honest benchmarking; the main caveat is the synthetic-training assumption and missing error bars on the headline TVD plot.","tokens_in":15410,"tokens_out":2156,"would_cite":true,"duration_ms":18502,"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":"Two machine-learning classifiers let transition-edge sensors resolve photon numbers at 800 kHz, four times their intrinsic thermal recovery rate, while preserving accurate photon-number assignment up to at least five photons.","keywords":["transition-edge sensor","photon-number resolution","machine learning","K-nearest neighbours","HDBSCAN","principal component analysis","coherent states","squeezed light"],"falsifier":"Measure the raw voltage traces at 800 kHz from a source whose photon numbers are independently known (e.g., a heralded single-photon source that provides a ground-truth tag per pulse), then compare the KNN labels to the heralded truth. If the KNN labels systematically overcount or degrade for exactly those traces that show the flattening seen in the supplementary material, the training emulation missed a nonlinearity that undermines the reported accuracy.","tokens_in":14357,"feed_emoji":"⚛️","tokens_out":2805,"duration_ms":26535,"temperature":0.7,"pith_summary":"This paper claims that machine-learning signal processing can overcome the main practical limitation of transition-edge sensors (TESs) for photon-number-resolving detection: their slow thermal recovery, which normally forces operation below a few hundred kilohertz. The authors propose a supervised classifier (K-nearest neighbours, KNN) trained on low-rate voltage traces that are computationally overlapped to emulate high-rate pile-up, and an unsupervised clustering method (HDBSCAN) that groups PCA-reduced traces. Benchmarking with coherent and squeezed light, they report that the supervised method maintains accurate photon-number assignment up to at least five photons at 800 kHz, a four-fold improvement, and that the unsupervised method resolves up to ten photons for thermal states at the same rate. A sympathetic reader would care because this is a hardware-agnostic route to faster PNR detection for quantum information and metrology applications.","feed_headline":"Machine learning quadruples photon-counting rate to 800 kHz","feed_subtitle":"Supervised and unsupervised classifiers keep photon-number accuracy up to five photons without waiting for thermal recovery.","key_machinery":"The load-bearing objects are (1) the voltage trace, a digitized time series of the SQUID-amplified TES response to one light pulse; (2) principal component analysis, which reduces each trace to a small set of factor scores and separates photon-number clusters in two dimensions; and (3) the two classifiers: KNN, which labels a new trace by majority vote of its five nearest labelled neighbours in Euclidean distance, and HDBSCAN, a hierarchical density-based clustering algorithm that identifies high-density regions and labels low-density points as noise. The critical training construction is the overlap of 100 kHz traces to emulate higher repetition rates, which transfers reliable low-rate labels to the high-rate regime.","core_discovery":"The central claim is that photon-number assignment for TES signals does not require waiting for full thermal recovery if the classifier accounts for the influence of preceding pulses. The supervised method builds a labelled training set by taking 100 kHz traces, whose photon numbers are reliably assigned by an inner-product method, and overlapping them to synthesize traces at the target repetition rate; a KNN classifier then labels new high-rate traces. The unsupervised method projects traces onto their first two principal components and clusters them with HDBSCAN, which works well for squeezed-light thermal distributions even at 800 kHz. Tomographic reconstruction of the detector's POVM confirms that the KNN method's diagonal terms remain accurate up to at least five photons at 800 kHz, and the authors state the fidelity of the reconstructed POVM is on par with the low-rate reference.","pith_inferences":["The training-emulation assumption—that overlapping low-rate traces reproduces high-rate responses—sets a practical ceiling: the paper itself notes that at 900 kHz and 1 MHz the traces flatten, suggesting nonlinear thermal saturation begins around 800 kHz; a similar approach might work for other slow thermal detectors (e.g., kinetic inductance detectors) if the same additivity holds.","The reported 93.3% detection efficiency is calibrated against an uncalibrated power meter, so absolute photon-number accuracy inherits that uncertainty; the POVM comparison against the 100 kHz reference sidesteps this, but absolute efficiency claims should be treated with caution.","If the method holds for other TES geometries and materials, it could increase the data-acquisition rate in photon-starved applications such as bio-imaging and astrophysics, where the detector's photon-number resolution is currently rate-limited."],"forward_implications":["TES-based experiments that currently demultiplex or discard pulses to stay within the thermal recovery window could instead run at 800 kHz without adding hardware overhead.","The methods are hardware-agnostic, so they can be combined with faster TES designs that trade photon-number resolution for speed, potentially pushing the useful rate further.","Accurate PNR at 800 kHz enables fast active feed-forward of detection outcomes, a requirement for non-Gaussian state preparation and other conditional quantum operations.","The unsupervised HDBSCAN method provides training-free classification, which is useful when calibration data at low rate is unavailable or noisy, at the cost of leaving some traces unclassified.","Reproducible code is released with the paper, so other groups can test the classifiers on their own TES data."],"supporting_citations":[{"why":"Provides the state-of-the-art TES experiment limited to 375 kHz despite a 6 MHz source, the baseline this work aims to beat.","marker":"[16]"},{"why":"Supplies the inner-product (IP) pulse-filtering method used to label low-rate 100 kHz calibration traces.","marker":"[26]"},{"why":"Gives the tomographic method for reconstructing the detector's POVM from coherent probe states, used to benchmark accuracy.","marker":"[31]"},{"why":"Describes the POVM confusion-matrix representation and its reconstruction from coherent states, the basis for the fidelity comparison.","marker":"[27]"},{"why":"Defines the K-nearest-neighbour algorithm and its Scikit-Learn implementation used as the supervised classifier.","marker":"[32,33]"},{"why":"Define the HDBSCAN hierarchical density-based clustering algorithm used for unsupervised classification.","marker":"[42,43]"},{"why":"Provides the fidelity definition used to compare reconstructed POVMs against the 100 kHz reference.","marker":"[46]"}],"fun_headline_variants":["Machine learning quadruples photon-counting speed to 800 kHz","ML boosts photon-number detection rate fourfold","800 kHz photon counting achieved with machine learning","Photon-counting rate quadrupled to 800 kHz via ML"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The supervised method's training assumes the TES response is linear and additive, so that overlapping 100 kHz voltage traces faithfully emulates the detector's response at 800 kHz; the paper's own supplementary data show that at 900 kHz and 1 MHz the traces flatten, indicating this assumption breaks down at higher rates.","fun_headline_variants_meta":{"raw":{"variants":["Machine learning quadruples photon-counting speed to 800 kHz","ML boosts photon-number detection rate fourfold","800 kHz photon counting achieved with machine learning","Photon-counting rate quadrupled to 800 kHz via ML"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000308,"raw_usage":{"total_tokens":1719,"prompt_tokens":863,"completion_tokens":856,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":479,"completion_tokens_details":{"reasoning_tokens":792}},"tokens_in":479,"tokens_out":856,"duration_ms":7331,"temperature":1.0,"reasoning_tokens":792,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:22:43.861765+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the raw voltage traces at 800 kHz from a source whose photon numbers are independently known (e.g., a heralded single-photon source that provides a ground-truth tag per pulse), then compare the KNN labels to the heralded truth. If the KNN labels systematically overcount or degrade for exactly those traces that show the flattening seen in the supplementary material, the training emulation missed a nonlinearity that undermines the reported accuracy.","supporting_citations":[{"cited_title":"Algorithm for finding clusters with a known distribution and its application to photon-number resolution using a superconducting transition-edge sensor,","cited_arxiv_id":null,"evidence_quote":"Supplies the inner-product (IP) pulse-filtering method used to label low-rate 100 kHz calibration traces."},{"cited_title":"Tomography of photon-number resolving continuous-output detectors,","cited_arxiv_id":null,"evidence_quote":"Gives the tomographic method for reconstructing the detector's POVM from coherent probe states, used to benchmark accuracy."},{"cited_title":"Precisely determining photon-number in real time,","cited_arxiv_id":null,"evidence_quote":"Describes the POVM confusion-matrix representation and its reconstruction from coherent states, the basis for the fidelity comparison."},{"cited_title":"Mapping coherence in measurement via full quantum tomography of a hybrid optical detector,","cited_arxiv_id":null,"evidence_quote":"Provides the fidelity definition used to compare reconstructed POVMs against the 100 kHz reference."}],"review_version":1}