{"id":"ca112c48-2b3e-4675-889d-0842eea3c132","arxiv_id":"2507.15032","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A Gaussian mixture model clustering of 1,769 ULX observations yields 85 new candidate pulsating ULXs with properties similar to the six known pulsating ULXs.","lead":"This paper applies an unsupervised clustering algorithm to XMM-Newton observations of ultraluminous X-ray sources (ULXs) and identifies 85 new candidate pulsating ULXs that resemble the known pulsators in flux and variability properties. The work is a practical demonstration that machine-learning triage can prioritize follow-up observations of ULXs that may harbor neutron stars.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The PULX cluster selection is driven by peak flux and variability, which are the pulsation-detectability selection function, not by spectral or physical accretor-type indicators; the 85 candidates may be bright variable ULXs rather than preferentially neutron-star accretors.","rationale":"The reader's weakest_assumption exactly identifies the load-bearing concern: the two-component GMM is assumed to separate NS-ULXs from BH-ULXs, but the separation may be driven by brightness, distance, or observation count. Internal evidence strongly supports this concern: the decision tree in Fig. 3 selects on broadband flux, Fpeak, and VAR_FLAG, while HRHard has a negligible Silhouette Index (0.039, Table 4). The paper's own caveat in Sect. 4.1 admits that all candidates have the highest Fpeak values, and Sect. 4.2 lists several sources with literature evidence for BH accretors among the candidates. Together these indicate that the cluster labeled 'PULX' corresponds to a bright, variable subset of ULXs rather than to a physically distinct NS population. If a computational ablation test removing Fpeak and VAR_FLAG collapses the candidate list or makes the known PULXs inseparable, the central claim would be unsupported. Because the paper's candidate list may still be useful as a target list for future high-count pulsation searches, the verdict remains CONDITIONAL, with the explicit condition that the authors demonstrate (or clearly caveat) that the clustering is not dominated by the pulsation-detectability selection function. If the ablation test fails, the paper should be revised to present the 85 sources as observationally favorable ULX targets rather than as candidate pulsating ULXs.","tokens_in":27539,"tokens_out":6658,"duration_ms":75787,"concrete_test":"Re-run the complete pipeline (Sect. 2.3) after removing Fpeak and VAR_FLAG from the input feature set, retaining the same hyperparameter grid and the PR>=0.99 constraint. If the known PULX observations no longer form a coherent cluster at PR>=0.99 (e.g., the threshold cannot be satisfied), or if the overlap between the newly recovered candidate list and the 85 sources in Table B.1 drops by more than ~75%, then the original clustering is driven by the observability features (peak flux and variability) rather than by any physical property indicative of a neutron-star accretor. This would validate the concern that the 'candidate PULX' label is a selection-effect artifact rather than evidence for hidden pulsars.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that 85 ULXs are candidate pulsating ULXs depends on the GMM components corresponding to NS vs BH accretors. However, the internal evidence indicates the separation is driven by observability. The decision tree (Fig. 3) isolates the PULX cluster using EP_8_FLUX > -12.387, Fpeak > -11.962, and VAR_FLAG, while HRHard has a Silhouette Index of only 0.039 (Table 4), so hardness does not separate the clusters. The authors themselves state in Sect. 4.1 that all candidate PULXs have the highest Fpeak values in the dataset, and that a source needs roughly three observations to have a ~70% chance of detecting pulsations. This is precisely the selection function that determines whether pulsations are detectable in a ULX, independent of accretor type. Moreover, Sect. 4.2 reports that literature-classified BH ULXs (Holmberg II X-1, NGC 5408 X-1, NGC 7793 P9, M31 ULX-1) fall in the PULX cluster, confirming that the cluster does not uniquely trace neutron-star accretors. The preliminary timing search found no new pulsations, consistent with the candidates being bright, non-pulsating ULXs. The load-bearing assumption that the two GMM components correspond to accretor type is therefore not supported by the data; the separation is at least plausibly a brightness/variability selection effect.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper applies an unsupervised Gaussian Mixture Model (GMM) to XMM-Newton observations of ultraluminous X-ray sources (ULXs) drawn from the Walton et al. (2022) catalogue, updated with 4XMM-DR13. The algorithm clusters the data into two components, and the cluster containing the known pulsating ULXs (PULXs) is identified by setting a probability threshold that recovers at least 99% of known PULX observations while maximizing the fraction of unknown ULXs placed in the other cluster. The pipeline selects hyperparameters and a threshold on the same dataset. The resulting PULX cluster contains 85 previously unknown candidate PULXs across 355 observations, with a decision tree explaining the classification through high broadband flux, high peak flux, and variability. A preliminary timing search finds no new pulsations in these candidates.","tokens_in":27838,"tokens_out":3509,"duration_ms":40693,"significance":"If the cluster assignment genuinely traces neutron-star accretors, the 85 candidates constitute a valuable, ranked target list for deep pulsation searches and would represent a major step toward quantifying the PULX fraction. The manuscript is transparently written, the pipeline is clearly specified, and the catalogue (Tables B.1 and C.1) is a useful community resource. However, the physical interpretation of the clusters is not established: the separation is driven by brightness and variability, which are the same factors that determine whether pulsations are detectable, and the paper's own internal checks do not rule out that the candidates are simply bright, variable, non-pulsating ULXs. The absence of any new pulsations in the timing search is consistent with this alternative interpretation.","major_comments":[{"comment":"The central claim that the PULX cluster contains candidate neutron-star ULXs rests on the assumption that the two GMM components separate NS accretors from BH accretors. The evidence presented does not support this. The decision tree in Fig. 3 and the silhouette indices in Table 4 show that the separation is driven by EP_8_FLUX, Fpeak, and VAR_FLAG, while HRHard has a silhouette index of only 0.039, meaning hardness does not separate the clusters. Furthermore, Sect. 4.2 reports that literature-classified BH ULXs (Holmberg II X-1, NGC 5408 X-1, NGC 7793 P9, M31 ULX-1) fall inside the PULX cluster. The authors themselves note in Sect. 4.1 that all candidate PULXs have the highest Fpeak values in the dataset. These features (peak flux and variability) are exactly the selection function that determines whether pulsations are detectable, independent of accretor type. The paper therefore does not demonstrate that the 85 candidates are preferentially NS accretors; they may be bright, variable ULXs of any accretor type. To make the central claim load-bearing, the authors should show that the cluster separation persists when controlling for flux, distance, and observation count, or that physical spectral indicators (e.g., hardness, spectral shape) contribute beyond the detection-selection features. The current null result from the timing search, while preliminary, is fully consistent with the null hypothesis that the candidates are non-pulsating sources.","section":"Section 2.3.2, Table 4, Fig. 3, Sect. 4.2"},{"comment":"The probability threshold is selected on the same dataset that is used to evaluate the method: the pipeline (steps 2c and 3) chooses the threshold that maximizes UR subject to PR>=0.99 computed on the known PULX observations. Consequently, the reported PR is a design constraint rather than an independent measure of recall. The leave-one-out robustness checks in Sect. 4.1 only show that this imposed constraint is stable when a small fraction of PULX observations is removed; they do not test whether the cluster assignment generalizes to genuinely unseen sources. For a claim of predictive power and a candidate list that is meant to prioritize future observations, a proper validation scheme is required, for example holding out a random subset of known PULX observations for threshold selection and evaluating on the rest, or performing cross-validation. Without such a validation, the 85 candidates are the result of a model tuned to fit the known PULXs, and their status as robust predictions is unproven.","section":"Section 2.3 (Pipeline), Eq. (2), Sect. 3"}],"minor_comments":[{"comment":"The statement that PR is 'always ≥0.5 by definition' is correct only because the maximum is taken over clusters; this is worth stating more explicitly, since the reader may otherwise interpret PR as a genuine recall metric rather than a constrained optimization objective.","section":"Sect. 2.2, Eq. (2)"},{"comment":"The timing search is described as using HENaccelsearch with a first period derivative, but no sensitivity estimate (e.g., minimum detectable pulsed fraction as a function of photon count) is given. Reporting this would help the reader interpret the null result.","section":"Sect. 4.2, last paragraph"},{"comment":"Several z-scores are given as '>8.00' or '-8.13' with no explanation of the clipping convention. It would be helpful to specify the range of z-scores considered and what values outside ±8 indicate.","section":"Table B.1"},{"comment":"The statement that 'all candidate PULXs have the highest Fpeak values in the dataset' is a key caveat and should appear earlier, in the Results section, rather than only in the Discussion.","section":"Sect. 4.1, first paragraph"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely to be of interest to the A&A readership as a methodology demonstration and a source catalogue. The main concern is that the candidate PULX list, which is the principal product, is not convincingly shown to trace neutron-star accretors rather than a brightness/variability selection effect. If the authors can add a validation that controls for observability selection (e.g., matching on flux or using physical spectral features) or an external test that demonstrates the candidates preferentially pulsate, the paper would be publishable. As it stands, the physical claim is defensible only as a 'targets worth observing' list, which is weaker than the abstract suggests."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper is a clear, well-documented application of GMM clustering to build a ranked list of ULX observations that look like known PULXs in a few observable dimensions. The 85-source candidate list is new and will be useful for follow-up. But the load-bearing interpretation—that the two GMM components separate neutron-star from black-hole accretors—is not supported by the paper's own data.\n\nWhat the paper does well: the pipeline is fully specified, the metrics are clearly defined, and the caveats in Sect. 4.1 are written honestly. The decision-tree explainability is a nice touch, and the result that Fpeak is the key ingredient is a concrete, falsifiable finding. The QPO coincidence is suggestive.\n\nSoft spots: the decision tree isolates the PULX cluster with cuts on broadband flux, peak flux, and a variability flag; hardness has almost no separating power (Silhouette 0.039). The authors themselves note that all candidates have the highest Fpeak values in the dataset, and they show that known black-hole ULXs (Holmberg II X-1, NGC 5408 X-1, NGC 7793 P9, M31 ULX-1) fall in the PULX cluster. That is exactly the pulsation-detectability selection function, not a physical accretor-type signature. The preliminary timing search found no new pulsations, consistent with the cluster being bright, variable ULXs rather than hidden neutron stars. The PR>=0.99 target plus a threshold chosen on the same data means the recovery of known PULXs is partly enforced by design; the leave-one-out test only re-derives the threshold, not the cluster model; there is no held-out validation; and the dataset is available only on request. None of these are fatal for a candidate-list paper, but they should be stated more prominently.\n\nWho this is for: observers planning high-statistics pulsation searches, and anyone working on ML applications to high-energy catalogs. I would send it to a serious referee, but I would expect major revisions: reframe the candidates as \"PULX-like in observability space,\" add a validation set, and be explicit that the cluster does not identify neutron-star accretors.","headline":"Useful ranked list of ULX observations that resemble known PULXs, but the clusters track brightness and variability, not accretor type.","tokens_in":28403,"tokens_out":3380,"would_cite":false,"duration_ms":35651,"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 clustering analysis singles out 85 ULXs that share the signature of known pulsating ULXs, making them prime targets for pulsation searches.","keywords":["ultraluminous X-ray sources","pulsating ULXs","Gaussian mixture models","unsupervised clustering","XMM-Newton","neutron stars","super-Eddington accretion","X-ray variability"],"falsifier":"A coherent pulsation search on the 85 candidates using exposures with at least 10,000 photons per source, the known threshold for PULX detection, would settle the claim: if spin signals appear in a substantial subset, the cluster is accretor-selected; if none appear despite statistics equal to the confirmed PULX detections, the cluster is not identifying hidden pulsars. A cheaper cross-check is to run the same two-component Gaussian mixture on a luminosity- and distance-matched sample of known black-hole ULXs and see whether they too fall in the PULX cluster.","tokens_in":27368,"feed_emoji":"🔭","tokens_out":8082,"duration_ms":77208,"temperature":0.7,"pith_summary":"The paper tries to establish that many ultraluminous X-ray sources (ULXs) that have never shown pulsations are nevertheless hidden pulsars, and that their identity can be predicted from archival catalogue data alone. The authors apply an unsupervised two-cluster Gaussian mixture model to 1769 XMM-Newton observations of 640 ULXs, then use the observations of the six known pulsating ULXs to label the cluster that should contain PULXs. With a threshold that captures at least 99% of known PULX observations, that cluster also contains 355 observations of 85 previously non-pulsating ULXs, about 85% of them observed more than once. The authors conclude that these 85 sources share the multi-dimensional phase-space signature of confirmed PULXs and are therefore the best candidates for deep pulsation searches.","feed_headline":"Clustering finds 85 hidden pulsar candidates among ULXs","feed_subtitle":"AI clustering of XMM-Newton observations flags bright, variable ULXs that match known pulsars.","key_machinery":"The engine of the analysis is a two-component Gaussian mixture model, run on scaled ULX observations of unknown accretor type; confirmed PULX observations are not used in the fit but are projected onto it to set the probability threshold (PULX ratio at least 0.99). A shallow decision tree then translates the cluster boundary into three observable cuts: broadband flux brighter than about $4\\times10^{-13}\\,\\mathrm{erg\\,cm^{-2}\\,s^{-1}}$, peak flux $F_{\\mathrm{peak}}$ brighter than about $1\\times10^{-12}\\,\\mathrm{erg\\,cm^{-2}\\,s^{-1}}$, and a variability flag indicating intra-observational variability. $F_{\\mathrm{peak}}$ is the load-bearing feature: the separation is much weaker without it, and it is what lets low-flux observations of known PULXs still be classified by the brightness of other epochs of the same source.","core_discovery":"The central claim is that the known PULXs are not isolated outliers: a two-component Gaussian mixture, fitted only on ULXs of unknown accretor type and thresholded so that at least 99% of confirmed PULX observations fall on one side, places 355 observations of 85 additional ULXs in the same cluster. Those candidates include sources already suspected to be neutron-star accretors, such as NGC 7793 ULX-4 and NGC 4559 X7, as well as some sources previously interpreted as black-hole ULXs. The decisive feature is the maximum observed flux Fpeak: including it raises the uncertain-ratio metric from 0.10–0.11 to 0.79, while luminosity adds nothing. The authors also report that 19 of 22 QPO-bearing observations fall in the PULX cluster, which they read as supporting a link between mHz quasi-periodic oscillations and neutron-star accretors.","pith_inferences":["Extension: if most of the 85 candidates are real PULXs, then the fraction of ULXs powered by neutron stars is far larger than the six confirmed cases suggest, and the super-Eddington neutron-star channel is a dominant, not exceptional, ULX mode.","Extension: because the cluster boundary is set by brightness and variability rather than by spectral hardness, the same recipe could be transferred to Chandra and Swift ULX catalogues to build a larger cross-mission candidate list.","Extension: the paper does not quantify how many of the 85 are bright black-hole ULXs; a matched control sample of spectroscopically identified black-hole ULXs, clustered in the same feature space, would give a direct contamination estimate."],"forward_implications":["The 85 candidates become a prioritized sample for high-statistics pulsation searches; detecting a spin period in any of them would confirm that the cluster boundary tracks accretor type.","The method stays stable when known PULX observations are removed one at a time or source by source (average PULX ratio around 97–99%), so new XMM-Newton ULX observations can be classified without refitting the mixture from scratch.","Because most candidates have multiple observations and PULX pulsations are seen in only 31 of 95 known-PULX observations, roughly three well-spaced observations per source should give about a 70% chance of catching a pulse if the candidate population behaves like the confirmed one.","The concentration of QPO-bearing ULXs in the PULX cluster implies that mHz quasi-periodic oscillations are a practical marker for identifying neutron-star accretors even before pulsations are found."],"supporting_citations":[{"why":"It supplies the base ULX catalogue, cross-matched with XMM-Newton, Chandra, and Swift data and host-galaxy distances, from which the dataset is built.","marker":"Walton et al. 2022"},{"why":"It provides the 4XMM serendipitous source catalogue from which fluxes, peak flux, and variability flags are taken.","marker":"Webb et al. 2020"},{"why":"It is the first discovery of pulsations in an ULX, establishing the PULX class that the clustering labels are meant to recover.","marker":"Bachetti et al. 2014"},{"why":"It supplies the known PULX observations, including M51 ULX-7, and the photon-statistics threshold used to benchmark pulsation searches.","marker":"Rodríguez Castillo et al. 2020"},{"why":"It provides the NGC 5907 ULX-1 low-flux, propeller-regime observations used to test that Fpeak recovers faint PULX epochs.","marker":"Fürst et al. 2023"},{"why":"It contributes confirmed PULX observations, notably NGC 1313 X-2, that define the target cluster.","marker":"Israel et al. 2017a,b"},{"why":"It reports mHz quasi-periodic oscillations in ULXs, used to check whether QPO-bearing observations fall in the PULX cluster.","marker":"Imbrogno et al. 2024"},{"why":"It argues that hard spectra select hidden PULXs, a prior used to interpret the spectral similarity of the candidates.","marker":"Pintore et al. 2017"}],"fun_headline_variants":["Clustering reveals 85 hidden pulsar candidates in ULXs","AI clustering spots 85 pulsating ULX candidates","Machine learning flags 85 new pulsar candidates in ULX data","Unsupervised clustering finds 85 ULX pulsar candidates","85 more ULX pulsar candidates found by AI clustering"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole candidate list rests on the assumption that the two-cluster split separates neutron-star ULXs from black-hole ULXs rather than separating bright, well-observed sources from faint, poorly observed ones.","fun_headline_variants_meta":{"raw":{"variants":["Clustering reveals 85 hidden pulsar candidates in ULXs","AI clustering spots 85 pulsating ULX candidates","Machine learning flags 85 new pulsar candidates in ULX data","Unsupervised clustering finds 85 ULX pulsar candidates","85 more ULX pulsar candidates found by AI clustering"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000278,"raw_usage":{"total_tokens":1716,"prompt_tokens":1072,"completion_tokens":644,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":688,"completion_tokens_details":{"reasoning_tokens":560}},"tokens_in":688,"tokens_out":644,"duration_ms":6988,"temperature":1.0,"reasoning_tokens":560,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:42:01.916137+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A coherent pulsation search on the 85 candidates using exposures with at least 10,000 photons per source, the known threshold for PULX detection, would settle the claim: if spin signals appear in a substantial subset, the cluster is accretor-selected; if none appear despite statistics equal to the confirmed PULX detections, the cluster is not identifying hidden pulsars. A cheaper cross-check is to run the same two-component Gaussian mixture on a luminosity- and distance-matched sample of known black-hole ULXs and see whether they too fall in the PULX cluster.","supporting_citations":[{"cited_title":"J., Mackenzie, A","cited_arxiv_id":null,"evidence_quote":"It supplies the base ULX catalogue, cross-matched with XMM-Newton, Chandra, and Swift data and host-galaxy distances, from which the dataset is built."},{"cited_title":"E., Amato, R., et al","cited_arxiv_id":null,"evidence_quote":"It reports mHz quasi-periodic oscillations in ULXs, used to check whether QPO-bearing observations fall in the PULX cluster."},{"cited_title":"2017, ApJ, 836, 113","cited_arxiv_id":null,"evidence_quote":"It argues that hard spectra select hidden PULXs, a prior used to interpret the spectral similarity of the candidates."}],"review_version":1}