{"id":"20dd22b4-2b80-4fa4-aa66-3d8a80585027","arxiv_id":"2605.05318","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":4.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":2,"one_line_summary":"A convLSTM classifier identifies lensed SNe Ia in simulated LSST-like time series, reaching ~60% true-positive rate at O(10^{-4}) false-positive rate by the seventh epoch even after adding realistic PSF variations and foreground SN contaminants.","lead":"The paper extends a convolutional LSTM deep-learning model to detect strongly lensed Type Ia supernovae from realistic multi-band, multi-epoch image sequences that include PSF variations and an extra false-positive class of supernovae in the lens galaxy. If the reported performance holds on real data, it could enable timely alerts for rare events that serve as cosmological probes in surveys like LSST.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.3","headline":"HSC PDR3 simulations with injected arcs and modeled PSF/noise variations may not fully capture LSST-specific systematics and false-positive distributions","rationale":"The reader's weakest assumption matches the load-bearing point exactly. The full text (methods and simulation sections) would need to be inspected for any additional validation against real LSST data or cross-checks with alternative simulators; absent that, the performance numbers remain conditional on simulation fidelity rather than unconditionally demonstrated.","tokens_in":1844,"tokens_out":449,"duration_ms":43800,"concrete_test":"Re-generate a 10% hold-out subset of the test time series using an independent LSST-like simulator (e.g., with LSST OpSim cadence, actual filter throughputs, and measured early LSST PSF models) while keeping the same lensing and light-curve parameters; retrain or evaluate the published ConvLSTM weights on this new set and report the change in TPR at FPR=10^{-4}. A shift >15% in TPR indicates the original HSC-based realism is insufficient.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central performance claim (TPR ~60% at FPR O(10^{-4}) by epoch 7, ~80% by epoch 10) is measured exclusively on a test set constructed from HSC PDR3 single-epoch images by adding epoch-to-epoch PSF variations, variance-map corrections, Poisson noise, simulated lensed arcs, SN light-curve variations, and one extra negative class (foreground SN Ia in the lens). This construction implicitly assumes that the resulting statistical properties (noise correlations, PSF ellipticity distributions, host-galaxy morphologies, and contaminant rates) match those of real LSST multi-epoch difference imaging. If unmodeled LSST effects (e.g., correlated read-noise, filter-dependent depth variations, or additional variable-source classes such as AGN or stellar flares) are present at levels that alter the decision boundary, the quoted ROC numbers will not translate. The paper's own addition of realism makes this assumption the single most load-bearing step for the claim that the approach is 'well suited for real-time LSN searches in LSST'.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"This paper extends a prior convolutional LSTM framework for detecting strongly lensed Type Ia supernovae (LSNe Ia) from multi-band, multi-epoch image cutouts. It adds realism to HSC PDR3-based simulations by introducing epoch-to-epoch PSF variations with variance-map corrections, Poisson noise, simulated lensed host arcs, SN light-curve variations, and an extra negative class of foreground SN Ia in the lens galaxy. Despite these complexities, the model shows rapid improvement in ROC performance, reaching a true-positive rate of ~60% at a false-positive rate of O(10^{-4}) by the seventh observation and ~80% by the tenth, with additional analysis of confusion from sibling SNe in LRGs. The work concludes that the approach is robust and suitable for real-time LSST searches.","tokens_in":2097,"tokens_out":583,"duration_ms":23287,"significance":"If the simulations prove representative, the results provide concrete evidence that a spatiotemporal deep-learning classifier can maintain useful detection efficiency under realistic multi-epoch conditions, which is valuable for prompt follow-up of rare LSNe Ia in LSST alert streams. The explicit addition of PSF variations, variance corrections, and a challenging foreground contaminant class strengthens the practical relevance beyond the earlier study.","major_comments":[{"comment":"The central performance claim (TPR ~60% at FPR O(10^{-4}) by epoch 7, ~80% by epoch 10) is obtained exclusively on a test set constructed from HSC PDR3 single-epoch images with added epoch-to-epoch PSF variations, variance-map corrections, Poisson noise, injected lensed arcs, SN light-curve variations, and foreground SN Ia contaminants. This construction implicitly assumes that the resulting statistical properties match those of real LSST difference imaging; unmodeled effects such as correlated read noise, filter-dependent depth variations, or additional variable-source classes (AGN, stellar flares) could alter the decision boundary. Because this assumption is load-bearing for the claim that the method is 'well suited for real-time LSN searches in LSST', a quantitative sensitivity test or direct comparison against more comprehensive LSST mocks is required.","section":"Abstract and dataset construction"}],"minor_comments":[{"comment":"The abstract employs both ~ and mathcal{O} notation; ensure identical usage and explicit definitions appear in the main text and figure captions.","section":"Abstract"},{"comment":"A table listing the exact simulation parameters (PSF variation model, noise levels, number of injected arcs, etc.) would improve reproducibility.","section":"Methods"}],"recommendation":"major_revision","confidential_remarks":"The manuscript fits the scope of astro-ph.IM. The citation pattern appropriately references the prior part of the series without over-citation."},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for their careful and constructive review of our manuscript. The concern about the degree of realism in our simulations relative to full LSST difference imaging is valid, and we have revised the text to acknowledge limitations while preserving the core results.","responses":[{"response":"We agree that our simulations, although they incorporate PSF variations, variance-map corrections, Poisson noise, lensed host arcs, SN light-curve variations, and foreground SN Ia contaminants, do not capture every possible effect present in real LSST difference imaging. Unmodeled contributions such as correlated read noise, filter-dependent depth variations, and additional variable-source classes (e.g., AGN or stellar flares) could in principle shift the decision boundary. In the revised manuscript we have added a new subsection (Section 5.3) that explicitly discusses these limitations and qualitatively assesses their likely impact on classifier performance. We have also revised the abstract and concluding paragraph to replace the phrase 'well suited for real-time LSN searches in LSST' with the more cautious statement that the approach 'shows promise under LSST-like conditions and merits further validation with more complete simulations.' A quantitative sensitivity test or direct comparison against comprehensive LSST mocks would require generation or access to substantially more advanced mock datasets and is outside the scope of the present study; such work is planned for a follow-up investigation.","revision_made":"partial","referee_comment":"The central performance claim (TPR ~60% at FPR O(10^{-4}) by epoch 7, ~80% by epoch 10) is obtained exclusively on a test set constructed from HSC PDR3 single-epoch images with added epoch-to-epoch PSF variations, variance-map corrections, Poisson noise, injected lensed arcs, SN light-curve variations, and foreground SN Ia contaminants. This construction implicitly assumes that the resulting statistical properties match those of real LSST difference imaging; unmodeled effects such as correlated read noise, filter-dependent depth variations, or additional variable-source classes (AGN, stellar flares) could alter the decision boundary. Because this assumption is load-bearing for the claim that the method is 'well suited for real-time LSN searches in LSST', a quantitative sensitivity test or direct comparison against more comprehensive LSST mocks is required."}],"tokens_in":1605,"tokens_out":525,"duration_ms":45451,"standing_objections":["Quantitative sensitivity tests to unmodeled effects (correlated read noise, AGN, stellar flares, filter-dependent depth variations) or direct comparison against full LSST mocks cannot be performed within the current revision."]},"desk_editor":{"model":"grok-4.3","letter":"The core advance here is the simulation pipeline that turns single-epoch HSC PDR3 cutouts into multi-epoch time series with epoch-to-epoch PSF changes, variance-map corrections, and an explicit foreground SN Ia contaminant class. That is a concrete step beyond the 2026 paper. They keep the same convLSTM architecture and show the ROC curve improving to roughly 60% true-positive rate at 10^{-4} false-positive rate by the seventh observation and 80% by the tenth, even with the extra realism and the new negative class. They also run a quick check on sibling SNe in LRGs to see which configurations are most likely to confuse the classifier. Those numbers are reported clearly and the incremental realism is documented without overclaiming.","headline":"This follow-up adds realistic PSF variations and a foreground SN negative class to their earlier convLSTM pipeline, but performance is still measured only on HSC-derived mocks.","tokens_in":2620,"tokens_out":227,"would_cite":false,"duration_ms":25035,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A deep learning model using convolutional LSTM detects strongly lensed Type Ia supernovae from realistic multi-band time-series images, achieving approximately 60 percent true-positive rate at a false-positive rate of order 10 to the minus ","keywords":["strongly lensed supernovae","Type Ia supernovae","LSST","deep learning","time series","gravitational lensing","transient detection","image classification"],"falsifier":"Comparing the model's receiver operating characteristic curves on actual LSST observations against the simulated performance curves would confirm or refute the reported true-positive and false-positive rates.","tokens_in":2773,"feed_emoji":"🔭","tokens_out":659,"duration_ms":40360,"temperature":0.7,"pith_summary":"This paper extends a previous deep-learning approach for finding rare strongly lensed supernovae Ia in upcoming LSST data. It tests the model on more realistic simulations that include varying point-spread functions between epochs, variance corrections, and supernovae occurring in the foreground lens galaxy as contaminants. The core result is that the classifier still performs well, updating its classification as new observations arrive. Sympathetic readers would care because prompt detection is needed to trigger expensive follow-up observations that can measure time delays and study the lens system. The work shows the method is robust enough to handle real survey conditions.","feed_headline":"Lensed supernova detector reaches 60% true positives by seventh observation","feed_subtitle":"Realistic simulations with PSF variations and foreground contaminants show the time-series classifier identifies candidates for prompt LSST","key_machinery":"The convolutional LSTM architecture that processes multi-band, multi-epoch image cutouts to capture spatiotemporal correlations and update classifications with each new observation.","core_discovery":"We extend the previous convolutional LSTM framework by constructing realistic image time series from HSC PDR3 observations, introducing epoch-to-epoch PSF variations with variance-map corrections, simulated lensed arcs, SN light-curve variations, Poisson noise, and foreground SN Ia contaminants. Despite these additions, the model reaches a true-positive rate of ~60% at a false-positive rate of O(10^{-4}) by the seventh observation and ~80% by the tenth. We also examine confusion with sibling SNe in LRGs and identify mimicking configurations.","pith_inferences":["The robustness suggests the model is ready for deployment on early LSST data without extensive additional tuning.","Similar time-series approaches could be tested for detecting other lensed transients such as core-collapse supernovae or quasars.","If the low false-positive rate holds on real data, it would greatly reduce the resources needed for spectroscopic follow-up of candidates."],"forward_implications":["The classifier supports real-time LSN searches in LSST alert streams.","Detection performance improves rapidly with additional epochs of observation.","Foreground lens-galaxy supernovae form an important false-positive class that must be distinguished.","Specific configurations of sibling supernovae can be used to improve model robustness."],"fun_headline_variants":["Lensed SN Ia detector hits 60% true positives by seventh observation","Realistic HSC observations test lensed supernova detection framework","Time series model identifies lensed SNe Ia amid PSF and contaminant noise","Model reaches 60% true positive rate by seventh epoch in complex data"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The HSC PDR3-based simulations with injected lensed arcs, SN light-curve variations, Poisson noise, and PSF variations sufficiently capture the statistical properties of real LSST observations and all relevant false-positive classes.","fun_headline_variants_meta":{"raw":{"variants":["Lensed SN Ia detector hits 60% true positives by seventh observation","Realistic HSC observations test lensed supernova detection framework","Time series model identifies lensed SNe Ia amid PSF and contaminant noise","Model reaches 60% true positive rate by seventh epoch in complex data"]},"model":"grok-4.3","cost_usd":0.009237,"raw_usage":{"total_tokens":4129,"prompt_tokens":815,"num_sources_used":0,"completion_tokens":73,"cost_in_usd_ticks":92365500,"prompt_tokens_details":{"text_tokens":815,"audio_tokens":0,"image_tokens":0,"cached_tokens":64},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":3241,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":815,"tokens_out":73,"duration_ms":28179,"temperature":1.0,"reasoning_tokens":3241,"cache_read_input_tokens":64,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-05-08T16:19:46.757886+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Comparing the model's receiver operating characteristic curves on actual LSST observations against the simulated performance curves would confirm or refute the reported true-positive and false-positive rates.","supporting_citations":[],"review_version":1}