pith. machine review for the scientific record. sign in

arxiv: 2604.08196 · v2 · submitted 2026-04-09 · 🌌 astro-ph.IM · astro-ph.GA· astro-ph.HE

Recognition: unknown

A Statistical-AI Framework for Detecting Transient Flares in SDSS Stripe 82 Quasar Light Curves

Atal Agrawal
Authors on Pith no claims yet

Pith reviewed 2026-05-10 17:30 UTC · model grok-4.3

classification 🌌 astro-ph.IM astro-ph.GAastro-ph.HE
keywords quasar flareslight curve analysisanomaly detectionSDSS Stripe 82damped random walktransient eventsaccretion physicsmachine learning
0
0 comments X

The pith

The FLARE framework identifies 51 quasars with distinct flaring activity in SDSS Stripe 82 by modeling baseline variability and scoring anomalies.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Quasars normally vary in brightness in a way that follows a damped random walk pattern. Occasional flares stand out as sharp departures from this pattern and could reveal sudden changes in how material falls onto the central black hole. The paper introduces FLARE, a three-stage process that first fits the expected variability, then flags statistical outliers, and finally verifies them with a recognition step. When run on r-band light curves for more than nine thousand quasars spanning roughly ten years, and checked against g-band data, the method yields 51 clear flare candidates. This separates rare physical events from the constant background of stochastic changes.

Core claim

The paper presents FLARE as a modular pipeline that models baseline quasar variability with either a physics-informed gated recurrent unit trained on simulated damped random walk light curves or an iterative Ornstein-Uhlenbeck process fitted to the data with outlier masking, applies extreme value theory to score anomalies, and uses vision language models to verify candidate events. Applied to the r-band light curves of 9,258 spectroscopically confirmed quasars in the SDSS Stripe 82 dataset with g-band cross-checks, the framework detects 51 quasars that exhibit distinct flaring activity as significant departures from the modeled baseline.

What carries the argument

FLARE, a three-stage pipeline that models baseline stochastic variability with damped random walk or Ornstein-Uhlenbeck processes, applies extreme value theory for anomaly scoring, and employs vision language models for final verification.

Load-bearing premise

The damped random walk model together with outlier-masked Ornstein-Uhlenbeck fitting fully describes normal quasar variability, so that the remaining anomalies are genuine physical flares rather than unmodeled changes or data artifacts.

What would settle it

Reprocessing the 51 candidate light curves with a more flexible variability model that includes additional parameters for non-stationary behavior and finding that most or all of the flagged events fall within the revised baseline distribution.

Figures

Figures reproduced from arXiv: 2604.08196 by Atal Agrawal.

Figure 1
Figure 1. Figure 1: —: The FLARE framework. The pipeline consists of three stages: (1) Baseline Modeling, where a physics￾informed probabilistic GRU or an iterative OU process models the DRW variability of each quasar light curve; (2) Anomaly Scoring, where standardized residuals are analyzed using Extreme Value Theory with a peaks-over￾threshold GPD fit to flag candidate events; and (3) Recognition Engine, where Vision Langu… view at source ↗
Figure 2
Figure 2. Figure 2: —: Architecture of the recognition engine. Two VLMs act as independent classifiers, each providing a flare/non-flare classification for the candidate light curve. A third VLM serves as an evaluator, flagging misclassi￾fications and providing feedback to the classifiers. This evaluation–feedback cycle is repeated for two iterations to allow the classifiers to refine their predictions [PITH_FULL_IMAGE:figur… view at source ↗
Figure 3
Figure 3. Figure 3: —: Example simulated DRW light curve with Gaussian noise injected based on per-epoch Stripe 82 photometric errors. Axes and labels are intentionally omitted as these images represent the direct morpholog￾ical inputs fed to the VLMs. We inject three morphologically distinct flare types. For all three, the peak time tpeak is drawn randomly from the observed MJD timestamps, ensuring that the flare peak coinci… view at source ↗
Figure 5
Figure 5. Figure 5: —: Example simulated DRW light curve with an injected Gamma flare. Axes and labels are intentionally omitted as these images represent the direct morpholog￾ical inputs fed to the VLMs [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: —: Example simulated DRW light curve with an injected Gaussian flare. Axes and labels are intentionally omitted as these images represent the direct morpholog￾ical inputs fed to the VLMs. Figures 3, 4, 5, 6, and 7 show examples of simulated light curves for each of the five classes. The plots use a white background with a faint grid and no axis labels, so that the VLMs classify based solely on the morpholo… view at source ↗
Figure 7
Figure 7. Figure 7: —: Example simulated DRW light curve with an injected single-point spike. Axes and labels are in￾tentionally omitted as these images represent the direct morphological inputs fed to the VLMs. 4.1.1. Physics-Informed Probabilistic GRU We choose a GRU architecture over alternative se￾quence models for several reasons. Standard RNNs lack gating mechanisms, making them less effective at se￾lectively retaining … view at source ↗
Figure 8
Figure 8. Figure 8: —: Training loss curves for the physics-informed probabilistic GRU over 100 epochs. (a) Data loss (negative log-likelihood), which converges within the first ∼ 20 epochs. (b) Physics loss (drift regularizer), measuring deviation from the OU conditional mean. (c) Variance loss, measuring deviation from the OU conditional variance. All three components are shown unweighted; the annealed weights λphys and λva… view at source ↗
Figure 9
Figure 9. Figure 9: —: EVT anomaly scoring for the GRU baseline. (a) GPD fit to the empirical exceedances above the 95th percentile threshold. (b) Distribution of calibrated maximum |z|-scores across 9,258 simulated quasars, with the detection boundary at 8.69σ (dashed line). Objects exceeding this threshold yield 51 candidates. 4.1.2. Iterative OU Process The second baseline fits an OU process directly to the observed r-band… view at source ↗
Figure 10
Figure 10. Figure 10: —: EVT anomaly scoring for the iterative OU baseline. (a) GPD fit to the empirical exceedances above the 95th percentile threshold. (b) Distribution of maximum |z|-scores across 9,104 converged quasars, with the detection boundary at 3.73σ (dashed line), yielding 92 candidates. bution (GPD) (Coles 2001): P(z > u + y | z > u) =  1 + ξ y β −1/ξ , (23) where ξ is the shape parameter and β is the scale para… view at source ↗
Figure 11
Figure 11. Figure 11: —: Comparison of r-band DRW parameters estimated in this work using the iterative OU process with those derived by MacLeod et al. (2010) (M10). (a) Damping timescale τ (Spearman ρ = 0.86), showing strong agreement along the 1:1 line (dashed). (b) Variability amplitude ˆσ (ρ = 0.39), showing larger scatter and a systematic offset above the 1:1 line, indicating that our estimates are generally higher. 4.3. … view at source ↗
Figure 12
Figure 12. Figure 12: —: Binary flare detection precision versus recall for all 12 benchmarked VLMs, obtained by collapsing the five classes into flare (FRED, Gaussian, Gamma) and non-flare (DRW, Spike). Qwen-3.5-plus achieves the highest precision (∼ 88%), while Grok-4.1-fast achieves the highest recall (∼ 70%). Based on these results, we select Grok-4.1-fast as the high-recall classifier (Classifier A), Qwen-3.5-plus as the … view at source ↗
Figure 14
Figure 14. Figure 14: —: Confusion matrix for the recognition engine applied to the 51 flare candidates from the GRU base￾line. True labels are determined by independent human verification of both r-band and g-band light curves. The recognition engine achieves a precision of 100% and a re￾call of 75.9%, with 22 true positives, zero false positives, 7 false negatives, and 22 true negatives. For the recognition engine, performan… view at source ↗
Figure 13
Figure 13. Figure 13: —: Confusion matrix for the recognition engine applied to the 92 flare candidates from the iterative OU baseline. True labels are determined by independent hu￾man verification of both r-band and g-band light curves. The recognition engine achieves a precision of 55.2% and a recall of 59.3%, with 16 true positives, 13 false posi￾tives, 11 false negatives, and 52 true negatives. Of the 51 confirmed flares, … view at source ↗
Figure 15
Figure 15. Figure 15: —: Stability of the GPD fit for the iterative OU baseline across threshold percentiles from the 90th to the 98th. (a) The shape parameter ξ remains stable around ∼ 0.10 (dashed red line). (b) The detection boundary varies by less than 0.06σ across percentiles. (c) The number of tail exceedances decreases linearly with increasing percentile, as expected [PITH_FULL_IMAGE:figures/full_fig_p013_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: —: Stability of the GPD fit for the GRU baseline across threshold percentiles from the 90th to the 98th. (a) The shape parameter ξ remains stable around ∼ 0.35 (dashed red line). (b) The detection boundary varies by less than 0.15σ across percentiles. (c) The number of tail exceedances decreases linearly with increasing percentile, as expected. B. VLM ACCURACY PLOT 0 10 20 30 40 50 Accuracy (%) GPT-5 GPT-… view at source ↗
Figure 17
Figure 17. Figure 17: —: Five-class classification accuracy for all 12 benchmarked VLMs on the test set of 4,630 light curves (926 per class). GPT-5 achieves the highest accuracy at 42.8%. All models exceed the 20% random baseline [PITH_FULL_IMAGE:figures/full_fig_p014_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: —: Confusion matrix for the base Qwen2.5VL-7b model on the five-class test set (926 light curves per class). The model predicts non-flare for the vast majority of inputs across all classes, achieving 97% accuracy on the non-flare class but correctly identifying only ∼ 3% of flare light curves. non_flare spike gaussian gamma fred Predicted Label non_flare spike gaussian gamma fred True Label 722 3 1 200 0 … view at source ↗
Figure 19
Figure 19. Figure 19: —: Confusion matrix for the QLoRA fine-tuned Qwen2.5VL-7b model on the five-class test set (926 light curves per class). Fine-tuning improves flare detection—the fraction of flare light curves correctly classified as a flare type increases from ∼ 3% to ∼ 15%—but introduces a systematic bias toward the Gamma class, with most flare predictions collapsing into this single category regardless of the true morp… view at source ↗
Figure 20
Figure 20. Figure 20: —: r-band (red) and g-band (green) light curves for the 51 confirmed flaring quasars identified by the FLARE framework. Each light curve is titled with the quasar’s database identifier (dbID) from the MacLeod et al. (2010) catalog. The horizontal axis shows the time elapsed since the first observation in Modified Julian Days (∆MJD), and the vertical axis shows the apparent magnitude. The correlated variab… view at source ↗
Figure 20
Figure 20. Figure 20: —: continued. 0 500 1000 1500 2000 2500 3000 3500 ¢MJD (days) 20.0 20.2 20.4 20.6 20.8 Magnitude dbID 697882 r-band g-band 0 500 1000 1500 2000 2500 3000 ¢MJD (days) 20.25 20.50 20.75 21.00 21.25 21.50 21.75 Magnitude dbID 914271 r-band g-band [PITH_FULL_IMAGE:figures/full_fig_p017_20.png] view at source ↗
Figure 20
Figure 20. Figure 20: —: continued. 0 500 1000 1500 2000 2500 3000 ¢MJD (days) 19.50 19.75 20.00 20.25 20.50 20.75 Magnitude dbID 1527057 r-band g-band 0 500 1000 1500 2000 2500 3000 ¢MJD (days) 19.8 20.0 20.2 20.4 20.6 20.8 Magnitude dbID 1568428 r-band g-band [PITH_FULL_IMAGE:figures/full_fig_p018_20.png] view at source ↗
Figure 20
Figure 20. Figure 20: —: continued. 0 500 1000 1500 2000 2500 3000 3500 ¢MJD (days) 20.0 20.2 20.4 20.6 20.8 21.0 21.2 Magnitude dbID 2069433 r-band g-band 0 500 1000 1500 2000 2500 3000 3500 ¢MJD (days) 19.8 20.0 20.2 20.4 20.6 20.8 21.0 Magnitude dbID 2084710 r-band g-band [PITH_FULL_IMAGE:figures/full_fig_p019_20.png] view at source ↗
Figure 20
Figure 20. Figure 20: —: continued. 0 500 1000 1500 2000 ¢MJD (days) 19.6 19.8 20.0 20.2 20.4 20.6 20.8 Magnitude dbID 2957688 r-band g-band 0 500 1000 1500 2000 2500 3000 ¢MJD (days) 19.8 20.0 20.2 20.4 Magnitude dbID 3047173 r-band g-band [PITH_FULL_IMAGE:figures/full_fig_p020_20.png] view at source ↗
Figure 20
Figure 20. Figure 20: —: continued. 0 500 1000 1500 2000 2500 3000 ¢MJD (days) 19.25 19.50 19.75 20.00 20.25 20.50 20.75 Magnitude dbID 3785650 r-band g-band 0 500 1000 1500 2000 2500 3000 ¢MJD (days) 20.00 20.25 20.50 20.75 21.00 21.25 Magnitude dbID 3801729 r-band g-band [PITH_FULL_IMAGE:figures/full_fig_p021_20.png] view at source ↗
Figure 20
Figure 20. Figure 20: —: continued. 0 500 1000 1500 2000 2500 ¢MJD (days) 19.8 20.0 20.2 20.4 20.6 20.8 Magnitude dbID 4989108 r-band g-band [PITH_FULL_IMAGE:figures/full_fig_p022_20.png] view at source ↗
read the original abstract

Quasars exhibit stochastic variability across wavelengths, typically well described by a Damped Random Walk (DRW). Occasionally, however, they undergo extreme luminosity changes--known as flares--that represent significant departures from this baseline behavior and provide valuable probes of accretion disc dynamics and the physics of supermassive black hole fueling. Although modern transient surveys have spurred growing interest in flare detection, no systematic search has yet been conducted within the legacy SDSS Stripe 82 dataset, which contains 9,258 spectroscopically confirmed quasars observed over a ~10-year baseline. The principal statistical challenge is distinguishing these rare events from the ever-present stochastic variability. To address this, we present FLARE (Flare detection via physics-informed Learning, Anomaly scoring, and Recognition Engine), a modular three-stage framework for detecting transient flares in quasar light curves. FLARE models baseline DRW behavior, applies statistical anomaly scoring to flag candidate events, and employs a recognition engine to verify detections. For the Stripe 82 implementation, we deploy two complementary baselines--a physics-informed probabilistic Gated Recurrent Unit (GRU) trained on simulated DRW light curves, and an iterative Ornstein-Uhlenbeck (OU) process fitted directly to observed data with outlier masking--followed by Extreme Value Theory (EVT) for anomaly scoring. We benchmark twelve open-weight and proprietary Vision Language Models (VLMs) as recognition engines for final candidate verification. Detection is performed on r-band light curves, with candidates cross-checked against g-band data to rule out instrumental artifacts. Applying this framework, we identify 51 quasars exhibiting distinct flaring activity.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces the FLARE framework, a three-stage pipeline for detecting transient flares in SDSS Stripe 82 quasar light curves. It models baseline variability using a physics-informed GRU trained on simulated DRW light curves and an iterative Ornstein-Uhlenbeck process fitted to observed data with outlier masking, applies Extreme Value Theory for anomaly scoring, and uses Vision Language Models for final verification, ultimately claiming the detection of 51 quasars with distinct flaring activity in r-band (cross-checked in g-band).

Significance. If the pipeline's ability to separate genuine flares from stochastic variability is demonstrated with quantitative metrics, the work would provide a systematic, reproducible approach to identifying rare accretion events in legacy photometric datasets, offering new probes of supermassive black hole fueling and disk physics. The modular combination of physics-informed simulation, statistical anomaly detection, and AI verification is a constructive direction, but the absence of performance benchmarks currently limits its assessed impact.

major comments (3)
  1. [Abstract] Abstract: The central claim of 51 detections is presented without any reported false-positive rates, precision/recall on simulated DRW-only light curves, recovery efficiency for injected flares, or comparison against a control sample of known non-flaring quasars. This validation is load-bearing for the claim that the two-stage baseline plus EVT scoring cleanly isolates physical flares.
  2. [Methodology (OU fitting)] Methodology section on the iterative OU process: Fitting the Ornstein-Uhlenbeck parameters directly to the observed light curves while iteratively masking outliers creates circularity, as the baseline model is partly defined by the same data used to score anomalies. This risks either excising real flares during masking or retaining unmodeled red-noise components that EVT then flags as candidates.
  3. [Results] Results section: No quantitative comparison is provided between the 51 candidates and either previously reported flaring quasars in the literature or the output of simpler DRW-residual methods, leaving the added value of the GRU+EVT+VLM stages unquantified.
minor comments (2)
  1. [Abstract] Abstract: The acronym FLARE is expanded only after first use; define it on first appearance for clarity.
  2. [Methodology (VLM stage)] The description of the twelve VLMs benchmarked lacks details on prompt engineering or decision thresholds used for verification; this should be expanded in the methods for reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed review of our manuscript on the FLARE framework. We have addressed each major comment point by point below, providing clarifications and committing to revisions that strengthen the validation and methodological transparency without misrepresenting our results.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim of 51 detections is presented without any reported false-positive rates, precision/recall on simulated DRW-only light curves, recovery efficiency for injected flares, or comparison against a control sample of known non-flaring quasars. This validation is load-bearing for the claim that the two-stage baseline plus EVT scoring cleanly isolates physical flares.

    Authors: We acknowledge that the abstract and main text do not explicitly report quantitative performance metrics such as false-positive rates, precision/recall on DRW-only simulations, or recovery efficiency for injected flares. The manuscript describes training the GRU on simulated DRW light curves and VLM-based verification of the 51 candidates (with g-band cross-checks), but these do not constitute the full suite of benchmarks requested. We will revise the manuscript by adding a dedicated validation subsection that includes these metrics, derived from our existing simulations and a control sample of non-flaring quasars, to be placed before the results section. revision: yes

  2. Referee: [Methodology (OU fitting)] Methodology section on the iterative OU process: Fitting the Ornstein-Uhlenbeck parameters directly to the observed light curves while iteratively masking outliers creates circularity, as the baseline model is partly defined by the same data used to score anomalies. This risks either excising real flares during masking or retaining unmodeled red-noise components that EVT then flags as candidates.

    Authors: The iterative OU fitting with outlier masking is a deliberate design choice to estimate the underlying DRW baseline while mitigating the influence of transients, following established practices in time-series modeling. We recognize the potential circularity concern and will expand the methodology section with a detailed algorithmic description, including the specific masking threshold, iteration limits, and convergence criteria. We will also add simulation experiments demonstrating that injected flares are preserved and that residual red-noise components are not systematically passed to the EVT stage. revision: partial

  3. Referee: [Results] Results section: No quantitative comparison is provided between the 51 candidates and either previously reported flaring quasars in the literature or the output of simpler DRW-residual methods, leaving the added value of the GRU+EVT+VLM stages unquantified.

    Authors: The SDSS Stripe 82 quasar sample has not previously undergone a systematic flare search, which limits direct literature comparisons. We agree that explicit benchmarking against simpler DRW-residual methods would better quantify the incremental value of the full pipeline. In the revised manuscript, we will add a comparison subsection in the results that applies a basic DRW residual outlier detection to the same light curves, reports overlap with our 51 candidates, and discusses any matches to known flaring events from the literature. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation or claims

full rationale

The paper presents an applied statistical-AI pipeline (FLARE) for flare detection rather than a mathematical derivation chain. Baseline modeling uses a GRU trained on external simulated DRW light curves plus iterative OU fitting with outlier masking on the target data, followed by EVT scoring and VLM verification. No equations, self-citations, or steps are described that reduce any claimed result (such as the count of 51 flaring quasars) to its inputs by construction; the anomaly detection operates on deviations from a fitted model, which is a standard non-circular statistical procedure. The framework remains self-contained against external benchmarks like simulations and cross-band checks.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 1 invented entities

The central claim rests on the domain assumption that DRW describes baseline variability, on fitted parameters of the OU process and GRU, and on the new FLARE pipeline itself; no independent evidence is given for the 51 events beyond the framework output.

free parameters (2)
  • DRW parameters
    The Damped Random Walk model parameters are implicitly fitted or simulated to represent baseline behavior.
  • OU process parameters
    Iterative Ornstein-Uhlenbeck process is fitted directly to observed data with outlier masking.
axioms (1)
  • domain assumption Quasar light curves are typically well described by a Damped Random Walk (DRW).
    Stated explicitly in the abstract as the baseline behavior from which flares depart.
invented entities (1)
  • FLARE framework no independent evidence
    purpose: Modular three-stage detection pipeline combining physics-informed learning, anomaly scoring, and VLM recognition.
    Newly introduced system whose output produces the 51 detections.

pith-pipeline@v0.9.0 · 5602 in / 1567 out tokens · 75578 ms · 2026-05-10T17:30:17.267254+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. PDRS : A Linear $\mathcal{O}(N)$ Algorithm for Segmentation of High-Activity Regions in Irregularly Sampled Time Series

    astro-ph.IM 2026-05 unverdicted novelty 7.0

    PDRS is an O(N) algorithm for identifying high-activity regions in time series by seeding at local maxima and using gradient-aware search, achieving performance comparable to Bayesian Blocks.

Reference graph

Works this paper leans on

26 extracted references · 23 canonical work pages · cited by 1 Pith paper

  1. [1]

    author author M. J. \ Graham , author S. G. \ Djorgovski , author A. J. \ Drake , author D. Stern , author A. A. \ Mahabal , author E. Glikman , author S. Larson , \ and\ author E. Christensen ,\ 10.1093/mnras/stx1456 journal journal Monthly Notices of the Royal Astronomical Society \ volume 470 ,\ pages 4112 ( year 2017 ) NoStop

    work page doi:10.1093/mnras/stx1456 2017

  2. [2]

    L., Ivezi´ c,ˇZ., Kochanek, C

    author author C. MacLeod , author Z . Ivezi \'c , author C. S. \ Kochanek , author S. Koz owski , author B. Kelly , author E. Bullock , author A. Hope , author R. J. \ Assef , author R. Becker , author P. B. \ Hall , author P. Harding , author M. Juri \'c , author R. M. \ Kratzer , author M. R. \ Meade , author W. Ricketts , author B. Sesar , author D. P....

    work page doi:10.1088/0004-637x/721/2/1014 2010

  3. [3]

    H., & Saban, D

    author author C.-H. \ Chan , author T. Piran , author J. H. \ Krolik , \ and\ author D. Saban ,\ 10.3847/1538-4357/ab2b40 journal journal The Astrophysical Journal \ volume 881 ,\ pages 113 ( year 2019 ) NoStop

    work page doi:10.3847/1538-4357/ab2b40 2019

  4. [4]

    author author A. J. \ Drake et al. ,\ 10.1088/0004-637X/735/2/106 journal journal Astrophys. J. \ volume 735 ,\ pages 106 ( year 2011 ) ,\ http://arxiv.org/abs/1103.5514 arXiv:1103.5514 [astro-ph.CO] NoStop

    work page doi:10.1088/0004-637x/735/2/106 2011

  5. [5]

    Zheng , author Y

    author author Z. Zheng , author Y. Shi , author S. Jin , author H. Dannerbauer , author Q. Gu , author X. Li , \ and\ author X. Yu ,\ 10.1093/mnras/stae1036 journal journal Monthly Notices of the Royal Astronomical Society \ volume 530 ,\ pages 3527 ( year 2024 ) NoStop

    work page doi:10.1093/mnras/stae1036 2024

  6. [6]

    The Astrophysical Journal Supplement Series282(1), 13 (2025) https://doi.org/10.3847/1538-4365/ae1d64 8

    author author L. He , author Z.-Y. \ Liu , author R. Niu , author M.-S. \ Zhou , author P.-R. \ Zou , author B.-Z. \ Gao , author R.-D. \ Liang , author L.-G. \ Zhu , author J.-M. \ Wang , author N. Jiang , author Z.-Y. \ Cai , author J. an Jiang , author Z.-G. \ Dai , author Y.-F. \ Yuan , author Y.-J. \ Chen , \ and\ author W. Zhao ,\ 10.3847/1538-4365/...

    work page doi:10.3847/1538-4365/ae1d64 2025

  7. [7]

    author author S. A. J. \ McLaughlin , author J. R. \ Mullaney , \ and\ author S. P. \ Littlefair ,\ 10.1093/mnras/stae721 journal journal Monthly Notices of the Royal Astronomical Society \ volume 529 ,\ pages 2877 ( year 2024 ) NoStop

    work page doi:10.1093/mnras/stae721 2024

  8. [8]

    author author K. L. \ Suberlak , author Z . Ivezi \'c , \ and\ author C. L. \ MacLeod ,\ 10.3847/1538-4357/abc698 journal journal The Astrophysical Journal \ volume 907 ,\ pages 96 ( year 2021 ) NoStop

    work page doi:10.3847/1538-4357/abc698 2021

  9. [9]

    author author B. C. \ Kelly , author J. Bechtold , \ and\ author A. Siemiginowska ,\ 10.1088/0004-637X/698/1/895 journal journal The Astrophysical Journal \ volume 698 ,\ pages 895 ( year 2009 ) NoStop

    work page doi:10.1088/0004-637x/698/1/895 2009

  10. [10]

    author author B. C. \ Kelly , author A. C. \ Becker , author M. Sobolewska , author A. Siemiginowska , \ and\ author P. Uttley ,\ 10.1088/0004-637X/788/1/33 journal journal The Astrophysical Journal \ volume 788 ,\ pages 33 ( year 2014 ) NoStop

    work page doi:10.1088/0004-637x/788/1/33 2014

  11. [11]

    Yu , author G

    author author W. Yu , author G. T. \ Richards , author M. S. \ Vogeley , author J. Moreno , \ and\ author M. J. \ Graham ,\ 10.3847/1538-4357/ac8351 journal journal The Astrophysical Journal \ volume 936 ,\ pages 132 ( year 2022 ) NoStop

    work page doi:10.3847/1538-4357/ac8351 2022

  12. [12]

    Fagin , author J

    author author J. Fagin , author J. H.-H. \ Chan , author H. Best , author M. O'Dowd , author K. E. S. \ Ford , author M. J. \ Graham , author J. W. \ Park , \ and\ author V. A. \ Villar ,\ 10.3847/1538-4357/addabc journal journal The Astrophysical Journal \ ( year 2025 ),\ 10.3847/1538-4357/addabc NoStop

    work page doi:10.3847/1538-4357/addabc 2025

  13. [13]

    Tachibana , author M

    author author Y. Tachibana , author M. J. \ Graham , author N. Kawai , author S. G. \ Djorgovski , author A. J. \ Drake , author A. A. \ Mahabal , \ and\ author D. Stern ,\ 10.3847/1538-4357/abb9a9 journal journal The Astrophysical Journal \ volume 903 ,\ eid 54 ( year 2020 ) NoStop

    work page doi:10.3847/1538-4357/abb9a9 2020

  14. [14]

    AJ , volume =

    author author D. Foreman-Mackey , author E. Agol , author S. Ambikasaran , \ and\ author R. Angus ,\ 10.3847/1538-3881/aa9332 journal journal The Astronomical Journal \ volume 154 ,\ pages 220 ( year 2017 ) NoStop

    work page doi:10.3847/1538-3881/aa9332 2017

  15. [15]

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations

    author author M. Raissi , author P. Perdikaris , \ and\ author G. E. \ Karniadakis ,\ 10.1016/j.jcp.2018.10.045 journal journal Journal of Computational Physics \ volume 378 ,\ pages 686 ( year 2019 ) NoStop

    work page doi:10.1016/j.jcp.2018.10.045 2018

  16. [16]

    In Proceedings of the 26th Annual International Conference on Machine Learning (Montreal, Quebec, Canada) (ICML ’09)

    author author Y. Bengio , author J. Louradour , author R. Collobert , \ and\ author J. Weston ,\ in\ 10.1145/1553374.1553380 booktitle Proceedings of the 26th International Conference on Machine Learning \ ( year 2009 )\ pp.\ pages 41--48 NoStop

    work page doi:10.1145/1553374.1553380 2009

  17. [17]

    Koz owski ,\ 10.1051/0004-6361/201629890 journal journal Astronomy & Astrophysics \ volume 597 ,\ pages A128 ( year 2017 ) NoStop

    author author S. Koz owski ,\ 10.1051/0004-6361/201629890 journal journal Astronomy & Astrophysics \ volume 597 ,\ pages A128 ( year 2017 ) NoStop

    work page doi:10.1051/0004-6361/201629890 2017

  18. [18]

    Springer Series in Statistics

    author author S. Coles ,\ 10.1007/978-1-4471-3675-0 title An Introduction to Statistical Modeling of Extreme Values \ ( publisher Springer ,\ year 2001 ) NoStop

    work page doi:10.1007/978-1-4471-3675-0 2001

  19. [19]

    Dettmers , author A

    author author T. Dettmers , author A. Pagnoni , author A. Holtzman , \ and\ author L. Zettlemoyer ,\ in\ c booktitle Advances in Neural Information Processing Systems ,\ Vol. volume 36 \ ( year 2023 )\ pp.\ pages 10088--10115 NoStop

    2023

  20. [20]

    author author C. R. \ Harris et al. ,\ 10.1038/s41586-020-2649-2 journal journal Nature \ volume 585 ,\ pages 357 ( year 2020 ) NoStop

    work page doi:10.1038/s41586-020-2649-2 2020

  21. [21]

    Oliphant, Matt Haberland, Tyler Reddy, David Cournapeau, Evgeni Burovski, Pearu Peterson, Warren Weckesser, Jonathan Bright, Stéfan J

    author author P. Virtanen et al. ,\ 10.1038/s41592-019-0686-2 journal journal Nature Methods \ volume 17 ,\ pages 261 ( year 2020 ) NoStop

    work page doi:10.1038/s41592-019-0686-2 2020

  22. [22]

    Paszke , author S

    author author A. Paszke , author S. Gross , author F. Massa , author A. Lerer , author J. Bradbury , author G. Chanan , author T. Killeen , author Z. Lin , author N. Gimelshein , author L. Antiga , et al. ,\ in\ https://proceedings.neurips.cc/paper/2019/file/bdbca288fee7f92f2bfa9f7012727740-Paper.pdf booktitle Advances in Neural Information Processing Sys...

    2019

  23. [23]

    Yu \ and\ author G

    author author W. Yu \ and\ author G. T. \ Richards ,\ https://ascl.net/2201.001 title EzTao: Easier CARMA Modeling , \ howpublished Astrophysics Source Code Library, record ascl:2201.001 ( year 2022 ) NoStop

    2022

  24. [24]

    Transformers: State-of-the-Art Natural Language Processing , url =

    author author T. Wolf et al. ,\ in\ 10.18653/v1/2020.emnlp-demos.6 booktitle Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations \ ( year 2020 )\ pp.\ pages 38--45 NoStop

    work page doi:10.18653/v1/2020.emnlp-demos.6 2020

  25. [25]

    author author J. D. \ Hunter ,\ 10.1109/MCSE.2007.55 journal journal Computing in Science & Engineering \ volume 9 ,\ pages 90 ( year 2007 ) NoStop

    work page doi:10.1109/mcse.2007.55 2007

  26. [26]

    2010 , keywords =

    author author W. McKinney ,\ in\ 10.25080/Majora-92bf1922-00a booktitle Proceedings of the 9th Python in Science Conference \ ( year 2010 )\ pp.\ pages 56--61 NoStop

    work page doi:10.25080/majora-92bf1922-00a 2010