FLARE: Physics-Informed Quasar Flare Detection

• FLARE is a generalized framework that models quasar variability using a physics-informed GRU to establish a DRW baseline for light curves.
• It employs Extreme Value Theory to statistically identify rare, significant deviations as candidate flares with a calibrated threshold of 8.69σ.
• A recognition engine using vision language models further validates candidates by differentiating true astrophysical flares from instrumental artifacts via cross-band checks.

FLARE, short for “Flare detection via physics-informed Learning, Anomaly scoring, and Recognition Engine,” is a generalized three-stage statistical-AI framework for detecting rare flaring events in stochastic quasar light curves. It was introduced to address a central time-domain inference problem: quasars exhibit persistent intrinsic variability that is typically well described by a Damped Random Walk (DRW), yet astrophysically interesting flares appear as rare, extreme departures from that baseline. In its SDSS Stripe 82 implementation, FLARE combines physics-informed probabilistic sequence modeling, Extreme Value Theory (EVT), and Vision LLM (VLM) verification to separate genuine multi-epoch flares from stochastic variability and instrumental artifacts, yielding a final sample of 27 flaring quasars among 9,258 confirmed objects (Agrawal, 9 Apr 2026).

1. Definition and scientific setting

FLARE was developed for the systematic detection of quasar flares, defined here as extreme luminosity changes that depart significantly from the stochastic DRW baseline of quasar variability. The scientific motivation is that such departures can constrain accretion disc dynamics and the physics of supermassive black hole fueling (Agrawal, 9 Apr 2026).

The framework is explicitly modular. Its three stages are:

1. Baseline Modeling: learning the usual DRW variability for each quasar.
2. Anomaly Scoring: identifying statistically significant deviations.
3. Recognition Engine: morphologically verifying candidates to filter true flares from artifacts (Agrawal, 9 Apr 2026).

This modularity is central to the framework’s design. FLARE is not a single classifier or a single thresholding rule; rather, it is a pipeline in which the statistical model of normal variability, the rare-event detector, and the final recognizer can be improved independently. For SDSS Stripe 82, these roles are instantiated by a physics-informed probabilistic GRU, EVT, and VLMs, respectively (Agrawal, 9 Apr 2026).

A notable methodological feature is the rejection of ad hoc significance cuts in favor of a calibrated rare-event model. That choice aligns with broader evidence from flare-statistics research that arbitrary thresholds can systematically bias inferred flare distributions, as shown for GOES and LYRA solar-flare catalogs (Ryan et al., 2016).

2. Stripe 82 implementation and baseline DRW modeling

The Stripe 82 application uses 9,258 spectroscopically confirmed quasars, each with approximately 60–80 epochs over approximately 10 years. Detection is performed on $r$-band light curves, and surviving candidates are then checked against $g$-band data to rule out instrumental artifacts (Agrawal, 9 Apr 2026).

The baseline model is a physics-informed probabilistic Gated Recurrent Unit (GRU) trained to represent expected DRW variability. For each epoch, the model ingests the mean-centered magnitude $m_i$ and the time step

$\Delta t_i = t_i - t_{i-1},$

and predicts both a predictive mean $\mu_i$ and an uncertainty $\sigma_i$ (Agrawal, 9 Apr 2026).

The physical prior is the Ornstein–Uhlenbeck/DRW kernel

$$k(\Delta t) = \hat{\sigma}^2 \exp\left(-\frac{\Delta t}{\tau}\right),$$

with damping timescale $\tau$ and variability amplitude $\hat{\sigma}$. Under the DRW model, the conditional mean and variance are

$$\mu_{\mathrm{OU},i} = \bar{m} + (m_{i-1} - \bar{m}) e^{-\Delta t_i/\tau},$$

$g$0

Training combines a negative log-likelihood term with drift and variance regularizers that penalize deviations from these DRW expectations, and the regularization weights are linearly annealed during training (Agrawal, 9 Apr 2026).

The training set is not drawn from observed flares. Instead, the GRU is trained on large sets of simulated DRW light curves constructed from each object’s measured DRW parameters, cadence, and photometric error. This design makes the first stage a model of normal stochastic quasar behavior, not a direct flare classifier (Agrawal, 9 Apr 2026).

3. Extreme-value anomaly scoring

After baseline prediction, FLARE converts each observed point into a standardized residual,

$g$1

and then calibrates these residuals using the global mean $g$2 and standard deviation $g$3 from validation data,

$g$4

For each light curve, the relevant summary statistic is the maximum absolute calibrated residual, i.e. the largest event observed in that object (Agrawal, 9 Apr 2026).

The key empirical finding is that the distribution of these maxima across the simulated DRW population is heavy-tailed, with large kurtosis. Consequently, classical sigma-clipping is treated as inadequate. FLARE instead uses an EVT Peaks-over-Threshold (POT) construction. A high percentile $g$5 of the reference distribution—here the 95th percentile—is chosen, and the excesses above $g$6 are modeled with a Generalized Pareto Distribution (GPD):

$g$7

where $g$8 is the shape parameter and $g$9 the scale parameter (Agrawal, 9 Apr 2026).

For a target false alarm probability,

$m_i$0

with $m_i$1 the total number of objects and $m_i$2 the number above $m_i$3. In the Stripe 82 analysis, using $m_i$4 produced a threshold of $m_i$5. Only objects with

$m_i$6

were flagged as flare candidates (Agrawal, 9 Apr 2026).

This thresholding strategy is significant because it is explicitly tied to the tail behavior of the null population rather than to an arbitrary fixed-$m_i$7 convention. A plausible implication is that the framework is better aligned with the actual rare-event geometry of DRW maxima than standard outlier rules.

4. Recognition engine and artifact rejection

The third stage addresses a different failure mode: a statistically extreme point sequence need not be a genuine astrophysical flare. It may instead be a cosmic-ray spike, a photometric error, or another non-flare artifact. FLARE therefore introduces a recognition engine based on Vision LLMs (Agrawal, 9 Apr 2026).

Each candidate light curve is rendered as an unlabeled image, specifically to avoid bias from axis values or timescales. Several VLMs are benchmarked on simulated data containing DRW-only light curves, flares with several morphologies—FRED, Gaussian, and Gamma—and single-point spikes. Operationally, the recognition task collapses five morphology classes into a binary flare/non-flare decision (Agrawal, 9 Apr 2026).

The recognition engine uses an ensemble structure:

• Two VLMs serve as primary classifiers, one optimized for high recall and one for high precision.
• A third VLM acts as evaluator, resolving disagreements, flagging misclassifications, and providing feedback.
• The process is repeated for 5 cycles (Agrawal, 9 Apr 2026).

The Stripe 82 implementation reports model choices such as Grok-4.1-fast for high recall, Qwen-3.5-plus for high precision, and GPT-5 as evaluator (Agrawal, 9 Apr 2026).

VLM verification is not the final safeguard. Surviving $m_i$8-band candidates are cross-checked against $m_i$9-band light curves, and candidates without corresponding events in $\Delta t_i = t_i - t_{i-1},$0 are rejected as likely instrumental. The accepted objects are therefore required to display correlated multi-epoch variations in both bands (Agrawal, 9 Apr 2026).

5. Empirical results in SDSS Stripe 82

Applied to Stripe 82, FLARE identified a rare but non-negligible population of flaring quasars (Agrawal, 9 Apr 2026).

Pipeline stage	Output	Criterion or note
EVT candidate selection	51 candidates	Maximum calibrated residual above $\Delta t_i = t_i - t_{i-1},$1
VLM recognition engine	30 candidates	Classified as genuine flares
Cross-band vetting	27 final flares	3 rejected for lacking $\Delta t_i = t_i - t_{i-1},$2-band flare

The final sample consists of 27 astrophysical flares among 9,258 confirmed quasars (Agrawal, 9 Apr 2026). The study also reports a systematic flare rate of $\Delta t_i = t_i - t_{i-1},$3 in Stripe 82, stated to be higher than rates found by previous CRTS- and ZTF-based surveys, likely because of Stripe 82’s longer temporal baseline and the tailored, data-driven EVT threshold rather than ad hoc sigma cuts (Agrawal, 9 Apr 2026).

The VLM stage shows a marked difference between fine-grained and binary performance. On simulated five-class recognition, the best VLM accuracy is 42.8%, which is above the 20% random baseline. However, for binary flare/non-flare detection, the best models achieve approximately 88% precision and approximately 70% recall (Agrawal, 9 Apr 2026). This suggests that morphology subclassification remains difficult, whereas binary rejection of obvious non-flares is materially stronger.

The paper further notes that parameter-efficient fine-tuning of open-weight VLMs improved recall, but that bias toward certain flare profiles remained (Agrawal, 9 Apr 2026).

6. Interpretation, scope, and limitations

FLARE’s principal contribution is methodological: it decomposes flare detection into physically and statistically distinct subproblems. The baseline model handles stochastic quasar variability; EVT calibrates rarity under the null; the recognition engine checks morphological plausibility; and cross-band agreement supplies a final astrophysical consistency test (Agrawal, 9 Apr 2026).

This structure matters because quasar flare searches are vulnerable to at least three distinct confounders: intrinsic DRW variability, heavy-tailed extremes under the null, and non-astrophysical image- or photometry-level artifacts. FLARE addresses each confounder with a different mechanism rather than attempting to collapse them into a single discriminant.

The framework is also explicitly described as generalized and modular, with intended applicability to ongoing and future time-domain surveys such as ZTF and LSST, and robustness to differences in cadence, noise, and DRW parameter regime (Agrawal, 9 Apr 2026). By automating candidate verification, it is presented as a way to mitigate the bottleneck of manual inspection in large survey pipelines (Agrawal, 9 Apr 2026).

Several limitations are equally clear. First, the recognition engine is substantially better at binary filtering than at five-way morphological recognition. Second, the remaining profile bias in VLMs indicates that morphological priors are not yet neutral across flare families. Third, the need for $\Delta t_i = t_i - t_{i-1},$4-band corroboration shows that purely single-band inference is not considered sufficient in this application (Agrawal, 9 Apr 2026).

A common misconception would be to treat FLARE as a purely deep-learning system. In fact, its distinctive feature is the coupling of physics-informed recurrent modeling with extreme-value statistics and multimodal recognition. Another misconception would be to view the $\Delta t_i = t_i - t_{i-1},$5 threshold as a generic constant; it is instead the outcome of a survey-specific EVT calibration at 1% false alarm probability on the simulated DRW maximum-residual distribution (Agrawal, 9 Apr 2026).

7. Scientific role and future prospects

The Stripe 82 deployment demonstrates that legacy survey data can support a systematic flare census when rare-event statistics and automated verification are explicitly integrated. The identification of 27 validated flaring quasars provides a basis for follow-up work on physical origin, event rates, and environmental dependence, with the stated goal of probing accretion phenomena, black hole fueling, and extreme quasar variability (Agrawal, 9 Apr 2026).

The framework’s modularity also defines its likely research trajectory. The GRU baseline could be replaced or refined, the EVT component could be re-estimated for other survey cadences, and the recognition engine could improve as VLMs mature. The paper explicitly identifies room for improvement in the recognition stage and anticipates gains from further physics-informed fine-tuning (Agrawal, 9 Apr 2026).

Within quasar time-domain analysis, FLARE therefore occupies a specific methodological niche: a survey-scale pipeline for detecting events that are simultaneously rare under a DRW null, morphologically flare-like, and cross-band consistent. Its Stripe 82 results establish that such events can be isolated in a statistically controlled and largely automated way from a background dominated by ever-present stochastic variability (Agrawal, 9 Apr 2026).