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

TL;DR

This paper introduces FLARE, a modular framework combining statistical models and anomaly detection to identify transient flares in quasar light curves from SDSS Stripe 82, revealing 51 flaring quasars.

Contribution

The paper presents a novel, systematic approach for flare detection in quasar light curves using physics-informed models and anomaly scoring, applied to a large legacy dataset.

Findings

Identified 51 quasars with flaring activity in SDSS Stripe 82.

Developed a three-stage detection framework combining DRW modeling, anomaly scoring, and recognition.

Benchmarking of vision-language models for flare verification.

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.