🔭 astrophysics

Anomaly detection in Fink. I. Discovery, follow-up, and classification of unusual sources

This paper presents the first-year performance of the Fink broker's anomaly detection pipeline, which successfully identified and facilitated the discovery and follow-up of rare astrophysical phenomena—including a WZ Sge-type AM CVn system, UX Ori-type stars, and numerous supernovae and dwarf novae—by combining automated Isolation Forest ranking with expert validation and an active feedback loop.

Original authors: M. V. Pruzhinskaya, M. V. Kornilov, A. V. Dodin, A. Baluta, T. A. Pshenichniy, A. M. Zubareva, E. E. O. Ishida, J. Peloton, I. Beschastnov, I. Ippolitov, A. A. Belinski, P. Golysheva, N. P. Ikonnikova

Published 2026-04-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: M. V. Pruzhinskaya, M. V. Kornilov, A. V. Dodin, A. Baluta, T. A. Pshenichniy, A. M. Zubareva, E. E. O. Ishida, J. Peloton, I. Beschastnov, I. Ippolitov, A. A. Belinski, P. Golysheva, N. P. Ikonnikova, V. A. Kiryukhina, V. V. Krushinsky, A. M. Tatarnikov, S. G. Zheltoukhov, D. A. Buckley, A. Kniazev, S. V. Karpov, A. Möller, Y. Tampo

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the night sky as a massive, bustling city that never sleeps. Every night, a giant, automated camera (the Zwicky Transient Facility, or ZTF) takes millions of photos of this city, looking for anything that moves, flickers, or changes.

The problem? The camera takes so many pictures that it generates a flood of data—like a firehose spraying millions of water droplets every night. Most of these droplets are just normal rain (common stars), but hidden inside are rare, magical gems: exploding stars, strange cosmic collisions, or alien-like phenomena. Finding these gems by looking at every single droplet is impossible for humans.

This is where the paper comes in. It describes a new "smart filter" called Fink, and specifically, a special module within it designed to find the "weird stuff."

Here is the story of how they do it, explained simply:

1. The "Smart Bouncer" (The Anomaly Detection Pipeline)

Think of the Fink system as a nightclub bouncer at the entrance of the cosmic city.

The Job: Every night, the bouncer has to sort through millions of "guests" (astronomical alerts).
The Trick: Instead of checking every guest's ID, the bouncer uses a smart algorithm (an Isolation Forest model) trained on years of past data. This algorithm knows what a "normal" star looks like.
The Selection: If a guest looks slightly off—maybe they are wearing a neon suit when everyone else is in black, or they are dancing to a rhythm no one else is—the bouncer flags them.
The Shortlist: The bouncer picks the top 10 weirdest guests of the night and immediately sends a text message (via Slack and Telegram) to a team of expert astronomers.

2. The "Detective Squad" (Expert Follow-up)

Once the experts get the text, they become detectives. They don't just look at the photo; they grab their telescopes (the "magnifying glasses") to investigate the top 10 suspects.

The Goal: To see if these "weird" lights are just glitches in the camera or something truly new and exciting.

3. The Big Discoveries (The "Gems" Found)

In the first year of this system running, the team found some incredible things:

The Cosmic Vampire (AM CVn Star): They found a rare type of star system where one star is stealing material from its partner. It's like a cosmic vampire. This specific one, named Fink J062452, was behaving strangely, having two massive "burps" of energy in a row. It's so rare that only a handful like it are known in the whole universe.
The Star with a "Pre-Show" (SN 2023mtp): They found a supernova (an exploding star) that did something unexpected. Before the main explosion, it had a massive, bright "pre-show" that was brighter than almost any other known explosion. It's like a firework that glows brightly for weeks before the big bang, and even the experts couldn't quite figure out exactly what kind of firework it was.
The Chameleon Star (UX Ori-type): They found a young star that changes color and brightness dramatically. It's like a chameleon that hides behind a curtain of dust. When the dust moves, the star looks different. This helps astronomers understand how stars are born and how they are surrounded by dusty disks.
The Flaring M-Dwarf: They found a small, cool star that throws tantrums, shooting out massive flares of energy. This helps us understand how magnetic storms work on stars like our Sun, but much more intense.
The "Ghost" Supernovae: They found 30 new exploding stars that no one had reported before. Some were so bright they might be "Super-Luminous" (the cosmic equivalent of a supernova going to a party and outshining everyone). Some were "hostless," meaning they exploded in the empty space between galaxies, far away from any home star system.

4. The "Feedback Loop" (Learning from Mistakes)

The system isn't perfect. Sometimes, the bouncer flags a "normal" star that just happens to be moving fast (a high proper motion star) or a glitch in the camera.

The Solution: The team built a Telegram bot where experts can give a "Thumbs Up" (this is weird!) or a "Thumbs Down" (this is normal).
The Result: The computer learns from these thumbs. If the experts say, "Hey, eclipsing binary stars are actually normal, stop flagging them," the computer updates its rules. It's like teaching a dog new tricks; the more you correct it, the better it gets at finding the real weird stuff.

The Bottom Line

This paper shows that while computers are great at sorting through mountains of data, they still need human intuition to find the truly special things. By combining a smart computer filter with human experts who can look at the data and say, "Wait, that's interesting!", astronomers are turning a flood of data into a stream of new discoveries.

It's like having a robot that scans a million books a day, but it needs a human librarian to pull out the one book that contains a secret map to a new world.

1. Problem Statement

Modern wide-field time-domain surveys, such as the Zwicky Transient Facility (ZTF), generate massive alert streams containing hundreds of thousands of candidates nightly. While automated pipelines are effective at filtering for known classes of transients (e.g., standard supernovae), they often struggle to identify rare, unusual, or previously unknown astrophysical phenomena ("edge cases"). The scientific potential of these surveys lies heavily in these rare events, but manual inspection of every alert is impossible. The challenge is to develop an automated system that can efficiently rank anomalies, distribute them to human experts for rapid validation, and facilitate follow-up observations before the transient fades.

2. Methodology

The authors present the first-year performance of an Anomaly Detection (AD) pipeline integrated into the Fink broker, a community-driven system for processing ZTF alerts. The methodology consists of four main components:

Feature Extraction (AD Features Module):
- Light curves from $g$ and $r$ bands are transformed into a compact set of 26 numerical features (e.g., amplitude, mean, standard deviation, Stetson K coefficient) using the light-curve package.
- Features are derived from either difference magnitudes (for transients) or apparent magnitudes (for variable stars), depending on the presence of a reference source.
Anomaly Detection Model:
- The core algorithm is an Isolation Forest (IF) model, trained on archival ZTF data (841,200 alerts from April–July 2025).
- Two separate models are employed (one per passband) to handle cases with sparse data.
- The final anomaly score for an object is the minimum of the scores from both passbands. Lower scores indicate higher anomaly.
Notification and Dissemination:
- The pipeline ranks all alerts nightly and selects the top-10 most anomalous candidates.
- These are distributed to experts via Slack and Telegram bots and exposed via a REST API. Notifications include alert IDs, coordinates, light curves, image cutouts, and the anomaly score.
Active Anomaly Discovery (AAD) & Citizen Science:
- To refine the model, the authors implemented a feedback loop using a public Telegram bot.
- Experts label candidates as "anomalous" (thumbs up) or "nominal" (thumbs down).
- The Active Anomaly Discovery (AAD) algorithm re-trains the Isolation Forest using this feedback to iteratively adjust decision boundaries and reduce false positives (e.g., eclipsing binaries).

3. Key Contributions

Operational AD Pipeline: The first year of operation (starting Jan 25, 2023) of a fully functional anomaly detection system within a major astronomical broker.
Hybrid Human-AI Workflow: A demonstrated framework coupling machine learning ranking with rapid expert validation and automated dissemination channels (Slack/Telegram).
Active Learning Implementation: The integration of expert feedback via the AAD algorithm to dynamically improve the model's ability to distinguish between true anomalies and known "nominal" classes.
Discovery of Rare Objects: The pipeline successfully identified and triggered follow-up for multiple rare astrophysical systems that would likely have been missed by standard classification filters.

4. Key Results

During the first year of operations, the pipeline identified 33 supernova candidates (30 previously unreported) and 9 new dwarf novae. Notable individual discoveries include:

Fink J062452.88+020818.3 (AM CVn System): A rare WZ Sge-type helium dwarf nova. The system exhibited a complex double superoutburst structure. Follow-up spectroscopy (SALT, 2.5-m CMO) confirmed it as the third reliably classified object of this type, characterized by helium-dominated spectra and superhumps.
SN 2023mtp (Supernova with Precursor): A transient showing a luminous precursor activity ( $M \approx -16.3$ ) followed by a main explosion. Spectroscopy revealed features inconsistent with a single standard supernova type (showing characteristics of Type IIn, IIP, and IIb), suggesting unusual progenitor physics or interaction with a dense circumstellar medium.
Fink J222324.32+744222.0 (UX Ori-type Star): A G2-type star exhibiting sharp brightness increases followed by a 4-year plateau. Multi-wavelength analysis showed an anticorrelation between optical and infrared brightness, indicative of variable circumstellar dust obscuration.
Fink J042203.10+362318.7 (Flaring M-Dwarf): A magnetically active M-dwarf exhibiting recurrent short-duration flares, demonstrating the pipeline's ability to detect stellar activity.
Supernova Statistics:
- Identified 7 candidates for Superluminous Supernovae (SLSNe).
- Found 2 candidates for "Pure" SNe Ia (exploding in elliptical galaxies/halos), which are valuable for cosmology due to low dust extinction.
- Detected 6 hostless transients, potentially originating from ejected progenitors or intergalactic star formation.
Model Refinement: The citizen science feedback loop successfully lowered the anomaly ranks of known "nominal" classes (specifically eclipsing binaries), demonstrating the efficacy of the AAD algorithm in reducing false positives.

5. Significance

This work establishes that broker-level anomaly detection, when coupled with rapid dissemination and expert assessment, serves as an effective bridge between large-scale survey data and domain expertise.

Scientific Impact: It transforms raw "anomaly scores" into concrete astrophysical discoveries, enabling the study of rare phenomena (like AM CVn systems and unusual SNe) that are critical for understanding stellar evolution and explosion mechanisms.
Methodological Impact: It highlights that automated classification alone is insufficient for edge cases; a "human-in-the-loop" approach is mandatory for adding value to time-domain astronomy.
Future Direction: The paper argues for the development of personalized anomaly detection models tailored to specific scientific interests, acknowledging that the definition of an "anomaly" is inherently subjective and expert-dependent.

In conclusion, the Fink AD pipeline proves that combining unsupervised machine learning with active expert feedback creates a powerful discovery engine for the next generation of time-domain surveys.