🔭 astrophysics

Anomaly Hunter for Alerts (AHA): Anomaly Detection in the ZTF Transient Alert Stream

This paper presents "Anomaly Hunter for Alerts" (AHA), an unsupervised pipeline utilizing three distinct autoencoders to effectively identify exotic transients and anomalous supernovae within the high-volume ZTF alert stream, demonstrating its efficiency and low data requirements for future application to the Rubin Observatory.

Original authors: Leyla Iskandarli, Chris J. Lintott, Steve Croft, Heloise Stevance, Joshua Weston

Published 2026-02-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Leyla Iskandarli, Chris J. Lintott, Steve Croft, Heloise Stevance, Joshua Weston

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 you are the head librarian of the world's biggest library. Every single night, a conveyor belt dumps one million new books onto your desk. Most of these books are boring, repetitive manuals about "How to Build a Standard Brick Wall." But hidden inside that mountain of manuals are a few rare, magical storybooks about dragons, time travel, or alien civilizations.

Your job is to find those magical books. But you can't read a million books a night. You need a robot assistant to do the heavy lifting.

This paper introduces that robot assistant, called AHA (Anomaly Hunter for Alerts). Here is how it works, explained simply:

The Problem: Too Much Noise, Too Little Signal

The Zwicky Transient Facility (ZTF) is a giant telescope that scans the sky every night. It acts like a security camera for the universe, snapping photos and sending out "alerts" whenever something changes (like a star exploding).

The Challenge: Most of these changes are just normal supernovae (exploding stars). They are the "brick wall manuals."
The Goal: We want to find the "dragons." These are rare, weird events like black holes eating stars, stars that shouldn't exist, or signals from alien technology.

The Solution: Three Different Sets of Glasses

Instead of building one giant, complicated robot brain that tries to look at everything at once, the authors built three simpler robots. Each robot looks at the incoming data through a different pair of glasses. If you try to combine all the data into one giant pile, you might lose the details. By looking at them separately, they catch different things.

Here are the three "glasses" (or data types) they used:

1. The "ID Card" Glasses (Object Features)

What it sees: This robot looks at the "stats" of the object. Is it bright? Is it getting brighter or darker? How far away is it from its home galaxy?
The Analogy: Imagine a bouncer at a club checking IDs. If a guest says they are 25 but looks 50, or if they are standing 10 miles away from the VIP section they claim to belong to, the bouncer flags them.
What it found: This robot is great at spotting objects that are in the wrong place (like a star exploding far outside its galaxy) or changing brightness in weird, rapid ways.

2. The "Photo Album" Glasses (Image Cutouts)

What it sees: This robot looks at the actual pictures. It sees three snapshots: the sky before the event, the sky during the event, and the "difference" image (what's new).
The Analogy: This is like a detective looking at a crime scene photo. If the photo is blurry, or if the "new" object is floating in a dark, empty void with no house nearby, the detective gets suspicious.
What it found: This robot is excellent at finding explosions that happen in the middle of nowhere, far from any visible galaxy, or in very faint, fuzzy galaxies that are hard to see.

3. The "Storyline" Glasses (Light Curves)

What it sees: This robot looks at the "story" of the object over time. It draws a graph of how the object's brightness changed from day 1 to day 300.
The Analogy: Imagine a normal supernova is like a firework: it shoots up, explodes, and fades away quickly. This robot is looking for fireworks that keep glowing for years, or fireworks that flicker on and off like a broken lightbulb.
What it found: This robot caught objects that refused to fade, or had strange "double peaks" in their brightness, suggesting they aren't normal stars at all.

How They Learned (The Training)

The robots were trained using a "school of normal students." They were shown thousands of examples of normal exploding stars (Supernovae).

The Lesson: "This is what a normal star looks like. If you see something that doesn't fit this pattern, raise your hand."
The Result: The robots learned to ignore the boring "brick wall manuals" and only flag the weird stuff.

The Results: Catching the Dragons

The team tested these robots on a live feed of the sky for 25 days.

The Magic: Because the three robots looked at the data differently, they didn't all flag the same things. They were like three different detectives solving a mystery; one found a clue the others missed.
The Success: They found 87 very strange candidates that needed human experts to look at them closely.
- Some were likely Cataclysmic Variables (dying stars stealing food from neighbors).
- Some were Active Galactic Nuclei (supermassive black holes eating matter).
- Some were Orphan Transients (explosions with no home galaxy).
- Some might be Pair-Instability Supernovae (stars so massive they blow themselves apart completely).

Why This Matters for the Future

The next big telescope, the Rubin Observatory, will be like a firehose compared to ZTF's garden hose. It will send out millions more alerts every night. Humans will never be able to read them all.

This paper proves that you don't need a super-complex AI to find the needles in the haystack. You just need a few simple, specialized tools working together. This "AHA" system is ready to be upgraded and used on the Rubin telescope to find the most bizarre, exciting, and potentially universe-changing events in the cosmos.

In short: They built a team of three simple robots to scan the sky. Each robot looks at the data in a different way, and together, they found 87 weird cosmic mysteries that humans would have missed in the noise.

1. Problem Statement

Modern time-domain astronomical surveys, such as the Zwicky Transient Facility (ZTF) and the upcoming Vera C. Rubin Observatory (LSST), generate alert streams at rates ( $\sim 10^6$ per night for ZTF) that exceed human vetting capacity. Traditional rule-based filtering often misses rare, exotic, or previously unseen phenomena because they do not fit predefined categories.

The Challenge: Identifying "anomalous" transients (e.g., TDEs, SLSNe, AGN, or misclassified supernovae) within a sea of "normal" supernovae (SN Ia, SN II, SN Ib/c) without relying on labeled training data for every possible anomaly class.
The Constraint: Alert packets are often incomplete; some lack image cutouts, others lack light curves, and some lack derived features. A single unified model often discards alerts with missing modalities, reducing discovery potential.

2. Methodology

The authors propose AHA (Anomaly Hunter for Alerts), an unsupervised anomaly detection pipeline implemented within the Lasair alert broker. The core innovation is an ensemble approach using three independent, simple autoencoders, each trained on a distinct data modality. This ensures maximum coverage even when specific data products are missing.

A. Data Preparation

Training Set: Constructed from 5,294 spectroscopically classified ZTF objects.
- Normal Class (Majority): SN Ia, SN II, SN Ib/c.
- Anomalous Class (Minority/Exotic): AGN, TDEs, SLSNe, CVs, nuclear transients.
- Imbalance Handling: The minority normal classes (SN II, SN Ib/c) were upsampled using SMOTE (Synthetic Minority Over-sampling Technique) to match the majority class (SN Ia), ensuring the model learns the full spectrum of "normal" supernova behavior.
Data Modalities:
1. Object Features: 22 statistical descriptors (magnitudes, color, host galaxy separation, redshift, photometric rates).
2. Triplet Image Cutouts: 3-channel (Science, Template, Difference) $64 \times 64$ pixel FITS stamps.
3. Light Curves: Resampled, fixed-grid (176 time steps) light curves in $g$ and $r$ bands, aligned to peak brightness.

B. Algorithm Architecture

Three distinct autoencoders were trained to reconstruct "normal" data. High reconstruction loss indicates an anomaly.

Feature Autoencoder: A fully connected symmetric network (Encoder: 88-56-27 $\to$ 8 latent dim; Decoder mirrors). Uses Huber loss to handle extreme values robustly.
Image Autoencoder: A convolutional autoencoder (CNN) with stride-2 downsampling and transposed convolutions for upsampling. Uses Mean Squared Error (MSE) loss.
Light Curve Autoencoder: A fully connected network (Encoder: 256-128-64 $\to$ 8 latent dim). Uses weighted MSE loss, where weights are based on the effective support (number of observations) at each time step to prevent sparse data from artificially inflating anomaly scores.

C. Evaluation Metrics

Exotic Recall: The fraction of known exotic objects (AGN, TDE, SLSN, etc.) flagged as anomalous.
Purity: The fraction of flagged objects that are genuinely interesting (either exotic or supernovae with anomalous properties, e.g., wrong host, weird light curve).
Thresholding: Anomaly thresholds were defined based on the reconstruction loss distribution of the top-ranked candidates in a held-out test set.

3. Key Contributions

Modality-Specific Ensemble: Demonstrated that treating object features, images, and light curves independently yields better coverage and distinct anomaly sets compared to a single combined model.
Definition of Anomaly: Expanded the definition of "anomaly" beyond just exotic classes to include supernovae with anomalous properties (e.g., incorrect host associations, missing hosts, or non-standard light curves).
Data Efficiency: Proved that a curated training set of only a few thousand objects is sufficient to train effective anomaly detectors for live streams.
Live Stream Application: Successfully deployed the pipeline on the live ZTF alert stream for 25 days, moving beyond static dataset testing.

4. Results

A. Performance on Held-Out Test Set

Object Features:
- Recall: 2.3% (Top 10).
- Purity: 70%.
- Behavior: Best at flagging objects with incorrect/missing host associations and extreme photometric rates.
Triplet Images:
- Recall: 10% (Top 10).
- Purity: 40%.
- Behavior: Best at identifying sources with faint, unresolved, or undetected host galaxies (often associated with SLSNe or TDEs).
Light Curves:
- Recall: 8.3% (Top 10).
- Purity: 60%.
- Behavior: Best at identifying objects with non-standard temporal evolution (e.g., multi-year rises, double peaks, or non-declining light curves).

B. Live Alert Stream Application (25 Days)

The pipeline processed alerts from Dec 1–25, 2025, identifying 87 Rank 3 candidates (highest priority) for follow-up.

Total Flagged: 206 (Features), 127 (Images), 359 (Light Curves).
Overlap: The overlap between algorithms is minimal.
- Max overlap between any two algorithms: 11 objects.
- Zero overlap for supernova-labeled objects across all three algorithms.
- Only one object (an AGN) was flagged by all three.
Key Findings:
- The Light Curve Autoencoder achieved the highest purity (97%) in the live stream, identifying 121 interesting supernova-labeled objects.
- The Object Features Autoencoder was highly effective at finding AGN and nuclear transients driven by extreme photometric rates.
- The Image Autoencoder successfully isolated candidates with faint hosts, a signature of rare luminous transients.

C. Notable Candidates

The pipeline identified several high-priority candidates, including:

Long-lived transients: Objects with 600–800 day light curves inconsistent with standard supernova evolution.
Host-less candidates: Orphan transients or those with large offsets from host galaxies.
Misclassified objects: Sources labeled as supernovae by automated classifiers but exhibiting AGN or CV-like behavior.

5. Significance and Future Outlook

Scalability for Rubin Observatory: The framework is designed to be data-efficient and modular, making it ideal for the Vera C. Rubin Observatory, which will produce alert rates an order of magnitude higher than ZTF. The independent modality approach ensures no alerts are discarded due to missing data products.
Discovery Space: By focusing on reconstruction error rather than specific class labels, the system can detect novel classes of transients that have never been seen before, not just known exotic types.
Technosignatures: The authors note that this unsupervised approach is naturally suited for searching for technosignatures (unusual variability patterns) that do not fit standard astrophysical models.
Conclusion: The AHA pipeline demonstrates that unsupervised autoencoders, applied to specific data modalities in an ensemble, provide a robust, scalable, and highly effective method for triaging massive transient alert streams, significantly enhancing the discovery potential of time-domain astronomy.