Searching for Anomalies in the ZTF Catalog of Periodic Variable Stars

Ho-Sang Chan, V. Ashley Villar, Siu-Hei Cheung, Shirley Ho, Anna J. G. O’Grady, M. R. Drout, Mathieu Renzo

This paper, published in The Astrophysical Journal in June 2022, represents a significant step forward in "discovery-oriented" astronomy. As wide-field surveys like the Zwicky Transient Facility (ZTF) and the upcoming Vera C. Rubin Observatory (LSST) flood us with data, the bottleneck is no longer finding stars, but finding the interesting ones.

TL;DR

Researchers developed an unsupervised machine-learning pipeline to scan 730,000 periodic variable stars and flag the most "anomalous" objects. By combining a Convolutional Variational Autoencoder (C-VAE) with an Isolation Forest, they identified rare, dusty red giants and young stellar objects that defy standard classification, providing a roadmap for targeted spectroscopic follow-up.

What Problem Does This Paper Address?

The universe is full of "periodic variables"—stars whose brightness changes in a repeating pattern due to pulsations or eclipses. While we have catalogs of millions of these stars, traditional classification focuses on the "average" member of a group (e.g., a typical RR Lyrae).

However, the most scientifically valuable objects are often the outliers—stars undergoing rare evolutionary transitions, interacting with hidden companions, or exhibiting physics we don't yet understand. Manually searching through nearly a million light curves for these "black swans" is impossible. The authors sought to build an automated system that doesn't just classify stars into known boxes, but asks: "Which of these looks nothing like the others?"

Methods

The authors built a multi-stage unsupervised pipeline:

  1. Preprocessing (MGPR): Astronomical data is notoriously "gappy" due to weather and daylight. The team used Multivariate Gaussian Process Regression to interpolate the raw data into a smooth, regular grid of 160 data points across two color bands ($g$ and $r$).
  2. Feature Learning (C-VAE): They used a Convolutional Variational Autoencoder. Unlike a standard autoencoder, a VAE learns a smooth, continuous "latent space" (a compressed 10-dimensional representation). By treating the light curves like a 1-dimensional image, the convolutional layers capture temporal correlations (shapes) in the brightness changes.
  3. Anomaly Detection (Isolation Forest): Once the stars were mapped into this 10D latent space, the team used an Isolation Forest algorithm. This identifies anomalies by seeing how "easy" it is to isolate a data point from the rest of the pack. They also tested a spherical coordinate transformation of the latent space to better separate clusters of star types.

Datasets

  • ZTF Catalog of Periodic Variable Stars (CPVS): The primary dataset, containing ~730,000 light curves.
  • Gaia DR2/EDR3: Used for "ground truth" physical parameters like absolute magnitude (brightness) and color to validate what the anomalies actually were.
  • WISE & Pan-STARRS: Used to construct Spectral Energy Distributions (SEDs) for the top anomalies to check for the presence of dust.

Key Insights and Ideas

  • The Latent Space is Physical: Even though the VAE was unsupervised (it wasn't told what the star types were), it naturally clustered similar stars together. Eclipsing binaries (EW) lived in one area, while pulsating giants (Mira) lived in another.
  • Anomalies are "Dusty": A major finding was that the most anomalous stars weren't just noisy data; they were physically distinct. Most were located in the Milky Way's galactic disk and showed signs of being shrouded in dust.
  • Phase Lagging: The pipeline flagged stars where the "peak" brightness in the $g$-band happened at a different time than the $r$-band. This "phase lag" is a hallmark of complex atmospheric physics in evolved stars.
  • Beyond Cartesian: The authors found that transforming the latent space into spherical coordinates helped the Isolation Forest find more meaningful outliers by accounting for the "annular" (ring-like) structures that naturally form in VAE latent spaces.

Key Results

  • Top 100 Anomalies: The paper provides a list of the most unusual periodic variables in the ZTF catalog.
  • Mira and SR Dominance: Semi-regular (SR) and Mira variables were found to be the most "anomalous" classes. These are red giants at the end of their lives, often losing mass and creating dusty envelopes.
  • Discovery of YSOs: The pipeline successfully identified several Young Stellar Objects (YSOs). These stars are still forming and are surrounded by accretion disks, causing irregular but somewhat periodic variability that often confuses standard classifiers.
  • Validation: By fitting the anomalies with the DESK and 2DUST codes, the authors confirmed that their top hits were indeed cool, luminous stars with significant mass-loss rates—exactly the kind of rare objects they hoped to find.

Limitations and Future Work

  • Unphysical Interpolation: The MGPR step occasionally creates "unphysical" light curve shapes if the data is too sparse, which the Isolation Forest might flag as a false-positive anomaly.
  • Bias Toward Complexity: The model tends to flag Mira and SR variables because their light curves are inherently more complex and irregular than simple eclipsing binaries.
  • The Need for Spectra: While the ML pipeline says that a star is weird, it can't always say why. The authors emphasize that the next step is "spectroscopic follow-up"—using large telescopes to take a chemical fingerprint of these 100 stars to see if they represent new stages of stellar evolution.

This work serves as a template for the "Big Data" era of astronomy, showing how we can use deep learning to keep the spirit of serendipitous discovery alive in an ocean of automated data.

Summary updated

Abstract

Abstract Periodic variables illuminate the physical processes of stars throughout their lifetime. Wide-field surveys continue to increase our discovery rates of periodic variable stars. Automated approaches are essential to identify interesting periodic variable stars for multiwavelength and spectroscopic follow-up. Here we present a novel unsupervised machine-learning approach to hunt for anomalous periodic variables using phase-folded light curves presented in the Zwicky Transient Facility Catalogue of Periodic Variable Stars by Chen et al. We use a convolutional variational autoencoder to learn a low-dimensional latent representation, and we search for anomalies within this latent dimension via an isolation forest. We identify anomalies with irregular variability. Most of the top anomalies are likely highly variable red giants or asymptotic giant branch stars concentrated in the Milky Way galactic disk; a fraction of the identified anomalies are more consistent with young stellar objects. Detailed spectroscopic follow-up observations are encouraged to reveal the nature of these anomalies.
BibTeX
@article{chan2022searching,
  author = {Ho-Sang Chan and V. Ashley Villar and Siu-Hei Cheung and Shirley Ho and Anna J. G. O’Grady and M. R. Drout and Mathieu Renzo},
  title = {{Searching for Anomalies in the ZTF Catalog of Periodic Variable Stars}},
  journal = {The Astrophysical Journal},
  year = {2022},
  doi = {10.3847/1538-4357/ac69d4},
}
Figure