Searching for Anomalies in the ZTF Catalog of Periodic Variable Stars

H. S. Chan, V. Ashley Villar, S. H. Cheung, Shirley Ho, Anna J. G. O’Grady, M. R. Drout, Mathieu Renzo

TL;DR

Researchers have developed an unsupervised machine learning pipeline to find "needles in the cosmic haystack"—anomalous stars hidden within the massive Zwicky Transient Facility (ZTF) catalog. By combining a Convolutional Variational Autoencoder (C-VAE) with an Isolation Forest, the team identified 100 high-priority anomalies, most of which are rare, dusty evolved stars or Young Stellar Objects (YSOs) that defy standard classification.

What Problem Does This Paper Address?

Astronomy has entered the era of "Big Data." Wide-field surveys like the ZTF and the upcoming Vera C. Rubin Observatory (LSST) generate millions of light curves—plots of a star's brightness over time. While most stars follow predictable patterns (like the steady pulse of a Cepheid variable), the most scientifically interesting objects are often the "weird" ones that don't fit the mold. These anomalies can reveal rare stages of stellar evolution or even new physics.

The challenge is that there are too many stars for humans to inspect manually. Standard automated methods often rely on pre-defined classes, meaning they might miss something truly unique because they weren't "trained" to look for it. This paper addresses the need for an unsupervised approach that can flag anomalies without knowing what they are beforehand.

Methods

The authors built a multi-stage machine learning pipeline:

  1. Data Pre-processing: They took raw brightness data and "phase-folded" it (aligning observations based on the star's period). They then used Multivariate Gaussian Process Regression (MGPR) to interpolate the data, creating a smooth, two-channel "image" (representing the $g$ and $r$ spectral bands) with 160 data points.
  2. Feature Extraction (C-VAE): They used a Convolutional Variational Autoencoder. The encoder compresses the 2x160 "image" into a low-dimensional "latent space" (a 20-element vector). The decoder then tries to reconstruct the original light curve. By forcing the network to compress the data, it learns the most essential physical features of stellar variability.
  3. Anomaly Detection (Isolation Forest): To find the outliers, they fed the latent vectors—along with the star's period and color—into an Isolation Forest. This algorithm works by "isolating" points; anomalies are easier to isolate (requiring fewer splits in a random tree) than points in dense clusters.
  4. Coordinate Transformation: The authors found that the latent space had an annular (ring-like) structure. They experimented with a spherical coordinate transformation to better rank anomalies based on their distance from the "normal" clusters.

Datasets

  • ZTF Catalog of Periodic Variable Stars (CPVS): The primary dataset, containing ~730,184 periodic variable stars.
  • Gaia EDR3: Used for cross-referencing absolute magnitudes and colors to understand the physical nature of the identified anomalies.
  • WISE & 2MASS: Infrared data used to construct Spectral Energy Distributions (SEDs) for the top anomalies.

Key Insights and Ideas

  • Latent Space as a Classifier: Even though the VAE was unsupervised, its latent space naturally clustered stars into known astronomical types (e.g., Mira variables, Eclipsing Binaries). This proves the model successfully captured the underlying physics of light curves.
  • The "Messiness" of Evolution: The top anomalies weren't just random noise; they were primarily Red Giants and Asymptotic Giant Branch (AGB) stars. These stars are in the late stages of their lives and exhibit irregular, "messy" variability due to thermal pulses and massive dust production.
  • Dusty Envelopes: Many anomalies showed significant infrared excess. This suggests these stars are "shrouded" in shells of dust they've blown off themselves—a key indicator of high mass-loss rates.
  • Young Stellar Objects (YSOs): A subset of anomalies were identified as stars still in the process of forming. Their "anomalous" nature comes from irregular accretion of material from surrounding disks.

Key Results

  • Top 100 Anomalies: The pipeline successfully flagged the most unusual 0.01% of the catalog.
  • Super-AGB Candidates: The study highlighted two specific stars (ZTFJ200229.70+295140.8 and ZTFJ191700.77+100441.0) with exceptionally high luminosities, suggesting they might be "Super-AGB" stars—rare objects on the verge of exploding as electron-capture supernovae.
  • Spatial Clustering: The identified anomalies are tightly concentrated in the Milky Way's galactic disk, supporting the idea that they are relatively young, massive, and highly evolved stars.
  • Phase Lags: Some anomalies exhibited a phase lag between different color bands, a complex behavior that standard Fourier-based models often fail to capture but the C-VAE flagged easily.

Limitations and Future Work

  • Interpolation Artifacts: While rare (<0.01%), the MGPR interpolation can sometimes create unphysical "wiggles" in the light curve, which the Isolation Forest might mistake for a real astronomical anomaly.
  • Need for Spectroscopy: Photometry (brightness) can only tell us so much. The authors emphasize that detailed spectroscopic follow-up is required to confirm the chemical composition and exact evolutionary state of these weird stars.
  • Scalability: While tested on ZTF, the real test will be applying this to the billions of sources expected from the LSST, which will require even more efficient computational pipelines.

Summary updated

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 multi-wavelength 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 \citet{Chen_2020}. 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 = {H. S. Chan and V. Ashley Villar and S. H. 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