Abstract
The periodic pulsations of stars teach us about their underlying physical process. We present a convolutional autoencoder-based pipeline as an automatic approach to search for out-of-distribution anomalous periodic variables within The Zwicky Transient Facility Catalog of Periodic Variable Stars (ZTF CPVS). We use an isolation forest to rank each periodic variable by its anomaly score. Our overall most anomalous events have a unique physical origin: they are mostly highly variable and irregular evolved stars. Multiwavelength data suggest that they are most likely Red Giant or Asymptotic Giant Branch stars concentrated in the Milky Way galactic disk. Furthermore, we show how the learned latent features can be used for the classification of periodic variables through a hierarchical random forest. This novel semi-supervised approach allows astronomers to identify the most anomalous events within a given physical class, significantly increasing the potential for scientific discovery.
Introduction
Astronomy is entering an era of unprecedented data volume. Upcoming surveys like the Legacy Survey of Space and Time (LSST) are expected to observe over 40 billion objects, making manual inspection of periodic variable stars (PVSs) impossible. This transition necessitates automated methods to address two primary challenges: the efficient classification of known stellar types and the detection of rare, out-of-distribution anomalies that could lead to new physical insights.
Methodology
The researchers developed a multi-stage pipeline that treats stellar light curves as a computer vision problem. The process begins with preprocessing, where raw telescope data is phase-folded and interpolated using a Multivariate Gaussian Process (MGPR) to account for irregular sampling and asynchronous multi-band observations (g and r filters).
Following preprocessing, a Convolutional Variational Autoencoder (C-VAE) is used for feature extraction. The interpolated light curves are stacked into a $2 \times 160$ image format and compressed into a 10-dimensional latent space. This latent representation captures the essential morphological features of the star's pulsations. For anomaly detection, an Isolation Forest algorithm ranks objects based on their uniqueness in the latent space. Finally, for classification, these latent features are combined with metadata (period and color) and fed into a Hierarchical Random Forest to categorize stars into 13 distinct physical classes.
Results and Physical Validation
The pipeline was applied to the ZTF Catalog of Periodic Variable Stars, containing approximately 781,602 objects. A key finding was the physical consistency of the identified anomalies. When plotted on a Hertzsprung-Russell (HR) diagram using multiwavelength data from the Gaia mission, the top 100 anomalies clustered in a specific region corresponding to cool, luminous, and highly irregular evolved stars, such as Red Giants or Asymptotic Giant Branch stars located in the Milky Way's disk.
The learned latent space exhibited an annular structure, suggesting the model naturally captures periodicity. Despite significant class imbalances in the training data, the hierarchical classifier achieved robust performance across both broad and specific stellar categories. The entire training process was computationally efficient, requiring roughly 6 hours on a single NVIDIA RTX2080-Ti GPU.
Conclusion
This convolutional autoencoder-based approach provides a scalable and robust framework for the analysis of massive astronomical datasets. By effectively compressing complex light curves into physically meaningful latent features, the pipeline enables the discovery of rare astrophysical phenomena and the automated classification of millions of stars. Future work will aim to extend this methodology to non-periodic transients and incorporate spectroscopic follow-up for deeper physical characterization.
Summary updated