New benchmark tests how well AI foundation models handle astronomical light curve data

Researchers have released StarEmbed, a novel evaluation framework designed to measure how well time series foundation models (TSFMs)—neural networks trained on vast amounts of sequential data—perform on real astronomical observations. The benchmark centers on light curves, measurements of stellar brightness over time, drawn from ~40,000 observed variable stars.

Why astronomy needs a new benchmark#

Existing TSFM training datasets typically rely on regularly-spaced measurements, but real astronomical data tells a messier story. Observations of distant stars arrive at uneven intervals due to weather, instrument availability, and orbital mechanics. Light curves also exhibit multiple simultaneous measurements and variable noise levels across observations.

By introducing a dataset of expert-labeled stellar observations spanning seven distinct classes, StarEmbed addresses a key gap: testing whether models designed for tidy financial or sensor data can transfer effectively to the complexities of astronomy.

How the models performed#

The evaluation covered three core tasks: grouping unlabeled light curves by similarity, classifying stars into their known types, and detecting unusual observations that don't fit standard patterns. The Chronos family of models—trained originally on regularly-sampled, non-astronomical sequences—emerged as the strongest performer in clustering and out-of-distribution detection, demonstrating that strong foundation models can generalize across domains.

However, when it came to precise classification, no foundation model fully matched the results of established domain-specific methods developed specifically for stellar astronomy. Yet the TSFMs showed promising flexibility, suggesting they could serve as universal building blocks for downstream astronomical tasks.

Implications for star science#

The work represents an early step toward developing light curve embeddings that could standardize how astronomers represent and analyze stellar data. As surveys like the Vera Rubin Observatory generate millions of new light curves, having models that can efficiently process this deluge of information becomes increasingly valuable.