Two views of variable light sharpen astronomical classification

A source’s changing brightness can be described as a sequence of observations or as a set of candidate rhythms. Combining those two views helped an AI system learn useful representations of variable astronomical sources. On a Zwicky Transient Facility catalogue, a simple classifier then separated 14 classes with 77.1% balanced accuracy.[1]

M. Madarász, G. Marton and M. Audard describe the work in a 21 October preprint listed as accepted by Astronomy & Astrophysics. A Transformer processes raw measurements and a convolutional network reads a periodogram. The branches learn representations that remain consistent across two randomly selected halves of a source’s observations while preserving useful differences between sources.[1]

Two complementary views of changing light

Raw brightness points contain timing, amplitude, gaps and non-repeating structure. A periodogram answers a different question: which candidate periodicities seem to explain the observed variation? Each view can miss information available to the other. A star with the same dominant period as another may differ in asymmetry or flare behavior, while irregularly sampled observations complicate obvious interpretations of the original sequence. Combining the views gives the model two complementary descriptions of the same physical process.[1]

The evaluation catalogue contained 37,006 ZTF sources from 14 classes, including stars and active galactic nuclei, arranged into four broader categories. With the learned 64-dimensional representations frozen, a linear classifier achieved 90.5% balanced accuracy on those broad groups. In a separate cross-validation comparison, combining the branches raised recall for Cepheids from 60.2% with the better single branch to 75.1%, a gain of 14.9 percentage points. Class labels were absent from the representation-learning objective. Labels did help balance which classes entered training batches, select model configurations and evaluate classification.[1]

The researchers found that dimensions in the learned representations correlate with properties such as amplitude, periodicity and asymmetry. The representation also reveals variation within familiar categories. Those associations help explain what the classifier can use, while giving astronomers a way to inspect structure that a single catalogue label compresses.[1]

Reusing the representation across observations

The team retrained the framework on ZTF’s other wavelength bands without changing its design. In five-fold cross-validation on 17,442 sources observed in all three bands, combining the representations reached 86.1% class balanced accuracy. That tests reuse across bands within ZTF; it does not by itself demonstrate transfer to a different telescope, cadence or calibration pipeline.[1]

The advance is a reusable description of variability. A representation that retains amplitude, periodicity and asymmetry can support questions beyond the original classification task. How well those features survive a different survey remains an empirical question.[1]

Report an error or suggest a correction