Data-driven characterisation of variable source light curves with multimodal representation learning
Year: 2026
Type: article
Abstract: Large-scale photometric surveys such as the Transient Facility (ZTF) and the Observatory Legacy Survey of Space and Time (LSST) are producing light curves for hundreds of millions of variable sources. Characterising these sources has traditionally relied on supervised classifiers trained with hand-crafted features and curated label sets, limiting their applicability to predefined taxonomies and individual surveys. Zwicky Vera C. Rubin We present a general, survey-agnostic framework for the characterisation of variable sources from photometric time series that operates entirely without labelled training data and produces low-dimensional representations capturing light curve properties such as amplitude, periodicity, and asymmetry directly from the observations. Our method combines two complementary views of each light curve in a multimodal architecture. Raw time-domain observations are processed by a Transformer encoder, and a generalised Lomb--Scargle (GLS) periodogram is processed by a convolutional neural network. Training uses a split-half augmentation strategy, where each light curve is randomly divided into two temporal halves that serve as independent views of the same underlying variability. The two branches are trained jointly using the Variance-Invariance-Covariance Regularisation (VICReg) objective to produce embeddings that are invariant across the two halves while maintaining high information content per dimension. No labels enter the training procedure at any stage. We evaluate the framework on 37,006 ZTF sources spanning 14 variable star classes across four broad physical categories. A linear classifier trained on the frozen 64-dimensional embeddings achieves 77.1% balanced accuracy at the individual class level and 90.5% at the broad category level. Combining the light curve and periodogram branches improves balanced accuracy by up to 14.9 percentage points over the best single modality for individual classes, confirming that the two branches capture complementary information. The embedding encodes interpretable light curve properties on distinct embedding dimensions, and within-class analysis reveals physically interpretable substructure that the reference labels do not resolve. Retraining on ZTF g and i bands without modification reproduces the main results, and concatenating multi-band embeddings improves balanced accuracy to 86.1%, demonstrating complementary wavelength-dependent information capture. The self-supervised framework produces physically interpretable representations without requiring labelled training data, phase folding, or survey-specific pre-processing. The framework is survey-agnostic by design and directly applicable to other ground-based and space-based photometric surveys with different cadences and wavelength coverage. The trained model and the source code with tutorials are publicly released.
Source: Astronomy and Astrophysics
Authors: M. Madarász, G. Marton, M. Audard
Language: English
Subfield: Astronomy and Astrophysics
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $1,865
OpenAlex ID: W7223211909