Preservation of low-signal long-period transit signatures by pretrained representations

Determine whether self-supervised and pretrained light-curve representations preserve the signatures of low-signal-to-noise, long-period transits observed only a few times, particularly for habitable-zone candidate searches.

Background

Self-supervised methods such as DTST and ASTROMER learn representations from unlabeled light curves, while related pretrained models have shown transfer benefits for variable-star classification, stellar-parameter estimation, and flare identification. Reliable labeled examples of long-period, small-radius planets remain scarce, motivating the use of such representations for exoplanet detection.

The unresolved issue is whether representations learned from general astronomical variability retain the weak, sparse transit morphology needed for habitable-zone searches. The review identifies task-specific injection–recovery tests and evaluation on independent real examples as necessary evidence before these models are deployed for this purpose.