Gaia EB classification Cone Search

The systems are classified into detached and overcontact configurations, followed by the identification of starspot signatures within both morphological classes. The classification was performed using a hierarchical computer-vision pipeline based on a fine-tuned ResNet-18 convolutional neural network trained on synthetic light curves generated with the ELISa code.

The phase-folded Gaia G-band light curves are represented as 3-channel 128×128 pixel images encoding the flux distribution, its polar transformation, and the flux gradient. A tailored augmentation scheme calibrated to the Gaia cadence distribution was applied to reduce the synthetic-to-real domain gap (in prep). Because the morphological classification is based on single-passband Gaia G photometry alone, overcontact and ellipsoidal systems cannot be reliably distinguished.