Speaker
Description
Gaia DR3 provides an unprecedented sample of more than two million eclipsing-binary candidates, enabling population studies on a Galactic scale but also posing challenges due to sparse and irregular photometric sampling. We developed a deep-learning framework for the morphological classification of Gaia DR3 eclipsing binaries using physically motivated synthetic light curves. The synthetic data were modified to reproduce Gaia-like observational conditions, including sparse sampling, photometric noise, phase shifts, and outliers, and transformed into image-based representations for convolutional neural networks.
We investigated phase, polar, and gradient-based representations and found that the choice of input representation strongly affects classification reliability; high validation accuracy alone may still lead to systematic biases when models are applied to real survey data. The final classifier was applied to more than 2.18 million Gaia DR3 eclipsing binaries, producing populations of detached and overcontact systems. Their orbital-period distributions show clearly different structures and recover the known short-period cutoff of contact binaries, while their spatial distributions trace Galactic structure.
These results demonstrate the potential of combining physically motivated simulations, machine learning, and Gaia survey data for large-scale studies of eclipsing-binary populations.