Convolutional Autoencoders to estimate open cluster astrophysical parameters

Abstract

Deriving open cluster parameters through traditional isochrone fitting is a slow process, difficult to scale to large photometric surveys. In this thesis, we develop a Convolutional Autoencoder (CAE) that infers cluster distance, age, binary fraction, and total stellar mass from open cluster members photometry, astrometry and member star counts. We train the CAE on synthetic clusters simulated with PARSEC stellar evolutionary models mimicking Gaia DR3 observations, and apply it to estimate these parameters for 42 well-characterised real open clusters. On the validation set with simulated data, the network recovers all four parameters with good accuracy and precision, except the binary fraction. Applied to real clusters, mass predictions remain robust, while distance, age, and binary fraction are consistently overestimated. The reduced accuracy on real clusters reflects an expected generalisation gap between idealised simulations and real observational data, rather than a structural failure of the architecture. Upgrading the training simulations with observational effects remains the primary direction for future improvement