The thesis investigates the use of machine learning methods to infer stellar radial velocities from the Gaia DR3 catalogue. The goal is to examine the extent to which the astrometric and photometric measurements, which are available for a far larger sample of stars compared to spectroscopic measurements, can be used to predict the missing third component of velocity, radial velocity. We use Bayesian neural networks and gradient-boosted decision trees to predict radial velocities. The models are trained on a clean sample of stars from the Gaia DR3 catalogue with measured radial velocities. We use the mean absolute error as a measure of the accuracy of the models' predictions and investigate how well the kinematic structure of the stellar halo is preserved. The most successful model is a Bayesian neural network that uses a combination of astrometric and photometric features and achieves a mean absolute error of 27.75 km/s on the test sample. We observe that the errors are on average significantly larger in the region of the sky corresponding to the Galactic bulge, and for faint stars. The model is not successful at predicting high radial velocities and is better suited to stars with lower radial velocities, such as the thin disc stars in the solar neighbourhood. However, the model is quite successful at preserving the kinematic structure of the stellar halo, correctly identifying approximately 75 % of the stars classified as halo stars using their measured radial velocities.
|