Testing deep learning and machine learning methods to estimate the metallicity of RR Lyrae stars from their light curves

RR Lyrae stars (RRLs) are pulsating variable stars that play a crucial role in stellar astrophysics. They are abundant in the old stellar populations Milky Way (MW) halo, globular clusters, and dwarf spheroidal galaxies, but they are present also among the young populations in the MW disc. The estimate of their metallicity is fundamental to use such stars to study the age, formation, and structure of their parent stellar systems and shed light on the MW origin. The metallicity of such stars can be estimated from the properties of their light curves. Traditional approaches rely on fitting linear relations models to parameters obtained from the Fourier decomposition parameters of their light curves, but they often suffer from limited accuracy due to various sources of noise and uncertainty. Recent advancements in deep learning techniques, particularly recurrent neural networks (RNNs), have shown promise in capturing temporal dependencies and extracting meaningful patterns from sequential data (see Dekany & Grebel. 2022, https://ui.adsabs.harvard.edu/abs/2022ApJS..261...33D). The objective of this thesis is to design and evaluate deep learning networks (fully connected, RNNs) for analyzing RR Lyrae light curves and improving photometric metallicity estimation. The student will implement and train deep learning network using a subsample of RRL light-curves from the public Gaia catalogue for which high-resolution metallicity estimates are available (Crestani et al. 2021, https://ui.adsabs.harvard.edu/abs/2021ApJ...908...20C). The student will also compare the performance of the tested methods against conventional approaches. Finally, the trained networks can be used to improve the photometric metallicity estimate of the whole dataset of RRLs in the Gaia catalogue.