Spectral classification of young stars in Trumpler 14 using conditional invertible neural networks: comparison of deep learning method and classical method
Da Eun Kang (DIFA - Alma Mater Studiorum Univ. di Bologna)
Tuesday 23/04/2024 @ 14:00, Sala Antonio Sollima (IV piano Battiferro)
Low-mass stars account for the majority of the stars formed in star-forming regions and about half of the total stellar mass. Living longer than massive stars, low-mass stars remain in the early phases of stellar evolution even when massive stars are dead, providing important information for studying stellar evolution and planet formation. Correctly characterising the stellar parameters of these stars is an essential and fundamental first step in understanding the stellar system. As the volume of accumulated observations has ever expanded in recent days, it has become important to develop time-efficient tools that analyse large amounts of data in a faster and more consistent way. In this talk, I will introduce a new deep-learning tool for the spectral classification of young low-mass stars and its application to young stars in Trumpler 14 in the Carina Nebula Complex. Our novel spectral classification tool based on a conditional invertible neural network (cINN) architecture estimates stellar parameters (effective temperature, surface gravity, extinction, and veiling factor) from optical stellar spectra. cINN is one of the deep learning architectures adopting a supervised learning approach and it is especially specialised for degenerate inverse inference problems. The advantage of our network is not only the time-efficient nature of the machine learning technique but also that it can overcome instrumental limitations (low spectral resolution) and methodological limitations of the classical classification method. After pretesting the applicability of cINN in the study of low-mass stars by using template stars well-classified by literature, we analyse about 2000 young, low-mass stars in Trumpler 14, observed with VLT/MUSE. We compare stellar parameters measured by our cINN with literature values measured from the same data by using the classical template fitting method.