Exploring the Infrared Variable Sky with Machine Learning

Abstract

The last decade in astronomy has seen the growth of time-series data and with it, the emergence of large surveys. Surveys such as PTF (Law et al., 2009), ZTF (Bellm et al., 2019), CoRoT (Auvergne et al., 2009), HOYS (Froebrich et al., 2018) and VVV(Minniti et al., 2010) provide large amounts of large-area, multi-epoch data. Such surveys bring a multitude of new issues, many of which are in the form of ‘unknown-unknowns’. From this, novel techniques are required to properly analyse these data. Manual analysis is unfeasible and hence, efforts have been taken to develop tools that seek to automate large portions of the data analysis. The new dimension of study afforded to us by these surveys allows us to probe the formation, evolution and death of stars in unique ways. A fundamental issue arises “How can we completely and robustly extract information from modern astronomical time series data?” – Answering this question requires the development of novel methods and the improvement of those already established. In doing so, I aim to further expand and explain the demographics of variable stars in the Milky Way. By coupling more sensitive and robust identification methods with more thorough and complete analysis, I aim to identify and characterise new and known stellar classes. These actions seek to provide a more complete and accurate view of the Milky Way, its structure and demographics. Key contributions of this thesis include the development of a neural network-based false alarm probability (NN FAP) method, which significantly improves the identification of periodic variables in large-scale surveys like VVV, LSST, and TESS. This method generates a universally comparable and unbiased FAP, making it applicable across various types of variable stars, leading to a more complete view of the demographics of periodic variable stars. The creation of the PeRiodic Infrared Milky-way VVV Star-catalogue (PRIMVS) underscores the effort to identify periodic variable stars comprehensively and without bias. Utilising the VVV survey’s depth and breadth, PRIMVS processed over 86 million candidate variable sources using multiple period-finding methods and a novel neural network-based false alarm probability, leading to the identification of approximately 5 million periodic variables. Moreover, the thesis introduces a contrastive learning approach based on the SimCLR framework with a gated recurrent neural network (GRU) backbone, specifically designed to handle stochastically sampled time-series data. This method improves variable star classification by creating semantically meaningful embeddings, enabling more nuanced and accurate analysis. Additionally, the integration of VVV data with Gaia astrometry enhances distance measurements to star forming regions, while the use of Denoising Diffusion Probabilistic Models (DDPMs) for generating synthetic light curves provides a novel solution for developing extensive training sets.