Neural Network Approaches for Classifying Variable Star Light Curves: Advancing Astrophysical Data Analysis

This thesis explores the application of two neural network architectures—Long Short-Term Memory (LSTM) and two-dimensional Convolutional Neural Networks (2D-CNN)—to the problem of classifying variable stars based on their light curves. Variable stars, such as Cepheids, BYDraconis, Delta Scuti, and Algol-type eclipsing binaries, provide fundamental information about stellar evolution and distance measurement, yet the manual analysis of their photometric variations is often slow and prone to errors. Using a large dataset from the Zwicky Transient Facility (ZTF), LSTM and 2D-CNN models were developed, trained, and evaluated to automatically identify and distinguish different types of variable stars. The LSTM approach focuses on learning temporal dependencies in one dimensional sequential data, while the 2D-CNN model leverages transformed representations of light curves to capture distinguishing features in a two-dimensional space. The results show that both models achieve high classification accuracy, significantly reducing the need for human intervention and accelerating the identification of new candidate variable stars. Moreover, this study highlights how deep learning methods can make the study of stellar variability more efficient and objective compared to traditional techniques. In conclusion, the findings demonstrate the potential of data-driven approaches in astronomy: by combining astrophysical knowledge with advanced machine learning techniques, we obtain a powerful framework for analyzing large amounts of photometric data and deepening our understanding of the fundamental processes governing variable stars.