Reconstruction of the Equations of State (EoSs) of Compact Stars using machine and deep learning regression techniques
Abstract
This dissertation focuses on the reconstruction of Equations of State (EoSs) describing the interior of compact stars, using modern machine learning and deep learning methods. The pipeline is based on data from mass-radius (M-R) curves, obtained by numerically solving the Tolman-Oppenheimer-Volkoff equations for a wide range of admissible EoSs. The manuscript is divided into a Theoretical Part (Chs. 1-4) and a Computational Part (Chs. 5-7). The theoretical chapters analyze the properties of neutron and quark stars, the physical constraints of viable EoS models, and introduce regression algorithms (Decision Tree, Random Forest, Gradient Boosting, XGBoost) and neural networks with normalization and dropout techniques. The computational part presents the generation of artificial EoSs for hadronic and quark stars (MIT bag, CFL), the numerical solution of the TOV equations, data preparation, and hyperparameter tuning. Results include training and evaluation of models using MSE/MSLE metrics, learning curves for neural networks, and reconstruction of 21 hadronic and 20 quark star EoSs. Source code and tools for reproducibility and future research are provided. The work aims to establish a reusable and scalable framework, strengthening the connection between theoretical astrophysics and computational science.