Neural Network and Machine Learning Methods for Distinguishing Hadronic and Quark Compact Stars

Compact stars offer a powerful window into the behavior of matter at extreme densities, although their internal structure is still not fully understood. Distinguishing hadronic neutron stars from quark stars is particularly difficult, since the two classes can produce similar mass–radius relations and may therefore be indistinguishable using standard observables alone.
In this study, we employ machine-learning techniques and neural networks to classify compact stars on the basis of their mass, radius, tidal deformability, Love number, and central pressure.
The analysis is based on 39,920 stellar configurations derived from 3,992 equations of state, comprising 2,048 hadronic models and 1,944 quark-matter models described within the MIT bag and Color–Flavor-Locked approaches. Several supervised learning algorithms, including Random Forest, XGBoost, Decision Tree, and Support Vector Machine, are tested and compared with a feedforward neural network. Under ideal noise-free conditions, all models show excellent performance, with XGBoost, SVM, and the neural network achieving accuracy, F1-score, and AUC values close to one. Feature-importance and ablation studies indicate that the Love number plays a central role in the classification process, emphasizing the diagnostic value of tidal-response observables. The robustness of the models is further examined by introducing Gaussian uncertainties in mass and radius and by reconstructing tidal deformability through a compactness–deformability relation including intrinsic scatter. While observational uncertainties degrade the performance of some classifiers, XGBoost and the neural network retain strong predictive capability. These results demonstrate that machine learning can provide an effective framework for identifying the nature of compact stars and underline the importance of accurate tidal-response measurements for constraining the dense-matter equation of state.