Machine-Learning Classification of Neutron-Star Matter Composition from Macroscopic and Oscillation Observables

TL;DR

The paper tackles the problem of inferring neutron-star interior composition from multimessenger observables by framing it as a four-class EOS-family classification using seven macroscopic and oscillation observables derived from TOV-based stellar models. It trains a compact two-hidden-layer MLP on ~770 labeled models spanning nucleonic, hyperonic, dark-matter-admixed, and strange-matter EOSs, achieving a held-out accuracy of 97.4% with strong class discrimination. Analysis shows oscillation descriptors, especially the $f$-mode frequency and damping time $\tau$, drive separability and that mass–radius information provides supplementary, but weaker, constraints. The work provides a reproducible baseline for EOS-family inference and highlights pathways for incorporating observational posteriors and uncertainty quantification to further exploit multimessenger neutron-star data.

Abstract

The microscopic composition of neutron star interiors remains uncertain, with possible scenarios including nucleonic matter, hyperonic matter, dark matter admixed cores, and strange self-bound matter. Traditional constraints on the equation of state rely on Tolman Oppenheimer Volkoff modelling and comparison with multimessenger observations, but machine learning provides a complementary pathway by learning composition dependent patterns directly from astrophysically accessible observables. This work presents a compact supervised learning framework for EOS family classification using stellar properties derived from TOV modelling, asteroseismology, and gravitational wave descriptors. A labelled dataset of neutron star configurations spanning four EOS families (nucleonic, strange matter, dark matter admixed, and hyperonic) is constructed using seven input features: gravitational mass, radius, fundamental f mode frequency, quadrupole moment, redshift, damping time, and characteristic strain. A multilayer perceptron is trained to infer the underlying matter composition. On a held out test set, the classifier achieves an accuracy of 97.4 percent with strong class wise precision and recall. Permutation based feature importance analysis shows that oscillation related quantities, especially the f mode frequency and damping time, dominate the discriminatory power, while mass and radius provide secondary support. Residual misclassifications occur in physically intuitive regions where different EOS families produce partially overlapping macroscopic signatures. These results show that lightweight neural models can reliably identify EOS family fingerprints from a modest set of observables, providing a reproducible baseline for future extensions incorporating Bayesian uncertainty and observational posteriors from NICER and gravitational wave events.

Figures (7)

• Figure 1: Representative mass-radius relations for the four EOS families used to construct the machine-learning dataset. While the curves partially overlap in certain mass ranges, composition-dependent trends persist, particularly at intermediate and high masses, motivating the supervised classification approach adopted in this work.
• Figure 2: Training loss as a function of iteration for the multilayer perceptron classifier. The monotonic decrease and subsequent stabilisation indicate robust convergence.
• Figure 3: Confusion matrix for the test dataset. Rows correspond to the true EOS class, and columns to the predicted class. Diagonal elements indicate correct classifications.
• Figure 4: Row-normalised confusion matrix for the test dataset. Each row sums to unity and represents the conditional probability of predicting a given class for a fixed true EOS family.
• Figure 5: Classification accuracy as a function of stellar mass, evaluated in discrete mass bins using the test dataset.