Inferring Neutron-Star Properties from Post-merger Gravitational-wave Spectra with Neural Networks

Abstract

We present a proof-of-concept study of the inverse problem of inferring neutron-star properties directly from the post-merger gravitational-wave spectrum of equal-mass binary neutron-star mergers. Using noise-free spectra from numerical-relativity catalogs, we train and compare three artificial-neural-network regression models and two multivariate linear-regression baselines to predict the stellar mass, $M$, the quadrupolar tidal deformability, $κ_2^τ$, and the slope of the mass--radius relation, $dR/dM$. Since the inverse mapping is nonlinear and cannot be obtained by analytically inverting the direct neural-network model, we construct inverse surrogates and train the networks with a two-stage procedure in which residuals from an initial pass define sample weights for a second pass, together with regularization via dropout, Gaussian-noise injection, and early stopping. We find that neural networks consistently outperform linear baselines, showing that nonlinear surrogates capture the inverse relation between post-merger spectra and source properties more effectively than algebraic inversion. The best performance is achieved by an ensemble of single-task networks, while a multi-task model gives comparable accuracy for predicting the mass--radius slope, and a mixture-of-experts architecture provides insight into spectral-region importance. We further show that the best model reproduces empirical relations between the dominant post-merger frequency and tidal deformability, and recovers equation-of-state-dependent mass--tidal-deformability trends, indicating physical consistency beyond pointwise accuracy. Although restricted to idealized noise-free spectra, the results show that neural-network surrogates provide a promising route for extracting neutron-star information from post-merger signals with future third-generation detectors.