Constraining the High-Density Equation of State with Present and Future NICER Observations Using Physics-Informed Regularized Machine Learning
The precise mass and radius measurements of neutron stars by NICER have significantly advanced our ability to constrain the properties of matter at supranuclear densities. In this work, we develop a physics-informed regularized conditional Invertible Neural Network (cINN) that bijectively maps mass--radius posterior distributions directly onto the corresponding central energy density and pressure, eliminating the need for explicit high-dimensional parameter sampling. The physics-informed regularisation guarantees that all inferred solutions satisfy causality and thermodynamic stability, ensuring physically consistent predictions without explicit forward modelling. We demonstrate that the framework accurately reconstructs central EoS posteriors for NICER-like observations while preserving the mapping between macroscopic stellar observables and the microscopic properties of dense matter. Exploiting the computational efficiency of the cINN, we perform a systematic optimisation study of 62,400 simulated mass--radius observations to identify the most informative targets for constraining the high-density EoS. We find that the constraining power depends strongly on the location of the observation in the mass--radius plane, with an optimal strategy that alternates between compact high-mass stars and extended intermediate-mass stars, reducing the uncertainty in the inferred EoS by up to $\sim 9\%-10\%$ relative to the current NICER baseline. These results establish physics-informed invertible neural networks as a powerful framework for rapid, physically consistent inference of dense-matter properties from present and future multi-messenger observations.