Speaker
Description
We present an amortized (train-once-use-many) simulation-based inference framework for constraining the nuclear equation of state of neutron stars within density-dependent relativistic mean-field models. The aim is to validate Neural Posterior Estimation (NPE) as a fast alternative to conventional nested-sampling analyses, while preserving full Bayesian posterior quantification. We benchmark two RMFs, one with density-dependent couplings (DDB) and the other with constant nonlinear couplings (RMF-NL), against a PyMultiNest reference analysis using recent empirical nuclear-matter constraints and the two-solar-mass neutron-star requirement. Training sets of 2--3 million simulator pairs generated with the CompactObject framework are used to train a noise-conditional Neural Spline Flow, allowing observational uncertainties to be incorporated directly during training. The key advantage of this approach is amortization: after a one-time training cost, the same trained network can be conditioned on future observations and produce posterior samples within seconds, without rerunning the full sampler. A fully differentiable JAX-GPU forward model further enables rapid posterior propagation to mass--radius--tidal-deformability predictions and importance-sampling diagnostics. Preliminary DDB results show close agreement with PyMultiNest, with $M_{\max}=2.176$ vs.\ $2.141,M_{\odot}$, $R_{1.4}=12.81$ vs.\ $12.65~\mathrm{km}$, and $\Lambda_{1.4}=512$ vs.\ 466, together with successful SBC and TARP calibration tests. To our knowledge, this is the first direct amortized-NPE framework for RMF-based neutron-star EoS inference benchmarked against a nested-sampling reference, offering a scalable route for rapid reanalysis as nuclear and multi-messenger constraints evolve, particularly in view of next-generation observatories such as the Einstein Telescope and other upcoming gravitational-wave detectors.