The use of machine learning for fast predictions of neutron skins and dipole polarizabilities of nuclei to use in Bayesian inference of neutron star properties.

We explain the use of a machine learning model called Multi-Layered Perceptron Regression (MLPR) and its application to predicting values of neutron skins and dipole polarization values for Lead 208 and Calcium 48. Neutron skins are an excess layer of neutrons that form over the core nucleus due to the nucleus being rich in neutrons. The dipole polarizability is the measurement of how susceptible the protons in the nucleus to respond to an oscillating electric field. Experiments such as PREX and CREX have measured the neutron skin using parity-violating electron scattering, and the dipole polarizability has been measured by proton scattering experiments. These two properties are strongly correlated with neutron rich matter, and hence can be used to extrapolate properties of neutron stars such as their radius. In the modern approach we use Bayesian analysis to infer neutron star properties from neutron skin and dipole polarizability measurements. Performing the calculations directly for this task has become too computationally expensive as we require hundreds of thousands of samples of neutron skins and dipole polarizabilities which require time consuming microscopic calculations. Here is where we use the MLPR, trained on a few hundred sets of neutron skin and dipole polarizability data, to accurately and quickly predict tens of thousands of samples. We will discuss the development of this model, the validation of its performance and the results from the initial test runs performed on the models.