Neutron star envelopes with machine learning: a single-hidden-layer neural network application

Abstract

Thermal and magneto-thermal simulations are an important tool for advancing understanding of neutron stars, as they allow us to compare models of their internal structure and physical processes against observations constraining macroscopic properties such as the surface temperature. A major challenge in the simulations is in modelling of the outermost layers, known as the envelope, exhibiting a drop of many orders of magnitude in temperature and density in a geometrically thin shell. This is often addressed by constructing a separate envelope model in plane-parallel approximation that produces a relation between the temperature at the bottom of the envelope, $T_b$, and the surface temperature, $T_s$. Our aim is to construct a general framework for approximating the $T_b$-$T_s$ relation that is able to include the dependencies from the strength and orientation of the magnetic field. We used standard prescriptions to calculate a large number of magnetised envelope models to be used as a training sample and employed single-hidden-layer feedforward neural networks as approximators, providing the flexibility, high accuracy, and fast evaluation necessary in neutron star simulations. We explored the optimal network architecture and hyperparameter choices and used a special holdout set designed to avoid overfitting to the structure of the input data. We find that relatively simple neural networks are sufficient for the approximation of the $T_b$-$T_s$ relation with an accuracy $\sim 3\%$. The presented workflow can be used in a wide range of problems where simulations are used to construct approximating formulae.