Modelling stars with Gaussian Process Regression: augmenting stellar model grid
Grid-based modelling is widely used for estimating stellar parameters.
However, stellar model grid is sparse because of the
computational cost. This paper demonstrates an application of a
machine-learning algorithm using the Gaussian Process (GP)
Regression that turns a sparse model grid on to a continuous
function. We train GP models to map five fundamental inputs
(mass, equivalent evolutionary phase, initial metallicity,
initial helium fraction, and the mixing-length parameter) to
observable outputs (effective temperature, surface gravity,
radius, surface metallicity, and stellar age). We test the GP
predictions for the five outputs using off-grid stellar models
and find no obvious systematic offsets, indicating good accuracy
in predictions. As a further validation, we apply these GP
models to characterize 1000 fake stars. Inferred masses and ages
determined with GP models well recover true values within one
standard deviation. An important consequence of using GP-based
interpolation is that stellar ages are more precise than those
estimated with the original sparse grid because of the full
sampling of fundamental inputs.