MAISTEP: A new grid-based machine learning tool for inferring stellar parameters

I. Ages of giant planet host stars

¹ Department of Physics, Faculty of Science, Kyambogo University, P.O. Box 1 Kyambogo, Kampala, Uganda
² Max-Planck-Institut für Astrophysik, Karl-Schwarzschild-Str. 1, D-85748 Garching, Germany
³ Instituto de Astrofísica e Ciências do Espaço, Universidade do Porto, Rua das Estrelas, 4150-762 Porto, Portugal
⁴ Departamento de Física e Astronomia, Faculdade de Ciências da Universidade do Porto, Rua do Campo Alegre, s/n, 4169-007 Porto, Portugal

^⋆ Corresponding author; kamulali@mpa-garching.mpg.de

Received: 2 December 2024
Accepted: 30 January 2025

Abstract

Context. Our understanding of exoplanet demographics partly depends on their corresponding host star parameters. With the majority of exoplanet-host stars having only atmospheric constraints available, robust inference of their parameters (including ages) is susceptible to the approach used.

Aims. The goal of this work is to develop a grid-based machine learning tool capable of determining the stellar radius, mass, and age using only atmospheric constraints. We also aim to analyse the age distribution of stars hosting giant planets.

Methods. Our machine learning approach involves combining four tree-based machine learning algorithms (random forest, extra trees, extreme gradient boosting, and CatBoost) trained on a grid of stellar models to infer the stellar radius, mass, and age using effective temperatures, metallicities, and Gaia-based luminosities. We performed a detailed statistical analysis to compare the inferences of our tool with those based on seismic data from the APOKASC (with global oscillation parameters) and LEGACY (with individual oscillation frequencies) samples. Finally, we applied our tool to determine the ages of stars hosting giant planets.

Results. Comparing the stellar parameter inferences from our machine learning tool with those from the APOKASC and LEGACY, we find a bias (and a scatter) of −0.5% (5%) and −0.2% (2%) in radius, 6% (5%) and −2% (3%) in mass, and −9% (16%) and 7% (23%) in age, respectively. Therefore, our machine learning predictions are commensurate with seismic inferences. When applying our model to a sample of stars hosting Jupiter-mass planets, we find the average age estimates for the hosts of hot Jupiters, warm Jupiters, and cold Jupiters to be 1.98 Gyr, 2.98 Gyr, and 3.51 Gyr, respectively.

Conclusions. Our machine learning tool is robust and efficient in estimating the stellar radius, mass, and age when only atmospheric constraints are available. Furthermore, the inferred age distributions of giant planet host stars confirm previous predictions – based on stellar model ages for a relatively small number of hosts, as well as on the average age-velocity dispersion relation – that stars hosting hot Jupiters are statistically younger than those hosting warm and cold Jupiters.