Predicting Stellar Parameters of Massive Stars from Light Curves with Machine Learning

Abstract

High-resolution spectroscopic measurements of OB stars are important for understanding processes like stellar evolution, but require labor-intensive observations. In contrast, photometric missions like the Transiting Exoplanet Survey Satellite (TESS) can monitor hundreds of thousands of stars with a range of temporal resolutions, but do not provide such detailed measurements. With surveys like the Legacy Survey of Space and Time promising unprecedented photometric coverage over the next ten years, it is increasingly important to develop methods that connect large-scale time-series photometry with the detailed stellar parameter measurements typically derived from spectroscopy. In this paper, we test whether machine learning can recover such parameters by combining TESS light curves with spectroscopic measurements from the IACOB project, using a sample of 285 light curves from 106 unique O stars. Using both multilayer perceptrons and convolutional neural networks, we demonstrate that (1) O star light curves contain sufficient information to meaningfully infer stellar parameters and (2) periodograms derived from light curves capture substantially more information than previously identified correlation parameters. Our best model achieves moderate success in predicting both spectroscopic luminosity ($R^2 = 0.641_{-0.167}^{+0.107}$) and effective temperature ($R^2 = 0.443_{-0.234}^{+0.056}$), key stellar parameters for determining positions of stars on the spectroscopic Hertzsprung-Russell diagram, despite the small dataset size. Further progress will require expanded datasets of matched photometric and spectroscopic observations.