Gaussian Process modeling of granulation and oscillations in red-giant stars

One of my grad students, Filipe Pereira, has recently published his very first, lead-author paper. Here, I provide an overview of his work (links to the article and accompanying open-source code are provided further below). May this be the first of many more articles to come, Filipe!

The analysis of photometric time series in the context of transiting planet surveys suffers from the presence of stellar signals, often dubbed "stellar noise". These signals, caused by stellar oscillations and granulation, can usually be disregarded for main-sequence stars, as the stellar contributions average out when phase-folding the light curve. For evolved stars, however, the amplitudes of such signals are larger and the timescales similar to the transit duration of short-period planets, requiring that they be modeled alongside the transit.

With the NASA TESS mission expected to deliver on the order of 10^5 light curves for stars along the red-giant branch, there is a need for a method capable of describing the "stellar noise" while simultaneously modeling an exoplanet's transit. In the work led by Filipe, a Gaussian Process (GP) regression framework (CELERITE; Foreman-Mackey et al. 2017, AJ, 154, 220) is used to model stellar light curves and the method validated by applying it to TESS-like artificial data (see Figs. 1 and 2). Furthermore, the method is used to characterize the stellar oscillations and granulation of a sample of well-studied Kepler low-luminosity red-giant branch stars. The parameters determined are compared to equivalent ones obtained by modeling the power spectrum of the light curve using the DIAMONDS software (Corsaro & De Ridder 2014, A&A, 571, A71). Results show that the method presented is capable of describing the stellar signals in the time domain and can also return an accurate and precise measurement of the frequency of maximum oscillation amplitude.

A link to the article (arXiv) is provided here.

The implementation of the method is publicly available and can be found here.

Figure 1: Predictive model output by the GP regression (mean and 1σ interval) when applied to an artificial TESS-like time series. The plot is zoomed in on the first ~3 days of simulated data.

Figure 2: Power spectral density (PSD) of the same (full) light curve depicted in Fig. 1. The PSD of the light curve is shown in light gray, with a slightly smoothed version overlapped in dark gray. The PSD of the GP regression output is shown as a solid red curve, with individual components shown in different line styles and colors (see legend).