How TESS and machine learning can mitigate stellar activity in radial velocity observations

TESS requires precise radial velocity (RV) measurements to achieve its key science goal of measuring the masses of a large sample of transiting sub-Neptunes and super-Earths, enabling studies of their compositions and atmospheres. RVs can also recover lost ephemerides and allow further characterization that transits alone cannot achieve. However, these science goals require extremely high RV precision, which is currently limited by stellar variability caused by inhomogeneities on the stellar surface like starspots and faculae. Here we show that machine learning techniques significantly improve RV precision by separating the stellar activity from real Doppler shifts, reducing the RMS by 40% for the Sun. We found a similar improvement in RMS for other stars and used this technique to add to TESS’s sample of planets with precise masses. We also find that adding lightcurves to our neural network enhances its ability to model stellar variability for the Sun. This means that TESS lightcurves could be a major ingredient in solving the stellar activity problem for other stars. Going forward, these techniques will help measure the masses of transiting planets and may eventually help us detect habitable-zone Earth-mass exoplanets.