Deep Learning–based Measurement of Planetary Radial Velocities in the Presence of Stellar Variability

Abstract We present a deep learning–based approach for measuring small planetary radial velocities (RVs) in the presence of stellar variability. We use neural networks to reduce stellar RV jitter in 3 years of HARPS-N Sun-as-a-star spectra. We develop and compare dimensionality-reduction and data-splitting methods, as well as various neural network architectures including single-line convolutional neural networks (CNNs), an ensemble of single-line CNNs, and a multiline CNN. We inject planet-like RVs into the spectra and use the network to recover them. We find that the multiline CNN approach is able to recover 50 day period planets with 0.2 m s −1 semiamplitude, with 8.8% error in the amplitude, compared to 80% error in the amplitude using a traditional cross-correlation function approach. This approach shows promise for mitigating stellar RV variability and enabling the detection of small planetary RVs with unprecedented precision.