Data-driven methods for mitigating stellar variability in NEID Sun-as-a-star observations

We developed machine learning models on solar spectra to characterise solar variability radial velocity (RV), with the goal of predicting stellar variability RV on Sun-like stars for detecting exoplanets in the presence of stellar noise. By studying the spectral line variability in NEID’s “Sun-as-a-star” observations, we utilised FIESTA, a Fourier-based parametrisation of the spectral cross-correlation function (CCF), to capture shape changes in solar spectra. We applied convolutional neural networks on the FIESTA parametrisation of NIED solar CCFs and obtained a significant reduction of 90% and 77% in the residual RV RMS squared between our predictions and the observations from 2022 and 2023, respectively. Additionally, our model determined a systematic RV offset of 10 m/s before and after the fire in 2022 at Kitt Peak National Observatory. In a planet-injection-recovery test, we successfully recovered the injected 1 m/s Doppler RV with a residual RMS of less than 0.3 m/s.