Since the discovery of the first extra-solar planet in 1995, Doppler spectroscopy proved to be one of the most successful methods in the search of exoplanets. With new high precision instruments like ESPRESSO and new data analysis methods, it will be possible to detect Earth-like planets on Sun-like stars with similar orbital parameters of Earth. Unfortunately, the search for exoplanets comes with several challenges. Stellar noise usually contaminates the RV measurements, due to the intrinsic activity star, that can hide or mimic planetary signals. Besides this problem, it is also difficult to ascertain the real number of planetary signals present in the data.
To deal with these two issues it was developed a tool called kima. This freely available software combines an MCMC algorithm known as diffusive nested sampling with Gaussian processes (GP) to model the stellar component of the signal and infer the number and properties of the existing planetary signals.
A new version of kima is being developed combining the previous MCMC algorithm with a Gaussian processes regression network (GPRN) that combines the properties of a Bayesian neural network with the flexibility of the GPs. The novelty of this regression network comes with its ability to take into account multiple inputs such as RV, bisector inverse slope, full width half maximum and activity indicators, e.g. log(R_hk).
This GPRN will be an adaptive mixture of GPs that accommodates the signal and noise correlations from various output variables. It will allow the full characterization of the stellar activity in a set of RV measurements and disentangle planetary signals, stellar activity, telluric contamination, and instrumental noise from the stellar spectra, bringing us ever closer to the
detection of Earth-like planets in the habitable zone of Sun-like stars.