# Deep-learning based measurement of planetary radial velocities in the presence of stellar variability

**Authors:** Ian Colwell, Virisha Timmaraju, Alexander Wise  
**arXiv:** [2304.04807](https://arxiv.org/abs/2304.04807)  
**Date:** 2023-04-10  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.04807v1/

## Key points

- The authors present a deep learning-based approach for measuring small planetary radial velocities in the presence of stellar variability.
- The primary objective is to develop a DL pipeline that can accurately separate planetary and stellar activity RVs, which is challenging for traditional EPRV approaches.
- Neural networks are used to reduce stellar RV jitter in three years of HARPS-N sun-as-a-star spectra.
- Spectra are used instead of CCFs as the dataset, leveraging differences between spectral lines' responses to stellar activity.
- The unprecedented Signal-to-Noise Ratio (SNR) and cadence of sun-as-a-star spectra allow them to evaluate the effectiveness and limitations of neural networks at separating stellar and planet-induced RVs in the wavelength domain at sub-m/s precision.
- The primary dataset used for this work is the HARPS-N Solar Spectra, which includes significant numbers of sunspots, faculae, and plage, as well as convective blueshift variation due to the solar magnetic cycle maximum and minimum that occurred in 2014 and 2019 respectively.
- To evaluate the ability of DL-based neural networks to separate stellar and planetary RVs, they utilize end-to-end injection and recovery of planet-like Doppler signals.
- They develop and compare dimensionality reduction and data splitting methods, as well as various neural network architectures including single line CNNs, an ensemble of single line CNNs, and a multi-line CNN.
- The multi-line CNN is able to recover planets with 0.2 m/s semi amplitude 50 day period with 8.8% error in amplitude and 0.7% error in period.
- This approach shows promise for mitigating stellar RV variability and enabling detection of small planetary RVs with unprecedented precision.
- Overall, this work demonstrates the potential of DL based approaches for improving EPRV measurements and pushing the limits of stellar activity correction.
- The authors provide a detailed description of their data sources, data preparation and preprocessing steps, machine learning approaches, results, future areas for improvement which further highlights its potential applications in astronomy research related fields such as exoplanet detection characterization habitability studies etcetera.

## Summary

In this paper, the authors present a deep learning-based approach for measuring small planetary radial velocities in the presence of stellar variability. The primary objective of this work is to develop a DL pipeline that can accurately separate planetary and stellar activity RVs, which is challenging for traditional EPRV approaches. The authors use neural networks to reduce stellar RV jitter in three years of HARPS-N sun-as-a-star spectra. They leverage the differences between spectral lines' responses to stellar activity by using spectra instead of CCFs as their dataset. The unprecedented Signal-to-Noise Ratio (SNR) and cadence of sun-as-a-star spectra allow them to evaluate the effectiveness and limitations of neural networks at separating stellar and planet-induced RVs in the wavelength domain at sub-m/s precision. The primary dataset used for this work is the HARPS-N Solar Spectra, which includes significant numbers of sunspots, faculae, and plage, as well as convective blueshift variation due to the solar magnetic cycle maximum and minimum that occurred in 2014 and 2019 respectively. The data are obtained from the University of Geneva Data & Analysis Center for Exoplanets website. To evaluate the ability of DL-based neural networks to separate stellar and planetary RVs, they utilize end-to-end injection and recovery of planet-like Doppler signals. They inject planet-like RVs into the spectra and use the network to recover them. The authors develop and compare dimensionality reduction and data splitting methods, as well as various neural network architectures including single line CNNs, an ensemble of single line CNNs, and a multi-line CNN. They find that the multi-line CNN is able to recover planets with 0.2 m/s semi amplitude 50 day period with 8.8% error in amplitude and 0.7% error in period. This approach shows promise for mitigating stellar RV variability and enabling detection of small planetary RVs with unprecedented precision. Overall, this work demonstrates the potential of DL based approaches for improving EPRV measurements and pushing the limits of stellar activity correction. The authors provide a detailed description of their data sources, data preparation and preprocessing steps machine learning approaches results future areas for improvement which further highlights its potential applications in astronomy research related fields such as exoplanet detection characterization habitability studies etcetera .