# Machine Learning for Exoplanet Detection in High-Contrast Spectroscopy: Revealing Exoplanets by Leveraging Hidden Molecular Signatures in Cross-Correlated Spectra with Convolutional Neural Networks

**Authors:** Emily O. Garvin, Markus J. Bonse, Jean Hayoz, Gabriele Cugno, Jonas Spiller, Polychronis A. Patapis, Dominique Petit Dit de la Roche, Rakesh Nath-Ranga, Olivier Absil, Nicolai F. Meinshausen, Sascha P. Quanz  
**arXiv:** [2405.13469](https://arxiv.org/abs/2405.13469)  
**Date:** 2024-05-22  
**Summary page:** https://www.summarizepaper.com/en/paper/2405.13469v1/

## Key points

- The emergence of advanced observatories and instruments like VLT/ERIS, JWST, and ELT has spurred the need for robust methods to identify faint and close-in exoplanets
- Traditional approaches such as molecular mapping and cross-correlation spectroscopy rely heavily on signal-to-noise ratio (S/N) metrics, which can sometimes lead to missed discoveries due to assumptions about noise distribution
- A novel approach called machine learning for cross-correlation spectroscopy (MLCCS) has been introduced to enhance detection sensitivity by leveraging weak assumptions about exoplanet characteristics
- MLCCS employs algorithms like perceptrons and unidimensional convolutional neural networks to operate in the cross-correlated spectral dimension, where molecular patterns can be identified
- Testing on mock datasets featuring synthetic planets inserted into real noise from instruments like SINFONI at K-band has shown remarkable improvements with MLCCS

## Summary

, , , , 

The emergence of advanced observatories and instruments like VLT/ERIS, JWST, and ELT has spurred the need for robust methods to identify faint and close-in exoplanets. Traditional approaches such as molecular mapping and cross-correlation spectroscopy rely heavily on signal-to-noise ratio (S/N) metrics, which can sometimes lead to missed discoveries due to assumptions about noise distribution. To address this limitation, a novel approach called machine learning for cross-correlation spectroscopy (MLCCS) has been introduced. This method leverages weak assumptions about exoplanet characteristics, such as the presence of specific molecules in atmospheres, to enhance detection sensitivity. MLCCS employs algorithms like perceptrons and unidimensional convolutional neural networks to operate in the cross-correlated spectral dimension, where molecular patterns can be identified. Testing on mock datasets featuring synthetic planets inserted into real noise from instruments like SINFONI at K-band has shown remarkable improvements with MLCCS. Results indicate that a perceptron can detect approximately 26 times more planets compared to an S/N metric at a false discovery rate of up to 5%. This detection factor increases significantly to 77 times with convolutional neural networks, leading to a substantial shift in statistical sensitivity from 0.7% to 55.5%. Furthermore, MLCCS methods exhibit significant enhancements in detection confidence and clarity in imaging spectroscopy. Once trained, these methods offer rapid and sensitive detection of exoplanets and their molecular compositions in the spectral dimension while effectively handling systematic noise and challenging observing conditions. The evaluation metrics used for assessing the performance of MLCCS models include receiver operating characteristic (ROC) curves on balanced datasets and precision-recall (PR) curves on imbalanced data. These metrics allow for exploring trade-offs between correctly detecting exoplanets with specific molecules like water while minimizing false positives. Overall, the introduction of MLCCS represents a significant advancement in exoplanet detection techniques, offering versatile capabilities across different spectroscopic instruments and atmospheric conditions. The potential impact of this work extends beyond current observations to enable identification of various planets in archival data sets as well as future astronomical surveys.