H2O molecular maps of real PZ Tel B data were created using spectroscopic cross-correlation. The figure shows a real case example where noise structures can reduce the detection power of the cross-correlation method. The brown dwarf was observed in good conditions (air mass: 1.11, start-to-end seeing: 0.77 – 0.72) and poorer conditions (air mass: 1.12, seeing: 1.73 – 1.54). For details on the observation conditions, see Appendix A. The top figure shows the molecular map of PZ Tel B, and the bottom figure shows the cross-correlation series along the radial velocity (RV) support of the object center and pixels within the object brightness region. In both cases, the brown dwarf should appear at the same spatial coordinates at the respective RV positions (see vertical lines), but it is clearly visible in good conditions, but barely visible at the same scale in poor conditions. — astro-ph.EP
A new generation of observatories and instruments (VLT/ERIS, JWST, ELT) will facilitate the development of robust methods to detect and characterize faint exoplanets and planets close to the solar system. Molecular mapping and cross-correlation spectroscopy uses molecular templates to separate the spectra of planets from their host stars.
However, relying on signal-to-noise (S/N) metrics can lead to missed discoveries due to the strong assumption of Gaussian independent and identically distributed noise. We present Machine Learning for Cross-Correlation Spectroscopy (MLCCS), a method that aims to improve exoplanet detection sensitivity by leveraging weak assumptions about exoplanet properties, such as the presence of specific molecules in their atmospheres. MLCCS methods, such as perceptrons and one-dimensional convolutional neural networks, operate on the cross-correlation spectral dimension, where patterns from molecules can be identified.
We tested on a simulated dataset of synthetic planets inserted into real noise from SINFONI in K-band. MLCCS results show a notable improvement. Results on a grid of faint synthetic gas giants show that with a maximum false positive rate of 5%, the perceptron is able to detect about 26 times more planets compared to the S/N metric. This factor increases up to 77 times with a convolutional neural network, shifting the statistical sensitivity from 0.7% to 55.5%. Moreover, the MLCCS method provides significantly improved detection reliability and conspicuity in imaging spectroscopy.
Once trained, the MLCCS method allows for sensitive and rapid detection of exoplanets and their molecular species in the spectral dimension, can cope with systematic noise and difficult observing conditions, is adaptable to many spectrometers and modes, and is versatile with respect to atmospheric properties, allowing it to identify a wide range of planets in archival and future data.
Emily O. Garvin, Marcus J. Bons, Jean Ayos, Gabriele Cugnot, Jonas Spiller, Polychronis A. Patapis, Dominique Petit-di de la Roche, Rakesh Nath Ranga, Olivier Absir, Nicolai F. Meinshausen, Sasha P. Quantz
Comments: 27 pages, 24 figures. Submitted for publication in A&A on January 2, 2024. Resubmitted on May 17, 2024 after first iteration with reviewers.
Subjects: Earth and Planetary Astrophysics (astro-ph.EP), Astrophysical Measurements and Methods (astro-ph.IM), Machine Learning (cs.LG), Applications (stat.AP)
Source: arXiv:2405.13469 [astro-ph.EP] (or arXiv:2405.13469v1 [astro-ph.EP] For this version
Submission History
Source: Emily Ommaya Garvin
[v1] Wednesday, May 22, 2024 09:25:58 UTC (2,774 KB)
https://arxiv.org/abs/2405.13469
Astrobiology, Astrochemistry,
