A data-driven approach for extracting exoplanetary atmospheric features

Ground-based high-resolution transmission spectroscopy has become a critical tool for probing the chemical compositions of transiting exoplanetary atmospheres. A well-known challenge in this scope lies in the detrending process, which consists in effectively removing contaminating stellar and telluric absorption features obscuring the planetary spectrum. Principal Component Analysis (PCA) is the current state-of-the-art method, but its effectiveness depends on selecting the correct number of components—a subjective choice that impacts how much of the planetary signal is preserved or lost, and the features to be removed are well represented by the linear combination of the principal components. Additionally, there is no quantitative framework for distinguishing between residuals from incomplete subtraction and those containing the true planetary signal.

In this work, we introduce a novel, computer vision-inspired approach to the task of detrending using Deep Convolutional Generative Adversarial Networks (DCGANs), combined with semantic image inpainting, able to overcome the limitations of PCA. In contrast to PCA, our proposed detrending method operates in a non-linear fashion, allowing for a scalable and robust separation of planetary atmospheric features from interfering signals and eliminating reliance on the manual selection of principal components. As a case study, we consider observations of the ultra-hot Jupiter KELT-9 b acquired by the HARPS-N spectrograph at the Telescopio Nazionale Galileo. Although further refinement is needed for full competitiveness with PCA, our method successfully produces realistic transit-free nights and promising residuals, paving the way for future machine learning-driven detrending methods.