Towards the Habitable Worlds Observatory: 1D CNN Retrieval of Reflection Spectra from Evolving Earth Analogs

Le résumé fourni par la source

Upcoming direct-imaging missions like the Habitable Worlds Observatory (HWO) aim to characterize dozens of Earth-like exoplanets by capturing their reflected-light spectra. However, traditional atmospheric retrieval frameworks are too computationally intensive to explore the high-dimensional parameter spaces such missions will generate. Here, we present a one-dimensional convolutional neural network (1D CNN), trained on over one million synthetic, noise-injected spectra simulating Archean, Proterozoic, and Modern Earth analogs, as observed by LUVOIR-B (0.2-2.0 $μ$m) and HabEx/SS (0.2-1.8 $μ$m). Our model simultaneously infers six molecular abundances (including biosignatures O$_2$ and O$_3$) along with radius, gravity, surface pressure, and temperature. Inference on unseen test data is performed via Monte Carlo Dropout, enabling uncertainty estimation across thousands of realizations within seconds. The network performs best where spectral features are prominent, accurately recovering CH$_4$ and CO$_2$ in Archean atmospheres and O$_2$ and O$_3$ in Modern cases, while avoiding false positives and outputting near-zero abundances in scenarios of true absence such as Archean O$_2$ and O$_3$. Interpretation via Integrated Gradients confirms that the model bases its predictions on physically meaningful features, including the Fraunhofer A band for O$_2$, and the Hartley-Huggins band for O$_3$. Credibility curve analysis indicates that O$_3$ remains retrievable across a wide range of stellar types and distances, while O$_2$ is detectable out to 12 pc around FG stars. These results elevate the CNN from proof of concept to a mission-ready retrieval engine, capable of processing direct-imaging spectra with HWO on an operational cadence.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.