| |
 |
| Paper: |
Convolutional Neural Networks for Characterizing Reflection Spectra of Evolving Earth-Like Planets |
| Monograph: |
11, HWO25 Proceedings Part II: Mission Framework, Technology, and Broader Contributions |
| Page: |
131 |
| Authors: |
Sarah G. A. Barbosa, Raissa Estrela, Paulo C. F. da Silva Filho, and Daniel B. de Freitas |
| DOI: |
10.26624/SPUM2984 |
| Abstract: |
Future direct-imaging missions such as the Habitable Worlds Observatory (HWO) will obtain reflected-light spectra of Earth-like exoplanets, demanding retrieval tools that scale beyond traditional, computationally intensive methods. We present a one-dimensional convolutional neural network (1D CNN) trained on over one million low-resolution synthetic spectra of Archean, Proterozoic, and Modern Earth analogs, simulating observations with LUVOIR-B (0.2–2.0 μm) and HabEx/SS (0.2–1.8 μm). The network infers six molecular abundances (including biosignatures O2 and O3) plus radius, gravity, atmospheric temperature, and surface pressure. Monte Carlo Dropout enables rapid uncertainty estimation, delivering thousands of realizations in seconds. Integrated Gradients highlight physically relevant features, such as the Fraunhofer A band (O2) and Hartley-Huggins band (O3). Credibility analyses show O3 remains robust across stellar hosts and distances, while O2 is detectable out to 12 pc for F- and G- stars. These results establish the CNN as a mission-ready framework for efficient exoplanet atmosphere retrieval. |
|
|
|
|
|
