Modeling and Retrieval of Exoplanetary Atmospheres using Autoencoders

The study of exoplanetary atmospheres is critical for understanding the formation, evolution, and potential habitability of planets beyond our Solar System. Transmission spectroscopy, especially with advanced instruments like the James Webb Space Telescope (JWST) and the upcoming ARIEL mission, has become a leading technique for probing these atmospheres. In this work, I develop a machine learning framework based on convolutional autoencoders to retrieve key atmospheric and planetary parameters—such as the mixing ratios of CH4, CO2, CO, and H2O, alongside the planet’s mass, radius, and isothermal temperature—from exoplanet transmission spectra. The model also reconstructs the full transmission spectrum over a wide wavelength range (0.3 to 50 microns), enabling an inpainting approach to recover missing spectral regions caused by instrumental limitations. This method offers a fast and efficient alternative to traditional Bayesian retrieval techniques by capturing complex nonlinear relationships within spectral data. While the approach demonstrates strong performance on synthetic data, particularly when broad spectral coverage is available, its accuracy decreases when dealing with limited spectral ranges or noisy observations. These challenges highlight the need for further refinements, such as including realistic noise in the training data and improved model architectures. Despite these limitations, the proposed framework provides a promising groundwork to accelerate atmospheric characterization and enhance our understanding of the diversity and formation of exoplanetary atmospheres.