Improving Exoplanet Atmospheric Retrievals with a Learning-Based Pressure-Temperature Profile Model

Atmospheric retrievals are commonly used to infer exoplanet properties (e.g., the chemical composition of the atmosphere) from an observed spectrum. A retrieval framework requires a forward model (calculates the spectrum corresponding to a set of model parameters) and a Bayesian inference scheme (samples the space of model parameters in search of the best combination). One important component of the forward model is the pressure-temperature (PT) profile, which describes the thermal structure of the atmosphere. Retrievals typically employ parametric functions to describe the atmospheric PT structure (e.g., polynomials in Konrad et al. 2022). While being versatile, such parametric forward models are not physical. Thus, we have no guarantee that the retrieved PT structure describes a physically possible atmospheric state. Furthermore, parametric forward models require large parameter numbers (more than 4), which slows down the atmospheric retrieval routine. We employ a new Machine Learning (ML) approach to parameterize the PT structure of exoplanet atmospheres in retrieval studies. This new approach requires fewer parameters (only two) and is trained on physically accurate PT profiles. We train our ML model on the PyAtmos data set (Bell at al. 2018, Chopra et al. 2018), which consists of more than 100000 physically and chemically self-consistent PT profiles of Earth-like planets around a solar-type star, which were calculated with Atmos (a one-dimensional coupled photochemistry-climate model; Arney et al. 2016, Meadows et al. 2018). To prove the applicability of the ML PT Model, we run atmospheric retrievals on a low resolution (R=50) MIR thermal emission spectrum of an Earth-twin exoplanet. We find that employing our ML PT model speeds up the retrieval significantly over the baseline retrieval with a polynomial PT model (by roughly 50%). The retrieved values for most planetary and atmospheric parameters are comparable. However, the retrieval that employs the ML PT model provides better estimates for the atmospheric PT structure than the polynomial baseline. These retrieval runs show the potential of ML-base PT models for atmospheric retrievals. Such models provide a physically accurate description of atmospheric PT profiles while requiring less parameters than the commonly used parametric PT models.

References:

Arney, G., Domagal-Goldman, S. D., Meadows, V. S., et al. 2016, Astrobiology, 16, 873

Bell, A., Chopra, A., Fawcett, W., et al. 2018, in 5th Workshop on Challenges in Machine Learning (NeurIPS)

Chopra, A. et al. 2018, About the FDL PyATMOS dataset

Konrad, B. S., Alei, E., Quanz, S. P., et al. 2022, A&A, 664, A23

Meadows, V. S., Arney, G. N., Schwieterman, E. W., et al. 2018, Astrobiology, 18, 13