Rainbow Connection: Exoplanet Cloud Scattering Via Neural Networks
Lead Research Organisation:
The Open University
Abstract
In recent years, the field of exoplanet science has moved decisively from the discovery phase into the realm of characterisation. The launch of JWST opened the possibility for atmospheric study across both a broad wavelength range and a wide target parameter space. JWST spectra of exoplanets obtained as they transit their parent stars can be analysed using retrievals – Bayesian computational frameworks that compare vast numbers of synthetic spectra with observational data, and thence provide a statistical inference of the planet’s atmospheric characteristics. Hubble and JWST transit spectra of hot Jupiter exoplanets have revealed that aerosols – solid or liquid particles suspended within an atmosphere – are as ubiquitous in the atmospheres of these objects as they are within solar system atmospheres.
Aerosols play a key role in determining both the thermodynamical behaviour of a planet’s atmosphere and its observed spectral characteristics. If not properly accounted for within the retrieval framework, the presence of aerosols can significantly limit the ability to accurately recover gas abundances and constrain the planet’s metallicity and carbon budget. Aerosols can be produced from condensation of atmospheric gases (such as water clouds on Earth, or ammonia clouds on Jupiter and Saturn); photochemical processes (such as tholin haze on Titan); or erosion of surfaces (such as dust devils on Mars). In contrast to the relatively cool solar system planets, aerosols on the hot Jupiters observed with JWST are thought to consist of silicates and minerals such as corundum and perovskite. Whilst the likely composition of aerosols on a given planet can be determined to first order from the conditions a particular species requires to condense, we know from detailed studies of solar system planets that cloud formation is complex and predictive models are rarely completely accurate; for example, the ammonia cloud decks on Jupiter and Saturn do not sit at the expected pressure level, and there is spectroscopic evidence of impurities in these clouds.
Until recently, most exoplanet retrieval frameworks adopted a parametric approach to aerosols, representing Rayleigh-scattering-type behaviour where the scattering efficiency follows a simple power law, and grey clouds where extinction is constant at all wavelengths. With the arrival of JWST, however, comes access to the mid-infrared where cloud species such as silicates have distinct absorption features, and power-law parameterisations are no longer adequate. To accurately represent the spectral behaviour of aerosols with a specific composition and size distribution would require a computationally intensive Mie scattering calculation to be performed for each retrieval iteration, which is not viable.
We propose to use neural networks to emulate the Mie scattering calculation within the well-established NEMESISPY retrieval framework. We will generate a training set of wavelength-dependent aerosol scattering properties, generated using Mie theory from a variety of compositions, both pure and mixed, over a wide range of particle sizes. The network will then be trained to infer the aerosol scattering properties without the need for a Mie scattering calculation. This innovation will enable us to accurately reproduce the contribution of aerosols to transit spectra, allowing us to break degeneracies between cloud properties and gas abundances, and also to infer aerosol composition and characteristics. This work is key to understanding the overall composition of gas giant planets and will be critical for interpreting JWST data, and for supporting future missions such as Ariel and Habitable Worlds Observatory.
Aerosols play a key role in determining both the thermodynamical behaviour of a planet’s atmosphere and its observed spectral characteristics. If not properly accounted for within the retrieval framework, the presence of aerosols can significantly limit the ability to accurately recover gas abundances and constrain the planet’s metallicity and carbon budget. Aerosols can be produced from condensation of atmospheric gases (such as water clouds on Earth, or ammonia clouds on Jupiter and Saturn); photochemical processes (such as tholin haze on Titan); or erosion of surfaces (such as dust devils on Mars). In contrast to the relatively cool solar system planets, aerosols on the hot Jupiters observed with JWST are thought to consist of silicates and minerals such as corundum and perovskite. Whilst the likely composition of aerosols on a given planet can be determined to first order from the conditions a particular species requires to condense, we know from detailed studies of solar system planets that cloud formation is complex and predictive models are rarely completely accurate; for example, the ammonia cloud decks on Jupiter and Saturn do not sit at the expected pressure level, and there is spectroscopic evidence of impurities in these clouds.
Until recently, most exoplanet retrieval frameworks adopted a parametric approach to aerosols, representing Rayleigh-scattering-type behaviour where the scattering efficiency follows a simple power law, and grey clouds where extinction is constant at all wavelengths. With the arrival of JWST, however, comes access to the mid-infrared where cloud species such as silicates have distinct absorption features, and power-law parameterisations are no longer adequate. To accurately represent the spectral behaviour of aerosols with a specific composition and size distribution would require a computationally intensive Mie scattering calculation to be performed for each retrieval iteration, which is not viable.
We propose to use neural networks to emulate the Mie scattering calculation within the well-established NEMESISPY retrieval framework. We will generate a training set of wavelength-dependent aerosol scattering properties, generated using Mie theory from a variety of compositions, both pure and mixed, over a wide range of particle sizes. The network will then be trained to infer the aerosol scattering properties without the need for a Mie scattering calculation. This innovation will enable us to accurately reproduce the contribution of aerosols to transit spectra, allowing us to break degeneracies between cloud properties and gas abundances, and also to infer aerosol composition and characteristics. This work is key to understanding the overall composition of gas giant planets and will be critical for interpreting JWST data, and for supporting future missions such as Ariel and Habitable Worlds Observatory.