Learning to Predict Clouds: A Neural Network Model for Predicting Exoplanetary Cloud Formation

Cloud formation modeling is a crucial frontier in understanding atmospheric compositions, dynamics, and potential habitability (i.e., biosignature) of exoplanets. It is pivotal in determining which gas species may be observable in exoplanet spectra. The formation of cloud particles is determined by the local gas temperature, gas density, and the local gas composition, and hence, traces the local thermodynamic conditions. The relevant cloud properties include mean particle size, cloud particle number density, material volume fractions, and depleted element abundances for the elements that participate in the cloud formation process, which are required for solving the radiative transfer. These cloud properties are computed along one-dimensional local gas pressure-temperature profiles obtained from a three-dimensional general circulation model (GCM). The computation of these cloud properties involves solving a complex time-dependent reaction-diffusion equation, which is computationally expensive. Additionally, increasing the vertical resolution of the one-dimensional profiles can add more computational burden while solving the reaction-diffusion equation. To overcome such computational expenses, we present an alternative approach based on machine learning (ML). This work develops a neural network regressor that learns the relationship between the input parameters, local gas pressure–temperature profiles, global planetary temperature, effective temperature, latitude, and longitude, and the output cloud properties (e.g., mean particle size and cloud particle number density) in a transformed latent space. Due to the inhomogeneous representation of cloud properties, the neural network regressor comprises multiple branches, each dedicated to a specific property. Each branch employs a specialized neural network to extract latent features for the corresponding output, while latent features are also created from the input parameters. A fully connected network then maps the latent input to latent output features. We apply this ML framework on a GCM grid comprising 60 inflated hot Jupiters orbiting A, F, G, K, and M-type host stars, modelled using ExoRad.The Preliminary results are promising, showing high prediction accuracy for cloud properties at lower global temperatures. At higher global temperatures, increased prediction errors are expected, reflecting the greater complexity of cloud formation in these regimes.