Faithful Neural Embeddings for 3D Exoplanet Climate Modeling

Abstract

With the rapid advancement of telescopes like JWST and Ariel, there is an urgent need for efficient 3D climate models to interpret observations of exoplanet atmospheres. Traditional 3D general circulation models (GCMs) are computationally intensive, prompting the development of machine learning (ML) emulators to accelerate simulations. Recent work, such as that by Plaschzug et al. 2026 \cite{plaschzug2026accelerating}, uses a dense neural network (DNN) to predict local gas temperatures and winds from input parameters, including local gas pressure, spatial coordinates (longitude and latitude), and global temperature. However, this model relies on predicting individual temperature values (points) at specific grid points, which can be limited by the resolution and constraints of the training grid. In this work, we investigate a couple of alternative frameworks based on latent-space representations of local gas temperature ($\text{T}_{\text{gas}}$) to obtain a faithful, low-dimensional representation of these profiles. This represents the first step toward developing a latent space regression model, offering a structurally cohesive alternative to the existing point-wise prediction method \cite{plaschzug2026accelerating}. By capturing the optimal embedding space of atmospheric data, our proposed framework can produce simulated profiles while maintaining computational efficiency, making it suitable for large-scale exoplanet ensemble studies.

Abstract: arXiv descriptive metadata (CC0). Thank you to arXiv for use of its open access interoperability. This service was not reviewed or approved by, nor does it necessarily express or reflect the policies or opinions of, arXiv.