Exoplanets: latent representations to speed up 3D climate models in the face of data from the JWST and Ariel

Rather than predicting the temperature point by point, researchers are compressing atmospheric temperature profiles into latent space – the first step towards more consistent and faster emulators.

With JWST today and the Ariel mission tomorrow, astronomers need fast three-dimensional climate models to interpret the atmospheres of exoplanets. Classical general circulation models are very computationally expensive, hence the rise of machine learning emulators.

From point to point to latent space

Recent emulators predict the temperature and winds of the gas at specific points on the grid, based on pressure, longitude, latitude, and overall temperature, making them dependent on the drive resolution. The study published on arXiv explores another path: learning a compact and accurate representation of the complete temperature profiles of the gas, on which to build a regression model.

The authors see it as a more structurally coherent alternative, capable of generating simulated profiles at low cost and suitable for overall studies of large populations of exoplanets.