Predicting the Atmospheric Composition of Exoplanets with Neural Networks

We developed neural network models to predict abundance profiles of ten separate molecules in exoplanet atmospheres. Each network or model takes carbon, oxygen, and sulfur abundances in a planet’s atmosphere as inputs and then predicts the abundance of a single molecule in 150 atmospheric layers. Training and evaluation were conducted using atmospheric data for the exoplanet WASP-39b, with total abundances measured by JWST and molecular abundance profiles generated by the VULCAN photochemical kinetics code. The neural networks required 43.5 minutes total for training and only 2.1 seconds to predict all molecular profiles - approximately 3% of the computational time needed by a full chemistry network. The models achieved a mean squared error of 0.096 on the test set, demonstrating that machine learning can efficiently and accurately reproduce atmospheric chemistry models for exoplanets, providing a much faster alternative to traditional computational models.