Machine-Learnt Ionospheric Modeling: Enhancing Ion Drag and Joule Heating Simulations with Machine Learning

The ionosphere plays a critical role in atmospheric dynamics, influencing the neutral atmosphere through ion drag and Joule heating, and thereby the circulation of the mesosphere and thermosphere on regional to global scales. Current state-of-the-art general circulation models (GCMs), like the Icosahedral Non-hydrostatic model (ICON), lack a detailed representation of the ionosphere, but rely on simplified parametrizations or external data for including ion drag and Joule heating.
We develop a machine learnt surrogate model for ion drag and Joule heating predictions for ICON, leveraging data from WACCM-X, which includes interactive electrodynamics. We tested various machine learning algorithms, including Random Forests, Fully Connected Neural Networks (FCNN), and Physics-Informed Neural Networks (PINNs), to predict ion drag and Joule heating from key inputs such as solar indices (F10.7), geomagnetic activity (Kp), latitude, longitude, time, and elevation.
The goal is to provide a scalable, accurate alternative to computationally expensive physics-based modules, enabling more efficient simulations of the whole atmosphere. This poster outlines the methodology, discusses the preprocessing steps for model training using WACCM-X data, shows first training results, and presents the potential advantages and challenges of integrating machine learning into atmospheric modeling.