Forecasting how the upper atmosphere (thermosphere) behaves is important for predicting satellite paths and avoiding collisions in low Earth orbit. This study introduces a forecasting system that uses machine learning models trained on past solar and space weather data. By combining the predictions from several models through different ensemble strategies, including a method that learns how to best weigh each model’s output, we improve the accuracy and reliability of forecasts.

We also create special models that are fine-tuned using data from geomagnetic storms to better predict conditions during those intense space weather events. This work is part of a larger effort to build fast, reliable, and uncertainty-aware tools that help improve satellite operations and space environment awareness.