Reproduce “Transformer-based Atmospheric Density Forecasting” on OpenResearch

About this paper

Julia Briden, Peng Mun Siew, Victor Rodriguez-Fernandez, Richard Linares

Researchers from MIT developed a transformer-based architecture for atmospheric density forecasting that consistently outperformed traditional linear methods like DMDc, achieving lower Mean Squared Error and Mean Absolute Error across various space weather conditions and models. This approach demonstrated superior robustness by effectively capturing nonlinear dynamics and long-term dependencies, particularly during high solar activity.

Key insights

• Transformer architectures, with their attention mechanisms and patching strategy, are highly effective at capturing the nonlinear dynamics and long-term dependencies critical for accurate atmospheric density forecasting.
• Utilizing a long look-back window (42 days) with the PatchTST architecture significantly improves forecasting robustness and accuracy, especially during dynamic space weather events.
• Combining reduced-order modeling with a transformer propagator offers a computationally efficient yet highly accurate solution, bridging the gap between physics-based and empirical models.

Summary by alphaXiv.