Background

Previous Work

Here we present some of the approaches developed in previous work to improve upon existing atmospheric density models. While prediction capabilities are not built into empirical models like NRLMSIS, developing that capability is essential for satellite orbit propagation, drag estimation, and collision avoidance, thereby enhancing the safety and efficiency of space operations. These newly-developed models utilize solar and geomagnetic indices, such as solar flux and geomagnetic activity, along with temporal features derived from sliding windows that capture historical and contextual patterns, and by leveraging advanced architectures like transformers and reduced-order models (ROM), they learn the intricate relationships between solar and geomagnetic activities and their impact on atmospheric density. Once trained, they generate global grids of atmospheric density at various altitudes and locations and can even be used to forecast future atmospheric conditions.

This approach forms the foundation of this challenge but is not the exact task participants are required to perform. What this work does provide participants is valuable context and insight into how space weather data are used to develop generalizable approaches to deriving atmospheric density. It is a prime example of how AI solutions can further space weather research and safer, more sustainable space operations, and for those new to the field, we hope that this work can provide a solid starting point for iterative improvements and experimentation with newer AI techniques. We encourage you to explore the model this competition is inspired by to gain a deeper understanding of the atmospheric density forecasting problem.

Choosing Your Approach

In addition to a spacecraft’s initial orbit, the provided input data includes a history of solar indices and magnetic flux data, which your model may use in part or in full. Valid outputs must contain time-series orbit-averaged densities corresponding to the predicted sequence of densities that the spacecraft will observe along it’s future trajectory. This output format allows for a more generalized challenge problem by focusing on the underlying objective – improving spacecraft propagation and drag models during periods of high solar activity – without requiring participants’ models to report global density grids or satellite trajectories. Limited data availability makes it difficult to validate global density predictions, and while many participants may choose a methodology that relies on a global density grid or orbit propagation, others may develop a solution that bypasses those intermediate steps. We leave the door open to either approach by utilizing orbit-averaged densities for model validation.

If you would like to use the methodology in the previous section to begin experimenting with the warmup dataset, you can generate atmospheric density information as follows: NRLMSIS 2.0, an empirical atmospheric density model, requires satellite time, altitude, longitude, latitude, and two additional space weather parameters: F10.7 solar flux and Ap geomagnetic index. These parameters are sourced from NASA’s historical OMNI dataset. The pymsis package is then used to generate atmospheric density values along an expected satellite orbit. For teams who are new to orbit propagation or do not have a preferred orbit propagation tool, a high-fidelity propagator is included in the STORM-AI devkit as a component of our baseline solution and tutorial. You may also incorporate additional SWARM data such as instantaneous densities and satellite coordinates for the purpose of self-validation. The ESA’s SWARM Data Access Manual contains an overview of the available SWARM data types and may be a helpful reference.