Speaker
Description
The key challenges in low Earth orbit (LEO) with space operations are tracking and catalogue maintenance for resident space objects (RSOs) including lethal non trackable (LNT) objects, collision and conjunction analysis, manoeuvre planning, re-entry prediction, etc. All these aspects are deeply dependent on drag, which, in turn, has its major source of uncertainty in the thermospheric density. The economy emerging around LEO operations has recently grown exponentially, and so have the associated risks and opportunities, pushing forward research and innovation in the field. Currently, we don’t have a baseline for drag modelling, hence one of our aims is to develop a globally recognized benchmark for drag and thermospheric modelling, allowing consistent operations and reliable decision-making.
With regard to this, we are developing the next generation probabilistic drag modeling framework. Its most significant parts are the forecasting of space weather drivers (typically $F_{10}$ and $k_p$), dynamic modeling of the thermospheric density, and the associated algorithms necessary to integrate the models with orbit determination and prediction. In this talk, we will focus on the development of a reduced order probabilistic emulator (ROPE) for physics-based models of thermospheric density.
To perform dynamical modelling, we reduce the dimensionality of the system either by principal component analysis (PCA) or via nonlinear methods (autoencoders). The thermospheric density system would otherwise be unmanageable with current computational resources due to the large number of degrees of freedom of the original system.
We are building upon previous work with Dynamic Mode Decomposition (DMD) with a framework for identifying nonlinear dynamics (SINDY) and neural networks (NN) both for dimensionality reduction and dynamic modelling (LSTM, GRU and transformer architectures). We can highlight some of the differences and features between these two classes of models.
On the one hand, SINDY offers interpretability and improved accuracy over DMD. Improved accuracy during geomagnetic storms and timely responsiveness of the system are crucial for operational purposes, and both have been achieved within our framework. SINDY also admits a continuous time representation of the dynamics. On the other hand, NNs offer a significantly more accurate alternative than SINDY although they currently lack a continuous time representation. Our aim is to converge toward an optimal set of models to be used for tracking, cataloguing and other uses, and that also admit a continuous time representation.
| Do you plan to attend in-person or online? | In-person |
|---|
