Simulation-Augmented Physics-Aware Neural Networks for Nonlinear Inverse Problems
Résumé
This study proposes a novel approach to inverse problems in the aerospace domain, with a particular focus on orbit determination. By leveraging deep learning, our framework infers the underlying physical parameters from observed data, even in the presence of highly nonlinear forward models. Crucially, we consider settings where the dimensionality of the physical parameters is substantially lower than that of the observations. Our method trains a neural network with a hybrid loss function that combines observational data and simulated data derived from an imperfect physical model. This joint utilization of actual and simulated data enables effective learning by pairing simulated observations with their corresponding parameter values. Experimental evaluations, including orbit determination tasks, highlight the advantages of our Simulation-augmented Physics-Aware Neural Networks (SimPANNs) over traditional methods that rely solely on observational data. Not only does our approach enhance both accuracy and robustness, but it also demonstrates a promising path for solving nonlinear inverse problems in reduced-parameter spaces by unifying physics-based modeling with data-driven techniques.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |