Speaker
Description
In the recent years, Physics-Informed Neural Networks (PINNs) have been identified as a promising approach for both forward and inverse modeling problems in physical systems. In particular, applied to Radiation Belt modelling, it has been shown that this methodology is able, to some extent, to reconstruct, from in-situ observation of particle fluxes, the diffusion coefficients which drive the dynamics of the radiation belts. However, because of the specificities of the physical system (including a very large temporal and spatial dynamics of the state, as well as very inhomogeneous and anisotropic coefficients), and the ill-posedness of the inverse problem, PINNs can sometime fail to accurately pick up the dynamics of the physical coefficients. Hence, a proper evaluation of the method is required before using them on in-situ data.
In this study, we present a quantitative evaluation of PINNs applied to the 1D Fokker-Plank equation, which correspond to a simplified model of the electron radiation belt dynamics. Using the framework of twin experiments, we show when the PINN method is able to accurately reconstruct the physical parameters of the system, and when it fails to do so. We also compare the use of PINNs to alternative approaches such as data assimilation. Using different scenarii, we try to provide a general understanding of the limitations of the different methods.
| Do you plan to attend in-person or online? | In-person |
|---|
