Physics-Informed Reconstruction of MHD Spacetime States from Simulated Spacecraft Measurements

Abstract:

Experimentally constraining global planetary magnetosphere models is challenging due to the sparseness of spacecraft observations. It is not yet clear how to make optimal use of this sparse information, but modern machine learning algorithms may be a way to efficiently utilize this information to reconstruct the global magnetosphere state. As a step towards this goal, we create a physics-informed neural network (PINN) in which expert knowledge in the form of the MHD equations is incorporated into the penalty function. We generate "spacecraft observations" from within high-resolution simulations of various one- and two-dimensional MHD benchmarks (e.g. Brio-Wu shock tube, GEM reconnection challenge) and test the PINN's ability to reconstruct the entire simulation using only the limited data sample. We find that the PINN is able to reconstruct the 1D simulations quite well, though shock waves and other steep discontinuities are a challenge to reproduce exactly. However, 2D simulations are much more difficult to recreate: the PINN requires a few orders more information and training time than in the 1D benchmark recreation. We conclude that if this technique scales well to three-dimensional spatial domains, it may be a way to recreate local spacetime domains from spacecraft observations. However, global reconstructions may require more advanced integrations of machine learning and spacecraft observations within traditional magnetosphere simulations.