Skip to main content
SearchLoginLogin or Signup

Physics-informed neural networks to model magnetic fields and velocity fields

Presentation #202.01 in the session Coronal Magnetic Fields and Solar Wind Formation.

Published onSep 18, 2023
Physics-informed neural networks to model magnetic fields and velocity fields

The Sun’s photospheric magnetic field is routinely measured, but its extent into the solar atmosphere (e.g., in the corona) remains elusive and is important to understand the initiation of solar eruptions and to predict energetic events. Similarly, characterizing photospheric and chromospheric motions is also crucial for understanding the formation of a plethora of atmospheric phenomena. The problem, however, is that direct observations provide limited estimates of flows, constraining only the line-of-sight velocity component.

We present ongoing work on two Physics-Informed Neural Networks (PINNs):

  1. Force-free magnetic field extrapolations through the optimization of multi-height magnetograms (including the photospheric lower boundary), while simultaneously minimizing residuals for the the force-free and divergence-free equations in the entire simulation volume. We utilize meta-learning concepts to perform realistic magnetic field extrapolations in quasi real-time.

  2. Magnetic field and velocity field reconstructions through the optimization of magnetograms and dopplergrams at the photospheric boundary, and the minimization of residuals of the magnetic induction equation.

We provide an outlook using analytical solutions, as well as synthetic data generated by MHD simulations, which both provide ground-truth references. We conclude with data-driven modeling based on a sequence of SDO/HMI and SOLIS/VSM data.

No comments here