Speaker
Description
First Name: Alexis
Last Name: Blaise
Affiliation: CEA-Saclay
All Authors: Alexis Blaise, Antoine Strugarek, Adam Finley, Miho Janvier, Eric Buchlin
Abstract: The Sun exhibits a cyclic magnetic activity that impacts our technological society on Earth. This activity originates in regions of concentrated magnetic flux called active regions. Active regions can produce impulsive eruptions, known as solar flares, when free magnetic energy that is built up by surface flows and magnetic field emergence is released via magnetic reconnection. Understanding the processes leading to this release is key to better anticipating these outbursts. However, to characterize the magnetic field configuration responsible for such eruptions, computationally expensive extrapolations are required and therefore limit extensive analyses of such regions. \ Non-linear force free field (NLFFF) extrapolation is one such method for reconstructing the topology of the magnetic field above active regions. Active regions that survive over several months can contribute significantly to the total flaring output of the Sun, and would require hundreds of extrapolations to be monitored correctly. Typically, NLFFF extrapolations are computationally expensive and so are not well suited to study the build-up of free magnetic energy in these long-lived active regions. \ To alleviate this limitation, we used a physics informed neural network (PINN), called NF2, to efficiently reconstruct the active region magnetic field. The PINN is driven by observations of the surface magnetic field and aims to satisfy the equations of the NLFFF extrapolation in a 3D domain, whilst obeying physical laws, like $\overrightarrow{\nabla} . \overrightarrow{\text{B} } = 0$. \ The PINN finds the best weights for the 8 layers of 256 neurons to compute magnetic field at a random point sampled in the 3D volume [Nx,Ny,Nz] as inputs, that satisfy the surface magnetic field, NLFF equations, and divergence-freeness of the magnetic field. To increase efficiency, the PINN is trained iteratively using the weights from the previous time interval. This allows the PINN to produce extrapolations at a much higher speed than standard methods with the same quality.\ For our project, we were able to extrapolate nearly 800 magnetograms (with one extrapolation every 3-4 minutes), allowing us to monitor the build-up of free magnetic energy over periods of weeks to months so as to characterize potential processes at the origin of the flaring activity of a sustained active region. \ PINNs are interesting new techniques that can help accelerate progress in solar physics. We may in a near future be able to reconstruct active region magnetic field efficiently enough to use it in real time as indicators for solar eruptions or as input for solar wind models in open field lines configurations.\