5:15 PM - 7:15 PM
[MGI30-P01] Experiment of Physics-Informed Neural Networks for Nonlinear Force-Free Field Extrapolation and its applicability
Keywords:Non linear force-free field, Solar Flares, Physics-informed Neural Networks, Space Weather
Fujitsu Limited and the Institute for Space-Earth Environmental Research (ISEE) at Nagoya University have been conducting joint research since September 2023 to enhance the safety of deep space exploration. In the previous fiscal year, we reported employing AI techniques called "causal discovery" to identify characteristic features of solar flares (SFs) as precursors to solar energetic particle (SEP) events (cf. 2024 JpGU, M-GI29, Kato et al.).
Existing physics-based methods, such as the κ-scheme (Ishiguro and Kusano, 2017; Kusano et al., 2020), utilize magnetohydrodynamic (MHD) relaxation method (Inoue et al., 2014) to extrapolate 3D coronal magnetic fields from photospheric vector magnetic field data observed by SDO/HMI. These Non-linear Force-Free Field (NLFFF) extrapolations are crucial for real-time space weather forecasting.
This study replicates the work of Jarolim et al. (2023, 2024), which uses Physics-Informed Neural Networks (PINNs) to derive NLFFFs. Their approach incorporates a loss function comprising force-free loss, divergence loss, and boundary conditions based on the photospheric magnetic field data observed by SDO/HMI. We focus on active region NOAA 11158 (which produced an X-class flare on February 15, 2011), comparing the PINNs-derived NLFFF with that obtained by the MHD relaxation method, both qualitatively and quantitatively. The predictive capabilities for flare occurrence will be discussed. Furthermore, we explore the potential of combining physics-based methods and AI approaches for broader applications in space, planetary, and Earth sciences.
Existing physics-based methods, such as the κ-scheme (Ishiguro and Kusano, 2017; Kusano et al., 2020), utilize magnetohydrodynamic (MHD) relaxation method (Inoue et al., 2014) to extrapolate 3D coronal magnetic fields from photospheric vector magnetic field data observed by SDO/HMI. These Non-linear Force-Free Field (NLFFF) extrapolations are crucial for real-time space weather forecasting.
This study replicates the work of Jarolim et al. (2023, 2024), which uses Physics-Informed Neural Networks (PINNs) to derive NLFFFs. Their approach incorporates a loss function comprising force-free loss, divergence loss, and boundary conditions based on the photospheric magnetic field data observed by SDO/HMI. We focus on active region NOAA 11158 (which produced an X-class flare on February 15, 2011), comparing the PINNs-derived NLFFF with that obtained by the MHD relaxation method, both qualitatively and quantitatively. The predictive capabilities for flare occurrence will be discussed. Furthermore, we explore the potential of combining physics-based methods and AI approaches for broader applications in space, planetary, and Earth sciences.