10:00 〜 10:15
[PEM10-05] Evaluation of horizontal flow velocity at the solar surface with machine learning and numerical simulation
キーワード：粒状斑
We develop a method to predict the horizontal velocity on the solar surface from snapshots of solar radiative intensity and vertical velocity using neural networks. The motion of solar plasma is related to many phenomena such as sunspot formation and flares. We can observe the line-of-sight (LOS) velocity field with the Doppler effect. In contrast, the velocity perpendicular to LOS is difficult to observe. Some methods estimate the horizontal velocity field by tracking the displacement of features on the solar surface, which require several images. Methods with a neural network for extracting the displacement are also proposed, but these methods also require images at two different times.
On the other hand, many numerical magnetohydrodynamic (MHD) simulations are carried out to explore the solar convection and magnetic field. They can reproduce the phenomena inside the sun with high accuracy thanks to recent improvements in computer performance. In this study, we develop a code for predicting the horizontal velocity field from the corresponding intensity and vertical velocity field by combining numerical MHD simulation and neural networks. We train the neural network using the results of our own MHD simulations. The networks output two components of the horizontal velocity field with the input of corresponding intensity and the vertical velocity field. The network is trained efficiently since we only use the convolutional neural network. As a result, the networks achieve a high correlation coefficient of 0.9 between the predicted and simulated velocity field. The horizontal velocity field can also be predicted from the image with any number of pixels.
On the other hand, many numerical magnetohydrodynamic (MHD) simulations are carried out to explore the solar convection and magnetic field. They can reproduce the phenomena inside the sun with high accuracy thanks to recent improvements in computer performance. In this study, we develop a code for predicting the horizontal velocity field from the corresponding intensity and vertical velocity field by combining numerical MHD simulation and neural networks. We train the neural network using the results of our own MHD simulations. The networks output two components of the horizontal velocity field with the input of corresponding intensity and the vertical velocity field. The network is trained efficiently since we only use the convolutional neural network. As a result, the networks achieve a high correlation coefficient of 0.9 between the predicted and simulated velocity field. The horizontal velocity field can also be predicted from the image with any number of pixels.