Milne-Eddington Stokes Inversion of the NIRIS Magnetogram Data Achieved by Machine Learning Technique
Abstract
The Near InfraRed Imaging Spectrograph (NIRIS) at the Goode Solar Telescope produces Stokes I, Q, U, and V polarimetric profiles at a spectral resolution of 0.01 nm in 1564.8 nm band, with a typical range of -0.25 to +0.25 nm from the line center. This narrow band is achieved with a combination of a pre-filter and dual Fabry-Perot (F-P) etalon system. Typical line scan takes about 30 seconds. We use Milne-Eddington (ME) inversion technique to deduce physical parameters of an image pixel - such as total magnetic field strength, vertical/horizontal component of the magnetic field, Doppler shift of the line center and so on. However, due to many noise factors such inversion attempts are not always reasonable, especially when the profiles are complicated. Most of the ME fitting errors come from the initial guess of the longitudinal field strenghth calcuated from the center-of-gravity method.
We present our result of a new approach of inversion by using machine learning technique. Sagemaker, a new platform from Amazon Web Services, was adopted for training and modeling of the line profiles. We used principal component analysis (PCA) algorithm to deduce several physical parameters out of a trained model. This method not only reduces the dimension of the data but also enhances the speed of data processing. The result indicates that our model well fits into the actual measured line profiles as well as saving processing time. We present comparison of our new method to the ME inversion method in terms of accuracy and processing time.- Publication:
-
2018 Triennial Earth-Sun Summit (TESS)
- Pub Date:
- May 2018
- Bibcode:
- 2018tess.conf30818A