Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Out-of-Sample Validation of MagNet

Aryiadna Yesmanchyk, Yan Xu, Jason T. L. Wang, Haodi Jiang, Chunhui Xu, Haimin Wang

TL;DR

MagNet addresses the lack of vector magnetograms in historic SOHO/MDI LOS data by learning to map co-aligned $Hα$ images and LOS $B_z$ to transverse fields $B_x$ and $B_y$, with ground truth from SDO/HMI. The paper implements an out-of-sample (OOS) validation framework using Mees/IVM data not seen during training to assess generalization. The results show a strong correlation between $B_z$ observed by MDI and Mees/IVM ($r \approx 0.94$) and a satisfactory correlation for transverse-field strength $B_t=\sqrt{B_x^2+B_y^2}$ against Mees/IVM ($r \approx 0.78$), with $B_t$ avoiding the 180° ambiguity. This validates MagNet for application to the full SOHO/MDI archive and future analyses of AI-generated vectors.

Abstract

Machine learning is starting to be used in almost every industry and academic research, and solar physics is no exception. A newly developed machine learning model named MagNet helps us to tackle some of the most serious challenges in data mining by generating transverse fields of solar active regions. Being trained on line-of-sight magnetograms from Michelson Doppler Imager at Solar and Heliospheric Observatory (SOHO/MDI), Hα maps from Big Bear Solar Observatory and Kanzelhohe Solar Observatory and vector magnetograms from Helioseismic and Magnetic Imager at Solar Dynamic Observatory (SDO/HMI), this model provides vector magnetograms in active regions for SOHO/MDI data covering the strong solar cycle 23. In this study, we performed out-of-sample validation of the MagNet model with data from Imaging Vector Magnetograph (IVM) at Mees Solar Observatory, which was not included in the training process. Our results show good correlation between the AI generated data and the observed vector magnetograms and therefore strengthen the confidence of implementing MagNet to the entire SOHO/MDI archive and future scientific analysis of the AI generated data.

Out-of-Sample Validation of MagNet

TL;DR

MagNet addresses the lack of vector magnetograms in historic SOHO/MDI LOS data by learning to map co-aligned images and LOS to transverse fields and , with ground truth from SDO/HMI. The paper implements an out-of-sample (OOS) validation framework using Mees/IVM data not seen during training to assess generalization. The results show a strong correlation between observed by MDI and Mees/IVM () and a satisfactory correlation for transverse-field strength against Mees/IVM (), with avoiding the 180° ambiguity. This validates MagNet for application to the full SOHO/MDI archive and future analyses of AI-generated vectors.

Abstract

Machine learning is starting to be used in almost every industry and academic research, and solar physics is no exception. A newly developed machine learning model named MagNet helps us to tackle some of the most serious challenges in data mining by generating transverse fields of solar active regions. Being trained on line-of-sight magnetograms from Michelson Doppler Imager at Solar and Heliospheric Observatory (SOHO/MDI), Hα maps from Big Bear Solar Observatory and Kanzelhohe Solar Observatory and vector magnetograms from Helioseismic and Magnetic Imager at Solar Dynamic Observatory (SDO/HMI), this model provides vector magnetograms in active regions for SOHO/MDI data covering the strong solar cycle 23. In this study, we performed out-of-sample validation of the MagNet model with data from Imaging Vector Magnetograph (IVM) at Mees Solar Observatory, which was not included in the training process. Our results show good correlation between the AI generated data and the observed vector magnetograms and therefore strengthen the confidence of implementing MagNet to the entire SOHO/MDI archive and future scientific analysis of the AI generated data.
Paper Structure (5 sections, 4 figures)

This paper contains 5 sections, 4 figures.

Table of Contents

  1. Introduction
  2. Preparing OOS Data
  3. Generating Input Images for MagNet
  4. MagNet implementation
  5. Validation

Figures (4)

  • Figure 1: MagNet workflow. H$\alpha$ and LOS are MagNet's inputs. B'$_{x}$ and B'$_{y}$ are fields generated by MagNet. Credit: Jiang2023
  • Figure 2: AR extraction for AR 09643 on 2001-May-21. The left panel shows original full-disk H$\alpha$ image taken at 16:01:04 UT and the selected AR image. Similar maps of the MDI magnetogram taken at 15:59:41 UT are shown in the right panel.
  • Figure 3: AI generated transverse magnetograms, B$_y$(left) and B$_x$(right) of AR 09643 on May 21, 2001.
  • Figure 4: Top panels: Comparison between observed MDI B$_z$ with Mees/IVM B$_z$. Bottom panels: Comparison between AI generated vector field B$_t$ with "ground truth" Mees/IVM observations.