🔭 astrophysics

Out-of-Sample Validation of MagNet

This study validates the MagNet machine learning model's ability to generate accurate vector magnetograms for SOHO/MDI data by demonstrating strong correlations with out-of-sample observations from the Imaging Vector Magnetograph (IVM), thereby confirming its reliability for analyzing the full SOHO/MDI archive and future solar physics research.

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

Published 2026-02-27

📖 4 min read☕ Coffee break read

CC BY 4.0

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

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the Sun as a giant, chaotic power plant. Sometimes, it lets off massive bursts of energy called solar flares or coronal mass ejections (CMEs). To predict these "storms" before they hit Earth, scientists need to understand the Sun's magnetic fields. Think of these magnetic fields as the invisible wiring and pressure systems inside that power plant.

However, there's a problem with our historical data.

The Missing Puzzle Piece

For many years (specifically during Solar Cycle 23, roughly 1996–2011), we had a telescope called SOHO/MDI that took pictures of the Sun. But it was like a camera that could only see the Sun from one angle: straight on. It could tell us how strong the magnetic field was pointing toward or away from us (like looking at a pole head-on), but it couldn't see the magnetic field twisting sideways across the surface.

To predict solar storms accurately, we need to see the whole 3D picture, including those sideways twists. But for that old era of the Sun, those sideways measurements simply don't exist in the archives.

Enter "MagNet": The AI Artist

Scientists developed a clever Artificial Intelligence model called MagNet. Think of MagNet as a highly skilled artist who has been trained to "imagine" the missing sideways magnetic fields.

Here is how the artist was trained:

The Inputs: The artist was shown pairs of images: a standard "straight-on" magnetic photo (from SOHO/MDI) and a special "H-alpha" photo (which shows the Sun's surface activity in a different color).
The Teacher: The artist learned by looking at modern, high-quality 3D magnetic maps from a newer telescope (SDO/HMI). The teacher said, "Look at this old-style photo and this H-alpha photo; here is what the real 3D magnetic field looks like."
The Result: After studying thousands of examples, MagNet learned to look at the old, incomplete photos and "paint in" the missing sideways magnetic fields, creating a complete 3D map.

The Big Test: The "Pop Quiz"

The researchers knew MagNet was good at memorizing the examples it was trained on. But would it be smart enough to handle a new situation it had never seen before? This is called Out-of-Sample Validation.

Imagine a student who memorized the answers to a practice test. They get an A. But then, you give them a completely different test with new questions. Do they still get an A, or did they just memorize the answers?

To test MagNet, the scientists didn't use the data it was trained on. Instead, they used data from a different telescope on Earth (the Mees Solar Observatory) that observed a specific sunspot (AR 09463) back in 2001. This data was never shown to MagNet during its training.

The Results:

MagNet looked at the old, incomplete photos of that 2001 sunspot.
It generated a prediction of what the sideways magnetic fields should look like.
The scientists compared MagNet's prediction to the actual measurements taken by the Mees telescope.

The Verdict: MagNet got it right! The correlation was very high (0.78 out of 1.0). This proved that MagNet didn't just memorize the training data; it actually learned the underlying rules of how solar magnetic fields work.

Why This Matters

This is a huge deal because:

Unlocking History: Now, scientists can use MagNet to "fill in the blanks" for the entire archive of Solar Cycle 23. We can finally have complete 3D magnetic maps for the Sun's most active years in the past.
Better Predictions: With this complete historical data, we can better understand what causes massive solar storms. This helps us improve our ability to predict space weather, protecting our satellites and power grids on Earth.

In short: The researchers built an AI artist, taught it to imagine missing magnetic details, and then gave it a surprise test with a new sunspot. The AI passed with flying colors, proving it can help us rewrite the history of solar storms with much clearer pictures.

1. Problem Statement

Solar physics relies heavily on photospheric vector magnetograms to understand magnetic field topology, energy storage, and the prediction of solar eruptive events (flares and CMEs). While the Solar Dynamics Observatory (SDO) has provided consistent vector magnetograms since 2010 (Solar Cycle 24), the preceding Solar Cycle 23 (1996–2011) was characterized by stronger solar activity but lacked comprehensive vector data.

Data Gap: The Michelson Doppler Imager (SOHO/MDI) observed Solar Cycle 23 but only collected Line-of-Sight (LOS) magnetograms ( $B_z$ ), which are insufficient for analyzing complex active regions (ARs).
Need: There is a critical need to reconstruct the missing transverse magnetic field components ( $B_x$ and $B_y$ ) for the entire SOHO/MDI archive to enable retrospective analysis of Solar Cycle 23.
Challenge: Previous machine learning models (like MagNet) were trained and tested on overlapping data (MDI and SDO/HMI). There was a lack of Out-of-Sample (OOS) validation using independent data sources to ensure the model does not suffer from overfitting and can generalize to unseen data.

2. Methodology

The study focuses on validating the MagNet model, a deep learning framework designed to generate transverse magnetic fields from LOS magnetograms and H $\alpha$ maps.

Model Architecture & Inputs:
- Inputs: A pair of co-aligned images: an H $\alpha$ map (capturing chromospheric activity) and an MDI LOS magnetogram ( $B_z$ ).
- Outputs: Reconstructed transverse magnetic field components ( $B_x$ and $B_y$ ).
- Training Data: Previously trained on SOHO/MDI, SDO/HMI vector magnetograms (ground truth), and H $\alpha$ maps from Big Bear Solar Observatory (BBSO) and Kanzelhöhe Solar Observatory (KSO).
Data Preprocessing (OOS Setup):
- Source: The study utilized data from the Imaging Vector Magnetograph (IVM) at Mees Solar Observatory (Maui, Hawaii), which was never used during the MagNet training process.
- Alignment: H $\alpha$ images were truncated and aligned to the MDI reference images using Fourier transform-based cross-correlation (via IDL code align.pro) to match solar disk geometry.
- Region Selection: Active Region 09463 (observed on May 21, 2001) was selected. This date falls outside the MDI-HMI overlap period, ensuring a true OOS test.
- Cropping: Images were cropped to a $256 \times 256$ pixel field of view centered on the AR.
Validation Strategy:
- The MagNet-generated vector fields were compared against the "ground truth" Mees/IVM vector magnetograms.
- Metric: The study focused on the transverse field magnitude ( $B_t = \sqrt{B_x^2 + B_y^2}$ ) to avoid the 180-degree ambiguity inherent in vector magnetograms (where the direction of the transverse field is uncertain due to the Zeeman effect).
- Correlation Analysis: Pearson correlation coefficients were calculated between observed and generated data.

3. Key Contributions

First OOS Validation of MagNet: This is the first study to validate MagNet using an independent dataset (Mees/IVM) that was completely excluded from the training phase, moving beyond the standard in-sample validation.
Robustness Verification: The study demonstrates that MagNet can generalize its learned patterns to historical data (Solar Cycle 23) without relying on memorization of specific noise or fluctuations from the training set.
Methodological Framework: It establishes a rigorous workflow for aligning multi-instrument data (Space-based MDI vs. Ground-based IVM) and handling resolution differences (4" MDI vs. higher resolution H $\alpha$ ) for machine learning applications in solar physics.

4. Results

Longitudinal Component ( $B_z$ ): The correlation coefficient between the observed MDI LOS data and the Mees/IVM $B_z$ was 0.94, confirming the consistency of the observational data alignment.
Transverse Component ( $B_t$ ): The correlation coefficient between the AI-generated transverse field ( $B_t$ ) and the observed Mees/IVM $B_t$ was 0.78.
Interpretation: A correlation of 0.78 is considered satisfactory given the complexity of reconstructing vector fields from scalar LOS data and the inherent differences in instrument resolution and atmospheric effects. The use of $B_t$ successfully bypassed the 180-degree ambiguity, isolating the model's performance on field magnitude reconstruction.

5. Significance

Unlocking Historical Data: The successful OOS validation provides the confidence necessary to implement MagNet across the entire SOHO/MDI archive for Solar Cycle 23. This allows researchers to generate a complete vector magnetogram dataset for a period of intense solar activity that was previously limited to LOS data.
Scientific Impact: With reconstructed vector fields, scientists can now calculate crucial parameters like free magnetic energy and helicity for historical active regions, improving the understanding of solar eruption mechanisms and enhancing space weather prediction models.
Future Applications: The methodology paves the way for generating vector maps for other historical datasets and refining the model to better predict azimuthal angles, further bridging the gap between past observations and modern AI-driven solar physics.