Physically Motivated Deep Learning to Superresolve and Cross Calibrate Solar Magnetograms

Andrés Muñoz‐Jaramillo, Anna Jungbluth, Xavier Gitiaux, Paul Wright, Carl Shneider, Shane A. Maloney, Atılım Güneş Baydin, Yarin Gal, Michel Deudon, Freddie KalaitzisThe Astrophysical Journal Supplement Series 2024 · DOI

Abstract

Superresolution (SR) aims to increase the resolution of images by recovering detail. Compared to standard interpolation, deep learning-based approaches learn features and their relationships to leverage prior knowledge of what low-resolution patterns look like in higher resolution. Deep neural networks can also perform image cross-calibration by learning the systematic properties of the target images. While SR for natural images aims to create perceptually convincing results, SR of scientific data requires careful quantitative evaluation. In this work, we demonstrate that deep learning can increase the resolution and calibrate solar imagers belonging to different instrumental generations. We convert solar magnetic field images taken by the Michelson Doppler Imager (resolution ∼2″ pixel −1 ; space based) and the Global Oscillation Network Group (resolution ∼2.″5 pixel −1 ; ground based) to the characteristics of the Helioseismic and Magnetic Imager (resolution ∼0.″5 pixel −1 ; space based). We also establish a set of performance measurements to benchmark deep-learning-based SR and calibration for scientific applications.

Introduction: The Challenge of Multi-Generational Solar Data

To understand the Sun’s 11-year magnetic cycle and predict space weather events, heliophysicists require long-term, consistent data. While approximately 50 years of "magnetograms"—maps of the solar magnetic field—exist, they are a patchwork of data from various instruments. Older sensors, such as the space-based Michelson Doppler Imager (MDI) and the ground-based Global Oscillation Network Group (GONG), provide much lower resolution and different noise profiles than modern flagship instruments like the Solar Dynamics Observatory's Helioseismic and Magnetic Imager (SDO/HMI).

Timeline of solar magnetic field observations across different instrumental generations, showing varying pixel resolutions and coverage periods.

Standard image processing techniques like bicubic interpolation simply "stretch" old images, enlarging the blur without adding information. Furthermore, scientific data must preserve physical properties—such as magnetic flux balance and extreme values in sunspots—which standard AI enhancers often distort.

Methodology: Physically Motivated Deep Learning

The researchers adapted a convolutional neural network (CNN) architecture known as HighRes-net to bridge these instrumental gaps. The framework incorporates two critical innovations to ensure scientific validity:

  1. Location-Awareness: A "location channel" is added to the input. Since the Sun is a sphere, magnetic features near the limb (edge) appear differently due to projection effects. This channel provides the network with spatial context on the solar disk.
  2. Physically Motivated Loss Function: Moving beyond simple Mean Squared Error (MSE), which often results in blurry "average" outputs, the authors implemented a multi-term loss function including $\mathcal{L}_{l2}$ for pixel accuracy, $\mathcal{L}_{grad}$ for sharp polarity boundaries, and $\mathcal{L}_{hist}$ to ensure the statistical distribution of magnetic field strengths matches physical reality.
The adapted HighRes-net architecture featuring encoding, decoding with bilinear upsampling, and a final super-resolution stage.

Datasets and Evaluation

The study leveraged overlapping observation periods to create "ground truth" pairs, using MDI and GONG as source data and HMI as the high-resolution target. Evaluation focused on high-interest areas for solar flare prediction, known as Space Weather HMI Active Region Patches (SHARPs). Performance was benchmarked using scientific metrics including Pearson correlation, signed flux conservation, and the reproduction of extreme magnetic values.

Key Insights and Results

The deep learning approach demonstrated significant improvements over traditional methods:

  • Fidelity and Gradients: ML-based superresolution (SR) increased Pearson correlation with high-resolution targets by 5%–7% over bicubic upscaling. For MDI data, gradient errors were reduced by 30%.
  • Implicit Denoising and Error Correction: The network learned to distinguish instrumental noise from actual solar features. Remarkably, it also mitigated a known 24-hour systematic error in HMI data caused by satellite orbital velocity, as the input instruments (MDI/GONG) did not share this specific artifact.
  • Physical Consistency: The model successfully maintained the "divergence-free" nature of magnetic fields, ensuring that total positive and negative flux remained physically plausible even at higher resolutions.

Conclusion and Future Directions

While the current frame-by-frame processing shows great promise, future work could incorporate temporal information (multiframe SR) to further enhance detail. Additionally, implementing probabilistic machine learning could provide uncertainty quantification, informing scientists of the model's confidence in specific reconstructed features. This work establishes a robust path for creating a high-fidelity, 50-year record of the Sun's magnetic evolution.

Summary updated

Abstract

Abstract Superresolution (SR) aims to increase the resolution of images by recovering detail. Compared to standard interpolation, deep learning-based approaches learn features and their relationships to leverage prior knowledge of what low-resolution patterns look like in higher resolution. Deep neural networks can also perform image cross-calibration by learning the systematic properties of the target images. While SR for natural images aims to create perceptually convincing results, SR of scientific data requires careful quantitative evaluation. In this work, we demonstrate that deep learning can increase the resolution and calibrate solar imagers belonging to different instrumental generations. We convert solar magnetic field images taken by the Michelson Doppler Imager (resolution ∼2″ pixel −1 ; space based) and the Global Oscillation Network Group (resolution ∼2.″5 pixel −1 ; ground based) to the characteristics of the Helioseismic and Magnetic Imager (resolution ∼0.″5 pixel −1 ; space based). We also establish a set of performance measurements to benchmark deep-learning-based SR and calibration for scientific applications.
BibTeX
@article{munozjaramillo2024physically,
  author = {Andrés Muñoz‐Jaramillo and Anna Jungbluth and Xavier Gitiaux and Paul Wright and Carl Shneider and Shane A. Maloney and Atılım Güneş Baydin and Yarin Gal and Michel Deudon and Freddie Kalaitzis},
  title = {{Physically Motivated Deep Learning to Superresolve and Cross Calibrate Solar Magnetograms}},
  journal = {The Astrophysical Journal Supplement Series},
  year = {2024},
  doi = {10.3847/1538-4365/ad12c2},
}
Figure