ML in solar physics | Neelanchal Joshi

Over the last 100 years, the field of solar physics has seen an unprecedented amount of data being generated. Space and ground based telescopes provide long-term, continuous coverage of the sun. Naturally, with such an explosion of information, our ability to extract meaningful science is also enhanced. In parallel, the field of machine learning has also burgeoned in the last few years, allowing a harmonic amalgamation of these two fields.

The Kodaikanal Solar Observatory (KSO) has been taking daily images of the sun since 1904. Long term continuous coverage of the sun is essential in understanding the solar dynamo and studying the 22 year solar cycle. Studying this phenomenon is not only crucial in understanding solar dynamics, but also in understanding its effects on space weather and consequently, on Earth.

The project aimed to generate high resolution magnetograms from Ca-II-K images from the Kodaikanal Dataset.

Image-to-Image translation

Image to image translation is a machine learning problem that aims to learn the relation between a set of images and their corresponding output images and using these output images for various tasks such as style transfer etc.

Example of style transfer using an unsupervised generative adversarial network (Credits: Unpaired Image-to-Image Translation Using Cycle-Consistent Adversarial Networks(ICCV 2017))

For our purpose, we used a supervised conditional generative adversarial network developed by Nvidia, in collaboration with UC Berkeley, called the pix2pixHD. In the following sections, I will describe these jargons. Don’t worry.

*** Side note: Nobel prize in Physics for 2024 is awarded for the development of artificial neural networks!!!

Conditional Generative Adversarial Networks

Now, the first very fundamental question that might arise is - Wha… Whattt?
Yep. Good Question.

Well, cGANs basically are a type of generative adversarial networks that generate outputs based on a set of conditions. And GANs are special types of neural networks (well, two networks) that perform this task in a very special way. One of the networks, called the Generator, generates images using an input latent vector while the second network, called the Discriminator, tries to distinguish the fake image from the real one.

Example schematic of how a conditional GAN works vs how a standard GAN works (Credits: https://learnopencv.com/conditional-gan-cgan-in-pytorch-and-tensorflow/)

Now, a very special type of cGAN that we used for this project is the pix2pixHD cGAN developed by Nvidia and UC Berkeley. This is a high resolution cGAN that can generate images based on semantic maps. You can read the paper here.

How was this used for our benefit?

We use this high resolution network in order to translate a set of Ca II K solar images (basically the images of the sun when looked through a filter of a certain wasvelength). This Ca II K line is found to be a close proxy to the unsigned magnetic field in the sun. The figure below shows a schematic of how this network aims to translate these images into magnetograms.

The Pix2PixHD architecture consists of 3 separate networks: a generator and two discriminators. The generator is trained to generate a pseudo magnetogram of 512 x 512 pixels from a KSO Ca II K image, while the job of the two discriminators is to determine which of the two supplied pairs of images contains the real magnetogram. The loss generated by the discriminators is used to update the generator weights to output a better pseudo-magnetogram. This process is repeated until a Nash equilibrium is reached.

Using this technique, the network was able to learn the relationship between Ca II K images and magnetograms and could generate realistic magnetograms given a random Ca II K image.

Handling polarity

Because Ca II K images capture only the unsigned magnetic field, they carry no information about polarity (the sign of the magnetic field). This has to be reconstructed separately, using Hale’s polarity law: the leading polarity of active regions (the polarity on the side facing the direction of solar rotation) reverses from one solar cycle to the next. Polarity is therefore assigned systematically based on which solar cycle a given observation belongs to.

Results

The generated pseudo-magnetograms reproduce the large-scale distribution of magnetic flux across the solar disk with good qualitative agreement to real magnetograms. Performance is strongest in regions of strong magnetic field; the network is comparatively weaker in quiet-Sun regions, where the Ca II K signal is faint and its correspondence with the true magnetic field becomes less direct.

Comparison between the real MDI magnetogram and the generated pseudo-magnetogram given a Ca II K input image.

Why this matters

The real value of this approach lies in the temporal reach it unlocks. Direct magnetogram observations only became available from the 1970s onward, while the KSO Ca II K archive extends back to 1904 — spanning several additional solar cycles. By learning a reliable mapping from Ca II K images to magnetograms, this technique offers a path to extending the Sun’s usable magnetic record by decades, providing a substantially longer baseline for studying the solar dynamo and long-term solar magnetic activity.

More broadly, this project is an example of using machine learning not simply as a data-analysis tool, but as a way to bridge an observational gap — recovering physical information from an archival dataset that was never originally designed to provide it.