Virtual sunspots help AI find rare magnetic matches in vast solar archives

Research led by Southwest Research Institute (SwRI) has integrated three types of machine learning models to generate solar magnetic patches with physical properties and used those as a query to find matching patches in real observations. This elevates generative artificial intelligence (AI) from a means to produce artificial data to a novel tool for scientific data interrogation, supporting applicability beyond the heliophysics domain. The paper is published in The Astrophysical Journal Supplement Series.

“Modern astronomical observatories may produce millions of gigabytes of data during their lifetime,” said SwRI’s Dr. Subhamoy Chatterjee, first author of the paper. “Manually labeling and sifting through such a vast dataset is becoming impossible in a human lifetime. An even bigger problem is how to process these data and retrieve information hidden in such large datasets.”

Unraveling the rhythms of solar activity throughout its roughly 11-year cycle has fascinated scientists for over a century. Interpreting the patterns of solar active regions and their links to solar flares, coronal mass ejections, energetic particles and magnetic storms is crucial to protecting satellites and other Earth technology from space weather events. The active regions also carry information about the buildup of the sun’s polar magnetic field, which is pivotal to understanding solar processes and to forecasting future solar cycles.

“For example, rogue active regions of unusual size, tilt and location have been found to make substantial impacts on solar cycles,” said SwRI’s Dr. Andrés Muñoz-Jaramillo, the paper’s second author. “However, such regions are rare occurrences. To efficiently explore possible outcomes and their impacts on solar cycles, a scientist might want to create additional artificial examples.”

Deep generative models have immense potential in generating unseen artificial data exhibiting properties of real data. These models learn complex high-dimensional data-generating distributions starting from lower-dimensional hidden data. Connecting physical properties to the hidden data allows scientists to create virtual representations of regions with interesting properties and use these to re-analyze historical data to find equivalent features, without having to look at every active region in the prior data.

Such techniques build confidence in the accuracy of generative AI models through direct interaction with real data familiar to scientists.

“We used magnetic field measurements from Space-weather HMI Active Region Patches (SHARPs) to train a generative AI model,” Chatterjee said. “We then trained a model to connect the physical space and hidden generative space through ‘directions’ that correspond to different specific physical properties of active regions, including polarity, magnetic flux, complexity, flaring nature, etc.”

These derived connections allow a user to physically manipulate a generated active region image so that a particular property is varied. The team then trained another machine learning model to make queries with generated images and find matching real images.

“The generative and supervised model combination enables users to make generative model outcomes physically consistent,” said Dr. Anna Malanushenko, the paper’s third author, from the National Center for Atmospheric Research’s High Altitude Observatory. “Those outcomes can be used to retrieve real data that shares the same physical properties as the generated query.”

In heliophysics, this approach can serve as a generic framework for solving various problems such as instrument-to-instrument translation, artifact correction, reconstruction of far-side active regions and space weather forecasting.