NASA's COFFIES research team has developed a machine-learning model that can predict where sunspots will emerge on the Sun up to 12 hours before they become visible. The model analyzes subtle changes in acoustic waves and magnetic fields beneath the solar surface, giving forecasters a head start on predicting solar flares and coronal mass ejections that can threaten astronauts, disable satellites, and disrupt radio communications on Earth.
The work comes from COFFIES (Consequence Of Fields and Flows in the Interior and Exterior of the Sun), a NASA DRIVE Science Center that brings together researchers from the New Jersey Institute of Technology, Princeton University, and NASA's Ames Research Center. Their findings were published in the Journal of Geophysical Research: Machine Learning and Computation.
Finding sunspots before they appear
The Sun's churning interior generates intense concentrations of localized magnetic fields that can suddenly break through the solar surface, forming sunspots. Space weather forecasters currently track these sunspots only after they appear, then estimate the probability of flares based on the visible regions' characteristics.
The COFFIES model takes a different approach. Instead of waiting for sunspots to emerge, it looks for precursors - tiny reductions in the Sun's acoustic activity and magnetic field that occur as a magnetic structure rises from the interior toward the surface. These signals have been difficult for scientists to capture until now.
"We cannot directly see the magnetic structure while it is still rising through the solar interior. Instead, we must look for indirect effects - very small changes in the magnetic field and in the pattern of acoustic waves continually traveling through the Sun," said Alexander Kosovichev, a COFFIES co-investigator at NJIT. "The developed technique identifies precursors associated with an emerging active region in slight changes of the Sun's acoustic power - more like a slight change in rhythm within a very noisy orchestra."
How the AI model works
The team used a specialized AI architecture called a sliding-window transformer to analyze data from NASA's Solar Dynamics Observatory, processed on supercomputing resources at NASA Ames. Unlike earlier deep learning approaches that examined all solar surface activity at once, this model moves a fixed-size "viewing window" across a long timeline of solar activity, focusing on recent data while retaining patterns from earlier observations.
This approach allows the model to predict approximate locations of emerging sunspots rather than merely counting those already visible. For professionals working in AI for Science & Research, the technique demonstrates how transformer architectures can be adapted to handle very long sequences of time-series data in scientific applications.
The team emphasized that the model is not yet ready for operational real-time forecasting. They plan to validate the approach across more known solar events to fine-tune its predictions.
Why space weather forecasting matters
As NASA prepares for Artemis missions to the Moon and eventual crewed missions to Mars, predicting space weather has become a priority. Solar eruptions send waves of high-energy radiation and charged particles across space, creating storms that can endanger astronauts and the equipment they depend on.
Teams across NASA and NOAA collaborate to transition research capabilities into operational space weather monitoring tools, including NASA's Space Radiation Analysis Group, the Moon to Mars Space Weather Analysis Office (M2M SWAO), and NOAA's Space Weather Prediction Center. Predicting sunspot region emergence - especially on the Sun's far side - could supplement the models these teams currently use.
"The COFFIES AI model is exciting to our team because it could provide us with new capabilities towards predicting potential flaring locations ahead of time," said Michelangelo Romano, M2M SWAO deputy director. "With this heads up, we can provide additional support to NASA missions."
Why this matters for research scientists
For researchers working in data-intensive fields, the COFFIES model offers a concrete example of how machine learning can extract predictive signals from noisy, high-volume observational data. The sliding-window transformer architecture is a transferable technique - any scientific domain that relies on long time-series data, from climate monitoring to astronomical surveys, could apply a similar approach to detect subtle precursors before major events occur. Those interested in applying these methods to their own research can explore an AI Learning Path for Research Scientists.
Your membership also unlocks:
