AI Model Predicts Solar Storms 9 Hours Early, Offering Critical Warning Window

Beneath the Sun's visible surface, where no telescope can reach, magnetic activity begins its slow rise toward the light. By the time a sunspot darkens the solar disk—by the time we can see it—the forces that created it have already been working in the dark for hours. A team at the New Jersey Institute of Technology has built an artificial intelligence system that can detect those hidden stirrings nearly nine hours before they break through, a window of warning that could reshape how satellite operators and power grid managers prepare for space weather.

The system, called EarlyDetect, works by listening to the Sun in two ways at once. It monitors acoustic waves—sound traveling through the star's interior, the same principle geologists use to study earthquakes—and it tracks the magnetic field's subtle shifts. These signals have always been there, but they've been nearly impossible to read with enough clarity to make a forecast. The magnetic structures that will eventually become active regions develop out of sight, their presence betrayed only by faint disturbances in the acoustic rhythm and magnetic patterns constantly flowing across the star's surface. Alexander Kosovichev, a physics professor at NJIT and co-principal investigator on the project, described the challenge as detecting a single instrument's change in rhythm within an orchestra playing at full volume—the signal exists, but it's buried under relentless noise.

The research team, led by Jonas Tirona, an NJIT undergraduate, published their findings in August in the Journal of Geophysical Research: Machine Learning and Computation. They fed their model hourly maps of acoustic activity and magnetic field data collected by NASA's Solar Dynamics Observatory, which captures sound wave readings every 45 seconds through an instrument called the Helioseismic and Magnetic Imager. What happened next surprised them. When they tested the model, they discovered that filtering the data—a standard step meant to clean up noise and make patterns clearer, the way noise-canceling headphones work—actually made the forecasts worse. The filtering was averaging away the very faint fluctuations that contained the earliest warning signs. Once they removed that filtering step and trained EarlyDetect on raw, unfiltered data using a Transformer architecture (the same technology underlying systems like ChatGPT), the model's performance jumped. The strongest version detected precursor patterns an average of 9.24 hours before active regions became visible on the Sun's surface, outperforming both a standard Transformer model and earlier benchmark methods.

The practical stakes are real. Active regions take several hours just to emerge and can require one to several days to fully develop once they do. A nine-hour warning gives satellite communications companies and power grid operators time to take protective steps—to move critical systems offline, to reroute power, to prepare. That buffer could mean the difference between a solar storm that causes damage and one that doesn't.

But Tirona is careful about what this system can and cannot do. EarlyDetect still generates false alarms. Detecting an emerging active region does not guarantee that a solar flare or coronal mass ejection will actually follow. The model is not yet ready for real-time operational forecasting. Years of testing across far more solar events lie ahead before this approach moves from the laboratory into the control rooms where it might actually protect infrastructure. What the team has demonstrated, though, is that machine learning can find patterns in the Sun's hidden activity that human observers and traditional methods have missed—and that sometimes, the path to clearer vision runs through accepting more noise, not less.