Long before a sunspot darkens the Sun’s visible surface, the magnetic forces that create it may already be leaving a faint trail beneath the surface. A new artificial intelligence model developed by researchers at the New Jersey Institute of Technology (NJIT) has detected those subtle precursors and forecast the emergence of solar active regions nearly nine hours in advance on average. The system, called EarlyDetect, is designed to recognize changes in the Sun’s acoustic activity, magnetic field and surface brightness before an emerging active region becomes clearly visible. The result could mark an important step toward earlier warnings of the solar disturbances that can disrupt satellites, radio communications, navigation systems and electrical grids.
The research, published in the Journal of Geophysical Research: Machine Learning and Computation, focuses on active regions: magnetically complex areas of the Sun where sunspots form and powerful flares or coronal mass ejections may eventually originate. These regions do not appear instantly. Magnetic fields generated deep inside the Sun gradually rise through its turbulent interior and begin interacting with the visible surface over a period of several hours. Their complete development can take one or more days. During the earliest stage, the magnetic structure remains hidden from direct observation, making the emergence difficult to identify with conventional monitoring techniques. EarlyDetect attempts to solve that problem by learning how hidden magnetic evolution changes measurable signals at the surface.
The model was developed by NJIT undergraduate researcher Jonas Tirona with solar physicists and computer scientists at NJIT, Princeton University and NASA’s Ames Research Center. The team trained it using observations from NASA’s Solar Dynamics Observatory, or SDO, a spacecraft that continuously monitors the Sun. In particular, the researchers used data from SDO’s Helioseismic and Magnetic Imager, which records the Sun’s surface magnetic field and Doppler velocity every 45 seconds. Doppler measurements reveal motions toward and away from the spacecraft, allowing scientists to study the acoustic waves that travel through the solar interior. By processing these observations into hourly maps, the team created a time-dependent view of how a potential active region changes before it becomes visible.
The model does not simply look for a new dark spot. Instead, it analyzes several related signals and searches for patterns that precede emergence. One of the earliest signatures appears in acoustic power, a measurement of the strength of solar oscillations inferred from Doppler-velocity observations. As magnetic fields rise through the Sun, they can alter the propagation and behavior of acoustic waves. These changes may occur before the magnetic concentration is strong enough to produce a conspicuous surface feature. A later sequence of changes can appear in continuum intensity, which measures the Sun’s visible brightness, followed by more direct alterations in the magnetic field. EarlyDetect is designed to connect these weak, evolving signals rather than treating each observation as an isolated event.
To accomplish this, the researchers used a Transformer architecture, a class of machine-learning model best known for powering large language systems such as ChatGPT. Transformers are effective at analyzing sequences because they can compare information across widely separated points in time and determine which changes are most important. In EarlyDetect, the sequence is not made of words but of solar observations arranged chronologically. The model learns relationships among acoustic power maps, continuum-intensity data and magnetic-field measurements, then estimates whether the observed pattern is likely to develop into an emerging active region. This approach allows the system to consider both the strength of a signal and the order in which different signals appear.
One of the study’s most unexpected findings came from an attempt to make the data easier for the model to interpret. The researchers applied a filtering technique intended to suppress noise and emphasize broader trends. Instead, the filtering reduced the model’s performance. By averaging away rapid, low-amplitude fluctuations, it removed some of the faint variations that carried the earliest information about an emerging region. Alexander Kosovichev, distinguished professor of physics at NJIT and co-principal investigator, compared the problem to trying to detect a slight change in rhythm within a noisy orchestra. In this case, what appeared to be noise may have contained precisely the subtle short-timescale variations that distinguished an emerging region from ordinary solar turbulence.
After training on SDO/HMI observations, the researchers evaluated EarlyDetect using active regions that had not been included in the model’s training data. The best-performing version identified precursor signals an average of 9.24 hours before the regions became visible. It outperformed both a standard Transformer model and an earlier benchmark method used for comparison. The result suggests that advanced machine learning can extract predictive information from solar measurements that may be too faint, complex or variable for traditional analysis methods to identify consistently. However, the researchers emphasize that the reported average is not a guaranteed warning time for every event. Some regions may be detected earlier, while others may generate a late warning or no reliable warning at all.
The potential importance of earlier detection extends beyond predicting when a sunspot will appear. Active regions can produce solar flares, which release intense bursts of radiation, and coronal mass ejections, which propel clouds of magnetized plasma into space. When directed toward Earth, these events can disturb the planet’s magnetosphere, degrade high-frequency radio communication, interfere with satellite operations and induce electrical currents in long transmission lines. An alert that a magnetically active region is developing could give satellite operators, communication providers and power-grid managers additional time to review protective procedures. Yet the emergence of an active region is not itself a forecast that a major flare or coronal mass ejection will occur. Many active regions remain relatively quiet, and the relationship between emergence and later eruptions is not simple.
The researchers therefore describe EarlyDetect as a promising research prototype rather than an operational space-weather forecasting system. It was trained on known emergence events and still produces false alarms and late predictions. Its performance must be tested across a much larger and more diverse collection of solar events, including different phases of the Sun’s approximately 11-year activity cycle. The team has released the Solar Active Region Emergence Dataset, known as SolARED, to support that work. The dataset compiles observations from SDO and is accompanied by the Solar Active Region Portal, an interactive platform that allows researchers to explore the observations and develop alternative prediction methods. By making the data publicly available, the team hopes to create a shared benchmark for both heliophysics and artificial-intelligence research.
The project was among the first research efforts supported by NJIT’s Grace Hopper AI Research Institute, launched in 2025 to encourage interdisciplinary work in artificial intelligence. Additional support came from NASA heliophysics and space-weather research programs, including the NASA Science DRIVE Center Consequences Of Fields and Flows in the Interior and Exterior of the Sun. For Tirona and his collaborators, the study demonstrates how machine learning can turn continuous streams of solar measurements into early clues about events hidden beneath the Sun’s surface. A dependable operational system remains a distant goal, but the ability to identify active-region emergence hours before it becomes visible offers a new way to investigate the Sun—and could eventually provide an earlier line of defense against its most disruptive outbursts.
Subject of Research: Solar active-region emergence and machine-learning-based space-weather forecasting
Article Title: Forecasting Continuum Intensity for Solar Active Region Emergence Prediction Using Transformers
News Publication Date: 14-Aug-2026
Web References: https://doi.org/10.1029/2025JH001207; https://sun.njit.edu/sarportal/
References: Journal of Geophysical Research: Machine Learning and Computation, DOI: 10.1029/2025JH001207
Image Credits: Irina N. Kitiashvili, NASA Ames Research Center, and Spiridon Kasapis, Princeton University
Keywords
Artificial intelligence, machine learning, solar physics, solar active regions, space weather, solar magnetic fields, helioseismology, Transformer models, NASA Solar Dynamics Observatory, solar flares, coronal mass ejections, Sun
