AI detects faint signals from sunspots forming beneath the Sun’s surface, offering hours of advance warning before they become visible.
The version built to twitch at small changes gave more warning than the version built to be accurate.
Averaged across the test regions, an older, recurrent network was 0.14 hours late, which is essentially on time.
That’s the problem the team set out to fix.
That version, which they call EarlyDetect, flagged the five test regions a median of 9.40 hours before the surface changed, and 4.73 hours early on average.
AI detects faint signals from sunspots forming beneath the Sun’s surface, offering hours of advance warning before they become visible.
Satellite operators and power companies want more notice before a solar storm, but the earliest notice comes from a part of the Sun that nobody can look at directly.
Researchers have now trained a model to read the faint signals that run ahead of a new sunspot. The version built to twitch at small changes gave more warning than the version built to be accurate.
That warning is a narrow one. It means a magnetically active patch is about to appear. It doesn’t mean the patch will throw a flare or a coronal mass ejection at anyone, and plenty of active regions never do.
The model isn’t forecasting anything yet, either. It learned from emergences the team already knew about, and it was tested on five of them.
Sunspots start forming out of sight
A sunspot is the visible end of a longer process. Magnetic flux rises from deep inside the Sun to the photosphere, the surface layer that we see. By the time it dims a patch there, the process has been running for hours.
Flares and eruptions from those patches can inject energetic particles into Earth’s magnetic field and affect satellites and other technology that depends on it. So the warning signs have been worth hunting for a long time.
On the way up, the rising flux disturbs the sound waves traveling through the Sun’s interior.
NASA’s Solar Dynamics Observatory carries an instrument, the Helioseismic and Magnetic Imager, that records the up-and-down motion of the surface every 45 seconds and maps the magnetic field alongside it.
Alexander Kosovichev is a distinguished professor of physics at the New Jersey Institute of Technology (NJIT) and a co-principal investigator on the project.
“The main difficulty is that an active region begins developing beneath the Sun’s visible surface, where we cannot directly observe the magnetic structure,” he said.
Low error did not mean early warning
Jonas Tirona, an undergraduate researcher at NJIT and the study’s corresponding author, built the approach with colleagues at NJIT, Princeton University and NASA’s Ames Research Center.
The team worked from a public set of 50 tracked regions observed by that instrument.
Four of the regions came with data gaps and were dropped. The team trained on 41 of the rest and held five back for testing.
Each model read 110 hours of measurements at a time and predicted the next 12. Five channels went in: acoustic power in four frequency bands, plus the magnetic field along the line of sight.
Two scores mattered. One was ordinary error – how close the predicted surface brightness was to the real thing. The other was timing, and it split the models apart.
Averaged across the test regions, an older, recurrent network was 0.14 hours late, which is essentially on time.
A standard transformer, the architecture behind chatbots and much current weather forecasting work, scored better on error and came in 8.27 hours late.
A forecast that arrives after the event isn’t much of a forecast. That’s the problem the team set out to fix.
Smoothing erased the earliest signals
Their first idea was a filter. A convolutional layer in front of the model was supposed to pull short-timescale patterns out of noisy data, and it did the opposite.
“That surprised us most,” Kosovichev said. “We initially expected it to help isolate useful short-timescale patterns. Instead, it averaged away the very faint fluctuations that provided the earliest warning.”
Tirona put it in household terms. The filter, Tirona said, worked “kind of like noise canceling” but “was detrimental in almost every case.”
“The signals that the filter removed turned out to be really important in helping the model predict when an active region would emerge.”
Adding the filter to their best model pushed its timing from 4.73 hours early to 7.20 hours late, and grew it from 5.8 million internal parameters to 19.1 million.
The jumpy model warned earliest
What worked was building the impatience in. The team weighted the model’s attention toward the start of each window, then penalized it for late calls while letting early ones go free.
That version, which they call EarlyDetect, flagged the five test regions a median of 9.40 hours before the surface changed, and 4.73 hours early on average.
Three of the five came in early, against two of five for every other version they tried.
It also posted the lowest error of the group, about 11 percent better than that older baseline. And the margins held under a tougher bar than earlier work used.
A dip in brightness only counted as an emergence here if it lasted four straight hours instead of three. That rules out the brief flickers that would have flattered the numbers.
False alarms keep it out of forecasting
The sensitivity comes with noise. Timing varied by more than 14 hours from one region to the next.
On one region the model went off on individual patches as much as 70 hours ahead, missed another patch entirely, and raised a false alarm on a third.
Real-time use adds a delay of its own. Turning raw surface-motion data into the maps that the model reads takes about four hours, which leaves roughly eight of the 12-hour horizon.
“Machine learning hasn’t been widely applied to solar activity forecasting yet,” said Mengjia Xu, an assistant professor of data science at NJIT and the project’s principal investigator.
Her group released the code and the trained models, and the set of tracked regions behind them is public too.
Whether those five test regions were representative, nobody can say yet.
The researchers want the same design pointed at a much larger set, and at the magnetic flux itself, which starts changing before the surface does.
They also want to feed it neighboring patches of the Sun rather than one strip at a time.
The full study was published in the journal Journal of Geophysical Research: Machine Learning and Computation.
—–
Like what you read? Subscribe to our newsletter for engaging articles, exclusive content, and the latest updates.
Check us out on EarthSnap, a free app brought to you by Eric Ralls and Earth.com.
—–
