NASA Researchers Use AI to Spot Solar Active Regions Before They Become Visible
A COFFIES research model looks for subtle acoustic and magnetic precursors on the Sun, offering a potential earlier warning signal for space-weather forecasting — but NASA says it is not yet operational.
What happened
NASA described new research from the COFFIES collaboration — the Consequence Of Fields and Flows in the Interior and Exterior of the Sun DRIVE Science Center — that applies machine learning to the problem of detecting solar active-region emergence before the region is plainly visible on the solar surface. Active regions are concentrations of strong magnetic activity that can later produce sunspots, solar flares and…
Why it matters
- •Operational space-weather teams already monitor the Sun continuously, but much of the established workflow begins once…
- •The study uses Solar Dynamics Observatory observations, including acoustic-power measurements and line-of-sight…
- •The next phase is broader validation and refinement rather than immediate deployment. NASA says the team plans to test the…
Event timeline
- Published · Aug 16, 2026
- Updated · Aug 16, 2026
What to watch next
- The next phase is broader validation and refinement rather than immediate deployment.
- NASA says the team plans to test the approach across many more known solar events.
- The most important questions are whether the method maintains useful lead time across…
- Researchers also need to determine whether richer spatial inputs improve results without…
Sources
- NASA — COFFIES uses AI to predict solar active regionsprimary · Aug 14, 2026
- Journal of Geophysical Research: Machine Learning and Computationresearch
◷ 60-second read
- The study uses Solar Dynamics Observatory observations, including acoustic-power measurements and line-of-sight magnetic-field data, and tests a sliding-window Transformer architecture.
- The research paper evaluated 46 active regions and found that model timing performance varies from event to event, which is why the result should be treated as an experimental forecasting method rather than a finished service.
- NASA explicitly says the model is not ready for operational real-time forecasting and needs validation across more known solar events before any operational transition.
Related reading
Complete report
Background, mechanisms, consequences and uncertainty — beyond the dashboard.
What happened
NASA described new research from the COFFIES collaboration — the Consequence Of Fields and Flows in the Interior and Exterior of the Sun DRIVE Science Center — that applies machine learning to the problem of detecting solar active-region emergence before the region is plainly visible on the solar surface. Active regions are concentrations of strong magnetic activity that can later produce sunspots, solar flares and coronal mass ejections. Instead of waiting for a sunspot group to become visually established, the research looks for small changes in solar observations that may occur while magnetic structure is still rising through the Sun’s interior. NASA says the model can provide an early signal several hours before visible emergence, with forecasts extending as far as about 12 hours ahead in the study. The work combines solar-physics expertise with data-science methods and uses observations from NASA’s Solar Dynamics Observatory together with NASA Ames supercomputing resources.
What data the researchers used
The research paper focuses on time-series measurements derived from Solar Dynamics Observatory data, especially observations from the Helioseismic and Magnetic Imager. The model does not inspect a single photograph and simply guess where a sunspot will appear. It works on a sequence of measurements that represent how the local solar environment changes over time. The study uses acoustic-power information across multiple frequency bands together with line-of-sight magnetic-field measurements. Those features are intended to capture two different kinds of precursor behavior: changes in waves traveling through the solar interior and changes in magnetic structure at or near the surface. The paper reports experiments using 46 active regions, a useful research sample but still a relatively small data set for a machine-learning system expected to operate across the full range of solar conditions. The researchers normalize and track the data through time so the model can compare evolving patterns rather than isolated measurements.
How the sliding-window Transformer works
The core model is a sliding-window Transformer. In plain language, the system repeatedly examines a fixed-length segment of the recent time series, makes a forecast about how a target signal will evolve, then moves the window forward and repeats the process. The paper describes 12-hour-ahead forecasts that advance one hour at a time. This differs from approaches that attempt to treat the entire solar surface or an entire long sequence as one giant input. The sliding window lets the model focus on the most relevant recent period while still learning temporal relationships inside that window. The study also compares several configurations, including a standard Transformer, versions with a one-dimensional convolutional front end, and an Early Detection design that changes the attention and loss functions to reward earlier recognition of emergence-related behavior. The purpose of the comparison is not to show that one fashionable AI architecture automatically solves space weather, but to test which design choices actually improve timing and prediction quality.
What signal the model is trying to forecast
A key target in the study is the evolution of continuum intensity, which can change as an active region emerges. The researchers combine that target with precursor features drawn from acoustic and magnetic observations. NASA explains the physical idea this way: before magnetic structure becomes directly visible at the surface, it may slightly alter the local magnetic field and the behavior of acoustic waves traveling through the Sun. Those changes can be extremely small compared with the noisy background of normal solar activity. The machine-learning task is therefore closer to detecting a faint change in rhythm inside a very noisy orchestra than identifying an obvious object in an image. This distinction matters because the model is not predicting a flare directly. It is trying to recognize the earlier process of active-region emergence. A newly emerged active region may later become important for space-weather forecasting, but emergence and eruption are separate forecasting problems and should not be treated as the same event.
Why earlier active-region detection could matter
Operational space-weather teams already monitor the Sun continuously, but much of the established workflow begins once active regions are visible and can be numbered, tracked and characterized. Those regions are important because they are common source areas for solar flares and coronal mass ejections. Powerful solar activity can expose astronauts to elevated radiation, disturb satellite operations, interfere with some radio communications and navigation services, and in severe cases contribute to geomagnetic effects on infrastructure. An earlier indication that a magnetically active region is about to emerge could give forecasters an additional piece of information before the normal visible-region workflow is fully available. NASA specifically frames the COFFIES result as a possible supplement to existing monitoring rather than a replacement. For human exploration beyond low Earth orbit, even modest additional warning time can be valuable if it eventually helps mission teams decide when to increase monitoring, protect equipment, adjust operations or prepare for changing radiation conditions.
How this fits with today’s operational forecasting
NASA says current operational forecasting involves organizations including NOAA’s Space Weather Prediction Center and U.S. government partners that monitor active regions once they are visible and estimate the probability of solar flares from their observed characteristics. NASA also operates space-weather analysis capabilities for mission support, including teams focused on astronaut and spacecraft safety. The COFFIES research sits earlier in that chain: it asks whether the emergence itself can be anticipated from precursor measurements. If future validation shows that the signal is dependable, a system like this could become another input to existing forecasting workflows. That transition would require much more than a successful research paper. Operational systems need stable performance, clear false-alarm behavior, dependable data pipelines, reproducible results, monitoring for model failure and procedures that forecasters understand. The existence of a promising model therefore does not mean NASA or NOAA has replaced current methods, and it does not mean a new public warning product is available today.
What the study found — and where it struggled
The paper’s detailed results show why the research is promising but not yet mature. Performance differs substantially across the tested active regions and even across spatial tiles within the same event. Some cases show useful positive lead times, while other cases show late detections, missed detections, false positives or disagreement among model variants. In one test region, several Transformer configurations performed well enough to show clear early-detection behavior; in other regions, the timing results were weaker and an older LSTM baseline sometimes aligned better. That variability is scientifically important. A forecasting model can look impressive on an average metric and still be unsuitable for operations if it fails unpredictably on specific events. The paper also notes that experiments using full two-dimensional spatiotemporal data were outside the scope of the study because the 46-region data set creates computational and overfitting constraints. The team says more advanced spatial modeling is a direction for future work, not a capability already demonstrated here.
What the model does not prove
The result does not prove that solar storms can now be predicted 12 hours in advance, and it should not be summarized that way. The model addresses active-region emergence, not the complete chain from emergence to flare strength, coronal mass ejection trajectory, radiation environment and effects at Earth or on a spacecraft. It also does not establish that every active region will be detected with a 12-hour warning. The 12-hour figure is the forecast horizon used by the research, while actual onset lead time varies across events and model configurations. NASA explicitly says the approach is not ready for operational real-time forecasting. The small sample size, event-to-event variability, false detections and late detections all reinforce that caution. A useful operational system would need to demonstrate performance over many more known solar events and across different phases of the Sun’s activity cycle, while showing that added warning time is not bought at the cost of an unacceptable false-alarm rate.
Why the AI architecture is scientifically interesting
The research is a useful example of AI being applied to a physics problem where the important signal is indirect and temporal. The model is not being asked to generate text or classify a static image. It is learning from sequences produced by a physical system and looking for subtle relationships that may precede an observable event. The Early Detection design described in the paper modifies attention behavior and the training objective so the model is encouraged to identify the onset of meaningful change earlier. The study also uses an ablation framework, meaning the researchers deliberately turn architectural components on and off to test whether each one contributes measurable value. That is important because a more complex model is not automatically a better scientific model. By comparing a baseline Transformer, convolution-assisted variants, Early Detection variants and an LSTM reference, the team can separate the effect of architecture from the effect of simply adding more parameters. This makes the work more informative for future heliophysics machine-learning studies.
What happens next
The next phase is broader validation and refinement rather than immediate deployment. NASA says the team plans to test the approach across many more known solar events. The most important questions are whether the method maintains useful lead time across different types of active regions, how often it raises false alarms, how frequently it misses true emergence, and whether its performance remains stable as observing conditions change. Researchers also need to determine whether richer spatial inputs improve results without overfitting the limited number of labeled emergence events. For operational users, another key test is whether the model adds information that existing systems do not already provide and whether that information arrives early enough to change decisions. The long-term opportunity is a layered forecasting system in which precursor detection, visible-region analysis and eruption forecasting reinforce one another. The near-term reality is more modest: COFFIES has demonstrated a research path worth testing, not a finished early-warning product.
Full source trail
All linked evidence used by this report is preserved below; dashboard summaries are intentionally compact.
- NASA — COFFIES uses AI to predict solar active regions
primary · Aug 14, 2026
- Journal of Geophysical Research: Machine Learning and Computation
research · date not recorded
Verification & source notesClaim-level evidence, quick answers and update history
What you need to know
NASA says the COFFIES team developed a machine-learning approach that can identify signals associated with solar active-region emergence up to about 12 hours before the region becomes visible at the surface. The study uses Solar Dynamics Observatory observations, including acoustic-power measurements and line-of-sight magnetic-field data, and tests a sliding-window Transformer architecture.
Key facts
- NASA says the COFFIES team developed a machine-learning approach that can identify signals associated with solar active-region emergence up to about 12 hours before the region becomes visible at the surface.
- The study uses Solar Dynamics Observatory observations, including acoustic-power measurements and line-of-sight magnetic-field data, and tests a sliding-window Transformer architecture.
- The research paper evaluated 46 active regions and found that model timing performance varies from event to event, which is why the result should be treated as an experimental forecasting method rather than a finished service.
- NASA explicitly says the model is not ready for operational real-time forecasting and needs validation across more known solar events before any operational transition.
As of:
Current status: what is confirmed and what remains open
Confirmed in the source-backed record
- NASA says the COFFIES team developed a machine-learning approach that can identify signals associated with solar active-region emergence up to about 12 hours before the region becomes visible at the surface.
- The study uses Solar Dynamics Observatory observations, including acoustic-power measurements and line-of-sight magnetic-field data, and tests a sliding-window Transformer architecture.
- The research paper evaluated 46 active regions and found that model timing performance varies from event to event, which is why the result should be treated as an experimental forecasting method rather than a finished service.
Limits, uncertainty and next signals
- What happens next: The next phase is broader validation and refinement rather than immediate deployment. NASA says the team plans to test the approach across many more known solar events. The most important questions are whether the method maintains useful lead time across different types of active regions, how often it raises false alarms, how frequently it misses true emergence, and whether its performance remains stable as observing conditions change. Researchers also need to determine whether richer spatial inputs improve results without overfitting the limited number of labeled emergence events. For operational users, another key test is whether the model adds information that existing systems do not already provide and whether that information arrives early enough to change decisions. The long-term opportunity is a layered forecasting system in which precursor detection, visible-region analysis and eruption forecasting reinforce one another. The near-term reality is more modest: COFFIES has demonstrated a research path worth testing, not a finished early-warning product.
Quick answers
What happened?
NASA described new research from the COFFIES collaboration — the Consequence Of Fields and Flows in the Interior and Exterior of the Sun DRIVE Science Center — that applies machine learning to the problem of detecting solar active-region emergence before the region is plainly visible on the solar surface. Active regions are concentrations of strong magnetic activity that can later produce sunspots, solar flares and coronal mass ejections. Instead of waiting for a sunspot group to become visually established, the research looks for small changes in solar observations that may occur while magnetic structure is still rising through the Sun’s interior. NASA says the model can provide an early signal several hours before visible emergence, with forecasts extending as far as about 12 hours ahead in the study. The work combines solar-physics expertise with data-science methods and uses observations from NASA’s Solar Dynamics Observatory together with NASA Ames supercomputing resources.
What data the researchers used?
The research paper focuses on time-series measurements derived from Solar Dynamics Observatory data, especially observations from the Helioseismic and Magnetic Imager. The model does not inspect a single photograph and simply guess where a sunspot will appear. It works on a sequence of measurements that represent how the local solar environment changes over time. The study uses acoustic-power information across multiple frequency bands together with line-of-sight magnetic-field measurements. Those features are intended to capture two different kinds of precursor behavior: changes in waves traveling through the solar interior and changes in magnetic structure at or near the surface. The paper reports experiments using 46 active regions, a useful research sample but still a relatively small data set for a machine-learning system expected to operate across the full range of solar conditions. The researchers normalize and track the data through time so the model can compare evolving patterns rather than isolated measurements.
How the sliding-window Transformer works?
The core model is a sliding-window Transformer. In plain language, the system repeatedly examines a fixed-length segment of the recent time series, makes a forecast about how a target signal will evolve, then moves the window forward and repeats the process. The paper describes 12-hour-ahead forecasts that advance one hour at a time. This differs from approaches that attempt to treat the entire solar surface or an entire long sequence as one giant input. The sliding window lets the model focus on the most relevant recent period while still learning temporal relationships inside that window. The study also compares several configurations, including a standard Transformer, versions with a one-dimensional convolutional front end, and an Early Detection design that changes the attention and loss functions to reward earlier recognition of emergence-related behavior. The purpose of the comparison is not to show that one fashionable AI architecture automatically solves space weather, but to test which design choices actually improve timing and prediction quality.
What signal the model is trying to forecast?
A key target in the study is the evolution of continuum intensity, which can change as an active region emerges. The researchers combine that target with precursor features drawn from acoustic and magnetic observations. NASA explains the physical idea this way: before magnetic structure becomes directly visible at the surface, it may slightly alter the local magnetic field and the behavior of acoustic waves traveling through the Sun. Those changes can be extremely small compared with the noisy background of normal solar activity. The machine-learning task is therefore closer to detecting a faint change in rhythm inside a very noisy orchestra than identifying an obvious object in an image. This distinction matters because the model is not predicting a flare directly. It is trying to recognize the earlier process of active-region emergence. A newly emerged active region may later become important for space-weather forecasting, but emergence and eruption are separate forecasting problems and should not be treated as the same event.
Why earlier active-region detection could matter?
Operational space-weather teams already monitor the Sun continuously, but much of the established workflow begins once active regions are visible and can be numbered, tracked and characterized. Those regions are important because they are common source areas for solar flares and coronal mass ejections. Powerful solar activity can expose astronauts to elevated radiation, disturb satellite operations, interfere with some radio communications and navigation services, and in severe cases contribute to geomagnetic effects on infrastructure. An earlier indication that a magnetically active region is about to emerge could give forecasters an additional piece of information before the normal visible-region workflow is fully available. NASA specifically frames the COFFIES result as a possible supplement to existing monitoring rather than a replacement. For human exploration beyond low Earth orbit, even modest additional warning time can be valuable if it eventually helps mission teams decide when to increase monitoring, protect equipment, adjust operations or prepare for changing radiation conditions.
How this fits with today’s operational forecasting?
NASA says current operational forecasting involves organizations including NOAA’s Space Weather Prediction Center and U.S. government partners that monitor active regions once they are visible and estimate the probability of solar flares from their observed characteristics. NASA also operates space-weather analysis capabilities for mission support, including teams focused on astronaut and spacecraft safety. The COFFIES research sits earlier in that chain: it asks whether the emergence itself can be anticipated from precursor measurements. If future validation shows that the signal is dependable, a system like this could become another input to existing forecasting workflows. That transition would require much more than a successful research paper. Operational systems need stable performance, clear false-alarm behavior, dependable data pipelines, reproducible results, monitoring for model failure and procedures that forecasters understand. The existence of a promising model therefore does not mean NASA or NOAA has replaced current methods, and it does not mean a new public warning product is available today.
Update ledger
- PublishedInitial Atalk.TV publication.
- UpdatedLatest material update reflected in this durable article URL.
Claims and supporting sources
Each claim below is tied to the article's verified source set. When the wording names a publisher, Atalk.TV narrows the evidence to that publisher's linked source; otherwise the verification basis or complete article source set is shown.