EarlyDetect: The AI That Predicts Solar Active Regions Hours Before They Form

As magnetic fields rise toward the Sun's surface, they leave faint signatures in acoustic waves. A drop in acoustic power appears first — providing an early signature of active-region emergence — followed by changes in continuum intensity and magnetic field as the region becomes visible . The model uses a sliding-window approach to analyze these temporal sequences, detecting subtle changes in solar oscillations before sunspots appear .

Researchers describe the challenge as "detecting a slight change in rhythm within a very noisy orchestra" . The active regions begin developing beneath the Sun's visible surface, so the model must pick out very small changes in magnetic field and acoustic wave patterns amid the Sun's constant activity.

The Surprising Filtering Mistake

One of the most instructive findings from the research involved an unexpected failure. The team initially applied a filtering technique designed to help the AI identify important patterns — but it actually degraded performance .

"We initially expected it to help isolate useful short-timescale patterns. Instead, it averaged away the very faint fluctuations that provided the earliest warning," said co-principal investigator Alexander Kosovichev . Lead author Jonas Tirona noted the filter was "detrimental in almost every case" — the signals it removed were precisely the ones the model needed to predict when an active region would emerge .

This counterintuitive result highlights how AI models for heliophysics can behave differently than expected, requiring careful empirical validation of every preprocessing step.

Benchmark Performance and Critical Limitations

EarlyDetect's best-performing version achieved a 10.6% improvement in Root Mean Square Error (RMSE) compared to the previous LSTM baseline approach . It outperformed both a standard Transformer model and earlier benchmark methods .

The researchers emphasized several key limitations that are important for practical space weather applications:

• False alarms: The model sometimes predicts active region emergence that does not occur .
• Late predictions: Not all active regions are detected with sufficient lead time .
• No guarantee of eruptions: Detecting an active region does not guarantee it will produce flares or coronal mass ejections — the model predicts the emergence of the region itself, not its subsequent eruptive activity .

Public Dataset Released for Future Research

To accelerate work in this field, the team released SolARED (Solar Active Region Emergence Dataset), a machine learning-ready dataset derived from full-disk maps of Doppler velocity, magnetic field, and continuum intensity from SDO/HMI . The dataset includes time series of remapped, tracked, and binned data characterizing the evolution of acoustic power and other physical parameters . The data is publicly available through a dedicated portal to support future heliophysics and space weather research .

What This Means for Space Weather Forecasting

EarlyDetect represents a significant step toward operational early warning systems for space weather. By providing nearly nine hours of advance notice before active regions emerge, the model could help protect satellites, power grids, and astronauts from solar storms. The availability of the SolARED dataset also means other researchers can build on this work, potentially improving warnings further or extending the approach to predict actual eruptive events.

However, the model's current limitations — false alarms, late predictions, and the gap between active region emergence and actual eruptions — mean it is not yet a complete solution. The researchers' candid reporting of these shortcomings is itself valuable, laying out a clear roadmap for where the next generation of models needs to improve.