Overview
- The study, published Aug. 14, 2026, reports EarlyDetect, a Transformer‑based AI trained on NASA SDO/HMI acoustic power maps and magnetic measurements to spot precursor signs of active‑region emergence.
- EarlyDetect achieved an average 9.24‑hour lead time for detecting emergent active regions on held‑out events, outperforming a standard Transformer and prior benchmark methods.
- During development the team found a common data filter erased the faint fluctuations the model used, and removing that filter measurably improved forecast performance.
- The model is not yet operational because it still issues false alarms and late detections, and an emergence warning does not reliably predict a subsequent flare or coronal mass ejection.
- To speed community validation the researchers released the SolARED dataset and the SAR Portal so other teams can test models, extend event samples, and work toward linking emergence signals to eruption forecasts.