Reliability of supervised detectors for unlabeled event types

Determine whether supervised solar-wind event-detection methods remain reliable for event types absent from their training labels.

Background

The paper contrasts supervised methods, which learn from labeled catalogs, with unsupervised approaches that can identify anomalies without catalog labels. Because solar-wind event catalogs are incomplete, supervised detectors may encounter event types or signatures that were not represented during training. The reliability of such methods in that setting is explicitly unresolved.

References

Their reliability for event types absent from the labels is unknown, and they lack a physical decomposition of why a window is flagged.

— PISCES: Physics-Informed Solar-wind Convolutional autoEncoder for Space-weather Anomaly Detection and Early Warning  (2609.28022 - Lee et al., 23 Sep 2026) in Section 2, Background