Uncovering Anomalous Gamma-Ray Bursts beyond Duration-based Classification

Abstract Gamma-ray bursts (GRBs) are traditionally classified by duration, and most machine-learning approaches likewise retain T 90 as an input parameter. However, the observed duration T 90 does not reliably trace the physical origin, because observational effects such as the iceberg effect can cause only the brightest portion of an emission episode to be detected. Here, we classify a sample of 927 GRBs using a machine-learning dimensionality-reduction method based on fine prompt-emission variability and spectral properties, explicitly excluding T 90 . This duration-free representation reveals 46 anomalous bursts whose conventional labels are inconsistent with their locations in the reduced parameter space. Independent spectral-lag analysis shows that the anomalous long-duration bursts are consistent with merger-origin events, whereas the anomalous short-duration bursts are more consistent with collapsar-origin events, suggesting a broader population of duration–origin mismatch events not previously systematically recognized. These results show that physically meaningful GRB populations can emerge without using duration at all and suggest that multidimensional classification may offer a more physically grounded view of GRB diversity than the traditional short–long dichotomy alone.