Use of machine learning to enhance detection of transient astronomical phenomena in historical observatory images

Abstract

Transient, star-like point sources that appear and vanish over short timescales are described in astronomical images prior to launch of Sputnik. We have reported that transient numbers diminish significantly in Earth’s shadow (shadow deficit) and are more likely within (+/-) one day of nuclear testing (nuclear window). These findings remain debated; some argue transients identified via existing automated pipelines are simply plate defects. Machine learning (ML) was used to enhance transient identification accuracy and validate the phenomenon. The model was trained against 250 transient image pairs taken 30 min apart classified as real versus plate defect by expert visual review; it demonstrated good discrimination (out-of-fold AUC = 0.81; sensitivity = 0.71, specificity = 0.71). After deployment in a dataset of 107,875 previously-identified transients, the model assigned each a probability of being real. After controlling for ML-identified artifacts, transient counts were significantly elevated for dates within a nuclear window ( p = 0.024); transients with the highest probability of being real were more likely to occur within a nuclear window ( p <0.0001). The shadow deficit was significant ( p <0.001) and largest in the highest probability transients relative to lower probability transients ( p =0.003). Results support existence of an unrecognized population of transient objects in historical astronomical plates warranting further study.