A Machine Learning empowered search for Sub-Minute Optical Transient Events with the Deeper, Wider, Faster programme

TL;DR

This study develops and applies a machine-learning–accelerated pipeline to hunt for sub-minute optical transients within the Deeper, Wider, Faster program, leveraging minute-cadence, $m(g)\sim23$ DECam data. By enforcing a single-detection light-curve criterion and using a real/bogus CNN (robot) with a conservative threshold, the authors dramatically reduce an initial blob of candidates, revealing two high-confidence sub-minute transient candidates after manual vetting. Multiwavelength searches find no secure counterparts, while a Poisson-rate analysis yields a sky rate of $R_0 = 4.72^{+6.39}_{-3.28}\times10^5$ day$^{-1}$ (with a Rubin/LSST-normalised expectation of $\sim$7.6 events per night for typical survey parameters). The work also demonstrates robust checks against satellite debris glints, supporting a potential astrophysical origin for at least one candidate and highlighting the potential of sub-minute transient searches to inform future wide-field surveys such as LSST.

Abstract

Optical transient surveys continue to generate increasingly large datasets, prompting the introduction of machine-learning algorithms to search for quality transient candidates efficiently. Existing machine-learning infrastructure can be leveraged in novel ways to search these datasets for new classes of transients. We present a machine-learning accelerated search pipeline for the Deeper, Wider, Faster (DWF) programme designed to identify high-quality astrophysical transient candidates that contain a single detection. Given the rapid observing cadence of the DWF programme, these single-detection transient candidates have durations on sub-minute timescales. This work marks the first time optical transients have been systematically explored on these timescales, to a depth of m$\sim$23. We report the discovery of two high-quality sub-minute transient candidates from a pilot study of 671,761 light curves and investigate their potential origins with multiwavelength data. We discuss, in detail, possible non-astrophysical false positives, confidently reject electronic artefacts and asteroids, ruling out glints from satellites below 800 km and strongly disfavouring those at higher altitudes. We calculate a rate on the sky of $4.72^{+6.39}_{-3.28}\times10^5$ per day for these sub-minute transient candidates.

Figures (15)

• Figure 1: Example of selection criteria for the sub-minute optical transient candidates. Given a set of consecutive images (four images in this example; our data consists of $\sim$100 images per night), we define the following selection criteria: single detection, present anywhere except the first and last image.
• Figure 2: Misclassification performance of the robot CNN classifier as a function of the decision threshold. At the decision boundary of 0.06 (vertical black line), the algorithm performs with a 0.6% False Negative Rate (FNR) and 30.9% False Positive Rate (FPR). Since we are considering data with a large bias towards very low scores ($<$0.02) and are looking for a rare class of transient, our objective is to minimise the FNR as much as is feasible without diluting the results with too many contaminants.
• Figure 3: Light curves and detection images for the two high-confidence candidates. A) is candidate DWF041117.877-542554.144, and B) is candidate DWF040654.511-544056.411. The light curves, upper row, display the DECam $g$-band non-detection upper limits as blue triangles, and the single detection apparent magnitude as the black point. The detection images, bottom row (left to right), display the template image for that region of sky taken on the observational night, three minutes prior to the science detection, the science image of the single detection, and the subtraction of the science image from the template, leaving the residual flux from the source.
• Figure 4: Deep field $g$-band imaging of the sky region surrounding the sub-minute candidates, DWF040654.511- (left) and DWF041117.877 (right). These deep field images are a stack of one night of data and reach a limiting magnitude of $m(g)\sim25$.
• Figure 5: Point Spread Function (PSF) analysis of the two promising sub-minute optical transients candidates. This diagnostic plot shows the source extractorspread_model of sources found on promising candidate CCDs, with the 2 promising candidates highlighted as stars. spread_model is defined by Equation \ref{['eq:sm']} and is used as a star/galaxy classifier. A stellar locus (PSF-like sources) as a function of magnitude is featured at spread_model values close to 0. Sources with spread_model values greater than 0 outside of the stellar peak are classified as extended sources such as galaxies or nebulae, and those below 0 are classified as electronic artefacts or cosmic rays.