2025/06/11 by H. F. Stevance, K. Smith, K. W. Smith +14 · 2 voices · 4 citations
Computer Science · Physics and Astronomy · #Atlas (anatomy) #Computational Physics and Python Applications #Decision tree #Facility management #Metric (unit) #Particle Detector Development and Performance #Particle physics theoretical and experimental studies #Sky #Training set #Workload
paper · pdf · doi:10.3847/1538-4357/adf2a1
published in The Astrophysical Journal 990(2), 201 (IOP Publishing)
openalex publication_date 2025/09/09 · openalex created_date 2025/09/12 · openalex updated_date 2026/08/05
Abstract We present the Virtual Research Assistant (VRA) of the ATLAS sky survey, which performs preliminary eyeballing on our clean transient data stream. The VRA uses histogram-based gradient-boosted decision tree classifiers trained on real data to score incoming alerts on two axes: “Real” and “Galactic.” The alerts are then ranked using a geometric distance such that the most “real” and “extragalactic” receive high scores; the scores are updated when new lightcurve data is obtained on subsequent visits. To assess the quality of the training we use the recall at rank K , which is more informative to our science goal than general metrics (e.g., accuracy, F1-scores). We also establish benchmarks for our metric based on the pre-VRA eyeballing strategy, to ensure our models provide notable improvements before being added to the ATLAS pipeline. Then, policies are defined on the ranked list to select the most promising alerts for humans to eyeball and to automatically remove bogus alerts. In production the VRA method has resulted in a reduction in eyeballing workload by 85% with a loss of follow-up opportunity <0.08%. It also allows us to automatically trigger follow-up observations with the Lesedi telescope, paving the way toward automated methods that will be required in the era of LSST. Finally, this is a demonstration that feature-based methods remain extremely relevant in our field, being trainable on only a few thousand samples and highly interpretable; they also offer a direct way to inject expertise into models through feature engineering.