The ATLAS Virtual Research Assistant

TL;DR

The ATLAS Virtual Research Assistant uses machine learning to efficiently prioritize astronomical alerts, significantly reducing manual review workload while maintaining high detection quality, and enabling automated follow-up observations.

Contribution

This work introduces a novel ML-based alert ranking system for sky surveys that improves efficiency and integrates expert knowledge through feature engineering.

Findings

85% reduction in eyeballing workload

Less than 0.08% loss in follow-up opportunities

Effective alert ranking with minimal training data

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 "Extra-galactic" receive high scores; the scores are updated when new light curve 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 such as accuracy or 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. Finally, policies are defined on the ranked list to select the most promising alerts for humans to eyeball and to automatically remove the 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 to automated methods that will be required in the era of LSST. inally, 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