Needles in the Haystack: Unsupervised Methods of Anomaly Detection in Astronomical Surveys

Abstract

On 24th of February 2026, the Vera C. Rubin Observatory released 800,000 alerts on its very first night of observation, marking the early stages of operation of the Legacy Survey of Space and Time. The LSST survey will map the night sky of the southern hemisphere at a faster rate and a wider range than previously ever achieved. The system is predicted to record up to 15 terabytes of raw data in a single night, accumulating to 500 petabytes of processed data over the next 10 years. Pioneering the era of data-intensive astronomical surveys, the volume of LSST data makes visual inspection and manual follow-ups near to impossible, and necessitates the implementation of automated anomaly detection. In this thesis, we apply two models of unsupervised and semi-supervised methods of outlier detection: the partition-based Isolation Forest and the metric ensemble Distance Multi-Metric Anomaly Detection (DiMMAD). We train and evaluate these models on the reLAISS ZTF-based feature space of 17,000 supernova-like transients. The Isolation Model identifies a set of heterogeneous anomaly candidates including peculiar supernovae, Cataclysmic Variables, and spectroscopically unclassified transients. Incorporating a set of host-galaxy features into the feature space leads to the suppression of contaminants from the galactic plane and further reveals more astrophysically unusual supernovae among the anomaly candidates. Under the median-median configuration, DiMMAD detects a wider range of unclassified objects, with eight of the top ten lacking spectroscopic labels and represents interesting candidates for further inspection. The overlap of anomaly candidates in the top range of the two models is minimal and motivates the use of both in tandem, indicating that each model is sensitive to different regions of the feature space based on their difference in definitions of anomaly. In light of early data, each model is successfully applied to a small subset of 221 LSST objects. With no retraining required, the models are compatible with the new survey and preliminary results show the detection of confirmed supernova alongside Active Galactic Nuclei and image subtraction artefacts. These motivate additional filtering of data and further refinement on data preprocessing before deployment to the full LSST survey. This work demonstrates the potential of unsupervised machine learning approaches utilising astrophysically motivated feature representations to identify outliers and anomaly candidates in large astrophysical surveys. These results provide a feasible foundation for real-time anomaly detection pipelines for LSST.