Identifying Astrophysical Anomalies in 99.6 Million Cutouts from the Hubble Legacy Archive Using AnomalyMatch

TL;DR

The paper addresses the challenge of discovering rare astrophysical anomalies in massive image archives by reframing anomaly detection as an imbalanced binary problem and applying AnomalyMatch, a semi-supervised method with an active-learning loop. They process about 99.6 million Hubble Legacy Archive cutouts, achieving a full-scale scan in 2–3 days on ESA Datalabs and producing a catalog of 1,338 unique anomalies across 19 classes (e.g., mergers, gravitational lenses, jellyfish galaxies). Validation relies on morphology-based classifications and literature cross-checks (SIMBAD/ESASky), with a substantial portion of objects lacking prior references, illustrating the method’s capacity to uncover new systems and unseen morphologies, including candidates like lensed quasars. The work demonstrates the practicality and scalability of AnomalyMatch for upcoming survey data (Euclid, Rubin), and provides publicly accessible machine-readable catalogs and images to enable community-driven follow-up and refinement of anomaly classifications.

Abstract

Astronomical archives contain vast quantities of unexplored data that potentially harbour rare and scientifically valuable cosmic phenomena. We leverage new semi-supervised methods to extract such objects from the Hubble Legacy Archive. We have systematically searched approximately 100 million image cutouts from the entire Hubble Legacy Archive using the recently developed AnomalyMatch method, which combines semi-supervised and active learning techniques for the efficient detection of astrophysical anomalies. This comprehensive search rapidly uncovered a multitude of astrophysical anomalies presented here that significantly expand the inventory of known rare objects. Among our discoveries are 138 new candidate gravitational lenses, 18 jellyfish galaxies, and 417 mergers or interacting galaxies. The efficiency and accuracy of our iterative detection strategy allows us to trawl the complete archive within just 2-3 days, highlighting its potential for large-scale astronomical surveys. We present a detailed overview of these newly identified objects, discuss their astrophysical significance, and demonstrate the considerable potential of AnomalyMatch to efficiently explore extensive astronomical datasets, including, e.g., upcoming Euclid data releases.

Figures (12)

• Figure 1: The initial three images labelled anomaly used to train AnomalyMatch. These three images containing edge-on protoplanetary disks which we initially aimed to search additional instances of in the HLA. During active learning, this set was expanded to include sources with odd morphologies like mergers, lenses and jellyfish galaxies which were serendipitously discovered. Titles are the Source IDs of the objects found in 2023ApJ...948...40O.
• Figure 2: Fifty examples of the final training set used in applying $\texttt{AnomalyMatch}$ to the HLA. The top two rows, highlighted in red, show ten examples of the anomaly class. These are made up of mergers, lenses, edge-on proto-planetary disks as well as some galaxies showing odd morphology. The remaining 40 images are then examples of 'nominal' data. This is primarily isolated galaxies, star fields and artifacts.
• Figure 3: The workflow when using AnomalyMatch. We leverage both labelled and unlabelled data from a user dataset to train an EfficientNet architecture, and include an active learning loop. Here, the unlabelled data is ranked by anomaly score, the user can extract more examples of the object they are searching for, and add them to their training data. Once the desired model metrics are achieved, the model can be saved and then run across all images in their dataset.
• Figure 4: Exemplary anomaly scores on a random subsample of the data. Notably, artefacts are clearly isolated and increasing scores correspond well with increasingly interesting data. The model shows robustness against varying brightness, image noise or differing sizes of objects in the images.
• Figure 5: The anomaly score distribution obtained by applying our final trained $\texttt{AnomalyMatch}$ model on a random subset of the HLA of $\approx500,000$ cutouts. We find that the distribution is highly weighted to zero, as expected. The majority of our sources are not anomalous.