In early 2026 a machine-learning pipeline called AnomalyMatch swept through 99.6 million small image cutouts from the Hubble Legacy Archive in roughly two and a half days, producing a ranked shortlist of unusual sources and a human-curated catalogue of 1,339 unique candidates. That result is striking not because an algorithm suddenly replaced astronomers, but because it illustrates, in vivid detail, how a practical combination of curated data, semi-supervised learning and focused human attention can transform an intractable visual search into a tractable discovery workflow.

Data preparation: defining the unit of search

Any process analysis starts with the input. Here the inputs were not 100 million independent telescope exposures but nearly 100 million cutouts: small images centered on detected sources within pre-processed Hubble mosaics. Each cutout spanned a few dozen pixels and roughly seven to eight arcseconds, recorded through the Advanced Camera for Surveys Wide Field Channel using the F814W filter. By choosing a consistent instrument and filter the team limited heterogeneity, trading breadth for uniformity so morphological comparisons would be meaningful.

Why cutouts matter

Cutouts simplify the search problem in three ways. First, they reduce scene complexity: the algorithm sees an object-focused patch rather than an entire field. Second, they standardize scale and wavelength, making shape comparisons more robust. Third, they compress a massive archive into a manageable set of analysis units. But that convenience brings limits: important spectral information, temporal variability and context across filters were intentionally excluded. The process therefore targets morphology — visible shape — not every possible astronomical anomaly.

Preprocessing and artefact mitigation

Before any machine learning can begin, preprocessing cleans and normalizes images. Cosmic rays, detector noise and stitching artefacts can masquerade as anomalies. The Hubble Legacy Archive’s science-ready mosaics already addressed many such issues, but the team still needed to account for source shredding — a known catalogue problem where a single astrophysical object can produce multiple detections. Recognizing these failure modes up front influenced later design choices in ranking and de-duplication.

Model design: semi-supervised and active learning in practice

AnomalyMatch’s architecture exemplifies a pragmatic balance between labelled knowledge and large-scale discovery. Pure supervised learning requires extensive labelled datasets that don’t exist for rare cosmic oddities; fully unsupervised methods can highlight structure but may struggle to prioritize scientifically relevant deviations. The hybrid approach used here begins with a tiny labelled set and grows through active human-in-the-loop refinement.

From three examples to a working training set

Development began with only three labelled examples of edge-on protoplanetary disks. The model learned a notion of morphological distance and began assigning high anomaly scores to other unusual shapes, including merging systems and gravitational lenses. The researchers expanded the labelled pool during development up to roughly 1,400 images, of which 375 were labelled anomalous and 1,025 nominal. That incremental growth is a key process lesson: when rare classes dominate the desired signal, iterative labelling guided by model outputs is more efficient than exhaustive upfront annotation.

Active learning loop

Active learning lets a model query an expert for labels on the most informative examples. In this pipeline the model assigned anomaly scores across the 99.6 million cutouts, then the team inspected the top-ranked outputs and supplied additional labels. Those labels refined the model’s internal representation of what counts as ‘‘anomalous’’ in this specific dataset. The loop shrank the search space drastically while allowing human judgment to define scientific relevance rather than letting an algorithm alone decide what matters.

Ranking, review and catalog curation

A model ranking is not a catalogue; it is a prioritized queue. AnomalyMatch assigned a score to every cutout and the researchers retained the 5,000 highest-scoring images for manual inspection. Many of those were duplicates or fragmented detections of the same astrophysical source, so cross-matching and de-duplication reduced the list to 1,339 unique sources. From that set the authors judged 1,176 as scientifically interesting anomalies across 19 working categories.

Human judgment as the sieve

The process shows how automated systems amplify scarce human attention. The model concentrated promising candidates at the front of the queue; astronomers separated true morphological oddities from nominal sources, ambiguous blends and image artefacts. The catalogue therefore reflects a two-stage epistemic process: statistical triage followed by domain evaluation. Each step introduces different error modes — the model can miss signals, humans can misclassify ambiguous shapes, and cross-matching can fail to link multi-observation identifiers — but together they form an efficient discovery pipeline.

Cross-referencing and literature checks

To test novelty, the team cross-matched coordinates against SIMBAD, ESASky and associated publications and catalogues. They found 811 candidates had no literature reference in those systems. That is an important, carefully worded result: absence from those repositories does not prove an object was never noticed by a person nor does it imply novel physics. Hubble is a pointed observatory; an odd-looking object may have been in the background of a targeted observation without being the subject of a paper. The operational definition here was precise — no literature reference found via the chosen databases — and the authors released source identifiers, positions and provisional classifications to enable reproducibility and revision.

Classification outcomes: familiar oddities and genuine mysteries

Most high-ranked sources turned out to be rare morphologies of known phenomena. Galaxy mergers and interactions dominated the set, exhibiting distorted discs, multiple bright nuclei and tidal streams. The catalogue also contains gravitational lenses, jellyfish galaxies shedding gas, galaxies with large star-forming clumps, rings, arcs and jets, and two previously known edge-on protoplanetary disks. Forty-three images resisted morphological categorization and were released without labels for further study.

When ‘‘anomaly’’ is a relative term

Crucially, ‘‘anomaly’’ in this process meant differing from the dataset’s modal shapes, not that an object violates known physics. Many anomalies are atypical presentations of familiar processes or projections of blended sources. The modelling goal is therefore pragmatic: pull rare structures toward human attention so domain experts can decide which warrant follow-up. The value lies in enabling testing, not in autonomous proclamations of discovery.

False positives, artefacts and blended sources

Automated filters will always raise false positives and miss objects of interest. Prior work demonstrated similar trade-offs: an earlier unsupervised search reduced 176,808 objects to 1,100 candidates, and manual vetting rejected about 86 percent, yielding 147 outliers. AnomalyMatch scaled the idea by orders of magnitude while accepting that a non-zero nominal rate and imperfect recall are acceptable if the pipeline accelerates human inspection by hundreds of times.

Scaling to future surveys and the attention bottleneck

Hubble’s archive aggregates data from thousands of programs across decades, and upcoming missions will dwarf it. Euclid, the Vera C. Rubin Observatory and the Nancy Grace Roman Space Telescope will produce broader, deeper and repeatedly sampled views of the sky. The bottleneck in such regimes will not be computing anomaly scores; it will be allocating expert follow-up. Methods that learn from a handful of labels and present prioritized shortlists can therefore act as amplifiers of scarce specialist attention.

Integration with multiwavelength and follow-up strategies

One natural extension of this process is to embed multiwavelength context and to pipeline follow-up triggers. Morphological abnormality in a single optical band is a useful filter, but confirming scientific novelty often requires spectra, infrared or radio observations, or temporal monitoring. A process that couples anomaly ranking with prioritized follow-up — automated proposals for spectroscopic time or alerts to survey teams — would convert ranked lists into confirmed discoveries more efficiently.

Designing for reproducibility and community use

A vital process decision was openness: the authors released machine-readable catalogues, coordinates and images so other researchers can replicate cross-matches, apply different literature checks or perform targeted follow-up. That transparency allows the community to iterate on labels, refine categories and mine the same candidates for multiwavelength signatures. Open data makes the pipeline a platform rather than a one-off result.

Viewed as a process, the AnomalyMatch experiment emphasizes allocation of human attention more than autonomous discovery. A constrained model quickly sifted a vast archive and handed experts a focused set of candidates; astronomers applied domain knowledge to sort true astronomical oddities from artefacts and familiar but rare morphologies. The catalogue does not claim to have found new laws of nature, but it does show how modest, well-chosen automation can expose hundreds of interesting objects that escaped prior literature searches and prepare them for the next stage of scientific scrutiny. As survey archives grow larger and telescopes produce richer, multiwavelength data streams, workflows that combine semi-supervised learning, active human-in-the-loop refinement and open cataloguing will become essential tools for turning an ocean of pixels into a sustainable program of discovery.