Extragalactic Astrophysics and Computational Data Science Review Panel

Abstract:

This analysis details the application of a novel semi-supervised active learning framework, "Anomaly Match," developed by a team at the European Space Agency, to the Hubble Legacy Archive. The objective was to circumvent the overwhelming data volume challenge in modern astronomy by rapidly identifying previously unclassified anomalous objects. The AI tool successfully processed nearly 100 million individual Hubble image cutouts in under 72 hours, identifying approximately 1,400 anomalous objects, 800 of which were previously unexamined. These findings validate the "human-in-the-loop" approach for high-throughput anomaly detection and significantly augment existing catalogs of peculiar galaxies. The discovered objects primarily fall into five established categories of transient astrophysical phenomena, but also include objects that defy morphological classification, potentially offering insights into unknown physical processes.

AI Discovers Anomalies in Hubble Images We Never Knew Existed

• 0:00 The Data Challenge: Modern astronomy faces a significant challenge due to data saturation, specifically citing the Hubble Space Telescope (HST), which has accumulated nearly 100 million individual image cutouts in its Legacy Archive over 35 years.
• 1:01 Breakthrough Methodology: A team at the European Space Agency utilized a new AI tool, Anomaly Match, to sift through the data. This tool employs semi-supervised active learning, contrasting with standard supervised machine learning models.
• 4:10 Anomaly Match Framework: The semi-supervised approach involves training the neural network on a small number of known normal and anomalous examples. The network then assigns an "anomaly score" to unlabeled images, presenting the highest-scoring images to a human expert. This "human in the loop" feedback mechanism refines the AI’s performance.
• 1:15 Processing and Results: Anomaly Match processed the hundreds of millions of images in approximately 2.5 days (under 72 hours), identifying 1,400 objects deemed anomalous. Crucially, at least 800 of these objects had never been seen or previously examined.
• 1:40 Importance of Anomalies: The research is grounded in the historical precedent set by Halton Arp’s 1966 Atlas of Peculiar Galaxies, which emphasized that studying bizarre, transient structures (such as galactic mergers) is essential for understanding physical processes like dark matter effects, galactic collisions, and gas dynamics under extreme conditions.
• 6:00 Classification of Findings: The 1,400 verified anomalies were categorized into five main types:
  • Galactic Mergers and Tidal Tails: Common anomalous structures resulting from gravitational interaction between close galaxies.
  • Gravitational Lensing: 140 new examples found, representing the distortion of background light by massive foreground objects (per Einsteinian theory).
  • Jellyfish Galaxies: 35 examples found, formed by Ram Pressure Stripping as galaxies move through dense cluster gas, leading to star formation cessation.
  • Collisional Ring Galaxies: Only two new examples discovered, confirming their extreme rarity. These form when a smaller galaxy punches through the center of a larger disk galaxy.
  • Prot Planetary Discs: Objects within the Milky Way, representing young stars with surrounding dust and gas where planets are forming (some resembling "hamburgers").
• 9:21 Unclassified Objects: The AI also flagged objects with high anomaly scores that defy current morphological classification, referred to as "unknown objects" (e.g., the object shown at 9:39). These objects represent potential clues to new or currently unknown physics.
• 9:53 Future Applications: The successful implementation of Anomaly Match on HST data paves the way for applying the same strategy to forthcoming high-volume datasets from missions such as the Euclid mission (observing billions of galaxies) and the Vera Rubin Observatory (expected to generate 50 petabytes of data).