Abstract
We employ self-supervised representation learning to distill information from 76 million galaxy images from the Dark Energy Spectroscopic Instrument Legacy Imaging Surveys' Data Release 9. Targeting the identification of new strong gravitational lens candidates, we first create a rapid similarity search tool to discover new strong lenses given only a single labelled example. We then show how training a simple linear classifier on the self-supervised representations, requiring only a few minutes on a CPU, can automatically classify strong lenses with great efficiency. We present 1192 new strong lens candidates that we identified through a brief visual identification campaign, and release an interactive web-based similarity search tool and the top network predictions to facilitate crowd-sourcing rapid discovery of additional strong gravitational lenses and other rare objects.
Introduction
Strong gravitational lensing occurs when a massive foreground object, such as a galaxy or cluster, bends light from a distant background source, creating characteristic arcs or multiple images. These systems are invaluable for probing dark matter distributions and measuring the expansion rate of the universe. However, they are exceptionally rare, appearing in roughly 1 out of every 10,000 galaxies. As modern astronomical surveys scale to billions of objects, traditional manual inspection becomes unfeasible, and standard supervised deep learning models often struggle with the extreme class imbalance and the requirement for massive labeled datasets.
Data and Survey Coverage
This study leverages the DESI Legacy Imaging Surveys Data Release 9 (DR9), which provides a vast dataset of 76 million galaxy images across three optical bands ($g, r, z$). For initial representation learning, a curated subset of 3.5 million images was utilized. The model's performance was validated against a master database of approximately 6,000 known lens candidates.
Methodology: Self-Supervised Pipeline
The core of our approach is Self-Supervised Learning (SSL), which allows the model to learn morphological features directly from unlabeled data. The process follows a three-stage pipeline:
- Contrastive Pretraining: A ResNet50 encoder is trained to recognize that different augmented versions of the same galaxy image should have similar representations, while different galaxies should be distinct. This forces the network to capture essential semantic features of galaxy shapes.
- Similarity Search: By mapping galaxies into a high-dimensional vector space, we can identify potential lenses by finding objects with similar vectors to known examples using cosine similarity. This tool allows for rapid discovery from a single query image.
- Automated Classification: We demonstrate that a simple linear classifier trained on top of these frozen SSL representations can achieve state-of-the-art performance with minimal computational cost, requiring only minutes on a standard CPU.
Key Results and Discovery
Our methodology led to the identification of 1,192 new strong lens candidates. These discoveries were categorized into 404 "Grade A" (high-confidence) and 788 "Grade B" (likely) candidates.
Beyond raw discovery numbers, the efficiency gains were substantial. The similarity search tool enabled a 500x speedup in discovery compared to random inspection. Furthermore, the model demonstrated impressive generalization, successfully identifying lenses in noisier "North" survey data despite being trained primarily on higher-quality "South" data. We also found that training with realistic, imbalanced ratios (e.g., 1:4000) improved real-world precision over standard balanced training sets.
Conclusion and Future Directions
Self-supervised learning provides a powerful framework for mining rare objects in massive astronomical datasets. By reducing reliance on human labels and minimizing computational overhead, this approach democratizes large-scale discovery. Future improvements will focus on refining data augmentations to better handle common artifacts like foreground stars and integrating community-driven validation through interactive web tools to further expand the catalog of known gravitational lenses.
Summary updated