The discovery of multiple lensed galaxies is a cornerstone of modern cosmology, allowing astronomers to map dark matter and study the distant universe with unprecedented detail. However, as the volume of data from the James Webb Space Telescope (JWST) grows, the traditional method of manual visual inspection is becoming a significant bottleneck.
In the DARKLENS project, we developed a multiple-image identification method to automate this process, ensuring it remains feasible in the era of rapidly growing lensing datasets.
Gravitational lensing occurs when a massive foreground object, like a galaxy cluster, bends the light from a distant background galaxy, often creating multiple images of that same source. Identifying these images is traditionally time-consuming, taking hours of expert inspection per cluster field. Furthermore, many current methods rely on preliminary lens models—predictions of where images should be—which can introduce biases into the final results.
Our goal was to develop an identification method driven solely by the appearance (morphology) and color of galaxies, independent of their positions or the cluster's mass distribution.
AI-Driven Detection
In collaboration with the University of Ljubljana (UL FMF), we trained a convolutional neural network (CNN) to map galaxy images into high-dimensional vector representations, known as embeddings. By using contrastive learning, the model was taught to place multiple images of the same galaxy close together in this vector space, while pushing distinct galaxies further apart. After mapping all galaxy images into this vector space, a clustering algorithm was trained to mark nearby embeddings as candidates for representing the same galaxy.
Galaxy clusters figure Figure: Examples of predicted multiple-image candidate systems. Even when contaminated by other objects, these preselections significantly reduce the time needed for human experts to identify true systems.
Results
We found that using high-resolution images (20 milliarcseconds) in four short-wavelength color channels produced better results than using a full set of 10 filters at lower resolution of 40milliarcseconds. This highlights that a galaxy's detailed shape is often more important than its spectral energy distribution (color) for identification.
We qualitatively tested our method on the validation cluster MACS J1149+2223. The AI achieved 50% completeness score on this cluster, meaning it correctly grouped at least two lensed images of the whole lensed set, which is not precise enough to replace human expertise entirely, but enoudh to serve as a powerful accelerator. In a 20-minute inspection of the AI-predicted clusters, we successfully recovered 9 multiple image systems—roughly a quarter of the systems in that field. This level of preselection provides a sizable enough dataset to build preliminary lens models with a fraction of the usual time investment.
Looking Ahead: The Euclid Era
The lessons learned in this work are vital for upcoming missions like ESA’s Euclid. While JWST targets specific clusters, Euclid will observe the entire sky and is expected to discover thousands of new galaxy cluster lenses. At that scale, manual inspection of every image will be impossible. The DARKLENS project demonstrates that machine learning will be an essential tool for navigating this next frontier of space observation.
Partners:
- European Space Agency (ESA)
- Faculty of Mathematics and Physics, University of Ljubljana (UL FMF)
Similar projects
- Phot2Bin - Detection of multiple stellar systems from single-epoch photometry.