Deep-CEE, a deep learning technique developed by scientists at Lancaster University, can efficiently localize and classify objects in wide-field color images to identify galaxy clusters in deep space. Deep-CEE — which refers to Deep Learning for Galaxy Cluster Extraction and Evaluation — builds on the work of astronomer George Abell, who in the 1950s manually analyzed around 2000 photographic plates to look for visual signatures of galaxy clusters. Deep-CEE’s developers used Abell galaxy clusters as the ground-truth labels in color images to develop an artificial intelligence (AI) model based on neural networks. The researchers trained the AI model to identify galaxy clusters by showing it examples of known, labeled objects in images until the algorithm was able to learn to associate objects on its own. Image showing the galaxy cluster Abell 1689. Courtesy of NASA/ESA. The researchers ran a pilot study to test the algorithm’s ability to identify and classify galaxy clusters in Sloan Digital Sky Survey (SDSS) images that contained many other astronomical objects. They determined an 80% confidence score threshold to be optimal to balance precision and recall. At this threshold, their model achieved a precision of 70% and recall of 81% for all predictions in the test set, indicating that the training set and parameters for the Deep-CEE model could be suitable for generalized object detection of galaxy clusters. The Deep-CEE technique could be used in combination with traditional methods to confirm galaxy cluster candidates, the researchers said. By applying Deep-CEE to wide-deep imaging surveys, the researchers hope to discover many new higher redshift and lower mass galaxy clusters. “We have successfully applied Deep-CEE to the Sloan Digital Sky Survey,” researcher Matthew Chan said. “Ultimately, we will run our model on revolutionary surveys such as the Large Synoptic Survey telescope (LSST) that will probe wider and deeper into regions of the universe never before explored.” Deep-CEE could also be a powerful tool when combined with catalogs or imaging data from other wavelengths. A diagram showing a high-level overview of the Deep-CEE model architecture. Courtesy of M.C. Chan and J.P. Stott, MNRAS submitted and based on Ren et al. 2015. By automating the discovery process, scientists can quickly scan sets of images, and return precise predictions with minimal human interaction. “Data mining techniques such as deep learning will help us to analyze the enormous outputs of modern telescopes,” researcher John Stott said. “We expect our method to find thousands of clusters never seen before by science.” The research was presented at the Royal Astronomical Society National Astronomy Meeting 2019, June 30-July 4 at Lancaster University. The paper has been submitted to Monthly Notices of the Royal Astronomical Society (MNRAS) and can be found on Arxiv here: https://arxiv.org/abs/1906.08784.