StarcNet: Machine Learning for Star Cluster Identification

We present a machine learning (ML) pipeline to identify star clusters in the multi{color images of nearby galaxies, from observations obtained with Hubble Space Telescope as part Treasury Project LEGUS (Legacy ExtraGalactic Ultraviolet Survey). StarcNet (STAR Cluster classification NETwork) is multi-scale convolutional neural network (CNN) which achieves an accuracy 68.6% (4 classes)/86.0% (2 classes: cluster/non-cluster) for cluster nearly matching human expert performance. test performance by applying pre-trained CNN model galaxies not included training set, finding accuracies similar reference one. effect predictions on inferred properties comparing multi-color luminosity functions and mass-age plots catalogs produced human-labeling; distributions luminosity, color, physical characteristics are ML classified samples. There two advantages approach: (1) reproducibility classifications: algorithm's biases fixed can be measured subsequent analysis; (2) speed classification: algorithm requires minutes tasks that humans require weeks months perform. By achieving comparable classifiers, will enable extending classifications larger number candidate samples than currently available, thus increasing significantly statistics studies.