Neural networks sift 315,000 galaxy images for rare polar ring systems

The work

Technical · from the record

The authors assembled a visually inspected sample of polar ring galaxies (PRGs) and used 87 strong and good candidates with high-quality SDSS images to train convolutional neural network classifiers. Because the training sample was small, they used augmentation, central-object segmentation, model ensembling, and transfer learning from a CNN pre-trained on 1,000 synthetic PRG and 3,000 synthetic non-PRG images generated with GALFIT. Applying the final model to a catalog of 315,000 SDSS galaxies produced 3,246 galaxies with scores above 0.999, which were inspected by eye; the paper reports three PRGs discovered via the deep learning approach and four more found by visual inspection of about 2,200 ring galaxies. One discovered galaxy, SDSS J140644.42+471602.0, was fitted with CIGALE, giving a total stellar mass of 8.34 x 10^10 solar masses and a current star formation rate of 71 solar masses per year, which the authors note is limited by missing FUV data.

How AI was used

Galaxy images were prepared by background subtraction, Gaussian smoothing, threshold-based source detection and extraction of the central object, then resized to 80x80 pixels with bicubic interpolation and augmented with random flips and rotations. A CNN with two convolutional blocks, ReLU activations, dropout, pooling, a dense layer and a sigmoid output was trained with binary cross-entropy loss and the AdamW optimizer, with a plateau learning-rate scheduler and hyperparameters chosen by grid search. Two further architectures (one without convolutional layers, one including the segmentation step) were trained and combined into an equally weighted ensemble, and a separate shallow one-hidden-layer network was trained on a 25-fold augmented PRG set. For transfer learning, GALFIT was driven by a modified GalSim code to generate synthetic PRG and non-PRG light profiles, with the polar ring component's Sersic profile truncated by a hyperbolic tangent function and the ring position angle rotated; a CNN was trained on this synthetic set for eight epochs with Adam, most layers were then frozen, and a new output layer was retrained for 35 epochs on 87 real PRG images and 900 non-PRG catalog images. The trained models were run over an SDSS morphological catalog of galaxies at z < 0.1 to assign PRG scores, scores were thresholded to produce candidate shortlists, and the shortlisted galaxies were examined visually using SDSS and DESI Legacy Survey imagery. Photometry of one selected object was measured with the Aperture Photometry Tool and fitted with CIGALE using sfh2exp, bc03, nebular, dustatt_modified_starburst and dl2014 modules.