Paper detail

Deep Learning nearby galaxy peculiar velocities

We explore how information in images of nearby galaxies can be used to estimate their distance. We train a convolutional Neural Network (NN) to do this, using galaxy images from the Illustris simulation. We show that if the NN is trained on data with random errors added to the true distance (representing training using spectroscopic redshift instead of actual distance), then the NN can predict distances in a test dataset with greater accuracy than it was given in the training set. This is not unusual, as often NNs are trained on data with added noise, in order to increase robustness. In this case, however, it offers a route to estimating peculiar velocities of nearby galaxies. Given a galaxy with a known spectroscopic redshift one can use the NN-predicted distance to make an estimate of the peculiar velocity. Trying this using relatively low resolution (1.4 arcsec per pixel) simulated galaxy images we find fractional RMS distance errors of 7.7% for galaxies at a mean distance of 75 Mpc from the observer, leading to RMS peculiar velocity errors of 440 km/s. In a companion paper we apply the technique to 145,115 nearby galaxies from the NASA Sloan Atlas.