Using convolutional neural networks to predict galaxy metallicity from three-color images

Abstract

We train a deep residual convolutional neural network (CNN) to predict the gas-phase metallicity ($Z$) of galaxies derived from spectroscopic information ($Z\equiv 12+\log (O/H)$) using only three-band $gri$ images from the Sloan Digital Sky Survey. When trained and tested on $128\times 128$-pixel images, the root mean squared error (RMSE) of $Z_{\mathrm{pred}}-Z_{\mathrm{true}}$ is only 0.085 dex, vastly outperforming a trained random forest algorithm on the same data set (RMSE $=0.130$ dex). The amount of scatter in $Z_{\mathrm{pred}}-Z_{\mathrm{true}}$ decreases with increasing image resolution in an intuitive manner. We are able to use CNN-predicted $Z_{\mathrm{pred}}$ and independently measured stellar masses to recover a mass-metallicity relation with $0.10$ dex scatter. Because our predicted MZR shows no more scatter than the empirical MZR, the difference between $Z_{\mathrm{pred}}$ and $Z_{\mathrm{true}}$ can not be due to purely random error. This suggests that the CNN has learned a representation of the gas-phase metallicity, from the optical imaging, beyond what is accessible with oxygen spectral lines.

13 pages, 6 figures, accepted to MNRAS. Code is available at https://github.com/jwuphysics/galaxy-cnns