Connecting optical morphology, environment, and HI mass fraction for low-redshift galaxies using deep learning

A galaxy's morphological features encode details about its gas content, star formation history, and feedback processes which regulate its growth and evolution. We use deep convolutional neural networks (CNNs) to capture all of a galaxy's morphological information in order to estimate its neutral atomic hydrogen (HI) content directly from SDSS $gri$ image cutouts. We are able to predict a galaxy's HI mass fraction, $\mathcal M \equiv M_{\rm HI}/M_\star$, to within 0.25~dex accuracy using CNNs. The HI-morphology connection learned by the CNN appears to be constant in low- to intermediate-density galaxy environments, but it breaks down in the highest-density environments, i.e., for normalized overdensity parameter $\log(1+\delta_5) \gtrsim 0.5$ for the ALFALFA $\alpha.40$ sample, $\log(1+\delta_5) \gtrsim 0.1$ for the xGASS representative sample. This transition can be physically interpreted as the onset of ram pressure stripping, tidal effects, and other gas depletion processes in clustered environments. We also use a visualization algorithm, Gradient-weighted Class Activation Maps (Grad-CAM), to determine which morphological features are associated with low or high gas content. These results demonstrate that CNNs are powerful tools for understanding the connections between optical morphology and other properties, as well as for probing other latent variables, in a quantitative and interpretable manner.