Predicting the neutral hydrogen content of galaxies from optical data using machine learning

We develop a machine learning-based framework to predict the H_I content of galaxies from optical photometry and environmental parameters. We train the algorithm on z = 0-2 outputs from the MUFASA cosmological hydrodynamic simulation, which includes star formation, feedback, and a heuristic model to quench massive galaxies that yields a reasonable match to a range of survey data including HI.We employ a variety of machine learning methods (regressors), and quantify their performance using the slope of the predicted versus true relation, its root mean square error (RMSE), and Pearson correlation coefficient (r). Training on only Sloan Digital Sky Survey photometry, all regressors give r > 0.8 and RMSE ~ 0.3 at z = 0, led by random forests with r = 0.91, and a deep neural network (DNN) with comparable accuracy (r = 0.9). Adding near-IR photometry improves all regressors. All regressors perform worse with redshift, particularly at z ≲ 1. Slope values are generally sub-linear, so that we overpredict H_I in H_I-poor galaxies and underpredict H_I rich, because the regressors do not fully capture the scatter in the data. We test our framework on REsolved Spectroscopy Of a Local VolumE (RESOLVE) and Arecibo Legacy Fast ALFA (ALFALFA) survey data. Training on a subset of the observations, we find that our machine learning method can reasonably predict H Irichnesses in the remaining data (RMSE ~ 0.28 for RESOLVE and ~0.25 for ALFALFA). Training on mock data from MUFASA to predict observed data is worse (RMSE ~ 0.45 for RESOLVE and 0.31 for ALFALFA), with DNN well outperforming other regressors. Our method will be useful for making galaxy-by-galaxy survey predictions and incompleteness corrections for upcoming H_I 21 cm surveys on Square Kilometre Array precursors such as MeerKAT, over regions where photometry is already available.