Not yet reviewed by Pith; the record is open.

We develop a machine learning-based framework to predict the HI content of galaxies using more straightforwardly observable quantities such as 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 root mean square error ({\sc rmse}) and the Pearson correlation coefficient (r). Considering SDSS photometry, 3$^{rd}$ nearest neighbor environment and line of sight peculiar velocities as features, we obtain r $> 0.8$ accuracy of the HI-richness prediction, corresponding to {\sc rmse}$<0.3$. Adding near-IR photometry to the features yields some improvement to the prediction. Compared to all the regressors, random forest shows the best performance, with r $>0.9$ at $z=0$, followed by a Deep Neural Network with r $>0.85$. All regressors exhibit a declining performance with increasing redshift, which limits the utility of this approach to $z\la 1$, and they tend to somewhat over-predict the HI content of low-HI galaxies which might be due to Eddington bias in the training sample. We test our approach on the RESOLVE survey data. Training on a subset of RESOLVE data, we find that our machine learning method can reasonably well predict the HI-richness of the remaining RESOLVE data, with {\sc rmse}$\sim0.28$. When we train on mock data from Mufasa and test on RESOLVE, this increases to {\sc rmse}$\sim0.45$. Our method will be useful for making galaxy-by-galaxy survey predictions and incompleteness corrections for upcoming HI 21cm surveys such as the LADUMA and MIGHTEE surveys on MeerKAT, over regions where photometry is already available.