UHF-MRI offers increased scanning sensitivity, but suffers from pronounced B0 inhomogeneities. B0 maps are essential for shimming and reconstruction at UHFs, but suffer from scanning overhead and low measurement confidence. Predicting B0 maps from survey scans can effectively address these issues. We assessed predictability of brain B0 maps at 7T and 11.7T using deep learning on a large synthetic B0 map dataset created from 540 CT images. We found that deep learning can predict 1) spherical harmonics for shimming and 2) the complete B0 maps. Our results indicate that B0 predictions work as effectively as acquiring B0 maps for shimming.