B0 field inhomogeneities can negatively impact the image acquisition process, cause artifacts such as ghosting and blurring and introduce physiological noise in fMRI time series. It is common to address the static inhomogeneity components by using shim coils. Physiological motion during the acquisition can lead to temporal variations in the field configuration. Dynamic shimming necessitates fast, real-time estimation of B0 distortions. In this work, we augment projection-encoded FID readouts with a simple projection-based spatial encoding and train a neural network to learn the mapping from projection FIDs to field maps, which we obtain using a double-echo EPI sequence.