Hailstorms cause around 1 billion dollars in damage across the United States each year. At least a portion of this cost is associated with the inability to protect personal assets from damage in the short window of time offered by a severe weather warning. To address this problem, we developed a nowcasting model that uses UNet style convolutional neural networks (CNNs) to produce gridded severe hail forecasts for the next hour. One of the advantages of machine learning models is their ability to fuse large quantities of data from traditionally disparate sources such as ground observations and model output to produce a forecast. To exploit this hybrid predictor potential, these models are trained on the high-resolution (3 km spatial, 5 min temporal) output from the Warn-on-Forecast System (WoFS) numerical weather prediction (NWP) ensemble and remote sensing observations from Vaisala’s NLDN lightning detection system. Maximum expected size of hail (MESH) from the gridded NEXRAD WSR-88D radar (GridRad) dataset is used as the model’s truth labels. In addition to traditional machine learning optimization techniques such as hyperparameter searches and predictive feature selection, several different UNet architectures are compared to obtain a better machine learning model. The high-resolution nature of this data enables strategies such as using time as an additional dimension in a 3D UNet. This 3D model is compared against the effectiveness of a traditional 2D UNet. Finally, both models are compared against HAILCAST and simple logistic regression trained on 2 to 5 km updraft helicity to investigate their effectiveness.