We develop an interpretable deep learning model to forecast how the sizes of earthquakes evolve over time in geothermal fields. By combining physical data such as fluid injection, pressure, and past seismicity, the model predicts the full distribution of earthquake magnitudes—including the chances of larger, potentially hazardous events. Applied to real sites in California and Utah, it identifies early-warning signals and shows how operational changes influence seismic risk. This approach helps guide safer, data-driven decisions in managing geothermal energy.