Deep learning in semiconductor manufacturing depends on training data that captures rare defects and complex geometries. Simulation can help fill gaps in limited fab data, while GPU acceleration supports full-chip processing of curvilinear mask and wafer models.