Diffractive Optical Neural Network

In this optical setting, depth increases the complexity of the realizable linear field transformation rather than composing nonlinear activations. The system uses alternating learned phase modulation and free-space propagation operators. Each diffractive layer applies a learned phase-only modulation profile over the spatial grid. Local phase modulation becomes globally mixed after propagation because free-space propagation mixes spatial frequencies. Three diffractive layers are justified because early layers provide the largest redistribution of digit-field energy while later layers refine detector separation.