Free-space and on-chip diffractive deep neural network for next-generation optical computing

Diffractive deep neural networks (D 2 NNs), composed of multiple layers of trainable diffractive neurons, perform deep network operations entirely through the diffraction and interference of light. This approach has been widely studied in classification, 3D perception, and edge learning due to its zero-power consumption, ultrafast processing, and intrinsic parallelism of light. This review presents fundamental principles, design methodologies, and recent advances in D 2 NNs for next generation optical computing.