Coherent Optical Convolution Processor Based on MMI Structures for Deep Learning Applications

The neural networks in the all-optical domain can provide the advantages of high speed, energy-efficient deep learning algorithms, large bandwidth and high parallel computing. Optical neural networks have been proposed in the literature, but these structures are limited due to the requirement of hardware complexity. In this study, we present a new optical convolution processor based on coherent matrix multiplication using multimode interference (MMI) structures. This structure is suitable for complex deep learning tasks. In addition, the MMI is used, so it can provide the advantages of high bandwidth, large fabrication tolerance, accurate calculations and low loss. The silicon rib waveguide is used for optical waveguides, so it is compatible with the existing CMOS (Complementary Metal Oxide Semiconductor) technology. The structure be used for building real-value neural networks instead of only positive-value networks. Our structure is verified by using the numerical simulations based on the Finite-difference time-domain (FDTD) method. The Roberts operator for the edge detection of images is simulated with a footprint of 50x2100 μm2.