Development and applications of analog optical computing: A review

<p indent="0mm">Optical computing is a novel computing system that regards light as the basic information processing carrier. Compared with classical electronic computers, optical computing systems have the advantages of low latency, low power consumption, and parallel processing, which have the potential to solve the problems of computing capacity and energy consumption. Recently, analog optical computing systems represented by optical neural networks (PNNs) and coherent Ising machines (CIMs) have shown great potential in intelligent computing. This review summarizes the relevant research progress based on different PNN structures. Concretely, PNN systems are mainly divided into the planar integrated architecture and free space interconnected architecture. First, the planar integrated PNN is mainly based on Mach-Zehnder interferometer (MZI), micro-ring resonator (MRR), phase change material (PCM) structure and other types of unit structures. For the integrated PNN system, the research work on MZI-PNN, MRR-PNN and PCM-PNN related architectures is mainly reviewed. On the other hand, we also summarize the research progress related to the free-space interconnected PNN. Such non-integrated architectures are typically based on spatial light modulators (SLMs), digital micromirror devices, single-mode optical fibers, and electro-optic modulators. In addition, the training method and physical noise of PNN systems are further analyzed. Currently, most PNN architectures adopt an offline solution of “electrical domain training and optical domain testing”. Here, we present some improved PNN training methods. Meanwhile, several effective suppression schemes are summarized for the noise and error, for example, the MRR-PNN structure with convolution kernel pruning, the reduced rank PCM-PNN, and the binarized PNN, etc. In this review, we also summarize the related work on spatial optical CIM (SP-CIM) systems. Different from PNNs, the CIM system is used to solve combinatorial optimization problems. Especially, the SP-CIM system based on SLMs is one of the solutions to build large-scale CIMs, which has large computing scale, fast speed and low power consumption. Therefore, this review summarizes some research work on the improved SP-CIM structure. Finally, this review also analyzes several possible limitations of analog optical computing. For instance, the nonlinear transformation in PNNs is difficult to achieve via optical platforms. Meanwhile, the possible development trends of optical computing are summarized. This mainly includes: Further realizing the all-optical computing architecture and improving the noise robustness of systems, etc. In summary, by continuously improving the system architecture and algorithm performance, the optical computing platform can better exert its advantages in speed and energy consumption.