Significance Optical computing has been proposed for a few decades, although not yet widely applied in practice. This is partially because large-scale electronic circuits have been successfully developed as universal computing platforms. In recent years, it has been witnessed that the Moore’s law faces a bottleneck and photonic chips exhibit increasingly larger integrated arrays of tiny devices. Meanwhile, the emerging artificial intelligence has been inspiring a ubiquitous interest, which is featured by large amounts of matrix computation. This particularly triggers a renewed interest in optical neural networks for voice/image recognition, channel equalization in communications, and other data processing applications.

Compared with electronic neural networks, photonic neural networks potentially have the advantages of high speed and low power consumption. As a result, it has gradually attracted people's research interests in recent years. The photonic reservoir neural network is a kind of photonic recurrent neural network. Reservoir computing is expected to be suitable for processing sequence signals, and the training process is relatively simple. This could be greatly useful for optical fiber communications and wireless mobile communications.

This paper first introduces in detail the system configurations and technological characteristics of reservoir computing and presents the essential conditions to realize a reservoir. Then, the research progress in photonic reservoir computing is introduced through two different hardware implementations, which are called parallel and delay-based structures. Finally, the bottlenecks and corresponding solutions are discussed.

Progress Parallel photonic reservoirs are composed of optical node arrays. They have the potential to perform large-scale parallel computing. In 2008, researchers proposed a parallel photonic reservoir using semiconductor optical amplifiers. It outperformed traditional reservoirs on signal classification tasks at that time. This research team also reported a power-efficient experimental prototype in 2014, which only consists of passive waveguides, splitters, and combiners (Fig. 5). Microring resonators and photonic crystal cavities can also be used as nodes in parallel photonic reservoirs. Another kind of parallel optical reservoir is based on space optics. Maktoobi et al. demonstrated a reservoir with diffractively coupled nodes (Fig. 7), Rafayelyan et al. reported a reservoir based on multiple light scattering (Fig. 8), and Paudel et al. demonstrated a reservoir using speckles generated by mode interference in a multimode waveguide.

A delay-based photonic reservoir, also called a serial photonic reservoir, contains a single optical node with time-delayed feedback. Delay-based photonic reservoirs are easier to manufacture than parallel photonic reservoirs, but their parallel-computing capability is slightly poor. In 2012, Larger et al. demonstrated an optoelectronic delay-based photonic reservoir using a Mach-Zehnder modulator as the nonlinear node. Duport et al. reported the implementation of a photonic reservoir based on a semiconductor optical amplifier in the same year and it is the first delay-based all-optical reservoir (Fig. 11). In 2013, Brunner et al. used a semiconductor laser as the nonlinear node in a delay-based reservoir. Semiconductor lasers are power efficient, high-bandwidth, and widely used in modern fiber communications. In 2014, Dejonckheere et al. used a semiconductor saturable absorber mirror as a nonlinear node. It is the first photonic reservoir using fully passive nonlinearity.

There are schemes to improve the performance of reservoirs through a so-called hybrid configuration. In 2021, Nakajima et al. reported a photonic reservoir consisting of several delay-based reservoirs connected in parallel. The nodes are on-chip passive coherent cavities. This experiment realized the first image classification using an on-chip passive photonic reservoir.

By comprehensively comparing the recently proposed photonic reservoir computing schemes, we show a few features and evaluators, which can be used to estimate the capacity of new reservoir computing systems, including node type, nonlinearity mechanism, optical delay, array size, and input optical power. By organizing three tables (Tables 1--3), we clearly show the technical advantages and disadvantages of different reservoir neural network configurations, with an emphasis on the practical characteristics of nonlinear actuation functions. The nonlinear functional devices are based on semiconductor optical amplifiers, nonlinear microresonators, optical lasers with feedback, semiconductor saturable absorbers, and Mach-Zehnder modulators with a nonlinear transfer function.

Conclusions and Prospects Photonic reservoir neural networks can overcome some limitations of electronic neural networks, and their training processes are very simple. Photonic reservoirs have a broad development prospect. They have been used to implement speech recognition, chaotic time serie prediction, channel equalization, header recognition, and other functions.

We share some high-level perspectives on the future directions of photonic reservoir computing systems, by pointing out the potential technical issues and problems competing with an electronic version of reservoir computing. The array size, speed and accuracy of computation, and all-optical processing capability have been identified as three major tasks to advance the future development of photonic reservoir computing. According to these, some representative works recently published have been discussed, and the hybrid configuration of photonic reservoirs is particularly analyzed. We believe that all-optical input and output, hybrid configuration, on-chip implementation, and large-scale reservoirs are the future development directions of photonic reservoirs.

We believe this review would be of interest to the community of optical computing and neural networks as well as the community of integrated photonics.