Advances in photonic reservoirs: principles, architectures, and future directions (<italic>invited</italic>)

Significance Photonic reservoir computing (PRC), which couples neuromorphic computing with photonic information processing, is emerging as a key route toward high-bandwidth, low-power intelligent hardware in the post-Moore era. In contrast to conventional electronic reservoirs, photonic reservoirs exploit the intrinsic advantages of light—massive spatial and spectral parallelism, ultrafast carrier and cavity dynamics, and low-loss propagation—to realize complex dynamical mappings and real-time signal processing directly at the physical layer. Optical fields naturally operate at frequencies many orders of magnitude higher than those of electronic circuits, enabling sub-nanosecond state updates and femtojoule-level switching energies. As a result, PRC systems exhibit ultra-high computational throughput and very low latency, which is particularly attractive for demanding tasks such as high-speed signal classification, time-series prediction, adaptive filtering, and closed-loop control. Moreover, the reservoir computing paradigm decouples feature extraction from parameter learning: a fixed, often randomly configured dynamical network performs high-dimensional state expansion, while only a linear (or weakly nonlinear) readout layer is trained. This avoids heavy gradient-based backpropagation within the reservoir and greatly reduces training complexity and energy consumption. When implemented in photonic hardware, this idea realizes a concrete form of "machine learning in the physical layer", where the device dynamics themselves embody the recurrent computation that would otherwise require large electronic neural networks.Progress Recent advances in PRC can be organized along three tightly coupled dimensions: theoretical foundations, system architectures, and application demonstrations. On the theoretical side, the basic principles of reservoir computing have been mapped to photonic platforms with increasing clarity. The key requirements—nonlinear response, high-dimensional state embedding, and fading memory—have been analyzed for a variety of optical media and cavity configurations. Both linear reservoirs with nonlinear optical or electronic readout, and intrinsically nonlinear reservoirs based on modulators, lasers, or nonlinear waveguides, have been formulated within a unified dynamical systems framework. This has enabled systematic studies of memory capacity, separation properties, and trade-offs among nonlinearity, noise tolerance, and stability, providing design guidelines for photonic implementations. At the level of system architecture, PRC has evolved from proof-of-concept benchtop setups to increasingly integrated and diversified platforms. Time-delay reservoirs that use a single nonlinear node with optical or optoelectronic feedback have been extensively explored, exploiting temporal multiplexing to emulate large recurrent networks. In parallel, spatial-multiplexed reservoirs based on integrated waveguide meshes, microring arrays, diffractive optical elements, and scattering media realize high-dimensional state expansion through complex interference and diffraction patterns. Hybrid and multiscale architectures combine delay loops with on-chip photonic networks, or couple photonic reservoirs with electronic or material reservoirs, thereby enriching the accessible dynamics and enabling multi-rate processing. Alongside these structural developments, advances in programmable photonics and functional materials—such as phase-change media, electro-optic platforms, and light-responsive materials—are beginning to endow reservoirs with reconfigurable connectivity, tunable nonlinearity, and multi-timescale memory at the device level. On the application side, photonic reservoirs have been validated on a broad range of tasks spanning sensing, communication, and intelligent control. In information processing, PRC has been applied to pattern recognition problems, including speech and image recognition, where its ability to perform ultrafast, analog preprocessing reduces the burden on subsequent digital classifiers. In optical and wireless communications, photonic reservoirs have demonstrated effective equalization and modulation format identification for high-speed channels, compensating for dispersion and nonlinear distortions in real time. For dynamical modeling, PRC has been used to forecast chaotic and nonlinear time series, reconstruct system attractors, and approximate unknown dynamics. More recently, the combination of PRC with reinforcement-learning-style training of the readout layer has enabled hardware-accelerated decision-making and adaptive control, suggesting a route toward photonic controllers that interact with the environment and update their strategies on the fly.Conclusions and Prospects Overall, the rapid development of photonic reservoir computing is pushing neuromorphic concepts beyond electronics and opening a new pathway toward high-bandwidth, low-power intelligent information processing. PRC illustrates how analog device physics, rather than large digital models, can serve as the core of learning systems, effectively turning optical components into recurrent computational substrates. Looking ahead, several directions are expected to drive the field forward. On the materials side, integrating highly nonlinear, low-loss, and nonvolatile photonic materials will allow reservoirs to combine strong nonlinear mapping with controllable memory and weight programmability. At the architectural level, co-design of multi-scale delay networks, waveguide meshes, and heterogeneous photonic-electronic subsystems will be crucial for building large-scale, reconfigurable reservoir chips capable of processing multiple data streams and modalities in parallel. Algorithmically, deeper integration of PRC with modern machine-learning frameworks—such as deep learning, reinforcement learning, and probabilistic modeling—may yield hybrid schemes in which photonic reservoirs act as front-end dynamical encoders or physical priors inside larger learning pipelines. Finally, as fabrication technologies mature and system-level integration improves, photonic reservoirs are expected to move from laboratory prototypes toward application-oriented modules for real-time sensing, edge intelligence, and autonomous control. In this sense, the evolution of PRC marks a broader transition in optical computing: from static, logic-oriented operations toward dynamic, parallel, and self-organizing physical intelligence, laying an important foundation for next-generation computing systems beyond the limits of conventional CMOS scaling.