Recent Advances in Physical Reservoir Computing: A Review

• RC suits temporal/sequential processing because its reservoir transforms inputs into high-dimensional spatiotemporal patterns for readout analysis.RC derives from RNN models including echo state networks and liquid state machines.Source: §1 · p. 4
• Only the readout weights are trained, enabling simple, fast learning with substantially lower computational cost than standard RNNs.The reservoir input and recurrent weights remain fixed, while readout weights are trained using methods such as linear regression.Source: §1 · p. 4, Figure 1 · p. 6
• Physical RC replaces the RNN reservoir with nonlinear physical systems, substrates, or devices because adaptive training changes are unnecessary for reservoir implementation.This approach aims to realize fast information processing with low learning cost and is considered an unconventional computing paradigm based on novel hardware.Source: §1 · p. 5
• The review classifies recent physical RC advances by the physical phenomenon used for the reservoir and summarizes the characteristics, similarities, and differences of individual systems.It covers dynamical-systems, electronic, photonic, spintronic, mechanical, biological, and other physical RC systems.Source: §1 · pp. 5–6
• Physical RC could support low-cost real-time edge computing, but physical constraints make reservoir optimization difficult and leave practical efficiency issues unresolved.Hardware RC may reduce processing latency and machine-learning power consumption, while efficient physical systems and devices still require further development.Source: §1 · pp. 5–6

2.1. Basic framework

• RC fixes recurrent reservoir weights and trains only readout weights, using nonlinear high-dimensional mappings for efficient temporal pattern analysis.This differs from conventional RNN training, which uses gradient-based algorithms such as BPTT or RTRL to adapt recurrent models.Source: §2.1 · p. 7
• ESNs use discrete-time artificial-neuron reservoirs whose readout is trained under the echo state property, supporting fading-memory filters and universal approximation of bounded discrete-time inputs.The echo state property is empirically associated with reservoir spectral radius below unity and asymptotically removes dependence on initial conditions (Grigoryeva & Ortega, 2018).Source: §2.1 · pp. 7–8
• LSMs use biologically constrained spiking reservoirs, typically with excitatory and inhibitory LIF neurons, and train a memoryless readout to approximate fading-memory filters.Their liquid dynamics encode spike sequences into reservoir states, and universal approximation depends on point-wise separation and readout approximation properties (Maass et al., 2002; Maass & Markram, 2004).Source: §2.1 · pp. 8–9

2.2. Recent trends

• RC has been applied to classification, forecasting, generation, filtering, control, and system approximation because simple training supports low-cost, real-time processing of sequential data.Applications include spoken-digit recognition from MFCC features and image recognition after converting images into pixel sequences.Source: §2.2 · p. 9, Table 2 · p. 11
• Recent RC variants use multiple or evolving reservoirs, feature extraction, reinforcement learning, new learning algorithms, diverse reservoir elements, and theoretical analyses of reservoir dynamics.These developments extend the original models through architectural, algorithmic, and theoretical directions.Source: §2.2 · pp. 9–10
• Physical RC reservoirs require high dimensionality, nonlinearity, fading memory, and separation, while implementations increasingly use dynamical systems, physical substrates, devices, or neuromorphic hardware.Physical reservoirs target energy-efficient machine learning hardware or natural computing; an early fluidic system used water ripples for XOR and spoken-digit recognition.Source: §2.2 · pp. 10–12

3.1. Delayed dynamical systems

• Extended single-node delayed reservoirs achieved better performance, faster processing, and higher robustness than the original architecture while retaining simple physical implementation.Time-multiplexed inputs drive one nonlinear node, whose delay-line virtual-node states form the reservoir and support spoken-digit recognition and NARMA-10 prediction.Source: §3.1 · p. 13

3.2. Cellular automata

• Cellular-automaton reservoirs project encoded inputs through nonlinear evolution into expressive feature vectors, support long temporal memory with less computation than ESNs, and enable parallel hardware implementation.For MNIST, rule 90 gave the highest simulated accuracy and was implemented on an FPGA with competitive accuracy, speed, and power dissipation against other FPGA-based neural networks.Source: §3.2 · pp. 13–14
• Cellular-automaton architectures can combine chaotic or edge-of-chaos rules for input sensitivity with ordered rules for short-term memory.Parallel loosely coupled and layered deep architectures were also proposed, while binary operations and simple updates suit FPGA and GPU implementations.Source: §3.2 · pp. 14–15

3.3. Coupled oscillators

• Coupled oscillators provide reservoirs across mechanical, chemical, and phase-based systems, using interactions, synchronization, or biochemical dynamics for computation.Mechanical oscillators were evaluated on parity and spoken-digit tasks, DNA oscillators on temporal signal tracking, and phase oscillators on function approximation; mechanical systems can be compact and energy-efficiently fabricated with MEMS.Source: §3.3 · pp. 15–16

4. Electronic RC

• Electronic RC research targets low-training-cost machine-learning hardware using simpler reservoir configurations to reduce energy consumption, accelerate computation, and tolerate hardware imperfection and noise.The section covers analog circuits, FPGAs, VLSI designs, and memristive RC.Source: §4 · p. 17

4.1. Analog circuits

• Analog circuits implement single-node delayed-feedback reservoirs with time-multiplexed inputs, digital readouts trained by linear regression, and applications in recognition, memory-capacity estimation, and time-series prediction.The systems use Mackey-Glass or Chua nonlinear circuits, with 12-bit AD/DA interfaces; spike-based variants improve power efficiency, and optimal masks depend on the task.Source: §4.1 · pp. 17–18

4.2. FPGAs

• FPGA reservoirs support binary, stochastic, sigmoid, and spiking neurons, including online learning and LSM implementations that accelerate processing and reduce energy through parallelism, power gating, and approximate arithmetic.Applications include channel equalization, speech recognition, image recognition, and time-series prediction; FPGA single-node reservoirs also use Boolean or digitized Mackey-Glass dynamics.Source: §4.2 · pp. 19–20

4.3. VLSIs

• VLSI implementations include a 256-neuron, 33k-synapse ASIC whose temporal 3-bit parity performance peaks at critical dynamics, alongside pulse-based and hardware-efficient LSM designs.These approaches use asynchronous event-driven processing, hardware-friendly readouts, local learning, and adaptive rewiring to reduce implementation cost and support recognition or approximation tasks.Source: §4.3 · pp. 20–21

4.4. Memristive RC

• Memristive RC is organized into neuromemristive reservoirs combining neuron circuits with memristor synapses and reservoirs using memristive systems without neuron units.Memristors provide history-dependent nonlinear dynamics and are promising for energy-efficient computing.Source: §4.4 · pp. 21–22

4.4.1. Neuromemristive circuits

• Neuromemristive reservoirs use memristor conductance as synaptic weight and include cellular, crossbar, and analog nanosynapse designs for recurrent processing, readouts, and recognition tasks.A locally connected cellular model performed worse than a randomly connected ESN in time-series prediction, while crossbars support vector-matrix multiplication and adaptive conductance updates.Source: §4.4.1 · p. 22

4.4.2. Memristive systems and devices

• Neuron-free memristor networks nonlinearly map inputs into high-dimensional states, and simulations show element variability can improve performance when topology and input scaling are appropriately selected.Fabricated independent-memristor reservoirs performed well in image recognition and time-series prediction, while atomic-switch networks supported waveform generation.Source: §4.4.2 · pp. 23–24

5.1. Optical node arrays

• Optical node arrays range from chip-integrated SOA and microring networks to photonic-crystal and free-space systems, demonstrating logic, recognition, prediction, and imaging tasks.The experimentally demonstrated SOA reservoir used 11 of 16 nodes in a 4×4 swirl array and performed 2-bit XOR, header recognition, and spoken-digit classification. Digital masking reduced input rate by a factor of 40, while multimode circuits addressed loss accumulation.Source: §5.1 · pp. 25–26, Figure 8 · p. 26

5. Photonic RC

• Across photonic reservoirs, practical design must balance compactness and speed against short delays, accumulated optical loss, noise, phase sensitivity, and the need to optimize feedback and masking conditions.Proposed solutions include digital mask modulation, multimode low-loss circuits, modified readouts for phase changes, and gain located in optoelectronic neurons to support many physical nodes and multiple long delays.Source: §5.1 · p. 25, §5.2.1 · pp. 27, 29, §5.2.2 · pp. 29–30

5.2.1. Opto-electronic and optical feedback with gain

• Optoelectronic and all-optical delayed-feedback reservoirs use masked virtual nodes and offline readout, supporting NARMA-10 generation, nonlinear-channel equalization, and spoken-digit recognition at MHz rates.Noise degrades performance, but multi-valued preprocessing masks improve it; feedback-free operation can still support learning when response dynamics couple neighboring masked-input steps (Soriano et al., 2013; Ortín et al., 2015). Later systems reached million-words-per-second classification with approximately 17 GHz virtual-node spacing.Source: §5.2.1 · pp. 26–28

5.2.2. Optical feedback in a laser cavity

• Delayed laser-cavity reservoirs exploit the laser’s nonlinear response and achieve spoken-digit recognition, chaotic prediction, channel equalization, waveform classification, and optical packet-header identification.Mask performance improves when its frequency is near the ring-cavity relaxation oscillation, while packet-header recognition at 10 GB/s requires signal-to-noise ratio above 15 dB; two feedback loops improved prediction over one loop.Source: §5.2.2 · pp. 29–30

Overview

• Spintronics offers three reservoir approaches—spin oscillations, spin waves, and skyrmions—with potential for low-power, small-scale devices.Spintronics combines electron charge and spin in nanoscale electronics, including devices such as non-volatile storage.Source: §6 · p. 30

Spin oscillations

• Spin torque oscillator reservoirs use nonlinear current–frequency relations and transient spin dynamics; experimentally, they improved spoken-digit prediction over a system without an STO.The STO uses an MTJ with free and pinned ferromagnetic layers, taking preprocessed sound as current input and voltage as readout output.Source: §6 · p. 30
• STO-reservoir performance in short-memory and parity-check tasks depends on the duration of the input pulse voltage.The simulated reservoirs used single or multiple MTJs, random binary voltage inputs, time-varying resistance outputs, and LLG spin dynamics.Source: §6 · p. 31

Spin waves

• Spin-wave reservoirs exploit history-dependent transients and nonlinear interference, estimating input-pattern characteristics when electrode positions and computation duration are appropriately selected.Voltage stimulation changes magnetic anisotropy in a YIG layer, propagating phase differences through the material; output spin motions train the readout regression model.Source: §6 · pp. 31–32

Skyrmions

• Skyrmion reservoirs use current-induced nonlinear, history-dependent motion, with output voltage varying as a skyrmion moves toward and is annihilated at the drain.Interactions among multiple skyrmions can generate more complex behavior, making skyrmion fabrics potential reservoir implementations.Source: Figure 10 · p. 31, §6 · p. 32

Mechanical RC motivation

• Soft and compliant robots leverage complex body dynamics to generate the rich nonlinear behavior required for physical reservoir computing, an approach related to morphological computing.Their flexible bodies are harder to control than rigid robots, but the same complexity can support computation.Source: §7 · p. 32

Mass-spring networks

• Mass-spring and mass-spring-damper networks use externally forced nonlinear oscillators as reservoirs, supporting time-series approximation, robot-arm tasks, pattern generation, and locomotion learning.Inputs drive randomly selected nodes through external forces, while outputs are linear combinations of spring lengths; feedback enables pattern generation for robot and organism locomotion.Source: §7 · pp. 32–33

Tensegrity reservoirs

• Tensegrity reservoirs achieved stable gait generation and terrain classification, while ReCTeR used sensor-derived reservoir states to control a compliant planetary-rover prototype.ReCTeR contains 24 passive and 6 actuated spring-cable assemblies connecting non-parallel struts, and studies also examined learning rules and operating conditions.Source: §7 · p. 33

Muscular hydrostats

• Muscular-hydrostat reservoirs based on octopus-like soft arms learned motion in simulation and real silicone robots and demonstrated nonlinear approximation and body-dynamics control without an external controller.Their skeleton-free bodies have virtually unlimited degrees of freedom and highly complex, time-varying hydrodynamic dynamics.Source: §7 · p. 33

Related robotic systems

• Related RC-based robot-control studies explored pneumatically driven soft arms, spine-driven and dog-like quadrupeds, and less-compliant quadrupedal robots.These efforts extended the concept to controlling real robots and generating varied behaviors.Source: §7 · pp. 33–34

8.1. Brain regions

• Brain-region studies model prefrontal, visual, cerebellar, and striatal circuits as reservoirs and readouts for context-dependent, temporal, and cognitive computation.Prefrontal models use fixed recurrent cortical connections as reservoirs and modifiable striatal connections as readouts; visual cortex shows fading memory, while cerebellar and striatal models exhibit LSM properties.Source: §8.1 · pp. 35–37
• RC-based working-memory models address temporary information storage across tasks, but the biological plausibility and task-dependent switching of brain circuits require further investigation.Proposed models include cortical feedback networks, ESNs, multiple memory timescales, reward-modulated learning, and comparative RC analyses; evaluations should consider structural, dynamical, and functional properties.Source: §8.1 · p. 37

8.2. in-vitro cultured cells

• Cultured neuronal networks on microelectrode arrays implement hybrid biological-silicon reservoirs whose spike dynamics support stimulation classification, input reconstruction, and linear readout strategies.Studies varied stimulation patterns, window sizes, input coding, and output decoding; an early SVM readout was non-linear, leaving that reservoir’s effectiveness undetermined.Source: §8.2 · p. 38
• Optogenetically stimulated cultured neurons reproduced in-vivo short-term-memory results and achieved high classification performance when suitable MEA channels and multisecond spatiotemporal memory were used.Tasks included spike-template and musical-style classification, with channel selection performed before readout.Source: §8.2 · pp. 38–39
• Closed-loop cultured-neuron reservoirs controlled a mobile robot to maintain a target output, avoid obstacles, and traverse a maze using FORCE learning.The reservoir received feedback from the output in the robot-control setup.Source: §8.2 · p. 39
• Other living reservoirs include E. coli gene-regulatory networks and engineered bacterial cultures for classifying chemical inputs, with simulations showing partial XOR success and good separation performance.Perceptron readouts used measured mRNA and protein concentrations as reservoir states, while engineered cultures aggregated simple classifiers into more complex designs.Source: §8.2 · p. 39

Scope

• Novel physical reservoirs, including stimulation-dependent nano-scale substrates and unconventional materials, expand exploration of physical reservoir computing beyond the previously categorized types.The section identifies these candidates as contributions to unconventional computing methods.Source: §9 · p. 39

Nano-scale and material reservoirs

• A nano-scale quantum-dot and chemical-compound reservoir encodes environmentally induced absorption changes into emission patterns, with simulations confirming potential for image recognition.The compounds respond to pH or redox potential, altering signal transfer between randomly dispersed quantum dots.Source: §9 · pp. 39–40
• Material-based reservoirs using carbon nanotube–polymer mixtures achieved successful time-series prediction after evolutionary optimization of control voltages and electrode locations.The optimization algorithm determines input/output electrode locations and control voltage signals for the reservoir configuration.Source: §9 · p. 40

Quantum dynamics

• Quantum reservoir computing uses Hamiltonian evolution, ensemble measurements, and time multiplexing to expose many implicit hidden nodes, while spatial multiplexing can further enhance computational power; physical implementation remains challenging.For N qubits, the system has 2^N basis states, and multiple decoupled quantum reservoirs were proposed for spatial multiplexing.Source: §9 · p. 40

• Physical reservoirs exhibit diverse nonlinearity, transient response, transmission speed, and spatial dimension, making them appropriate for different tasks and data.Most systems are configured for properties including high-dimensionality, input history dependence, fading memory, and separation.Source: §10 · p. 40
• Physical reservoirs span network-type, time-delayed single-node, and excitable-medium architectures, each offering distinct scalability, hardware, or wave-based advantages with corresponding implementation challenges.Network reservoirs scale dimensionality but require massive recurrent interconnections; delayed-feedback reservoirs avoid those connections but are difficult to design; excitable media exploit waves but remain insufficiently understood.Source: §10 · pp. 40–41
• Effective physical RC requires suitable input preprocessing, temporal and spatial scaling, reservoir materials and hyperparameters, scalable architectures, and readout training compatible with physical properties.Readout training speed can limit total real-time computation even when the physical reservoir processes signals rapidly; a physical readout is one proposed solution.Source: §10 · p. 41
• Physical RC systems should be compared using task performance, processing speed, memory, power efficiency, and scalability, alongside measures such as kernel quality and generalization ability.Examples include classification accuracy, prediction error, words processed per second at low word error rate, memory capacity, task-specific power consumption, and chip-scale miniaturization.Source: §10 · pp. 41–42
• Physical RC hardware could enable low-power, high-speed online computation for dynamic edge-IoT data, while biological reservoirs may clarify real-time information processing in brain regions.Other systems can investigate information processing through natural phenomena in biology, chemistry, physics, and engineering.Source: §10 · p. 42
• Because physical RC remains at an early development stage, future work must improve practical evaluation, implementation technology, theoretical understanding, novel physical substrates, and hybrid systems.Direct technology comparisons remain premature because performance, speed, memory, power efficiency, and scalability depend strongly on implementation method.Source: §10 · p. 42