Neuromorphic photonics

Neuromorphic photonics is an interdisciplinary field that uses photonic or optoelectronic systems to implement models and processing principles associated with neuromorphic engineering. It combines the parallelism, bandwidth and propagation speed of optical systems with neural-network concepts such as weighted interconnection, nonlinear activation, recurrence, adaptation and spike-based information processing.^[1]^[2]

The term includes analog continuous-time photonic neural networks, spiking neural networks, reservoir computing, optical implementations of recurrent networks, and hybrid electronic–photonic processors. Some authors also group photonic accelerators for conventional artificial neural networks under the field, although a photonic matrix multiplier is not necessarily neuromorphic unless its architecture or dynamics implement neural or brain-inspired processing principles.^[3]

Research is motivated by applications requiring very low latency, high analog bandwidth, wavelength-level parallelism or direct processing of optical data. Demonstrated systems remain specialized prototypes. Their practical performance depends not only on the photonic core but also on lasers, modulators, photodetectors, electronic drivers, analog-to-digital converters, memory, thermal tuning and data movement.^[1]

Principles

File:Neuron.svg

A biological neuron. Neuromorphic photonics abstracts selected properties such as weighted inputs, integration, thresholding, spiking and interconnection rather than reproducing the full biological cell.

Neural abstraction

A common artificial-neuron model forms a weighted sum of inputs and applies a nonlinear function:

$y = f (\sum_{i = 1}^{N} w_{i} x_{i} + b),$

where $x_{i}$ are inputs, $w_{i}$ are synaptic weights, $b$ is a bias and $f$ is an activation function. In a photonic implementation, inputs may be encoded in optical intensity, phase, wavelength, polarization, time or spatial mode. Weighting can be performed with attenuators, interferometers, microring filters, diffractive elements, semiconductor gain, or programmable material states. The weighted signals are then combined and passed to an optical or optoelectronic nonlinear element.^[1]^[2]

Neuromorphic photonic hardware does not generally attempt a molecular replica of the nervous system. It instead selects computational features that are useful for a task. These can include continuous-time dynamics, event-driven spikes, leaky integration, refractory behavior, recurrent feedback, associative memory, adaptation and local learning.

Optical encoding and multiplexing

Optical information can be multiplexed across several physical dimensions:

Wavelength-division multiplexing assigns signals or neurons to different carrier wavelengths that share a waveguide.
Time-division multiplexing places values or virtual nodes in successive time slots.
Spatial multiplexing distributes signals across separate waveguides, free-space beams or pixels.
Mode and polarization multiplexing uses orthogonal waveguide modes or polarization states.
Coherent encoding represents signed or complex values in optical field amplitude and phase.
Intensity encoding represents non-negative values in optical power and can be detected directly.

These methods may be combined. Wavelength multiplexing is particularly important in integrated neuromorphic photonics because a single waveguide can carry many weighted channels, although each additional wavelength increases requirements for laser generation, resonance alignment, filtering and power control.^[4]

Analog and spiking operation

In rate- or amplitude-based systems, a neuron output varies continuously with its input. These systems are suited to matrix operations, dynamical systems and analog inference. In spiking systems, information is carried by discrete optical pulses whose timing, rate, wavelength or amplitude can encode data. A photonic spiking neuron may integrate optical or electrical inputs, emit a pulse when a threshold is reached and then enter a refractory state.^[5]

File:Action potential.svg

A biological action potential. Photonic spiking neurons reproduce abstract features such as threshold, pulse generation and refractory response on much shorter physical time scales.

The operating time scale of a photonic neuron is not tied to biology. Semiconductor lasers and nonlinear photonic devices can respond on picosecond or sub-nanosecond scales. The resulting machine may be called neuromorphic because of its information-processing model, not because it operates at biological speed.^[1]

Architectures

Broadcast-and-weight networks

The broadcast-and-weight architecture uses wavelength-division multiplexing to interconnect photonic neurons. Each neuron broadcasts its output on an assigned wavelength. At a receiving neuron, a bank of tunable wavelength-selective elements applies positive or negative weights, and a detector or balanced detector sums the weighted optical power. A nonlinear transmitter then generates the next optical output.^[4]

Microring-resonator weight banks are compact and compatible with silicon photonics, but their resonances are sensitive to fabrication variation and temperature. Calibration and thermal tuning consume energy, and dense wavelength channels can experience crosstalk. Broadcast-and-weight networks are therefore often optoelectronic: optical interconnection and weighting are combined with photodetection and electronic nonlinear processing.

Interferometer meshes

File:Mach-Zehnder interferometer+portlabels.svg

A Mach–Zehnder interferometer. Meshes of tunable interferometers can implement programmable linear transformations for photonic neural networks.

Meshes of Mach–Zehnder interferometers can implement unitary or general linear transformations by controlling phase shifts and coupler ratios. Singular-value decomposition permits a matrix to be represented using interferometer meshes plus diagonal attenuation or gain. Coherent optical neural networks use this approach for multiply–accumulate operations.^[6]

Interferometer meshes can be reconfigured and can represent signed and complex weights, but coherent operation requires phase stability. Imperfect couplers, loss, phase noise and thermal drift accumulate with mesh size. Many implementations use electrical phase shifters, so they are photonic or optoelectronic neural processors rather than strictly all-optical systems.

Reservoir computing

Reservoir computing maps an input into the nonlinear dynamics of a fixed recurrent system and trains only a linear readout. Photonic reservoirs can use delay loops, coupled lasers, interferometers, microring networks or disordered scattering media. Time-multiplexed reservoirs may create many virtual nodes from one nonlinear physical node, reducing hardware count at the cost of sequential processing.^[7]

Reservoir computing is well suited to temporal classification, prediction, channel equalization and control because the physical dynamics supply memory. Its limitations include task-dependent parameter tuning, difficulty interpreting internal states, and the energy and latency of input encoding and readout.

Diffractive and free-space networks

Diffractive optical neural networks encode transformations in spatially patterned layers. Light propagates through passive diffractive elements, and interference performs a trained linear transformation. Spatial light modulators and metasurfaces can provide reconfiguration. These systems offer large spatial parallelism but may be physically bulky, sensitive to alignment, and dependent on electronic cameras or detectors at their output.^[8]

Whether a passive diffractive classifier is described as neuromorphic varies. It implements a trained network topology, but generally lacks local state, recurrence, spiking or nonlinear activation within the passive optical layers.

Photonic in-memory computing

File:Wave guiding.gif

Light coupling between waveguides through an optical ring resonator. Related structures can combine routing, wavelength-selective weighting and optical memory elements.

Photonic in-memory computing co-locates storage and optical processing. Nonvolatile photonic devices can retain analog weights without continuous electrical power. Arrays of phase-change cells have been used for matrix–vector multiplication, synaptic weighting and spiking primitives.^[9]

Nonvolatile weights reduce static holding power, but programming phase-change materials requires heat and produces variability, drift and finite endurance. Optical absorption can also accumulate as networks scale.

Components

Photonic synapses and weights

An artificial synapse changes the influence of one signal on another. Photonic weights can be implemented with:

microring resonators and wavelength-selective filters;
Mach–Zehnder interferometers and phase shifters;
semiconductor optical amplifiers and variable optical attenuators;
phase-change, ferroelectric or electrochemical materials;
microelectromechanical photonic structures;
diffractive pixels, metasurfaces or spatial light modulators; and
optical interference between coherent fields.

Weights may be volatile or nonvolatile, binary or multilevel, and positive, negative or complex. Direct-detection systems require differential channels or balanced detection to represent negative values. Coherent systems can encode sign and phase directly but require stable optical phase references.^[1]

Photonic neurons and nonlinear activation

Linear optics efficiently performs addition and multiplication, but a multilayer neural network requires nonlinearity. Proposed neuron mechanisms include semiconductor-laser excitability, optical bistability, saturable absorption, carrier dispersion, the optical Kerr effect, phase-change transitions, stimulated Brillouin scattering, optomechanical effects and optical–electrical–optical conversion.^[2]^[5]

An optical nonlinearity alone is not necessarily a scalable neuron. A network element may also need gain or loss compensation, cascadability, fan-out, threshold stability, signal restoration and input–output isolation. Passive all-optical nonlinearities avoid photodetection but often require high optical intensity or resonant enhancement. Optoelectronic neurons can provide stronger and programmable nonlinear responses but incur conversion energy and electronic bandwidth limits.

Semiconductor lasers

File:Professor Ben Yoo - UC Davis Department of Electrical and Computer Engineering.jpg

A researcher handling a photonic integrated circuit. Neuromorphic photonic processors require co-integration of optical networks with sources, detectors and control electronics.

Semiconductor lasers can exhibit thresholding, excitability, refractory periods, gain and fast pulse generation. Devices used as photonic neurons include vertical-cavity surface-emitting lasers, distributed-feedback lasers, micropillar lasers and lasers with saturable absorbers. Optical injection, feedback or modulation can produce spiking, bursting and synchronization.^[10]

Laser neurons naturally provide optical output and gain, but they need pump energy and can be sensitive to bias, temperature, optical feedback and device variation.

Microring resonators

File:Double Optical Ring Resonator.png

Coupled optical ring resonators. Microrings can serve as wavelength-selective weights, nonlinear nodes, filters and delay elements.

Microring resonators are used as wavelength filters, tunable weights, modulators, nonlinear activations and recurrent elements. Their compact size and wavelength selectivity support dense wavelength-division multiplexing. The same resonance sensitivity that enables low-energy tuning also makes them vulnerable to fabrication mismatch and thermal drift. High-Q rings reduce operating power but narrow bandwidth and increase photon lifetime.

Phase-change materials

File:Graphene Lattice.svg

The lattice of graphene. Graphene and other two-dimensional materials have been investigated for fast nonlinear photonic neurons and saturable-absorption mechanisms.

Chalcogenide phase-change materials can switch between amorphous and crystalline states having different complex refractive indices. Intermediate states provide analog weight levels. Optical pulses can program some devices, enabling all-photonic memory and computation.^[11]

The principal limitations are programming energy, optical loss, endurance, drift, thermal crosstalk and device-to-device variability. Materials with lower absorption than conventional Ge₂Sb₂Te₅ are under investigation.

Photodetectors and optoelectronic neurons

Photodetectors convert weighted optical signals into current. Summation can occur naturally when several wavelengths illuminate the same detector. Electronics can then apply a nonlinearity, store state and drive an optical modulator or laser. This optical–electrical–optical approach is not all-optical, but it can offer gain, programmability and compatibility with electronic memory and control.

Learning and programming

Most experimental photonic neural networks are trained offline in conventional electronic computers. The resulting weights are mapped onto phase shifters, attenuators, rings or memory cells. Hardware-aware training can include measured loss, noise, quantization, drift and fabrication errors so the learned model is robust to the physical processor.

In situ training adjusts the physical hardware using measured errors or optical gradients. Proposed methods include adjoint-variable training, finite-difference perturbation, feedback alignment, equilibrium propagation and local spike-timing-dependent plasticity. Fully optical learning has been demonstrated in limited phase-change networks, but large-scale training usually retains electronic memory, control and optimization.^[9]

Calibration is distinct from learning. Calibration estimates how electrical controls map to optical weights and compensates for drift or component variation. A system may require frequent calibration even when its neural-network weights are fixed.

Applications

Telecommunications and radio-frequency processing

Neuromorphic photonic processors have been proposed for fiber-nonlinearity compensation, modulation-format recognition, channel equalization, header recognition, spectral analysis, radio-frequency filtering and cognitive radio. These tasks exploit high analog bandwidth and the ability to accept optical or microwave-photonic signals without first transferring every sample into a digital processor.^[2]^[7]

Artificial-intelligence acceleration

Photonic matrix processors can accelerate convolutional and fully connected layers. Frequency-comb systems exploit wavelength parallelism, while interferometer meshes and diffractive networks exploit coherent or spatial parallelism. Integrated photonic tensor cores have demonstrated high operation rates for convolution and matrix multiplication.^[12]^[13]

Operation counts for photonic cores do not by themselves establish system-level superiority. Comparisons must include data conversion, weight loading, memory access, laser efficiency and the precision required by the task.

Temporal inference and control

Recurrent and reservoir photonic networks can classify or predict time-dependent signals. Suggested applications include autonomous control, robotics, event-camera processing, radar, lidar and scientific instruments. Very low latency can matter more than arithmetic precision or throughput in feedback-control systems.^[1]

Sensing and edge processing

An optical processor can be placed near an optical sensor to reduce data conversion and movement. Examples include machine vision, spectroscopy, microscopy and fiber sensing. Practical edge systems must nevertheless supply stable light sources and operate across environmental variation.

Performance measures

Performance claims depend strongly on what is included in the accounting. Common measures include:

Latency, measured from input availability to usable output, including conversion and buffering;
Throughput, often expressed as operations per second, symbols per second or classifications per second;
Energy per operation or inference, including lasers, drivers, tuning, detection and electronic control;
Precision and noise, including shot noise, thermal noise, phase noise and quantization;
Accuracy, evaluated on a defined data set and compared with an appropriate electronic baseline;
Scalability, including the number of neurons, synapses, wavelengths, optical paths and independently programmable weights;
Cascadability and fan-out, describing whether one neuron can reliably drive later neurons;
Optical loss and gain, including splitter, waveguide, resonator and coupling losses;
Reconfiguration speed, distinct from the speed of optical inference;
Footprint, including control electronics, lasers and packaging where relevant; and
Stability, including sensitivity to temperature, fabrication variation, aging and vibration.

Reported tera-operations-per-second values often count multiply and accumulate operations within a photonic core. They should not be directly compared with complete processor power or application performance unless the same boundaries, precision and workload are used.^[1]

Advantages and limitations

Potential advantages include propagation at the speed of light within the circuit, very high carrier bandwidth, wavelength and spatial parallelism, low electromagnetic interference, and direct compatibility with optical communication and sensing signals. Analog physical computation can perform some linear operations with little incremental energy once light is present.

The principal limitations are:

Weak optical nonlinearity. Linear operations are natural in photonics, while compact, low-energy and cascadable nonlinear activation is difficult.
Data-conversion overhead. Electronic inputs and outputs may require high-speed digital-to-analog and analog-to-digital converters whose energy exceeds that of the optical core.
Laser efficiency and distribution. Optical loss requires source power, and wall-plug efficiency must be included in energy comparisons.
Analog noise and limited precision. Loss, detector noise, phase error and component variation accumulate across layers.
Thermal sensitivity. Resonant and coherent circuits require stabilization, trimming or frequent calibration.
Memory mismatch. Dense electronic memory is mature, whereas storing and updating large optical weight arrays remains difficult.
Packaging. Coupling, fiber attachment, electronic–photonic integration and heat removal can dominate cost and footprint.
Training. Backpropagation, optimizer state and weight updates are often executed electronically, limiting claims of fully photonic learning.
Benchmark inconsistency. Device-level throughput and energy figures may omit control, conversion and memory overhead.

Consequently, neuromorphic photonics is generally viewed as a complement to electronic computing for bandwidth- or latency-limited workloads rather than a universal replacement for digital processors.^[1]

Historical development

Optical neural-network research emerged in the 1980s from optical information processing, holographic associative memory and analog neural networks. Free-space optical systems used lenses, spatial light modulators and holograms to perform parallel matrix operations. Interest later declined as electronic digital computing and training algorithms advanced more rapidly.

The development of low-loss photonic integrated circuits, silicon photonics, wavelength-division multiplexing and compact semiconductor lasers renewed the field in the 2000s and 2010s. The term neuromorphic photonics became associated particularly with integrated systems combining neuron-like nonlinear dynamics with scalable photonic interconnection.^[2]^[3]

Major research directions during the 2010s included excitable laser neurons, silicon microring weight banks, integrated reservoirs, coherent interferometer meshes, diffractive neural networks and phase-change synapses. In the 2020s, work increasingly focused on system-level integration, frequency-comb tensor processors, in-memory photonics, hardware-aware training, nonlinear integrated neurons and electronic–photonic co-design.^[1]

Terminology and scope

Neuromorphic photonics, photonic neuromorphic computing, optical neural networks and photonic artificial intelligence overlap but are not identical. Neuromorphic photonics emphasizes neural dynamics or brain-inspired organization implemented with light. Optical neural networks emphasize the implementation of a neural-network model in optical hardware. Photonic AI is broader and can include non-neural accelerators, optical optimization and general matrix processors.

All-optical should be used only when the relevant signal-processing path does not require optical–electrical–optical conversion. A chip containing optical waveguides is not automatically all-optical if detection, nonlinear activation, memory or control occurs electronically. Conversely, an optoelectronic system may still be neuromorphic if its architecture implements neural processing.

References

↑ ^1.0 ^1.1 ^1.2 ^1.3 ^1.4 ^1.5 ^1.6 ^1.7 ^1.8 Shastri, Bhavin J.; Tait, Alexander N.; Ferreira de Lima, Thomas; Pernice, Wolfram H. P.; Bhaskaran, Harish; Wright, C. David; Prucnal, Paul R. (2021). "Photonics for artificial intelligence and neuromorphic computing". Nature Photonics. 15 (2): 102–114. doi:10.1038/s41566-020-00754-y.
↑ ^2.0 ^2.1 ^2.2 ^2.3 ^2.4 Ferreira de Lima, Thomas; Shastri, Bhavin J.; Tait, Alexander N.; Nahmias, Mitchell A.; Prucnal, Paul R. (2017). "Progress in neuromorphic photonics". Nanophotonics. 6 (3): 577–599. doi:10.1515/nanoph-2016-0139.
↑ ^3.0 ^3.1 Prucnal, Paul R.; Shastri, Bhavin J. (2017). Neuromorphic Photonics. CRC Press. doi:10.1201/9781315370590. ISBN 9781498725230 Check |isbn= value: checksum (help). Search this book on
↑ ^4.0 ^4.1 Tait, Alexander N.; Ferreira de Lima, Thomas; Zhou, Ellen; Wu, Allie X.; Nahmias, Mitchell A.; Shastri, Bhavin J.; Prucnal, Paul R. (2017). "Neuromorphic photonic networks using silicon photonic weight banks". Scientific Reports. 7: 7430. doi:10.1038/s41598-017-07754-z. PMC 5544723.
↑ ^5.0 ^5.1 Jha, Ashutosh; Huang, Chaoran; Peng, Hao-Tian; Shastri, Bhavin J. (2022). "Photonic spiking neural networks and graphene-on-silicon spiking neurons". Journal of Lightwave Technology. 40 (9): 2901–2914. doi:10.1109/JLT.2022.3149567.
↑ Shen, Yichen; Harris, Nicholas C.; Skirlo, Scott; Prabhu, Mihika; Baehr-Jones, Tom; Hochberg, Michael; Sun, Xin; Zhao, Shijie; Larochelle, Hugo; Englund, Dirk; Soljačić, Marin (2017). "Deep learning with coherent nanophotonic circuits". Nature Photonics. 11: 441–446. doi:10.1038/nphoton.2017.93.
↑ ^7.0 ^7.1 Vandoorne, Kristof; Mechet, Paul; Van Vaerenbergh, Thomas; Fiers, Martin; Morthier, Geert; Verstraeten, David; Schrauwen, Benjamin; Dambre, Joni; Bienstman, Peter (2014). "Experimental demonstration of reservoir computing on a silicon photonics chip". Nature Communications. 5: 3541. doi:10.1038/ncomms4541. PMC 3988801.
↑ Lin, Xing; Rivenson, Yair; Yardimci, Nezih T.; Veli, Muhammed; Luo, Yi; Jarrahi, Mona; Ozcan, Aydogan (2018). "All-optical machine learning using diffractive deep neural networks". Science. 361 (6406): 1004–1008. doi:10.1126/science.aat8084. PMID 30049787.
↑ ^9.0 ^9.1 Feldmann, J.; Youngblood, N.; Wright, C. D.; Bhaskaran, H.; Pernice, W. H. P. (2019). "All-optical spiking neurosynaptic networks with self-learning capabilities". Nature. 569: 208–214. doi:10.1038/s41586-019-1157-8.
↑ Peng, Hao-Tian; Angelatos, Gerasimos; Ferreira de Lima, Thomas; Shastri, Bhavin J. (2024). "Semiconductor lasers for photonic neuromorphic computing and artificial intelligence". APL Photonics. 9 (7): 070903. doi:10.1063/5.0212454.
↑ Ríos, Carlos; Stegmaier, Matthias; Hosseini, Peiman; Wang, Di; Scherrer, Torsten; Wright, C. David; Bhaskaran, Harish; Pernice, Wolfram H. P. (2015). "Integrated all-photonic non-volatile multi-level memory". Nature Photonics. 9: 725–732. doi:10.1038/nphoton.2015.182.
↑ Feldmann, J.; Youngblood, N.; Karpov, M.; Gehring, H.; Li, X.; Stappers, M.; Le Gallo, M.; Fu, X.; Lukashchuk, A.; Raja, A. S.; Liu, J.; Wright, C. D.; Sebastian, A.; Kippenberg, T. J.; Pernice, W. H. P.; Bhaskaran, H. (2021). "Parallel convolutional processing using an integrated photonic tensor core". Nature. 589: 52–58. doi:10.1038/s41586-020-03070-1.
↑ Xu, Xingyuan; Tan, Mengxi; Corcoran, Bill; Wu, Jiayang; Boes, Andreas; Nguyen, Thach G.; Chu, Siang T.; Little, Brent E.; Hicks, Damien G.; Moss, David J. (2021). "11 TOPS photonic convolutional accelerator for optical neural networks". Nature. 589: 44–51. doi:10.1038/s41586-020-03063-0.