Photonic Neuromorphic Computing

• Photonic neuromorphic computing is an emerging field that uses the ultrafast speed and parallelism of light to execute neural-inspired processing in integrated photonic circuits.
• It employs optical matrix–vector multiplication, nonlinear activation functions, and spiking neurons to perform high-speed AI and signal processing with reduced energy consumption.
• Recent advances integrate devices such as MRRs, VCSELs, and RTDs with on-chip learning algorithms, pushing computational performance beyond traditional electronic architectures.

Photonic neuromorphic computing is an interdisciplinary field that leverages the ultrafast and highly parallel nature of light to implement neural-inspired computing paradigms in photonic hardware platforms. These systems are designed to emulate the physics of neurons and synapses within integrated photonic circuits, with the aim of surpassing traditional electronic architectures in speed, energy efficiency, and bandwidth for AI and complex signal processing tasks.

1. Fundamental Principles and Photonic Architectures

Photonic neuromorphic systems exploit the unique characteristics of photons—namely, speed-of-light propagation, wavelength division multiplexing, and low crosstalk—to perform neural computations such as weighted summation, nonlinear activation, and spatiotemporal integration in the optical domain. Core computing primitives include:

• Optical Matrix–Vector Multiplication (MVM): This operation is realized using passive photonic networks (e.g., Mach–Zehnder interferometer (MZI) meshes, microring resonator (MRR) weight banks), in which tunable phase shifts or resonance conditions encode synaptic weights. The optical signal pathways inherently offer massive parallelism and high bandwidth (Shastri et al., 2020, Li et al., 2023).
• Nonlinear Activation Functions: Nonlinearities are implemented via devices such as Mach–Zehnder modulators, electro-absorption modulators, semiconductor lasers (VCSELs, DFBs) and photonic memristive elements. For example, a modulator’s electro-optic transfer function or the excitable spiking dynamics of a VCSEL directly emulate neuron-like activation (Tait et al., 2016, Skalli et al., 2021).
• Spiking and Integrate-and-Fire Neurons: Spiking is achieved through devices with “all-or-none” threshold dynamics, including saturable absorbers in lasers, RTDs, or vertical-cavity lasers under optical injection (Robertson et al., 2021, Xiang et al., 2022, Owen-Newns et al., 28 Jul 2025).
• Reservoir Computing: Recurrent photonic networks with delayed feedback, spectral slicing, or coupled laser arrays form analog reservoirs for processing temporal data, relying on the inherent memory and dynamics of photonic circuits (Sozos et al., 2022, Şeker et al., 19 Jun 2024, Foradori et al., 15 Sep 2025).

Wavelength-division multiplexing (WDM), spatial multiplexing, and time-division multiplexing are deployed to increase network dimensionality within a single photonic substrate (Skalli et al., 2021, Foradori et al., 15 Sep 2025).

2. Device Technologies: Photonic Neurons, Synapses, and Memories

A diverse array of photonic and optoelectronic devices underpin neuromorphic photonic networks:

• Microring Resonators (MRRs): Serve as programmable, thermally or electrically tunable synaptic weight elements. Their resonance properties allow scalable crossbar-style MVM and analog memory functions (Tait et al., 2016, Lam et al., 29 Jan 2024, Foradori et al., 15 Sep 2025).
• Modulator-Class Neurons: Mach–Zehnder and depletion-mode MRR modulators offer phase- or amplitude-based activation with sinusoidal nonlinearity. These devices are compatible with standard silicon photonics and support continuous activation functions (Tait et al., 2016).
• Vertical-Cavity Surface-Emitting Lasers (VCSELs): Intrinsically excitable and highly nonlinear, VCSELs can operate as ultrafast spiking neurons with sub-nanosecond pulse widths. Arrays enable both spatial and time-multiplexed architectures. Modulation bandwidths exceed 30 GHz, with energy per spike as low as 10 fJ (Skalli et al., 2021).
• Resonant Tunneling Diodes (RTDs): RTDs uniquely realize excitable spiking, thresholding, and refractoriness, mimicking biological neurons; their N-shaped I–V characteristics support GHz–THz spiking operation with compact footprints and dual-mode (electrical/optical) control (Zhang et al., 6 Mar 2024, Owen-Newns et al., 28 Jul 2025).
• Integrated Photonic Memories: All-optical memristive devices (“memlumors”) and electro-optic analog memory cells based on capacitive storage are introduced for dynamic weight storage and synaptic plasticity, enabling sub-millisecond retention and in-memory computing (Marunchenko et al., 2023, Lam et al., 29 Jan 2024).
• DFB Lasers with Saturable Absorbers: Used as nonlinear spike generators within programmable photonic neural networks, supporting both linear and nonlinear spike computations fully in the optical domain (Xiang et al., 9 Aug 2025).

3. Dynamical Systems and Mathematical Isomorphism

Photonic neuromorphic circuits are shown to be isomorphic to continuous-time recurrent neural networks (CTRNNs) described by differential equations:

$$\frac{d\vec{s}(t)}{dt} = W\,\vec{y}(t) - \frac{\vec{s}(t)}{\tau} + W_\text{in}\,u(t), \qquad \vec{y}(t) = \sigma[\vec{s}(t)]$$

where $\vec{s}(t)$ is the neuronal state determined by voltages or optical intensities, $W$ the synaptic weight matrix (embodied by MRR weight banks), and $\sigma$ the physical nonlinear transfer function (e.g., modulator nonlinearity, spiking threshold) (Tait et al., 2016).

Dynamical bifurcation analysis—including cusp, pitchfork, and Hopf bifurcations—has been experimentally confirmed in silicon photonic CTRNNs by tuning MRR weights and external inputs, demonstrating correspondence between network dynamics and physical photonic circuits (Tait et al., 2016).

Reservoir computing approaches use either multiple physical nodes, time-multiplexed “virtual” nodes, or spectral slicing to encode high-dimensional projection of input signals, with only simple linear readouts requiring training—a paradigm highly suited to optical implementation (Sozos et al., 2022, Owen-Newns et al., 2022, Parto et al., 28 Jan 2025, Foradori et al., 15 Sep 2025).

4. Learning Algorithms and Training Strategies

Photonic neuromorphic systems support a range of training approaches:

• Hardware–Algorithm Co-Design: Some experimental PSNNs combine hardware inference (spiking) with external or collaborative training of synaptic weights, including supervised algorithms inspired by methods such as Tempotron or ReSuMe, with weight updates governed by learning kernels $K(t)$ (Xiang et al., 2022).
• In-Situ Learning: On-chip learning is demonstrated using spike-timing dependent plasticity (STDP) mechanisms and supervised plasticity, where the photonic chip directly measures spike timing discrepancies for synaptic updates. Learning windows are often of the form $K(t) = V_0 \left( e^{-t/t_s} - e^{-t/t_m} \right)$ (Xiang et al., 17 Jun 2025).
• Reservoir and Extreme Learning Paradigms: For architectures where only output layers are trained, methods include ordinary least squares regression for analog outputs, binary node “significance” weighting to exploit the sparse nature of spiking reservoirs, and direct use of time-multiplexed spike patterns as features (Owen-Newns et al., 2022, Owen-Newns et al., 2022).
• Reinforcement Learning in Photonics: Recent work demonstrates a spiking proximal policy optimization (PPO) algorithm mapped to a photonic SNN as an actor network and conventional ANN as a critic. Collaborative training includes surrogate gradients in software, stochastic parallel gradient descent (SPGD) in hardware, and in-situ finetuning for hardware-aware inference (Xiang et al., 9 Aug 2025).

5. Experimental Demonstrations and Performance Metrics

Experimental advances span from single-neuron PSNN chips to large MRR arrays and integrated OPO-based processors, with key metrics as follows:

System Type	Speed	Energy/Power	Performance/Accuracy	Scalability
Silicon MRR CTRNN	48 ps loop delay, GHz bw	≈0.22 mW/neuron (modulator-class), ~106 mW for 24 nodes	294× CPU acceleration (Lorenz), precise bifurcation dynamics	Scaling limited by laser & thermal tuning (Tait et al., 2016)
VCSEL/DFB PSNN	Sub-ns spikes, up to 30 GHz	≈150 μW per neuron (VCSEL), ≈10 fJ/spike	>96% MNIST accuracy, ~100% (Iris), >94% on MADELON	Time and spatial multiplexing, 2D/3D arrays (Skalli et al., 2021, Robertson et al., 2021, Owen-Newns et al., 2022, Owen-Newns et al., 2022)
Photonic Reservoirs	47 ps delay/node, 10 GHz+ (OPO)	Passive, sub-mW–μW/node	~100% MFI accuracy, 93% time-series prediction	16-node physical, 1000s virtual (time/spectral), robust to imperfections (Sozos et al., 2022, Şeker et al., 19 Jun 2024, Parto et al., 28 Jan 2025)
RTD-based SNNs	300 ps refractory (1 GHz), <1 ns	100 pJ/spike	93–96.5% Iris, THz operation possible	Dual-mode (electrical/optical), arrays and feedback memory (Owen-Newns et al., 28 Jul 2025)
Photonic RL Chips	320 ps latency / layer	1.39 TOPS/W (linear), 987.65 GOPS/W (nonlinear)	CartPole convergence (reward=200), 98.5% accuracy	Full 16×16 MZI + DFB-SA, software/hardware co-optimization (Xiang et al., 9 Aug 2025)
Silicon PSNN Chips	4 GHz spiking, up to 15.93 MHz video	CMOS-compatible, low	80% KTH video accuracy, 100× speed-up vs frame-based	Event-driven, in-situ learning (Xiang et al., 17 Jun 2025)

6. System Integration, Memory, and Practical Challenges

Monolithic integration of photonic and electronic devices on silicon photonic platforms is a central focus, allowing for MRR synapses, analog memory, and photodetectors to be seamlessly co-located on chip. Advances in hybrid integration—such as flip-chip bonding and 2.5D/3D packaging—enable scalable photonic–electronic convergence (Shastri et al., 2020, Lam et al., 29 Jan 2024, Yoo et al., 28 Mar 2024).

Key engineering and application challenges include:

• Memory Integration: Co-locating analog or memristive memory with photonic circuits minimizes reliance on inefficient DAC/ADC conversions and reduces energy consumed in data movement (“memory wall”) (Lam et al., 29 Jan 2024, Marunchenko et al., 2023).
• Device Variability and Control: Fabrication-induced variation in device parameters (e.g., resonance, loss) is combated using fine-tuning of thermal/electrical control, and hardware-aware training or finetuning strategies.
• Nonlinearity and Power Efficiency: Many photonic nonlinear devices require higher-than-ideal optical powers or offer limited dynamic ranges; ongoing advances focus on engineering devices for lower thresholds, higher modulation efficiency, and energy-frugal operation (Li et al., 2023, Xiang et al., 1 Sep 2025).
• Interfacing and Scaling: Integration of on-chip light sources (III-V/Si lasers, frequency combs), photodetectors, and signal routing remains an active area, with progress in vertical coupling (VCSELs) and high-density integration for large-scale systems (Xiang et al., 1 Sep 2025, Skalli et al., 2021).

7. Applications, Impact, and Future Outlook

Photonic neuromorphic computing has established applicability across domains where speed, parallelism, and energy efficiency are paramount:

• High-Speed Signal Processing: RF signal processing, channel equalization in optical communications, and high-speed imaging benefit from picosecond-scale hardware (Tait et al., 2016, Sozos et al., 2022, Şeker et al., 19 Jun 2024).
• Scientific Computing and Control: Emulation of differential dynamical systems (e.g., Lorenz attractors, PDEs) and adaptive control for robotics and autonomous vehicles leverage the continuous and recurrent architectures of photonic CTRNNs (Tait et al., 2016, Xiang et al., 9 Aug 2025).
• Real-Time AI and Computer Vision: Ultrafast image classification, edge-feature detection, and event-driven processing (retina-inspired encoding) support real-time applications in autonomous navigation, surveillance, and smart sensors (Robertson et al., 2021, Xiang et al., 17 Jun 2025).
• Memory and Learning: Inline photonic memory (memlumor, analog storage, feedback loops in RTDs) enables dynamic, hybrid volatile/non-volatile synaptic functions critical for online and real-time learning (Marunchenko et al., 2023, Owen-Newns et al., 28 Jul 2025).

Outlook trends include the continued development of materials (lithium niobate, chalcogenide PCMs, perovskites, 2D materials), advances in all-optical learning and analog in-memory computation, the emergence of hybrid electronic/photonic/ionic 3D architectures for brain-scale systems, and the application of bio-realistic, local learning rules (e.g., contrastive attractor learning, STDP) (Yoo et al., 28 Mar 2024, Li et al., 2023).

The field is poised for expanded industrial adoption as scalable device integration, robust fabrication, and hardware-aware software co-design coalesce to deliver practical, energy-efficient photonic neuromorphic processors for post-von-Neumann and post-Moore AI computation (Xiang et al., 1 Sep 2025, Li et al., 2023).