Artificial Nano-Optoelectronic Neurons

• Artificial nano-optoelectronic neurons are nanoscale devices that merge optical, electronic, and mechanical processes to emulate neural spiking and synaptic functions.
• They employ diverse materials such as GaP, III–V compounds, graphene, and ScN to achieve excitable dynamics, weighted summation, and memory capabilities.
• Recent designs showcase applications in high-speed information processing, neuromorphic computing, and scalable hybrid optical–electronic networks.

Artificial nano-optoelectronic neuron denotes a nanoscale device or circuit in which neural computation is implemented through tightly coupled optical and electronic, or electro-optomechanical, degrees of freedom. Across the literature, the term covers full spiking neurons with threshold, latency, temporal summation, and refractory behavior, as well as neuron primitives such as weighted-summation cores, optoelectronic synapses, and passive nonlinear activation stages. Representative realizations include gallium-phosphide electro-optomechanical spiking cavities on silicon photonics, III–V nanowire photodiode–FET nodes, resonant-tunnelling-diode photodetectors and RTD–laser hybrids, Izhikevich-inspired CMOS–photonics neurons, graphene matrix–vector engines interpreted as dot-product neuron cores, and ScN-based optoelectronic synapses with persistent plasticity (Beltramo et al., 17 Jan 2026, Sestoft et al., 8 Sep 2025, Al-Taai et al., 2023, Lee et al., 2021, Gao et al., 2020, Rao et al., 2022).

1. Conceptual scope and taxonomy

Artificial nano-optoelectronic neurons fall into three recurring categories. The first comprises full neuron devices that implement excitable or spiking dynamics directly in hardware. This includes the GaP electro-optomechanical neuron, which exhibits all-or-none spiking, threshold tuning, temporal summation, and a refractory period (Beltramo et al., 17 Jan 2026); RTD-based O/E/O neurons and nanostructure RTD-photodetectors, which realize thresholded spiking through negative differential conductance (NDC) dynamics (Hejda et al., 2022, Al-Taai et al., 2023); and Izhikevich-inspired optoelectronic neurons combining photodetectors, transistor circuits, and a VCSEL or nanolaser (Lee et al., 2021).

The second category comprises linear neuron cores that realize synaptic integration but not the full nonlinear neuron. A central example is the graphene optoelectronic matrix–vector multiplication engine, where the row current

$I_j=\sum_i P_0\,T_i(V_{g,v,i})\,R_{ji}(V_{g,w,ji})$

maps to a weighted sum $I_j \propto \sum_i w_{ji} v_i$. In that architecture, synaptic weighting is encoded in graphene detector responsivity, input encoding is encoded in graphene spatial-light-modulator transmittance, and the nonlinearity is explicitly assigned to consecutive electronic circuits rather than to the optical hardware itself (Gao et al., 2020).

The third category comprises synaptic or activation building blocks that supply only part of neuronal function. ScN and Mg-doped ScN photoconductors emulate inhibitory and excitatory synaptic plasticity through persistent negative and positive photoconductivity, respectively, but do not by themselves implement spike generation (Rao et al., 2022). Passive all-optical nonlinear activation via pump-depleted second-harmonic generation in PPLN nanophotonic waveguides supplies a compact sigmoid-like activation stage, but not synaptic weighting or temporal integration (Fu et al., 25 Apr 2025). This broader usage implies that the phrase does not identify a single canonical device class; it identifies a design objective in which nanoscale photonic and electronic mechanisms realize the core primitives of neurons.

A common misconception is that the term necessarily implies a fully optical neuron. Multiple implementations are explicitly hybrid. The graphene platform performs linear weighted sums optoelectronically but delegates activation to electronics (Gao et al., 2020). The Izhikevich-inspired neuron uses optical input and optical output, but membrane, recovery, thresholding, and refractory dynamics are implemented with three transistors, resistors, and capacitors (Lee et al., 2021). The nanowire neuron built from two InP photodiodes and one InAs FET likewise converts optical excitation and inhibition into an electrical conductance state (Sestoft et al., 8 Sep 2025).

2. Materials platforms and nanostructures

The physical diversity of artificial nano-optoelectronic neurons is unusually large, but several platforms recur because they combine strong optical interaction with fast or memory-bearing electronic response. The GaP electro-optomechanical spiking neuron uses a 300 nm thick GaP one-dimensional optomechanical crystal nanobeam heterogeneously integrated on silicon-on-insulator photonic circuitry. The same nanobeam hosts an optical mode near 1550 nm with experimental optical quality factor $Q_o \approx 4.4\times 10^4$, a mechanical mode at $\Omega_m/2\pi \approx 3.078$ GHz with $Q_m \approx 1550$, and piezoelectric actuation through 1 μm-wide gold electrodes placed about 1.5 μm from the cavity (Beltramo et al., 17 Jan 2026). This is a genuinely co-localized optical, mechanical, thermal, and electrical nanostructure.

III–V nanowire platforms implement a different materials strategy. One nanowire neuron uses two InP p-i-n photodiodes with opposite doping polarity at the common gate node and one InAs nanowire FET separated from a Ti/Au gate by about 20 nm of HfO$_2$ on degenerately doped Si with 200 nm thermal SiO$_2$. The active on-chip area is $30\text{–}90~\mu\mathrm{m}^2$, and the device is explicitly described as CMOS-compatible, multiwavelength, and operable in the picowatt regime (Sestoft et al., 8 Sep 2025). A related III–V nanowire concept uses a T-shaped branched nanowire integrating two wavelength-selective npn phototransistors and a nanowire LED with a 5 nm InP quantum well, with a ~200 nm diameter main stem and ~50 nm diameter emitting branch, embedded in a shared quasi-2D waveguide (Winge et al., 2020). A later III–V nanowire design adds a floating capacitive memory node between photodiodes and a wrap-gate InAs FET, yielding a node with explicitly charge-based time-limited memory (Winge et al., 2023).

RTD-based neurons rely on quantum transport in double-barrier quantum wells. The nanostructure RTD-photodetector uses an In$_{0.53}$Ga$_{0.47}$As 5.7 nm quantum well between 1.7 nm AlAs barriers, together with a 250 nm InAlGaAs photoconductive spacer and a 500 nm diameter nanopillar top contact. The nanopillar restrains injection current while preserving a large optical window of $I_j \propto \sum_i w_{ji} v_i$0, yielding lower current per total area than earlier micrometer-scale RTD-PDs (Al-Taai et al., 2023). Related RTD–laser neurons use InGaAs/AlAs double-barrier structures on InP and commercial or modeled 1550 nm VCSELs or nanolasers to realize O/E/O spiking neurons (Hejda et al., 2022, Martins et al., 2024).

Graphene-based neuron cores exploit tunable electro-optics rather than excitability. One implementation uses graphene spatial light modulators and graphene photodetector arrays with extraordinary optical transmission metamaterials: periodic ring apertures with outer radius $I_j \propto \sum_i w_{ji} v_i$1 nm, gap $I_j \propto \sum_i w_{ji} v_i$2 nm, period $I_j \propto \sum_i w_{ji} v_i$3m, and resonance at $I_j \propto \sum_i w_{ji} v_i$4m. The same graphene monolayer supplies tunable absorption, transmittance, and photoresponsivity on a 1 μm pitch (Gao et al., 2020).

Scandium nitride supplies yet another route. Undoped ScN is degenerately n-type with Hall-measured $I_j \propto \sum_i w_{ji} v_i$5 and $I_j \propto \sum_i w_{ji} v_i$6, whereas Mg-doped ScN can become p-type with $I_j \propto \sum_i w_{ji} v_i$7, $I_j \propto \sum_i w_{ji} v_i$8, and resistivity about $I_j \propto \sum_i w_{ji} v_i$9. The films are tens to hundreds of nm thick and are grown by reactive DC magnetron sputtering at 800 °C on MgO(001), yielding a CMOS-process-compatible platform for excitatory and inhibitory photoconductive synapses (Rao et al., 2022).

A plausible synthesis is that nanoscale photonic confinement alone is not sufficient; the most successful platforms pair that confinement with a state variable that is either strongly nonlinear, as in RTDs and optomechanics, or history-dependent, as in ScN persistence and charge-storage nanowires.

3. Computational primitives and neuron mappings

The most direct hardware mapping of the artificial-neuron equation

$Q_o \approx 4.4\times 10^4$0

appears in the graphene optoelectronic MVM engine. Each input $Q_o \approx 4.4\times 10^4$1 is encoded as a graphene-SLM transmittance $Q_o \approx 4.4\times 10^4$2, each weight $Q_o \approx 4.4\times 10^4$3 as a detector responsivity $Q_o \approx 4.4\times 10^4$4, and row-wise Kirchhoff current summation yields the dot product. Signed arithmetic is implemented by differential coding, requiring four positive-only MVMs for $Q_o \approx 4.4\times 10^4$5 and $Q_o \approx 4.4\times 10^4$6, after which the results are recombined electronically (Gao et al., 2020). In neuron language, this is an analog dot-product core with explicit fan-in but external activation.

A different mapping appears in Izhikevich-inspired optoelectronic neurons. Their membrane node $Q_o \approx 4.4\times 10^4$7, recovery node $Q_o \approx 4.4\times 10^4$8, and laser-drive current are governed by

$Q_o \approx 4.4\times 10^4$9

$\Omega_m/2\pi \approx 3.078$0

$\Omega_m/2\pi \approx 3.078$1

Two photodetectors provide explicitly excitatory and inhibitory optical spiking inputs, while three transistors, two capacitors, and two resistors implement integration, thresholding, and refractory feedback. The result is an optoelectronic neuron whose optical output is generated by a VCSEL or nanolaser rather than by the membrane node itself (Lee et al., 2021).

Nanowire neurons encode excitation and inhibition structurally. In the InP/InAs three-nanowire neuron, illuminating the p-gate photodiode raises the InAs FET conductance and is interpreted as excitation, whereas illuminating the n-gate photodiode lowers it and is interpreted as inhibition. The two photocurrents sum at a common gate node, and the FET transfer curve supplies a sigmoid-like nonlinearity in optical power (Sestoft et al., 8 Sep 2025). In the charge-memory nanowire node, two wavelength-selective photodiodes of opposite polarity feed a floating capacitor–resistor memory node with

$\Omega_m/2\pi \approx 3.078$2

after which a nanowire FET and LED generate the optical output (Winge et al., 2023).

RTD devices implement the neuron differently: the nonlinearity is not a static transfer curve but an excitable trajectory. In O/E/O RTD–VCSEL neurons, optical fan-in is provided by a photodetector, the RTD biased near its NDC region serves as the excitable core, and a VCSEL transduces the spike back to optics (Hejda et al., 2022). In the nanostructure RTD-PD, optical input alone is sufficient to displace the I–V characteristic so that a bias point near the peak or valley crosses into NDC, yielding a deterministic spiking response (Al-Taai et al., 2023).

Not all neuron primitives include temporal state. Passive SHG-based PPLN nanophotonic waveguides instead implement only the activation stage. Under phase matching and low loss, the fundamental-harmonic amplitude obeys

$\Omega_m/2\pi \approx 3.078$3

with $\Omega_m/2\pi \approx 3.078$4 proportional to the input field amplitude; the resulting normalized FH output is sigmoid-like and can serve as a passive optical activation after an MZI-based linear layer (Fu et al., 25 Apr 2025). This suggests that neuron functionality can be decomposed across multiple nanophotonic devices even when no single element contains the full set of neuronal primitives.

4. Dynamical regimes: excitability, plasticity, and memory

Excitability is central to full spiking nano-optoelectronic neurons. The GaP electro-optomechanical neuron operates as a self-sustained electro-optomechanical oscillator that is injection-locked by an RF signal. Near the low-frequency edge of the locking range, optical perturbation pulses induce all-or-none spikes associated with a full $\Omega_m/2\pi \approx 3.078$5 phase excursion on the invariant circle, consistent with a saddle-node on invariant circle scenario. Experimentally, median spike amplitude becomes essentially input-independent above threshold, spike latency saturates around 15 μs for large perturbations, temporal summation is observed for two subthreshold pulses separated by a few microseconds, and the refractory period inferred from inter-spike statistics is about 30 μs (Beltramo et al., 17 Jan 2026).

RTD neurons realize excitability through NDC rather than phase locking. In the macroscopic O/E/O RTD–VCSEL neuron, suprathreshold perturbations generate about 100 ns optical spikes with a refractory period of about 90 ns, and paired-pulse experiments show a transition from absolute to relative refractoriness. The corresponding nanoscale model yields minimum reliable spike separation of about 300 ps, implying maximum spike rates of about 3.3 GHz (Hejda et al., 2022). The nanostructure RTD-PD provides a pure optically triggered variant: at the peak bias point the refractory period is about 120 ns, at the valley bias about 70 ns, and super-threshold optical pulses generate deterministic spikes of fixed shape (Al-Taai et al., 2023).

Delayed feedback turns RTD-based neurons into memory elements. In RTD–LD systems with optical delay, excitable pulses become temporal localized states that circulate as regenerative memory. The delayed FitzHugh–Nagumo reduction captures the slow-RTD limit, but a more realistic RTD–LD model shows that the RTD time scale controls whether pulse interactions are repulsive or attractive, thereby determining whether multi-pulse temporal localized states are stable or unstable (Romeira et al., 2017, Martins et al., 2024). This matters directly for memory capacity: slow-RTD regimes support stable multi-pulse storage, whereas fast-RTD regimes can destabilize multi-pulse states through attractive interactions induced by laser carrier dynamics.

History dependence also appears in non-spiking platforms. In Mg-doped ScN, persistent positive photoconductivity produces excitatory long-term potentiation with multi-exponential decay and a slow time constant $\Omega_m/2\pi \approx 3.078$6 s at low temperature, remaining large at room temperature. In undoped ScN, persistent negative photoconductivity produces inhibitory plasticity with decay constants differing by about an order of magnitude, for example $\Omega_m/2\pi \approx 3.078$7 s and $\Omega_m/2\pi \approx 3.078$8 s at 80 K. These devices explicitly demonstrate STM, LTM, STM-to-LTM transition, learning and forgetting, paired-pulse facilitation or depression, frequency-selective filtering, Hebbian learning, and logic operations (Rao et al., 2022).

Charge-based memory in III–V nanowire nodes operates on shorter, circuit-defined scales. In the nanowire neuron with two InP photodiodes and one InAs FET, a fast conductance change occurs within about 1 ms and recovery takes about 0.1–0.8 s, producing biologically relevant memory timescales (Sestoft et al., 8 Sep 2025). In the three-nanowire node with explicit $\Omega_m/2\pi \approx 3.078$9 storage, $Q_m \approx 1550$0 is engineered from 1 ns to 100 μs depending on network role (Winge et al., 2023). These results show that “memory” in artificial nano-optoelectronic neurons is not a single phenomenon: it can arise from persistence in semiconductor defects, charge stored on floating gates, slow carrier recovery in lasers, or delayed optical feedback.

5. Network architectures, learning, and demonstrated tasks

System-level organization is as important as the single neuron. One route emphasizes broadcast photonics and compact neuron nodes. A III–V nanowire network embeds neuron-like T-shaped nanowires in a single shared quasi-2D waveguide, with synaptic weight $Q_m \approx 1550$1 defined as the fraction of optical power emitted by node $Q_m \approx 1550$2 that is absorbed in node $Q_m \approx 1550$3. In that platform, an anatomically constrained insect central-complex navigation circuit preserves functionality while reducing connectivity footprint by at least an order of magnitude relative to conventional optical routing schemes (Winge et al., 2020). A related nanowire-node study shows that intentionally varying neuron memory times significantly improves reservoir-computing performance relative to single-$Q_m \approx 1550$4 reservoirs (Winge et al., 2023).

Another route uses programmable nanophotonic meshes. In scalable nanophotonic-electronic spiking neural networks, monolithic CMOS–silicon-photonics MZI meshes implement synaptic transforms, while co-integrated optoelectronic neurons supply rich spiking dynamics. Random backpropagation was demonstrated experimentally on-chip with a 6×6 MZI mesh and achieved 92/100 correct classifications on a simple Iris-derived task, close to the 94/100 of standard linear regression; contrastive Hebbian learning on MZI meshes improved RMSE by 11.21% on a random mapping task but remained below the ideal unconstrained network (Srouji et al., 2022).

A third route pursues high-throughput optical linear algebra rather than spike timing. The graphene MVM platform emulates GEMM by blocking and demonstrates singular-value-decomposition image reconstruction, support-vector-machine inference with loss 0.0626 on a GPP versus 0.0627 on graphene GEMM, and two-layer MLP inference with 92.3% versus 88.7% accuracy on MNIST10 and 78.7% versus 76.8% on Fashion-MNIST10 (Gao et al., 2020). Although that hardware lacks native nonlinearity, it directly realizes the multiply–accumulate-intensive part of neuron evaluation.

Some studies evaluate entire hybrid optical–electronic layers. Spatially varying nanophotonic neural networks implement a large-kernel spatially varying convolutional front-end in a flat metasurface camera, followed by a backend with about 2K parameters. The experimental system achieves 72.76% blind test accuracy on CIFAR-10, slightly above AlexNet’s 72.64%, thereby showing that nanophotonic weight banks can serve as a neuron layer rather than as a single neuron (Wei et al., 2023).

At the single-neuron or small-network scale, Izhikevich-inspired optoelectronic neurons were used in simulated fully connected and convolutional SNNs. Reported results include 90% accuracy on unsupervised MNIST learning and 97% accuracy on a supervised modified fully connected network, together with a nanoscale design point of 200 aJ/spike input, 10 fJ/spike nanolaser output, fanout of about 80 or 19 dB excess optical loss tolerance, and 10 GSpikes/s operation (Lee et al., 2021). Passive PPLN SHG activation, combined with silicon MZI linear optics, supports airfoil regression with $Q_m \approx 1550$5 and DermaMNIST classification at 82.66%, essentially on par with a ResNet-18 baseline at 82.50% (Fu et al., 25 Apr 2025). The diversity of these demonstrations indicates that artificial nano-optoelectronic neurons are being used both as event-driven spiking primitives and as analog activation elements in broader photonic learning systems.

6. Limitations, controversies, and emerging directions

Several limitations recur across implementations. In the graphene MVM engine, the nonlinearity is off-chip or off-core: “The generated photocurrents from the same row across different columns are added electronically and fed into consecutive circuits for nonlinear operations,” so the platform is not a self-contained spiking neuron (Gao et al., 2020). In ScN photoconductive synapses, the demonstrated devices are mesoscopic lateral structures rather than dense nanoscale arrays, and the long persistence that enables LTM is also a speed limitation for online plasticity (Rao et al., 2022). In the PPLN SHG activation, the process is passive and ultrafast but currently requires tens to hundreds of milliwatts of CW input to reach strong nonlinearity, and its phase-matching bandwidth is inherently narrow (Fu et al., 25 Apr 2025).

There are also trade-offs internal to the neuron dynamics themselves. In the RTD–LD delayed neuron, full-model analysis shows that shrinking the RTD time scale can reverse pulse interactions from repulsive to attractive, destabilizing multi-pulse temporal localized states and reducing memory capacity (Martins et al., 2024). In the GaP electro-optomechanical neuron, operation near the SNIC point yields desirable all-or-none excitability but also relatively long latencies and refractory times dominated by mechanical and thermal recovery (Beltramo et al., 17 Jan 2026). In the nanowire FET neuron, current prototypes prioritize biologically relevant millisecond and sub-second dynamics over maximum speed, even though related geometries suggest potential operation near 1 GHz (Sestoft et al., 8 Sep 2025).

A second misconception is that nanoscale devices automatically scale to large, accurate systems. Multiple studies emphasize that calibration, interconnect, and variability dominate at system scale. The graphene MVM work explicitly develops per-row calibration to absorb up to 20% device variation while keeping the MVM error standard deviation small (Gao et al., 2020). The insect-brain nanowire model tolerates some interconnect and memory variations, but threshold-voltage deviations above about 10 mV degrade navigation accuracy (Winge et al., 2023). MZI-mesh learning is limited not by the existence of an optical weight matrix but by heater crosstalk, calibration overhead, and the nonlocal relation between phase shifters and matrix entries (Srouji et al., 2022).

Current research directions are correspondingly concrete. One direction is fuller integration: O/E/O nanowire neurons by combining nanowire photodiodes, floating-gate memory, FETs, and LEDs (Winge et al., 2023) or integrating nanowire LEDs with the existing InP/InAs neuron node (Sestoft et al., 8 Sep 2025). A second is monolithic RTD plus nanolaser or VCSEL neurons that preserve RTD excitability while reducing parasitics to reach multi-GHz operation (Hejda et al., 2022, Al-Taai et al., 2023). A third is heterogeneous silicon–lithium-niobate integration to pair low-loss linear meshes with passive nonlinear PPLN activation (Fu et al., 25 Apr 2025). A fourth is architectural: tensorized or three-dimensional electronic–photonic integration to reduce the quadratic scaling burden of MZI meshes (Srouji et al., 2022). Taken together, these studies suggest that the field is converging not on one universal artificial nano-optoelectronic neuron, but on a set of interoperable nanoscale primitives—synaptic, dynamical, and nonlinear—that can be composed into neuromorphic photonic hardware.