GaP Electro-Optomechanical Spiking Neuron

• The paper introduces a GaP electro-optomechanical spiking neuron that integrates photonic, mechanical, and thermal functionalities via a SNIC bifurcation to yield all-or-none optical spikes.
• The device employs a CMOS-compatible nanobeam architecture with precise evanescent coupling and piezoelectric actuation, enabling tunable excitable thresholds and temporal summation.
• Measured characteristics reveal robust spike amplitudes with fast latencies (15–100 μs) and low energy consumption (~3 nJ per spike), paving the way for neuromorphic photonic applications.

A gallium-phosphide electro-optomechanical spiking neuron is a nanoscale device that implements excitable, neuron-like dynamics by integrating optical and electromechanical functionality within a single GaP nanobeam bonded to a silicon photonic circuit. The device is capable of generating all-or-none optical spikes in response to external perturbations, using photonic, mechanical, and thermal degrees of freedom coupled via optomechanical and piezoelectric mechanisms. Operating at telecommunications wavelengths (λ ≈ 1550 nm) with a mechanical resonance at 3.078 GHz, it is fabricated on a CMOS-compatible platform and supports features such as tunable excitable thresholds, temporal summation, and refractory periods, facilitating both neuromorphic and general-purpose optical computing applications (Beltramo et al., 17 Jan 2026).

1. Device Architecture and Material Properties

The neuron core is a 300-nm-thick GaP nanobeam, designed as a one-dimensional photonic crystal cavity with Gaussian-tapered elliptical holes to simultaneously confine an optical mode (λ₀ ≈ 1550.5 nm) and a high-frequency mechanical breathing mode (Ω_m/2π ≈ 3.078 GHz). The GaP nanobeam is heterogeneously bonded atop a silicon-on-insulator (SOI) photonic circuit. A standard Si waveguide (250 nm thick, ≈400–450 nm wide), separated from the GaP by a 280-nm SiN spacer, enables evanescent optical coupling.

Piezoelectric actuation is achieved via two gold electrodes (1 µm wide, positioned 1.5 µm from the nanobeam center), which flank the GaP beam. Gallium phosphide, with a refractive index n₀ ≈ 3.05 at λ = 1550 nm, also provides a substantial thermo-optic coefficient (κ_th ≈ 3.4×10^−5 K^−1) and high thermal conductivity (k_GaP ≈ 110 W·m^−1·K^−1). The material’s low two-photon absorption at telecommunications wavelengths facilitates high photon number operation without excess heating.

2. Coupled Optomechanical-Electrothermal Dynamics

The system is governed by three coupled ordinary differential equations for the state variables: the complex intracavity optical field $a(t)$, the normalized mechanical displacement $x(t)$, and the local temperature elevation $\Delta T(t)$. These are:

• Optical field:

$$\dot{a}(t) = \left[ i(\Delta\omega + g_0 x(t) + \kappa_{\text{th}} n_0 \Delta T(t)) - \frac{\kappa}{2} \right] a(t) + \sqrt{\kappa_\text{ex} s_\text{in}(t)}$$

• Mechanical motion:

$$\ddot{x}(t) + \Gamma_m \dot{x}(t) + \Omega_m^2 x(t) = 2\Omega_m g_0 |a(t)|^2 + F_0 \sin(2\pi f_d t)$$

• Thermal dynamics:

$$\dot{\Delta T}(t) = \Gamma_\text{th} |a(t)|^2 - \frac{\Delta T(t)}{\tau_\text{th}}$$

Here, $\Delta\omega$ is the laser-cavity detuning, $\kappa$ is the optical decay rate ($Q_o ≈ 4.4×10^4$, with $\kappa_\text{ex} ≈ \kappa/3$), $g_0/2\pi ≈ 600$ kHz is the vacuum optomechanical coupling rate, $\Omega_m/2\pi = 3.078$ GHz is the mechanical resonance ($Q_m ≈ 1550$, thus $\Gamma_m ≈ 2\pi\times2$ MHz), and $F_0$ and $f_d$ model the RF piezoelectric drive. The thermal time constant is $\tau_\text{th} ≈ 0.3$ μs.

3. Excitable Dynamics and the SNIC Mechanism

The neuron’s excitable behavior derives from a “saddle-node on invariant circle” (SNIC) bifurcation. When driven by an adequately strong blue-detuned optical pump ($P_b$), the system enters self-sustained, injection-locked oscillations with a stable phase equilibrium ($\theta_s$) and a nearby unstable saddle point ($\theta_u$) on the oscillator’s limit circle. Small perturbations cause only minor excursions; however, if a perturbation elicits a phase kick surpassing $\Delta\phi_\text{th} = |\theta_u - \theta_s|$, a full $2\pi$ phase slip, corresponding to a spike, is triggered before the phase relocks. This excitable threshold is directly linked to the SNIC scenario.

Temporal summation is observed: two consecutive subthreshold perturbations can combine to exceed the phase threshold, efficiently emulating coincidence detection principles. The system also expresses a well-defined refractory period, characterized by a lack of spike probability for input intervals $\Delta t \lesssim 30$ μs (as measured via interspike interval distributions under noise-driven conditions).

4. Measured Spiking Characteristics and Input–Output Nonlinearities

Spiking is detected as an all-or-none event—manifested as a sharp, polarity-defined pulse in the demodulated transmitted amplitude $S_T(t)$. Above threshold, the spike amplitude is robust ($\sim50\%$ of baseline modulation), invariant to further increases in perturbation strength, delineating the neuron-like all-or-none response.

Latency from perturbation to output spike is tunable from over 100 μs near threshold to $\sim$15 μs for stronger perturbations ($P_\text{pulse} \approx 0.75 P_b$), with spike temporal width (FWHM) of several microseconds, limited by the amplitude dynamics bandwidth ($\sim$100 kHz filter). The energy cost is approximately $$E_\text{spike} \sim P_b \times \tau_\text{spike} \approx 3$$ nJ (with $P_b = 1.51$ mW, $\tau_\text{spike} \sim 2$ μs), while the electrical actuation operates at $V_\text{RF} \lesssim 0.1$ V and negligible currents ($\ll 1$ mA; sub-nJ electrical energy cost).

The excitable threshold can be dynamically tuned by adjusting the RF drive detuning ($\Delta f = f_d - f_-$). For $\Delta f \approx 2$ kHz, the threshold in terms of optical perturbation is $\sim0.4 P_b$, shifting to $>1.0 P_b$ for $\Delta f > 10$ kHz. The output spike probability as a function of input amplitude follows a sigmoidal dependence, with the midpoint defining the operational threshold.

5. Device Metrics, Integration, and Compatibility

The neuron’s physical footprint is small: the GaP nanobeam is $\sim10$ μm long and $\sim700$ nm wide; the entire active area, including electrodes and waveguides, remains $\lesssim0.01$ mm². All fabrication steps (heterogeneous bonding, electron-beam lithography, dry etch, metal lift-off) are fully ME-compatible with standard SOI photonic foundry processes, supporting seamless integration into larger photonic circuits.

Latency is set by input amplitude and injection-locking parameters, within the 15–100 μs range for typical settings. The device's sub-nJ energy-per-spike consumption and fast recovery favor low-latency, energy-efficient processing suitable for edge computing.

6. Applications and Functional Significance

The integration of multimodal coupling within a compact, CMOS-compatible photonic platform enables several application domains:

• Edge neuromorphic photonics: The neuron’s μs-scale spiking and nJ energy budget make it suitable for spike-based optical processing systems operating at telecommunications wavelengths.
• On-chip optical pulse sources: The all-or-none excitable dynamics offer the ability to generate arbitrary, precise optical pulse trains for communication, metrology, or signal gating, without electronic bottlenecks.
• Cascadable spiking networks: Dual optical and RF ports allow construction of interconnected neurons, supporting advanced fan-in/fan-out and spike routing with pure photonic or hybrid signaling.
• Brain-inspired primitives: The system demonstrates key computation primitives—thresholding, coincidence detection (summation), and refractory gating—within an integrated platform directly analogous to excitable biological neurons and cardiac cells.

A plausible implication is that this device architecture, by combining optomechanical and electromechanical coupling in a versatile, standard-process-compatible form factor, lays a foundation for diverse on-chip neuromorphic and pulse-shaping photonic systems (Beltramo et al., 17 Jan 2026).