Integrated nano electro-optomechanical spiking neuron

TL;DR

The paper reports a CMOS-compatible nano-electro-mechanical neuron built from GaP integrated on a silicon photonic chip, operating at 1550 nm with a 3 GHz mechanical mode. It demonstrates excitable, SNIC-like spiking dynamics that are controllable via RF detuning and optical perturbations, including threshold behavior, temporal summation, and a refractory period. The work combines on-chip electromechanical injection locking with all-optical perturbations and supports on-chip optical spike generation, offering a scalable path toward edge neuromorphic computing. The integrated platform enables cascadable, low-latency, spike-based processing across optical and mechanical degrees of freedom, with potential for multiplexed, on-chip brain-inspired computation.

Abstract

Neuromorphic computing offers a pathway toward energy-efficient processing of data, yet hardware platforms combining nanoscale integration and multimodal functionality remain scarce. Here we demonstrate a gallium-phosphide electro-optomechanical spiking neuron that integrates optical and electromechanical interfaces within a single nanostructure on a silicon photonic chip operating at telecommunication wavelengths (1550 nm) and exploiting a 3 gigahertz-frequency mechanical mode. Our device displays excitable dynamics, generating optical spikes at its output, as in the spiking activity of neurons and cardiac cells and defined by the calibrated all-or-none response to external perturbations. This dynamic is consistent with the saddle-node on invariant circle scenario and associated features are demonstrated including control of excitable threshold, temporal summation and refractory period. Our device compact footprint and its CMOS-compatible platform make it well suited for edge-computing applications requiring low latency and establish a foundation for versatile brain-inspired optomechanical computing and advanced on-chip optical pulse sources.

Figures (4)

• Figure 1: a) Colored SEM image of the electro-optomechanical oscillator showing the piezoelectric Gallium Phosphide photonic crystal nanobeam cavity (in blue) suspended on top of a silicon-on-insulator bus waveguide (in red) and with gold electrodes (in yellow) positioned on both sides of the PhC nanobeam cavity. b) Experimental setup: tunable fibered IR laser (LASER), fibered polarization controller (FPC), $10$ GHz bandwidth electro-optic modulator (EOM), bias voltage supply (V$_{DC}$) and arbitrary function generator (AFG) constitute the excitation scheme. After the electro-optomechanical oscillator, the optical signal is split in two ports (50/50) and directed to an optical powermeter (PM) and a fast, 3.5 GHz bandwidth, photo-detector. The photodetected electrical signal is split 90% to the lock-in amplifier and 10% to the electric spectrum analyzer (ESA). The digital lock-in amplifier generates the RF driving at frequency $f_{d}$ at the gold electrodes. c) Power spectral density map of the photodetected signal recorded at the ESA as a function of the input power $P_{in}$. The highlighted power level $P_{in}=P_{b}=1.51$ mW corresponds to the one used for excitability measurements. d) Power spectral densities for $P_{in}=P_{b}$ and for three distinct frequencies offsets $f_{d}-f_{-}=-7, \ 10,\ 30$ kHz: before (bottom - yellow), inside (middle - orange) and after (top - red) the locking region, respectively. In c-d), $V_{RF} = 100$ mV. e) Full power spectral density map of the photodetected signal as a function of the driving frequency offset for the same parameters as in c-d).
• Figure 2: Spiking responses of electro-optomechanical oscillator subject to a perturbation. a) Temporal evolution of the optical input power sent to the device consisting of a bias, cw excitation $P_b = 1.51$ mW, plus a pulse train of rectangular perturbations with $1 \ \mu$s wide pulses of amplitude $P_\mathrm{pulse}$ and $200\ \mu$s period. b-d) Normalized time evolution of the output signal amplitude recorded by the digital lock-in amplifier after demodulation for $P_\mathrm{pulse}$ equals to (b) $0.1 \times$, (c) $0.45 \times$ and (d) $0.8 \times P_{b}$. The time from perturbation (TFP) is color-coded with a logarithmic scale from the perturbation pulse arrival (in yellow) to 200 $\mu s$ after the perturbation (in dark blue). e-g) Mumerical simulations of the normalized signal amplitude response after numerical demodulation for $P_\mathrm{pulse}$ equals to (e) $0.1 \times$, (f) $0.4 \times$ and (g) $0.5 \times P_{b}$ with the time evolution color-coded in the same way as in the experiment. In insets: phase space reconstructed from the demodulated quadratures $I(t)$ and $Q(t)$. The gray dotted circle represents the phase space dynamics of a pure phase oscillator in the unlocked regime. The dashed black lines in the experimental phase space diagram mark the average estimated positions of the stable and unstable phase equilibria and their difference $\phi_\mathrm{th}$, for the specific experimental conditions.
• Figure 3: Statistical analysis of the excitable response as a function of the perturbation amplitude. a) Median amplitude of the spiking response in absolute value, at $f_{d} - f_{-}=6$ kHz. The shaded region refers to the 30$^{th}$ and 70$^{th}$ percentiles. b) Median spike latency, at $f_{d} - f_{-}=6$ kHz. The shaded blue region refers to the 30$^{th}$ and 70$^{th}$ percentiles whereas the gray shaded one refers to the below-threshold perturbation amplitudes, $P_{pulse}\gtrsim0.44\times P_{b}$. c) Spike excitation probability as a function of the perturbation amplitude for three different values of $f_{d} - f_{-}$ : 6 kHz (yellow), 8 kHz (orange) and 10 kHz (dark red). d) Map of the spike excitation probability (color-scale) as a function of the perturbation amplitude and of the RF drive detuning $f_{d} - f_{-}$.
• Figure 4: Temporal summation and refractory period dynamics. a,d) Experimental time traces and b,e) phase spaces in the cases of two sub-threshold perturbations separated by a time delay $\Delta t=13 \ \mu$ s and $\Delta t=13 \ \mu$ s respectively. c,f) Numerical simulations of the dynamics in the phase space with two sub-threshold perturbations arriving with time delays $\Delta t=6 \ \mu$ s and $\Delta t=4 \ \mu s$ respectively. g) Experimental distribution of the noise excited inter-spikes time for $f_d-f_-=0$ kHz.