Resonate-and-Fire Photonic-Electronic Spiking Neurons for Fast and Efficient Light-Enabled Neuromorphic Processing Systems

TL;DR

This work introduces a photonic–electronic resonate-and-fire spiking neuron built around a light-sensitive resonant tunnelling diode, enabling excitability and spike generation in response to nanosecond optical inputs at telecom wavelengths. By exploiting the RTD’s valley-region dynamics and tuning the resonance via inter-pulse timing and bias, the authors demonstrate bandpass spike filtering, temporal pattern recognition, and multi-wavelength fan-in with VCSELs. Key results include impulse-response characterization, chirp-based filtering, temporal feature detection, and a 4-bit digital-to-spike encoding scheme achieving high decoding accuracy, all at nanosecond-scale operation. The approach offers a low-power, high-speed pathway for temporal information processing in light-enabled neuromorphic systems and suggests scalability toward integrated photonic neural networks with wavelength-division multiplexing.

Abstract

Neuromorphic computing seeks to replicate the spiking dynamics of biological neurons for brain-inspired computation. While electronic implementations of artificial spiking neurons have dominated to date, photonic approaches are attracting increasing research interest as they promise ultrafast, energy-efficient operation with low-crosstalk and high bandwidth. Nevertheless, existing photonic neurons largely mimic integrate-and-fire models, but neuroscience shows that neurons also encode information through richer mechanisms, such as the frequency and temporal patterns of spikes. Here, we present a photonic-electronic resonate-and-fire (R-and-F) spiking neuron that responds to the temporal structure of high-speed optical inputs. This is based on a light-sensitive resonant tunnelling diode that produces excitable spikes in response to nanosecond, low-power (100 microwatt) optical signals at infrared telecom wavelengths. We experimentally demonstrate control of R-and-F dynamics through inter-pulse timing of the optical stimuli and applied bias voltage, achieving bandpass filtering of both analogue and digital inputs. The R-and-F neuron also supports optical fan-in via wavelength-division multiplexed inputs from four vertical-cavity surface-emitting lasers (VCSELs). This electronic-photonic neuron exhibits key functionalities - including spike-frequency filtering, temporal pattern recognition, and digital-to-spiking conversion - critical for neuromorphic optical processing. Our approach establishes a pathway toward low-power, high-speed temporal information processing for light-enabled neuromorphic computing.

Figures (11)

• Figure 1: a) Optically-sensitive RTD neuron displaying excitable spike firing in response to input light stimuli. b) Schematic diagram of a biological neuron exhibiting resonate-and-fire spiking responses. c) SEM image of the light-sensitive RTD of this work. d) Experimentally-measured I-V curve of the RTD neuron when operated in dark conditions ($0$ mW) and under infrared illumination of CW light at $1550$ nm with optical power of $0.9$ mW. The NDR region of the RTD is shaded in green and the black dashed line marks a typical biasing point in the valley region.
• Figure 2: Impulse response test of the light-sensitive RTD. Sub-threshold optical stimuli are injected into the RTD biased either at (a) $V_{RTD} = 0.697$ V or (b) $V_{RTD} = 0.710$ V in its valley biasing region. Top plots (red time traces) show the optical input signals injected in the RTD, with 5-ns (sub-threshold) square pulses encoded in the optical input (average optical input power of $33.0$ µ W). Bottom plots (blue time traces) show the electrical RTD response when biased at $V_{RTD} = 0.697$ V (left) and $V_{RTD} = 0.710$ V (right), revealing the occurrence of underdamped or overdamped oscillations depending on the biasing case. (c) Zoomed in I-V curve of the RTD, marking the two bias voltage points applied to the device for the results in (a) and (b).
• Figure 3: Characterisation of the RTD response to injected optical stimuli (average optical input power of $45.0$ µ W). when reverse biased at the valley operation point. a) RTD is biased in the underdamped regime ($V_{RTD} = 0.699$ V) for changing input pulse amplitude. b) RTD is injected with input pulses of equal amplitude for changing RTD bias voltage.
• Figure 4: Demonstration of the resonate-and-fire effect where the RTD responds to optical stimuli in the form of doublets with different temporal separations when biased in the underdamped regime ($V_{RTD} = 0.699$ V).
• Figure 5: Characterisation of the resonate-and-fire effect where the RTD responds to optical stimuli in the form of doublets with different temporal separations for increasing RTD bias and injection power. a) Comparison of the wide and narrow filter for the same bias voltage ($V_{RTD} = 0.690$ V) for a tunable bandpass filter width. Map of how resonance changes with bias displaying the (b) the widest filter (highest power) and (c) the narrowest filter (lowest usable power).