Introduction

Neurons are the fundamental building blocks of the nervous system and are used to transmit and process electrochemical signals. They operate in a liquid electrolyte and communicate with each other via synapses which control the gradual and activity-dependent coupling between the axon of pre-synaptic neurons and the dendrite of post-synaptic neurons. Neuromorphic computing emulates synapses and neurons using hardware for efficient brain-inspired computation1. Synapses are typically emulated using memory devices2 (e.g. non-volatile) which can be binary, multistate or analogue. However, to emulate the physiology of neurons, and their oscillatory dynamics3, additional electronic4,5, and/or optoelectronic6 circuitry is required at the expense of increasing power consumption and footprint. The development of miniaturized hardware that replicates such spatiotemporal dynamics is essential since oscillatory phenomena plays a crucial role in neuronal activity3, and can be readily observed in local field potential and electroencephalogram signals7. In insects, the rhythmic burst firing in clock neurons sustains circadian timekeeping of locomotor rhythms in drosophila brain8, and in the dragonfly visual system neurons respond to moving targets with a series of spike bursts enhancing the robustness of target-evoked responses9. In the mammalian brain, oscillatory dynamics governs propagating neuronal activity in cortical-hippocampal circuits10, and provides amplification of sensory inputs11. Also, respiration-locked neuronal oscillations in the olfactory bulb have been found to control delta band neuronal oscillations in the somatosensory cortex of mice12, revealing the importance of oscillatory phenomena in biological organisms.

A considerable amount of neuromorphic devices have been shown recently aiming at realizing bioinspired nonlinear sensory neuronal responses to build a new class of intelligent sensory systems13,14,15,16,17,18,19. They include stochastic phase-change neurons20, bioinspired organic artificial afferent nerves21, and artificial nociceptors based on diffusive memristors22. These systems are responsive to electrical or other physical signals such as pressure or temperature. More recently, efforts have been made towards neuromorphic components that are responsive to optical stimuli to process visual information. These include artificial optic-neural synapses for color-mixed pattern recognition using h-BN/WSe2 heterostructures23, artificial sensory neurons consisting of InGaZnO4-based optical sensors, and NbOx-based oscillator neurons24. Instead of adopting new materials with inherent challenges, for future scalable network systems, III–V compound semiconductors (e.g., InP or GaAs) have been proposed as neuromorphic light-sensitive and emitting devices25, an approach already well-matched with mature optoelectronic devices (e.g. LEDs, lasers and photodiodes). Notable examples include excitable neuron using a distributed feedback laser with a saturable absorber26, vertical-cavity surface-emitting laser (VCSEL) sources as artificial spiking neurons27, among other approaches (review in28). Recently, III–V photosensitive and emitting nanowires have been proposed for insect brain computational circuits29, and quantum resonant tunnelling diode photodetectors (RTD-PDs) to enable voltage-controlled negative differential resistance (NDR)30, which emulates the FitzHugh–Nagumo neuron model31. The intrinsic electrical gain effect from the voltage-controlled NDR has been extensively studied to provide excitable spiking and other dynamic neural phenomena32,33. Taking advantage of the NDR effect and photosensitivity of III–V RTD-PDs, optically-triggered excitable spiking has been achieved34. However, miniaturized photonic devices with seamless integration of both sensory and oscillatory neurons remains unexplored. Since biophysical studies reveal that even single neurons are capable of complex oscillatory dynamics35, such as resonate at multiple frequencies, emulating oscillatory phenomena in compact optoelectronic neuron circuits is highly relevant for new applications in neuromorphic optical computing systems.

In this work, we present a III–V neuromorphic photonic micropillar RTD photodetector (µRTD-PD). The µRTD-PD emulates sensory oscillatory neurons capable of processing optical information. We demonstrate that near-infrared continuous-wave (CW) light, within a certain intensity range, induces a region of NDR, enabling voltage- and optically-controlled NDR. Subsequently, optic analogue information is encoded into spatiotemporal large-amplitude electrical oscillations resulting in amplification of sensory inputs. We demonstrate that a single oscillatory neuron under pulse-modulated light provides excitation and inhibition of burst firing patterns. The stability and temporal response of the spatiotemporal burst firing patterns are analysed in the context of breather-type oscillations which represent emulation of neural activity in networks. Such spatiotemporal oscillatory patterns (burst firing) form the basis for the combined sensing, pre-processing, and encoding abilities within the same neuromorphic device, while previous demonstrations of multi-functions require multi-integrated optoelectronic components, thus increasing the complexity, cost and size of the neuromorphic system. The neuromorphic photonic sensory oscillator neuron demonstrated here paves the way to miniaturized biological-inspired processors of visual information. The approach is already compatible with proven III–V semiconductor platforms which have matured to meet the demanding industrial requirements in 3D sensing and light detection and ranging (LiDAR), thus uniquely suited to operate over a wide-range of wavelengths and optoelectronic applications.

Results

Device design and fabrication

The biological sensory oscillator neuron and its optical semiconductor counterpart are displayed in Fig. 1. In Fig. 1a is shown an illustration of a visual (eye) and processing (neuron) biological system converting light signals into neuronal oscillatory electrical responses to process visual information. This behaviour is emulated by the μRTD-PD sensory oscillator neuron device shown in Fig. 1b, and its equivalent lumped electric circuit is displayed in Fig. 1c. The active element of the μRTD-PD sensory neuron consists of a III–V n-type micropillar RTD with photosensitive layers and double barrier quantum well (DBQW) layers, Fig. 1b. In the inset of the micropillar μRTD-PD (left hand side) is shown the energy band diagram (at zero voltage) depicting the AlAs/GaAs/AlAs (~ 10 nm wide) DBQW region (the epilayer stack of the entire device is described in Methods). Under an applied forward bias (current flowing from top contact to bottom contact) the resonant energy levels in the GaAs quantum well are swept through the emitter Fermi sea, resulting in an N-shape current–voltage (I–V) characteristic with a local current maximum (referred as peak) and a minimum (referred as valley). Since the device is unipolar (n-type), an anti-symmetric I–V is also expected under reverse bias conditions. In both forward and reverse applied voltage scenarios, the DBQW acts as an electron energy filter controlling the current flow via quantum resonant tunnelling, where the maximum and minimum current flow corresponds to the highest and lowest probability, respectively, of charge carriers tunnelling through the quasi-bounded states of the DBQW. The lumped circuit of the photosensitive µRTD-PD (purple dashed region) is shown in Fig. 1c (see circuit model in Supplementary). For voltage and light conditions where negative resistance exists, Fig. 1d, the electrical gain provided by the NDR effect, together with the resonant tank LC circuit, enables self-sustaining oscillations among other rich dynamic phenomena36.

Fig. 1
figure 1

Optical sensory oscillatory neuron concept. (a) Schematic showing a biological optical sensory oscillatory neuron system found in the visual system that senses and converts optical signals into electrical oscillations. (b) The µRTD-PD sensory oscillator neuron device mimicking the biological system. In the inset is displayed the energy band diagram of the semiconductor stacked epilayers forming the micropillar showing the dimensions of the intrinsic double barrier quantum well (DBQW) region and the surrounding collector and emitter n-type GaAs light absorption layers. (c) Lumped electrical circuit of the µRTD-PD sensory oscillator neuron device, where Ls is the µRTD-PD connecting circuitry equivalent inductance, Rs accounts for the µRTD-PD series resistance and remaining circuitry resistance, CRTD is the equivalent device’s and circuitry capacitance, and f(V, Pin) is the light and voltage controlled current source that mimics the RTD current-voltage (I–V) characteristic (which depends on the bias voltage and incident light intensity), Iph is the photocurrent, and In is the associated noise of the system. (d) I–V characteristic exemplifying the operation of the μRTD-PD oscillatory neuron under dark and illumination conditions. Under dark conditions (black dashed trace) only positive differential resistance regions (PDR) are present. Under illumination conditions (purple solid line), an N-shaped I–V with ‘peak’, ‘valley’ and negative differential resistance (NDR) region is obtained.

Full size image

The NDR region is key to achieve self-sustained oscillations in our photosensitive neuron. The NDR is quantified by the peak to valley current ratio (PVCR = Ip/Iv) and by the ratio ∆V/∆I, where ∆V = Vv-Vp is the difference in valley voltage, Vv, and peak voltage, Vp, and ∆I = Ip − Iv is difference between peak current, Ip, and valley current, Iv. In the DBQW AlAs/GaAs/AlAs material system studied here the PVCR is typically much lower, (Ip/Iv) < 5, than in other materials systems, such as in the DBQW AlAs/InGaAs/AlAs, (Ip/Iv) > 1037. The low PVCR (or absence of PVCR) at room-temperature for the AlAs/GaAs/AlAs system is due to the lower energy difference of the DBQW AlAs/GaAs/AlAs and the larger electron effective mass when compared to other heterostructures.

As discussed in the next sections, for the III–V epilayer stack shown in Fig. 1b only positive differential regions (PDR) have been measured under dark conditions (black dashed line of Fig. 1d). However, when the µRTD-PD is illuminated, (purple solid trace in Fig. 1d), pronounced NDR effect can be achieved enabling voltage- and light-control of the NDR effect in our neuron devices. The light-induced effect is related to the dynamics of electron-hole pair generation and charge accumulation in the DBQW and the surrounding regions. In this work, the NDR effect is controlled by the light intensity and modulation conditions, which is explored to switch on and off self-sustained oscillations and operate these devices as sensory oscillator neurons.

To study the optical sensing and oscillatory phenomena, n-type µRTD-PD devices were fabricated in shape of micropillars with diameter, d, ranging from 6 to 10 μm, with both top (collector) and bottom (emitter) contacts (assuming forward bias). Scanning electron microscope (SEM) images of fabricated µRTD-PDs with diameters of 6 μm, 8 μm, and 10 μm are shown in Fig. 2. A dielectric material, SiO2, protects and electrically isolates the micropillar structures. Further, when combined with chemical pre-treatment using sulphide ammonia38, the SiO2 provides a surface passivating material. To enable light receiving operation, a half-circle metal layer contact was covering the top part of each micropillar, while keeping the other half uncovered for illumination purposes. Details on the fabrication of electrically-connected GaAs-based micropillar devices can be found in the methods section and in39.

Fig. 2
figure 2

Scanning electron microscope (SEM) images of fabricated µRTD-PD sensory oscillator neuron devices with increasing micropillar diameters (d). (a) Overview of an entire device with d = 6 μm. Inset is displayed a magnified image of the micropillar region. Magnified images of micropillars of devices with diameters of (b) d = 8 μm, and (c) d = 10 μm.

Full size image

Light-activation of negative differential resistance

The static I-V characteristics of fabricated µRTD-PD sensory neuron devices were investigated under dark and light conditions. The CW illumination at λ ~ 830 nm wavelength was provided by a laser diode through a single-mode fibre, and the light was coupled to the top micropillar using a lensed fibre (see details in Methods). In Fig. 3 is shown the schematic of the characterization setup used to obtain the static and dynamic (next section) photoresponse.

Fig. 3
figure 3

Schematic of the static and dynamic photoresponse characterization setup of the μRTD-PD. The device is electrically probed and illuminated using a continuous wave (CW) diode laser (λ ~ 830 nm) coupled to an optical fibre. The current–voltage (I–V) static characteristic of dark and illuminated devices is measured in the DC port of the bias-T. For characterization of the dynamic response, the laser diode intensity is modulated via direct current modulation using an external RF signal generator, and the output electrical photoresponse of the μRTD-PD is measured from the RF port of the bias-T.

Full size image

In Fig. 4a the dark and illuminated I–Vs are shown under forward and reverse bias conditions for micropillar devices ranging from 6 to 10 µm. Under forward bias, the dark I–Vs (dashed traces in Fig. 4a) show positive differential resistance, without signatures of NDR. Under reverse bias voltage, nonlinear regions can be observed around VDC = − 3 V, showing evidence of signatures of NDR. When illuminated, Pin, of ~ 3.7 mW and λ ~ 830 nm, a pronounced increase of the photocurrent is observed in the so-called peak voltage, solid traces in Fig. 4a. This is attributed to the increase of electron density in the 500 nm GaAs light sensitive layer. This leads to a built-in electric field modulation caused by photogenerated charge accumulation in the regions surrounding the DBQW region, resulting in light-induced NDR regions (NDR 1–3 in Fig. 4a). Under forward bias, pronounced photoresponse is also observed, but only positive differential resistance is obtained. We relate this to the fact that under forward bias, there is insufficient band bending of the potential surrounding the DBQW region to obtain NDR effect. We have also measured the I–Vs under illumination with forward and downward voltage sweeps to verify hysteresis behaviour in our neuron (Supplementary S1), a property characteristic of memristor-based neurons. As shown in Fig. S1, our neuron devices do not exhibit hysteresis behaviour which enables stable bias conditions to operate our neuron devices for reliable neuromorphic applications.

Fig. 4
figure 4

Static characteristics of the µRTD-PD oscillatory neuron showing light-activation of the negative differential resistance region (NDR) and the evolution of the light-induced NDR and the peak to valley current ratio (PVCR) with effective input optical power. (a) Current–voltage (I-V) characteristics in reverse and forward bias voltage under dark (dashed-line) and near-infrared (NIR, λ ~ 830 nm) illumination (solid line) conditions for devices with d = 6 μm (NDR 1), d = 8 μm (NDR 2), and d = 10 μm (NDR 3). (b) Dark and illuminated I-V characteristics showing on and off switching of the NDR region with increasing effective input optical power Peff. Inset is shown the PVCR as a function of Peff. Here we selected a power range were the NDR region is clearly visible (in this case between 0.1 and 0.5 mW).

Full size image

To analyze the light-activated NDR in µRTD-PD devices of various dimensions we consider the effective optical power39, \({P}_{\text{eff}}={P}_{\text{in}}({A}_{\text{pillar}}/{A}_{\text{spot}})\), launched into the micropillars where Apillar is the active area of the micropillar, and Aspot is device light active area (uncovered surface of the pillar) corresponding to the illuminated area by the lensed fibre with a numerical aperture of 0.2. We have analysed the I–V curves under various optical excitation conditions, Fig. 4b. For Peff ranging from 0.1 to 0.5 mW, we observe light-induced NDR regions (solid line in Fig. 4b at Peff = 0.2 mW), however for Peff > 0.5 mW the NDR signature is switched-off (dotted line in Fig. 4b). This indicates the NDR region in µRTD-PD devices can be switched-on for specific ranges of light intensity conditions. In the inset of Fig. 4b, the PVCR values of the NDR are presented as a function of Peff. The PVCR increases gradually until reaching a maximum of PVCR = 1.11 at Peff = 0.25 mW, and then decreases for larger Peff. Thus, given a sufficient large input power (here a threshold level of Peff,th > 0.5 mW) the NDR is turned off (see Supplementary S2 and Fig. S2 for analysis of the neuron photoresponse).

Light-activation of oscillatory behaviour

To illustrate the working mechanism of our light-activated oscillatory neuron, our device is analysed considering the optoelectronic circuit of Fig. 1c, which emulates the FitzHugh–Nagumo oscillator model40, a prototype of an excitable neuron system (see Supplementary, section S3). Under dark conditions, Fig. 5a, the I–V curve exhibits a smooth negative resistance region. As a result, there is not sufficient gain to sustain self-oscillations, and the neuron system tends to evolve to its steady-state value. This is illustrated by the small limit cycle (purple curve in panel (a)) and by the damped oscillations after initial conditions, panel (b). When illuminated, panel (c), and DC biased within the negative resistance region, sufficient gain is provided by the negative resistance and oscillatory behaviour can be sustained after initial conditions, panel (d), which is characterized by a stable limit cycle (purple line in panel c), i.e., a closed-orbit attractor in the current–voltage phase space.

Fig. 5
figure 5

Illustration of the oscillatory behaviour of the artificial neuron under dark and illumination conditions. (a) I–V fitting of the experimental dark I–V curve (black line) showing in the current–voltage phase space a small perturbation of the stable steady state due to initial conditions (purple line). (b) Simulated output voltage when the bias voltage is 3.4 V showing the system evolving to its steady-state value after initial conditions. (c) I–V fitting of the experimental illuminated I–V curve showing a large limit cycle (purple line) characterized by a closed-orbit attractor in the current–voltage phase space. (d) Simulated output voltage when the bias voltage is 3.4 V showing the system evolving to stable self-sustained oscillations.

Full size image

In what follows, we present the results of generation of oscillatory behaviour when the NDR in the µRTD-PD sensory device is activated by CW NIR light. We employed an identical characterization setup described in the previous section (Fig. 3). Here, the µRTD-PD oscillatory response was measured through the RF port of the bias-T that was connected to an oscilloscope to characterize its oscillations in time domain (see Methods).

We focus our analysis on a µRTD-PD device with a diameter of d = 6 μm. Fig. 6a shows the evolution of the I–V characteristics under CW NIR illumination conditions ranging from Pin = 6.8 mW to Pin = 14.3 mW. To simplify the analysis in this section (and in the following sections), the input power parameter, Pin, is used instead of effective power, Peff. When the region of NDR is light-activated, the device exhibits large-amplitude voltage self-sustained oscillations, Fig. 6b. The self-oscillations correspond to a limit cycle in the I–V phase space, Fig. 6. In Fig. 6b(i)–(iii) is shown the corresponding voltage oscillations ~ 350 kHz under increasing optical input power Pin collected from the 6 μm µRTD-PD using a reverse bias of VDC = − 3.4 V. The self-oscillation frequency is mainly determined by the equivalent circuit’s resistance, Rs, equivalent circuit’s inductance, Ls (which include the length of electrical metal contact pads, coaxial cables length and bias-T), µRTD-PD’s intrinsic capacitance, CRTD, and the nonlinear I–V characteristic negative resistance, f(V, Pin) (see in Fig. 1c the equivalent lumped-circuit). The self-oscillations observed here are lower than what is feasible with RTD-based devices whose quantum resonant tunnelling effect persists well beyond GHz41. Optimisation of the circuit, namely reducing coaxial lines parasitic inductance in the setup in combination with a circuit for stabilizing high-speed RF oscillations will enable higher speed oscillations42. The self-oscillation frequency of the circuit under CW illumination does not change substantially, showing only around 5% tunability (Fig. S5 in Supplementary). Significant changes are expected for devices designed for larger PVCR. Methods to increase the PVCR of our neuron device includes to modify the DBQW material composition or the relative barriers and well thicknesses.

Fig. 6
figure 6

Continuous-wave (CW) near-infrared (NIR, λ = 830 nm) illumination and activation of a region of negative differential resistance (NDR) and voltage oscillations. (a) Static I–V characteristics under dark (black dots) and illuminated (red, green and blue dots) conditions for d = 6 μm µRTD-PD under different illumination (λ = 830 nm) with optical power Pin = 6.8 mW, 11.1 mW and 14.3 mW. The inset shows a SEM image of a fabricated 6 μm sized μRTD-PD device exhibiting self-oscillation with illumination. (b) Left panel: self-sustained oscillations at a constant reverse bias voltage (VDC = − 3.4 V) and increasing illumination, (i) Pin = 6.8 mW, (ii) Pin = 11.1 mW, and (iii) Pin = 14.3 mW. Right panel: Self-sustained oscillations at a constant optical power Pin = 11.1 mW for increasing reverse bias (iv) VDC = − 3.1 V (close to the peak), (v) VDC = − 3.4 V (at the NDR centre), and (vi) VDC = − 3.6 V (close to the valley).

Full size image

Regarding amplitude of the oscillations, the magnitude of the amplitude is determined by the NDR voltage difference (∆V = Vp − Vv), and the difference between peak and valley current (∆I = Ip − Iv), which is controlled by the illumination conditions. In Fig. 6b(i)–(iii), the self-oscillations have lower amplitude under lower intensity illumination (Pin = 6.8 mW), mainly because of the low ∆I. The maximum power output of self-oscillations can be estimated as \({P}_{RF max}=(3/16)(\Delta V\Delta I)\)43. Therefore, as the illumination intensity increases, ∆I is higher, which leads to larger amplitude of self-oscillations. This monotonous trend is seen until the PVCR ratio increases. Assuming constant NIR illumination conditions, the self-oscillations can also be controlled by the bias voltage, VDC. Panels (iv)–(vi) of Fig. 6b show the self-oscillation amplitude for three reverse bias conditions VDC for a fixed illumination, Pin = 11.1 mW. The observed self-oscillation amplitude changes can be explained by the limit-cycle trajectory30 taken by the oscillations in I–V phase space. Near the peak and valley voltage (Vp = − 3.1 V and Vv = −3.6 V, respectively), the amplitude (0.11 V and 0.33 V, respectively) is lower because of the smaller limit-cycle trajectory. The much lower amplitude at the peak voltage (VDC = − 3.1 V) can be explained by the unstable periodic orbits of the limit cycle around the peak bias point30. When biased near the centre of the NDR region (VDC = − 3.4 V), a larger limit cycle trajectory is obtained resulting in larger amplitude oscillations (0.56 V). Using the numerical model described in the Supplementary (Section S3), we have reproduced the experimental results of Fig. 6. Figures S2 and S3 show the simulations obtained for the self-oscillations under different illumination and bias voltage conditions, respectively, showing a good agreement with the experimental results.

Pulse-modulated light-induced excitatory and inhibitory oscillation response

In this section, we investigate the photo-induced excitatory and inhibitory control of the oscillation response in the µRTD-PD sensory neuron using pulse-modulated light, Fig. 7a. The goal is to demonstrate photo-switching between oscillatory and silent phase emulating dynamic neuronal-like signalling, a behaviour replicating visual-neuronal systems in nature44. In the experiment, the laser source was modulated by driving the laser current with electrical pulses with varying amplitude, Vin, at a frequency, fin, ranging between 20 and 90 kHz. The light modulated frequencies are below the μRTD-PD circuitry self-oscillating frequency at around 350 kHz, as shown in Fig. 7b. The average intensity of the pulse-modulated light is defined as <Pin>.

Fig. 7
figure 7

Experimental oscillatory firing signalling activated by incoming pulse-modulated light. (a) Scanning electron microscope (SEM) picture of a single µRTD-PD oscillatory neuron device. For optical activation of the µRTD-PD neuron, pulse-modulated light at λ ~ 830 nm was used employing direct modulation of the laser with an electrical pulse signal with frequency, fin and amplitude, Vin. (b) Electrical output showing neuron-like oscillatory signalling upon pulse-modulated light at frequency (fin = 20 kHz) as a function of the amplitude (Vin): (i) self-sustained oscillations at Vin = 20 mV and <Pin>  ~ 4 mW; (ii) oscillatory periodic firing activation at Vin = 40 mV and <Pin>  ~ 4 mW. (c) Neuron-like oscillatory signalling responses using (ii) a positive pulse input [(i) Vin = 100 mV, fin = 20 kHz, <Pin>  ~ 6.8 mW], and (iv) a negative pulse input [(iii) Vin = -100 mV, fin = 50 kHz, <Pin>  ~ 2.5 mW]. (d) Burst firing activation using positive pulses (Vin = 40 mV and <Pin>  ~ 4 mW) with variable frequency: (i) fin = 20 kHz, (ii) fin = 40 kHz, and (iii) fin = 90 kHz.

Full size image

Figure 7b, panels (i) and (ii), shows the all-or-nothing photo-switchable oscillatory behaviour of the µRTD-PD sensory device when reverse biased (VDC = − 3.6 V, valley). In the case of illumination with a subthreshold average optical power (weak modulation, Vin = 20 mV, providing an average light intensity of <Pin>~3 mW) only self-sustained oscillations are observed, Fig. 7b(i), similarly as in the previous analysed case for CW illumination. However, for average pulse-modulation optical power above threshold (that is, with Vin = 40 mV providing <Pin> ~ 4 mW), the µRTD-PD sensory neuron exhibits alternate periods of bursts of oscillations and silent response (without bursts), Fig. 7b(ii). This forms an inhibitory and excitatory oscillatory response which can be controlled by the optical power modulation frequency (fin) and amplitude (Vin) of the incoming pulse-modulated light signals. This behaviour is analysed in more detail as follows.

First, we consider the photo-switchable response dependence on the amplitude using either positive or negative electrical pulses driving the laser source input. In a first case scenario, Fig. 7c(i), (ii), the µRTD-PD is biased near the valley region (VDC = − 3.6 V) and perturbing positive pulses (increasing light intensity), panel (i), are injected in the µRTD-PD activating on/off firing of burst of oscillations. When the incoming pulse is at its baseline value (0 mV), the system remains biased in the NDR region forming the burst firing response. However, when the modulating pulse is at the high value (100 mV), the µRTD-PD neuron enters the 2nd positive differential resistance (PDC) region of the strong illuminated I–V curve (outside the NDR region), forming the silent response (without oscillations). In a second scenario, Fig. 7c (iii), (iv), the µRTD-PD is biased near the valley region (VDC = − 3.6 V) and negative pulses (decreasing optical intensity), panel (iii), are injected in the µRTD-PD. In this case, when the pulse is in its baseline value (0 mV), the system remains biased in the NDR region forming the burst firing response. However, when the perturbing pulse is at its low value (-100 mV), the µRTD-PD neuron enters the PDC region of the dark I-V curve (with no NDR region), forming the silent response (without oscillations). We note the input pulse voltage in panels (i) and (iii) corresponds to the voltage RF input signal driving the laser current using direct modulation, which changes the average input optical power (experimental setup in Fig. 3). As a result, our artificial neuron is responding to increasing or decreasing of light intensity. As a result, the smaller amplitude of the burst firing responses in panel (iv) is attributed to lower average input optical power of <Pin> = 2.5 mW as compared to panel (ii), <Pin> = 6.8 mW. Therefore, by tuning the modulated light intensity of the incoming signals, excitatory and inhibitory oscillation responses can be controlled within the same µRTD-PD sensory neuron.

Lastly, Fig. 7d depicts the dynamics of the neuron-like burst oscillations as a function of the modulation frequency of the injected optical signal when the modulation amplitude is set constant (Vin = 40 mV). As observed in Fig. 7d(i)–(iii), the number of oscillation periods (“spikes”) within each cycle can be fine-tuned. In this case the number of “self-oscillation periods” (“spikes”) decreases as the frequency of modulation increases from (i) 20 kHz, and (ii) 40 kHz to (iii) 90 kHz. This enables an extra degree of freedom to encode incoming optical pulse information into oscillatory bursting signals.

Stability and spatiotemporal response of the burst oscillatory patterns

In this section, the stability and spatiotemporal response of the burst firing patterns are analysed. To this end, we use a space-time representation of the time-domain signals. In this representation, the position of a temporal pattern that repeats itself is clearly evidenced on a timescale much larger than the timescale of the oscillatory signals. This enables to analyse any drift (as for example induced by noise) of the experimental time traces. This space-time like diagram is commonly used for describing time-delayed systems45.

Following a standard procedure, the recorded data is analysed in a spatially extended representation. The evolution of the burst firing patterns was recorded for a long time-period of about 0.2 s. The device was illuminated with pulse-modulated light (negative pulses, Vin = − 100 mV, fin = 50 kHz) providing an average power output <Pin >~ 2.5 mW. In Fig. 8 is shown the space-time diagrams for various bias conditions across the NDR region of the µRTD-PD sensory neuron. The time within each cycle is plotted in the vertical axis and here we assume each cycle covers 60 μs (a total of 3333 cycles in the vertical axis). Here, the repetition rate of the incoming pulse-modulating signal is used as each round-trip (cycle). Near the peak, panel (a), a clear noise-induced drift motion of the oscillating patterns is observed. This indicates that noise in the system affects the stability of the burst oscillations. We attribute this to the higher current operation which contributes to electrical noise. The larger responsivity in the peak also contributes to shot noise. By selecting two representative cycles of the spatiotemporal map of Fig. 8a (panels (i) and (ii)), we can observe suppression of the burst oscillations (panel (ii)) within one cycle. This indicates that noise-induced drift motion contributes to supress burst oscillations. We note for the case of the patterns, panels (b) and (c), no suppression is observed for the time-period analysed here indicating stable patterns. The µRTD-PD oscillatory neuron spatiotemporal patterns are analogous to periodic breather-like oscillations observed in other neuron-like systems with slow-fast scales44. This behaviour has been reported in the context of neural activity in large neural networks, thus the µRTD-PD oscillatory neuron could be employed to study complex neuronal-like oscillatory spatiotemporal patterns in brain circuits.

Fig. 8
figure 8

Spatiotemporal plots showing the evolution of the burst firing oscillations over a 0.2 s time-period. In this co-moving reference frame, the horizontal axis is a space-like coordinate that allows to localize the position of the burst oscillations within a given cycle while the vertical axis corresponds to the slow temporal evolution of the system over various cycles. (a) Space–time plot of burst firing oscillations when the µRTD-PD neuron is biased near the peak voltage, VDC = − 3.1 V. The insets (i) and (ii) represent the time traces at cycle 500 and 2600, respectively, showing suppression of oscillations in panel (ii). Space–time plot of burst firing oscillations when the µRTD-PD neuron is biased in: (b) the middle of the NDR region, VDC = − 3.4 V, and (c) in the near-valley region, VDC = − 3.6 V. All the measurements use the same contour plot scale denoting the amplitude of the µRTD-PD output, depicted on the right side of panel (c).

Full size image

Finally, since endurance is one important characteristics of neuron devices for reliable neuromorphic applications, in Fig. 9 we evaluate the stability and reliability of generated spike trains in response to the input. This was done by evaluating the spike trains on-off peak amplitude in response to a fixed input for a long acquisition time (>103 cycles, only limited by the oscilloscope acquisition time) and changing the DC bias point within the NDR region (panels a and b). The results suggest that the most stable operation is when the device is DC biased in the middle of the NDR. In this case, a clear on-off of the oscillations with small fluctuation of the spike amplitude is observed, panel a), while near the valley the spike train shows fluctuations of its peak amplitude, panel b). As a conclusion, for robust operation of the neuron as a spike train emitter, operation in the middle of the NDR provides the most reliable performance.

Fig. 9
figure 9

Experimental results which evaluate the spike trains on–off peak amplitude in response to a fixed input for a long acquisition time (> 103 cycles) in (a) middle of the NDR region and (b) near the valley region.

Full size image

Discussion

We demonstrated a III–V neuromorphic photonic oscillator neuron consisting of a micropillar quantum resonant tunnelling diode (RTD) with GaAs photosensitive absorption layers, the µRTD-PD sensory oscillator neuron. The µRTD-PD sensory device responds to incoming analogue optical signals, enabling light-activation of negative differential resistance (NDR) of relevance to emulate FitzHugh–Nagumo-type of neuron models. We fabricated µRTD-PD devices of various diameters (d = 6 μm, 8 μm and 10 μm), and investigated the static I-V characteristics under dark and near-infrared light conditions. A noticeable increase of the photocurrent was observed in the so-called peak voltage in reverse bias conditions, resulting in a significant increase of the devices’ optical responsivity and subsequent light-induced NDR region.

In this work, we focused on the analysis of bursting oscillatory patterns, a behaviour highly relevant in visual sensory neuronal biological systems. Taking advantage of the NDR effect, we demonstrated µRTD-PD functioning as sensory oscillatory neurons where oscillations can be activated using low intensity optical power. Using pulse-modulated light, excitatory and inhibitory control of spatiotemporal neural-like oscillations was demonstrated, emulating oscillatory and silent phase dynamic neuronal signalling and simulating neural activity in networks in the form of breather-type oscillatory phenomena.

The light-induced NDR and oscillatory phenomena were observed at an effective power as low as Peff ~ 100 µW. Considering the estimated lensed fibre coupling efficiency (ηc ~ 8 × 10−4) in the experiments39, and assuming a pulse frequency of the incoming pulse of 350 kHz (as fast as the self-oscillations), we obtain an optical pulse activation energy of ~ 230 fJ per burst oscillations, where the energy is calculated as \(E={\eta }_{c}{P}_{eff}/{f}_{in}\). Improvements include optimization of the light absorption in the micropillars as for example using nanostructuring of the micropillar. Also, optimization of speed operation could further reduce the optical energy requirements to activate optically the neuronal oscillatory functions. We note µRTD-PD devices have higher-speed oscillatory potential, since quantum resonant tunnelling effect in RTD-based devices persist well beyond GHz. Higher frequency operation is achievable via optimisation of the circuit, namely through the reduction of parasitic inductance from the experimental setup.

In Table 1, we provide a benchmark of recent examples of optically activated oscillator and spike-based neurons in various material platforms. Our neuron emulates the physiology of neurons and their oscillatory dynamics which goes beyond the simple leak-integrated-and-fire functions implemented in photonics46 and complementary metal–oxide–semiconductor (CMOS)47 approaches. Compared with electronics and electronic-photonics24,48,49, our approach enables monolithic integration of the photodetector and neuron within a single device, without additional circuitry which increases the power consumption and footprint of the neuron hardware. As a result, this provides potential for scalability since miniaturization from micropillar to nanopillar38 neurons could enable larger number of neurons in a single chip. Importantly, our neuron uses identical III–V materials that have matured to meet the demanding industrial requirements of 3D sensing and LiDAR systems. Lastly, regarding integrability, we have shown recently that a similar semiconductor architecture exhibits negative resistance and electroluminescence39, providing the possibility for artificial neuron emitters. This combined with interconnected links50, could enable emitter-receiver neurons for in-sensor neuromorphic edge computing applications.

Table 1 Optically activated oscillator and spike-based neurons in various material platforms including photonics, electronics and hybrid electronic-photonic.
Full size table

Methods

Epilayer design

The epilayer stack of the n–i–n type (Si-doped) GaAs μRTD-PD oscillatory neuron was grown by metal organic chemical vapour deposition (MOCVD) on a semi-insulating (SI) GaAs substrate. In Fig. 1b is depicted the schematic of the μRTD-PD semiconductor epilayer stack, including the double barrier quantum well (DBQW) intrinsic region. The DBQW is formed by a stack of AlAs/GaAs/AlAs (1.4 nm/7 nm/1.4 nm) layers, surrounded by a GaAs layer of 5 nm forming the emitter and collector spacer layers, all of which are unintentional doped. Surrounding the intrinsic region with DBQW we have lightly doped n-GaAs (Nd = 2 × 1016 cm−3) photoabsorption layers with 300 nm and 500 nm thicknesses on the emitter and collector sides respectively (in the scenario of forward bias operation). The emitter cladding layer (which is also the bottom contact layer) consist of a 300 nm thick graded AlxGa1−xAs layer (where x varies from 0 to > 0.3) moderately doped with Nd = 2 × 1018 cm−3. The collector cladding consists of a 300 nm thick Al0.3Ga0.7As layer with a moderate n-type doping concentration (Nd = 2 × 1018 cm−3). The collector top contact is formed by a highly doped 150 nm thick n-GaAs (Nd = 5 × 1018 cm−3).

Fabrication

The MOCVD grown 6-inch wafer is first diced into 1 inch square samples. The μRTD-PD devices were fabricated by defining micropillars by optical lithography by using direct laser writing, followed by etching using induced couple plasma reactive ion etching (ICP-RIE) until reaching the bottom n-contact AlxGa1−xAs layer (etch depth of ~ 1.3 μm). The micropillars were then passivated with a thick layer (~ 1 μm) of SiO2, and via openings for the bottom and top contacts were defined by optical lithography, followed by dry etching of the SiO2. Both the top and bottom metal contacts (using a coplanar waveguide transmission line design) were formed using Ge/Ni/Au (20 nm/10 nm/150 nm) metal alloys. Finally, samples of fabricated devices were subjected to rapid thermal annealing at 400 °C for 30 s.

Optoelectronic characterization

The μRTD-PDs were electrically connected using a ground-signal-ground (G-S-G) coplanar waveguide electrical probe connected to a bias-T (bandwidth of 0.1–4200 MHz). The photodetection setup consists of lensed fibre coupled to a pigtail continuous wave (CW) laser diode source in a single-mode fibre, λ = 830 ± 10 nm (Roithner Lasertechnik, model SPL830-5-4-PD), where the respective photocurrent was monitored via the DC port of the bias-T by the change in the I–V characteristics obtained from Keithley 2280S DC source meter upon optical injection of the μRTD-PD. The intensity from the laser diode was controlled by a current driver and temperature controller (Thorlabs, CLD1010LD) and the modulation was provided by an electrical signal generator (RIGOL, model DG1000Z). The input optical powers were calibrated by a commercial power meter (Newport, model 1918-R). The laser diode was electrically pumped to obtain an average optical power <Pin>. The pulse-modulated laser intensity was fed into the μRTD-PD top pillar and the electrical RF output was collected using an oscilloscope (R&S, MXO oscilloscope) which was connected to the AC port of the bias-T.