The human brain has a complex network, which contains ~10¹¹ neurons interconnected by ~10¹⁵ synapses^[1]. In the neurons network, a neuron receives action potentials from pre-neurons through the axons and pre-synapses, which integrated these signals and transferred to post-neuronal synapses^{[2, 3]}. The synaptic weight and synaptic plasticity can determine the working mechanism of neuronal systems for learning and memory in the human brain. In the past half century, traditional von Neumann computing machines have experienced rapid development in various areas of human work and life. However, inspired by the neuron network of human brain, next generation artificial intelligence based on neuromorphic computing has been proposed by researchers to meet faster data processing and lower power consumption^{[4, 5]}.

The performance of an artificial neuron network is usually both dependent on the semiconductor materials, the structure, and the working mechanism of synaptic device. Numerous types of neuromorphic devices applied in synaptic simulation have been reported, such as memristor^[6−8], ferroelectric memory^{[9, 10]}, optoelectronic^[4], and so on. Various semiconductor materials, including molybdenum disulfide (MoS₂)^[11−13], indium−gallium−zinc oxide (IGZO)^{[14, 15]}, perovskite^[16], gallium nitride (GaN)^[17−19], indium oxide (In₂O₃)^[20], and carbon nanotube^[21], have been integrated in fabricating the neuromorphic device for mimicking the function of artificial neuron network. As the fundamental characteristics, power consumption and the method for modulating synaptic weight have been the most discussed in previous publication. Furthermore, the research on main biological functions of neuromorphic device, including the excitatory post-synaptic currents (EPSC), the paired-pulse facilitation (PPF), the transition from short-term plasticity (STP) to long-term plasticity (LTP), and the simulation of the classical Pavlov’s dog experiment, is of greater significance.

In this research, a nanowatt level optoelectronic artificial synaptic device based on TiO₂/AlGaN/GaN heterostructure is demonstrated. The mechanism and functions of this artificial synapse are illustrated and discussed in the results and discussion section. The TiO₂ nanolayer gate and two-dimensional electron gas (2DEG) channel are used to mimic the synaptic weight and the synaptic cleft, respectively. A EPSC of 1 to 200 nA can be triggered by ultraviolet pulsed light. And under ultraviolet light spikes, the transformation from STP to LTP was realized in this synaptic device as the number of optical spikes increased. Meanwhile, a classical conditional experiment of Pavlov’s dog has been analyzed, the device can achieve "retraining" process to extend memory time through enhancing the intensity of synaptic weight.

The AlGaN/GaN epitaxial heterostructure was grown on a 1 mm thick〈111〉silicon substrate wafer using the metal−organic chemical vapor deposition technique. An undoped GaN buffer layer (4.7 μm), followed by a GaN channel layer (300 nm), an AlN interlayer (1 nm), an undoped Al_0.22Ga_0.78N barrier layer (21 nm), and a GaN cap layer (2 nm) was deposited. The electron mobility of the 2DEG was ~1500 cm²∙V⁻¹∙s⁻¹, with a sheet electron density of ~1 × 10¹³ cm⁻³. The chip fabrication started with a mesa etching to define the active area. Then, a Ti/Al/Ti/Au (20/110/40/50 nm) ohmic metallization multilayer film was deposited via electron beam evaporation and patterned by lift-off (LLO) process. After patterning, the ohmic contact multilayer was annealed at 870 °C for 45 s under ambient N₂. Next, an evaporated Ti/Pt (30/200 nm) was patterned by LLO process to form the micro-heater, followed by a 200 nm SiO₂ layer for isolation from the interconnect layer via plasma-enhanced chemical vapor deposition (PECVD). The evaporated Ti/Au (20/300 nm) layer stack was used to form a Schottky contact on AlGaN/GaN epitaxy layer. The topside of the wafer was passivated with a 200-nm PECVD SiO₂ layer and etched in buffered oxide etch solution to open the contact pads and gate windows. A 10-nm TiO₂ nanolayer was deposited on the AlGaN/GaN heterostructure surface to form the ultraviolet photosensitive floating gate via magnetron sputtering. The silicon substrate was polished down to 400 µm from the backside and etched by deep reactive ion etching to form a circular membrane with a diameter of 600 µm in the final step. After dicing, the 1 × 1 mm single chip was completely fabricated.

The spectral response of the TiO₂ floating-gate/AlGaN/GaN synaptic device was measured in a spectrometer testing DSR200 system (Zolix Instrument Co., Ltd, China). The UV (Ultraviolet) LED was driven by Keithley 2400 and Keysight E3631A. The I−V and I−t curves of drain to source channel in the synaptic device were measured by Keithley 2420 source measure unit. The pulse signal was generated by waveform generator (Rigol, DG1022U, China). An ultraviolet−visible spectrophotometer (Hitachi, UH4150, Japan) was used to obtain the absorption spectrum of the TiO₂ nanolayer. The structures of the devices were visualized using SEM (Hitachi, S-4800, Japan). The light intensity of the UVC LED was measured using a spectroradiometer system (JETI Technische Instrumente GmbH, specbos 1211UV, Germany).

Neurons are the fundamental units in the human brain’s complex neural network, which are connected by synapses, as shown in Figs. 1(a) and 1(b), respectively. Fig. 1(b) shows the structure of a synapse, including pre-synapse, post-synapse, synaptic cleft, synaptic vesicle, neurotransmitter receptor, and so on. In this work, the strength of EPSC is determined by the synaptic weight, which can be modulated by the external light stimuli spikes and pre-synapse voltage stimuli spikes. In Fig. 1(c), the cross-section schematic diagram of TiO₂ gate/AlGaN/GaN heterostructure synaptic neuromorphic device is illustrated. In this three-terminal synaptic phototransistor architecture, a Ti/Au schottky electrode without annealing and a Ti/Al/Ti/Au ohmic electrode with RTA annealing process are used as the presynaptic and postsynaptic terminals, respectively. The concentration of the AlGaN/GaN 2DEG channel is regarded as the synaptic weight, which is regulated by the potential of TiO₂ gate nanolayer under the illumination of ultraviolet light. Fig. 1(f) presents the I−V characteristics of the TiO₂ gate/AlGaN/GaN heterostructure synaptic neuromorphic device in the dark state. The scanning bias voltage ranged from −4 to 4 V and there was a good performance when the detector was operated under reverse bias voltage. For example in the curve, if the device was driven at a reverse bias not exceeding −2.5 V, the maximum drain to source current was 10 nA. And under −4 V bias, the dark current I_ds was 137 nA. The nA level working current can ensure that this neuromorphic device operates at extremely low power consumption similar to that of neurons. The inset shows the linear scale of current−voltage and semilogarithmic scale of resting power comsuption−voltage. In this synaptic device, the static power consumption maintains an extremely low level in resting mode without the stimuli of external UV light spike, as shown in the inset image of Fig. 1(f). According to the working mechanism of this synaptic device, the operation voltage will not exceed 3 V and the corresponding power consumption will not exceed 100 nW, respectively. Fig. 1(d) shows the optical image of the complete synaptic device with a chip size of 1 × 1 mm² and an active area of 340 × 340 μm². The device actually has ten pairs of 2DEG channel that mimic synaptic cleft, which are connected in parallel together using the same drain and source electrode. Fig. 1(e) illustrates the SEM image of active area and the structure distribution. The Schottky contact of drain terminal and the ohmic contact of source terminal are used to mimic the pre-synapse and post-synapse, respectively. And the characteristics of AlGaN/GaN 2DEG channel is similar to the synaptic cleft. As shown in Fig. 1(c), the TiO₂ gate AlGaN/GaN synaptic device integrated a micro-heater unit, wihch is suspended from the silicon substrate for thermal isolation. The temperature of the membrane can be modulated by the micro-heater based on Joule heating. And the transient optical characteristics can be enhanced by different temperatures. Meanwhile, the structure of the membrane is sensitive to the pressure from external stimuli, including air and mechanical pressure. Similar to the photoresponsive function of TiO₂-gated structure, as new methods for modulating synaptic weight, suspended epitaxy membrane and micro-heater will be studied and discussed in our future publication.

Fig. 2(a) shows the spectral response of the synaptic device at 3 V forward bias indicating high responsivity in the ultraviolet region (wavelength of 200−365 nm). The device had a peak responsivity of 784.2 A∙W⁻¹ at 3 V bias voltage under 240 nm UV illumination. A sharp cutoff wavelength of the detector, defined as the ratio between the maximum responsivity and the natural constant (e ≈ 2.718), is located at 365 nm. Fig. 2(b) shows the absorption spectrum of the 10 nm TiO₂ floating gate nanolayer, which has a cutoff wavelength edge of 365 nm, which covers the range of UVB and UVC light. In the inset of Fig. 2(b), we can clearly see the drain electrode, source electrode and TiO₂ nanolayer in the 2DEG channel area. And a 10 μm of artificial synaptic cleft can be observed.

As shown in Fig. 3(a), the voltage V_ds is fixed at −2 V and the basic dark current of the device is 6.7 nA. A typical EPSC can be obviously observed when triggerd by a UV light spike emitted from a 365 nm UVA LED with a 1 s duration at light intensity of 1.35 μW∙cm⁻². The peak EPSC can reach 200 nA, which is attributed to the TiO₂ floating gate, the AlGaN/GaN 2DEG channel and the Schottky contact. Fig. 2 and Fig. 3 demonstrate that our synaptic devices show strong prospect application for neuromorphic visual and computing in ultraviolet light range. Upon the removal of the UV light pulse, the EPSC ΔI_ds didn’t disappear immediately with a decay time of several minutes or even hours. TiO₂ is an n-type and ultraviolet photosensitive semiconductor material. The adsorption of oxygen molecules from the air and ultraviolet absorption from light illumination can alter the surface charges and states of the TiO₂ floating-gate nanolayer, thereby disrupting the gate potential and AlGaN/GaN 2DEG channel concentration. The photogenerated holes in the AlGaN and GaN layer can drift toward the Ti/AlGaN Schottky junction surface when the V_ds was fixed at reverse bias voltage (built-in electric field) and get captured at the interface centers or the deep energy centers. Fig. 3(b) inllustrates the schematic of the energy band diagrams of the synaptic device under UV light spikes. The holes trapping in the Schottky surface can induce height reduction of the Schottky barrier and width narrowing in the depletion region, which will enhance the reverse leakage current. After removing the UV illumination, a long recovery time was observed in the TiO₂-floating gate/AlGaN/GaN Schottky based synaptic device, as shown in Fig. 3(a). The phenomenon of long recovery time is mainly caused by the defects in GaN epitaxy materials, the photogenerated electrons in the 2DEG channel, and the trapping holes in the TiAu/AlGaN Schottky junction surface. Similar working principle also has been discussed in our previous publication^[22−26].

In nervous system, paired-pulse facilitation (PPF) is another essential behavior of synaptic plasticity in neuron network, which describes a fundamental feature in that the EPSC triggered by the second stimulating spike would be larger than that triggered by the first spike. The PPF of the synaptic device can be defined by the equation:

1PPF=100%×(A2/A1).

Here the A₁ and A₂ represent the EPSC value triggered by the first and second light spike, respectively. Fig. 4(a) shows a typical PPF behavior induced by two consecutive stimulating spike. The wavelength and intensity of ultraviolet light is 365 nm and 1.35 μW∙cm⁻², respectively. The pulse duration and interval are 100 and 300 ms, respectively. The value of A₂/A₁ is about 205% and the V_ds is fixed at −3 V. Fig. 4(b) displays the relationship between PPF index and light spike interval (Δt). It is can be seen that the PPF index decreases with the light spike interval increased. As shown in Fig. 4(b), PPF index results can be fitted by the dual exponential model:

2y=y0+y1e−Δt/τ1+y2e−Δt/τ2.

Here y₁ and y₂ are the initial facilitation magnitude, and τ₁ and τ₂ are the characteristic relaxation times of the respective phases. The PPF index decreases exponentially as Δt increases, which is similar to the behavior of biological synapse in human brain^{[27, 28]}.

Memorizing and forgetting are the fundamental neural activities in human brain. Fig. 5(a) illustrates the memory formation process from sensor memory (SM), STM and finally to LTM. The SM is obtained by external stimuli. The SM will be transferred to STM when more brain attention is involved. However, as time passes, the STM may be forgotten or lost. Before the STM complete losses, repeated stimulating can induce STM for a longer memory time, which is similar to the training or learning in the neuron network of brain^[29]. As shown in Figs. 5(b) and 5(c), the EPSC curves triggered by ultraviolet light with different numbers of pulses ranging from 1 to 20 are illustrated. The bias voltage was fixed at −2.5 V and the wavelength of the light pulse was 365 nm with a light intensity of 1.35 μW∙cm⁻². In Fig. 5(b), the pulse duration and period are 50 ms and 1 s, respectively. A obvious smaller EPSC was triggered by one light spike and decayed to a lower current after the removal of light spike, aligning with short-term plasticity (STP) behavior. As the number of optical spikes increased, a higher EPSC and retention time can be observed. While twenty light spikes are illuminated on the synaptic device, highest peak and memory currents were measured, which fits the typical psychological memory model. Hence, the transformation from STP to LTP was realized in the synaptic device. Ten ultraviolet light spikes at a frequency of 1 Hz is shown in Fig. 5(b). If the EPSC gain is defined as A₁₀/A₁ to characterize the amplitude enhancement induced by ten light spikes, where A₁₀ and A₁ can represent the EPSC value triggered by the tenth and first light spike. Here, the EPSC gain is 11.24 at 1 Hz.

In previous reports, in order to better understand the neuromorphic behavior of synaptic device, a classical conditioning experiment is usually presented, which is an associative learning similar to the interactive learning process in the human brain^[30−32]. In the typical Pavlov’s dog experiment, the key elements include neutral stimuli (NS), unconditional stimuli (US), conditional stimuli (CS), unconditional response (UR), conditional response (CR). In this research, the bell ringing and feeding food are defined as NS and US, respectively, as shown in Fig. 6(a). At the beginning, the bell ringing cannot induce any salivation (no response). However, salivation (UR) is easily achieved while the food (US) is provided on Pavlov’s dog. Then, if we want the bell ringing (NS) to be able to induce salivation, an association between NS and US should be established in the dog’s brain through training, which requires both feeding food and ringing bell at the same time. After several times, Pavlov’s dog begin to salivation while the bell is rung. Therefore, the bell ringing changes from NS to CS, and correspondingly, the salivation it triggered becomes CR. After training, the dog has learned to associate the bell ringing with the food feeding and form a memory, which is called acquisition.

As shown in Fig. 6, V_ds = −2.5 V represents the basic state of the synaptic device and a higher voltage V_ds = −2.8 V (NS) represents a neutral stimuli signal from preneuron. Synaptic weight of 30 nA is defined as the threshold current for dog’s salivation response. Before training, five bell stimuli (NS, −2.8 V voltage, 10 s width) were applied on the drain to source channel, which did not cause the current to exceed the salivation response threshold (red dash dot line, 30 nA) and receive no response from dog, as shown in Fig. 6(b). In Fig. 6(c), ten ultraviolet light stimuli (US, 365 nm, 0.5 Hz frequency, 1% duty cycle, 1.35 μW∙cm⁻² intensity) were applied on the gate area, which induced effective EPSC exceeding the threshold of 30 nA and triggered salivation of dog (UR). As shown in Fig. 6(d), during the training process, ten ultraviolet spike (US) and −2.8 V (NS) bias voltage were simultaneously applied on the synaptic device, which resulted in a higher EPSC current exceeding 30 nA and salivation of dog. Hence, an association between US and NS has been established. Meanwhile, a memory has been formed after the removal of light spike and V_ds voltage. As shown in Fig. 6(e), as time passes, once the current fell below the threshold current of 30 nA, dog began to forget this matter and stopped salivation. Then, PSC can also achieved the threshold current through applying −2.8 V bias voltage on drain to source channel. Here, the bell ringing changes from NS to CS through associating learning process, resulting the salivation of dog (CR) can be triggered by bell (CS). In Fig. 6(e), ten −2.8 V voltage pulses (CS) were applied, the acquisition can maintain a considerable period of time. Also, according to Fig. 5 and Fig. 6, we can achieve "retraining" process to extend memory time through enhancing the intensity of synaptic weight, which is similar to the working mechanism of human brain.

In summary, a TiO₂/AlGaN/GaN heterostructure-based neurmorphic synapse was successfully fabricated with the structure of phototransistor and Schottky diode. The Schottky contact of drain terminal and the ohmic contact of source terminal are used to mimic the pre-synapse and post-synapse, respectively. The AlGaN/GaN 2DEG is used to imitate the synaptic cleft due to its special characteristics. The PPF index decreases exponentially as Δt increases, which is similar to the behavior of biological synapse in human brain. A higher EPSC and retention time can be observed. The transformation from STP to LTP was realized in the synaptic device as the number of optical spikes increased. In classical Pavlov’s dog experiment, the device can achieve "retraining" process to extend memory time through enhancing the intensity of synaptic weight, which is similar to the working mechanism of human brain. Systematically analyzing and discussing the functions of synaptic plasticity can provide a significantly reference for future application, such as associative learning and neuromorphic computing.