With the vigorous development of big data and artificial intelligence, the demand for computing resources continues to grow, especially in the processing of massive data, which poses higher requirements for computing performance^[1−3]. It has led to the urgent task of implementing neuron computation in hardware^{[4, 5]}. One of the biggest challenges in this process is the development of neuromorphic devices and networks. In the recent years, many research works mainly focus on the simulating neural morphology at the software level^[6]. However, the required resource and energy consumption for real-time biologically inspired simulations through software are enormous^[7]. In contrast, the energy expenditure of a single synaptic event is extremely low in the human brain^[8], which is significantly lower than the energy levels of current artificial synapses. The human brain stores and processes signals in a parallel manner through neural networks, where biological synapses play a crucial role in simulating logic, memory and learning processes^[9−12]. Therefore, developing low-power and small-sized neuromorphic synapse devices from semiconductor materials is of significant importance in constructing intelligent neural systems^[13].

On the other hand, the wavelength range of deep-ultraviolet (UV) spectrum is from 100 to 280 nm^[14]. Due to the ultra-wide 4.9 eV bandgap, gallium oxide (Ga₂O₃) is a promising material for deep-UV photodetectors and artificial synapses^{[15, 16]}. It has a wide range of applications, including secure communication, radiation monitoring, flame sensing and environmental monitoring, etc.^[17−19]. Furthermore, nanowires (NWs) exhibit excellent light capture and absorption capabilities, outstanding carrier generation rates and photosensitivity due to their large surface area-to-volume ratio and abundant bio-capacity^[20−22]. Under optical stimulation conditions, synaptic nano-devices require characteristics such as high bandwidth, fast signal transmission and stable connections^[23−25]. In the hardware implementation of neuromorphic computing, the energy consumption of synaptic nano-devices is also a key area of focus^[26−30]. Hence, the fabrication of low-energy-consumption synaptic devices based on Ga₂O₃ nanowires is crucial for realizing efficient and secure neuromorphic computing systems^[31]. However, the high-temperature environment during the preparation process makes Ga₂O₃ extremely prone to decomposition and volatilization, which can easily generate a large number of defects such as embedded structures and spiral dislocations^[32−34]. Therefore, developing synaptic nano-devices based on high-quality Ga₂O₃ materials is not only necessary but also presents numerous challenges.

In our previous research, we have successfully fabricated GaN nanowires and the photodetectors^{[35, 36]}. However, up to now, very few works have been reported about the synaptic device based on Ga₂O₃ nanowires. In this work, we present a synaptic nano-device based on Ga₂O₃ nanowires, which can achieve the non-linear transmission characteristics and learning-memory behavior. We have also studied its features mimicking biological synapses, including short-term plasticity (STP), long-term potentiation (LTP) and spike-timing-dependent plasticity (STDP). Moreover, the stability of synaptic device can be achieved during the prolonged stimulation and storage, along with low energy consumption. Thanks to the device's learning capability, we can obtain a high accuracy by simulating a multi-layer perceptron for handwritten digit recognition.

Firstly, GaN vertical nanowire arrays (Fig. 1(a)) were fabricated on n-type Si(111) substrates using molecular beam epitaxy (MBE, Vecco G20) technology. In this process, nitrogen (N) atoms were supplied by the N plasma cell, while Ga atoms were produced by the Ga effusion cell. To remove the natural oxides within the growth chamber, the silicon substrate was preheated to 900 °C for 15 min. Subsequently, a layer of approximately 3-nm-thick AlN sacrificial layer was deposited on the Si surface. The growth of GaN nanowires continued for 350 min under a nominal Ga flux of 3.0 × 10^‒8 Torr, with the nitrogen flow rate and plasma power maintained at 4.8 sccm and 450 W, respectively.

After that, the GaN vertical nanowire arrays were placed in a quartz tube, exposed to oxygen gas at 800 °C for 10 min. Finally, the vertical nanowire arrays were cooled to room temperature, completing the preparation of Ga₂O₃ vertical nanowires (Fig. 1(b)). After fabricating the Ga₂O₃ nanowires, they were manually detached from the SiO₂ substrate. The lift-off Ga₂O₃ nanowires were then immersed in a mixed solution of isopropyl alcohol and water with a volume ratio of 3 : 2 (Fig. 1(c)). By using a pipette, this solution was deposited on the Ti/Au (20/60 nm) metal electrodes (Fig. 1(d)). Finally, the sample was annealed for 5 min in a pure N₂ atmosphere at 300 °C to improve the contact property.

The morphology and positioning of Ga₂O₃ nanowires were examined using a scanning electron microscope (SEM). Subsequently, samples for aberration-corrected transmission electron microscopy (AC-STEM) observation were prepared using a focused ion beam (FIB, Scios, FEI), enabling structural and elemental analysis of the Ga₂O₃ nanowires. The synaptic functionality of the device was evaluated using an Agilent B1500A semiconductor parameter analyzer. The response experiments utilized a 255 nm light-emitting diode (LED) as the lighting source.

We have developed a multilayer perceptron (MLP) neural network model to study the peak-time plasticity and spike-timing-dependent plasticity (STDP) of the synaptic nano-device. We utilized offline learning for end-to-end system training and testing. A complete hardware system with a three-layer array was constructed, with input layer dimensions of 784 × 1000, hidden layer dimensions of 1000 × 1000 and output layer dimensions of 1000 × 10. In the neural network, each synaptic nano-device is analogous to a neuron, with 28 × 28 image pixel information corresponding to 784 input neurons. We used a modified version of the Modified National Institute of Standards and Technology (MNIST) dataset as input, converting the ten-digit output (0‒9) into a vector for recognition. For images of different digits, the weight values at various image pixel positions vary for each type of digit. We continuously adjusted the overall weight values using the backpropagation algorithm, ultimately achieving accurate digit recognition.

From Figs. 1(e) and 1(f), the nanowires have a ~2 μm length with good verticality and uniformity. The Ga₂O₃ nanowires are connected well with the two metal electrodes. Fig. 1(g) shows the distribution of Ga and O elements within the Ga₂O₃ nanowire, demonstrating the effective oxidation of GaN nanowires. As illustrated in Figs. 1(h) and 1(i), biological synapses play a crucial role in the biological nervous system as key structures for signal transmission between two adjacent neurons. In the nanoscale synaptic device, the two electrodes correspond to the presynaptic and postsynaptic membranes in biological synapses. The Ga₂O₃ nanowires are proposed to play the role of synaptic cleft, while the internal photo-generated carriers act as the neurotransmitters in the biological synapses^[37].

A fundamental characteristic of biological synapses is short-term plasticity (STP), which plays a crucial role in cognition, short-term learning and memory. When using pulsed light to stimulate the Ga₂O₃ NWs synaptic device, the obtained results can undergo typical neural morphological processing. Upon applying a presynaptic light pulse, excitatory postsynaptic currents (EPSCs) can be triggered (Fig. 2(a)). When the light pulse disappears, the current gradually returns to its initial level. Paired-pulse facilitation (PPF) is an important characteristic in neural network learning, which can assess the impact of consecutive pulses^[38]. In the Ga₂O₃ synaptic device, following stimulation by two consecutive light pulses (0.16 mW·cm^‒2), the second EPSC is notably higher than the first one (Fig. 2(b)). With an increase in the number of light pulse stimulations, the EPSC value of the synaptic device continues to rise, corresponding to the continuous influx of Ca²⁺ and an increase of conductivity within a biological synapse. The facilitation effect is typically represented by the PPF index (η), which is determined by the time interval between the two pulse stimulations and is typically studied using an exponential decay model^[39−41].

1I=Im+C⋅exp(−t/τ).

2η=A2/A1.

I represents the current at time t. I_m represents the saturation current and C is the pre-factor. A₁ and A₂ correspond to the peak values of EPSC produced by the first and second light pulse stimulations, respectively. τ represents the decay time constant. The larger the decay time t, the longer it takes for the EPSC to recover to its original level, indicating the stronger memory ability. As shown in Fig. 2(c), t increases with the number of light pulses, indicating a transition from short-term memory (STM, ~2.5 s) to long-term memory (LTM, ~50 s). Therefore, the synaptic device based on Ga₂O₃ nanowires can possess the functionality to mimic biological synapses.

In Fig. 2(d), after 5 consecutive light pulse stimulations, the EPSC peak current of the synaptic device continues to increase, representing the "learning" process. The cessation of light pulse stimulation in a dark environment for 5 s leads to a gradual decrease in EPSC current, resembling the process of "forgetting"^[42]. Subsequently, another round of 5 consecutive light pulse stimulations occurs, representing the "relearning" process. In each "relearning" process, the number of pulses required to reach the preset current threshold gradually decreases. Within the same "forgetting" time frame, the decay amplitude of EPSC during each "forgetting" process is significantly less than in the previous one. Hence, the artificial synaptic device based on Ga₂O₃ nanowires can achieve a "learning−forgetting−relearning" functionality similar to the human brain. In general, the EPSC level can only be repeated after the EPSC returning to the initial level. Interestingly, the same sequence of light stimuli used in learning stages can not only recall the memory level, but also achieve the higher level (Fig. 2(d)). That means EPSC can be higher under the same stimulation in the relearning stage, leading to the excellent "learning−forgetting−relearning" functionality. Thanks to the transition from STM to LTM, the rapid and enhanced recovery is attributed to the retention of the memory level after the stepwise learning.

As shown in Fig. 3(a), when the synaptic device is stimulated by high-frequency pulses (5 Hz), the peak EPSC current is small, and it decays rapidly. Conversely, when stimulated by low-frequency pulses (0.1 Hz), the peak EPSC current is larger. It takes longer time to decay back to its initial level, indicating a transition from short-term memory (STM) to long-term memory (LTM) in the synaptic device under stimulation conditions. At different light power densities (Fig. 3(b)) under the same frequency, light pulses with the higher light power typically trigger larger EPSCs and require more time to return to the pre-stimulus level. Similarly, in Fig. 3(c), for the same light illumination frequency and power, more pulses of light stimulation generally result in higher EPSCs before reaching saturation current, while fewer pulses lead to faster decay of EPSCs. Overall, the ability of this synaptic nano-device to transition from STM to LTM typically depends on the frequency, number and light power of the stimulated light pulses, reflecting the synaptic plasticity of the device.

The energy consumption of a synaptic device is a key performance metric, which can be calculated as follows^[43]:

3E=Ipeak×V×td=∫t0t1V⋅I(t)dt.

I_peak represents the maximum response current in a single stimulation. t₀ and t₁ denote the start and end times of the optical pulse stimulation, while t_d is the duration of the optical pulse stimulation. V and I represent the bias voltage (8 V) and response current of the device generated by the light-pulse stimulation. In Fig. 3(d), when the pulse stimulation frequencies are 5 and 2 Hz, the calculated results of energy consumption are 2.39 × 10^‒11 and 8.01 × 10^‒11 J, respectively. It's worth noting that we could further reduce the energy consumption of the synaptic nano-device by adjusting the bias voltage and optical pulse stimulation time, thus maintaining it at a lower level.

The learning process of the human brain is exceptionally complex and diverse. When we learn new information or review old knowledge, the electrical signals and chemical substances within the brain undergo changes^[44]. The connection points between biological neurons, known as synapses (Figs. 1(h) and 1(i)), adjust their strength based on various external stimuli, a phenomenon referred as synaptic weight. When the device is stimulated by the first light pulses, synaptic weight increases, indicating the first learning process. Subsequent encounters with the same stimulus often result in even higher synaptic weights, reflecting memory reinforcement. Conversely, a decrease in synaptic weight indicates the process of forgetting. Therefore, synaptic weight can be regarded as a merit of our ability to memorize information. To study how artificial synaptic devices mimic the learning behavior of the human brain, researchers typically use the change in synaptic weight (Δw) to evaluate the performance level, which is defined as follows^[45]:

4Δw=In−I0I0.

I₀ is the initial current value and I_n is the current generated after the stimulation of the light pulse. n represents the number of stimulated light pulses. As shown in Figs. 4(a) and 4(b), after receiving the two cycles of consecutive 5-light pulses with 5 s interval, the synaptic weight of the synaptic nano-device increases from 18% to 60%. The synaptic weight increases rapidly before reaching saturation current, demonstrating the exceptional performance of the device. This process resembles the "learning" process of synaptic memory deepening in the biological synapses of the human brain. After cessation of stimulation, the EPSC current gradually decreases with the synaptic weight decreasing, corresponding to the process of "forgetting". Upon receiving the same continuous stimulation of 5 optical pulses again, only 2 optical pulses are needed in the second round to reach the maximum EPSC level achieved during the first time, indicating the "relearning" effect after the initial stimulation. With a decrease in frequency and an increase in optical power (Figs. 4(c) and 4(d)), the maximum value of synaptic weight continuously increases, deepening the "memory". The maximum synaptic weight can reach about 32 after stimulation (Fig. 4(d)). Hence, the frequency and power of optical pulse stimulation can also affect the maximum synaptic weight of device. In other words, during the relearning process, the previous cognitive level can be reached with a shorter duration of optical pulse stimulation and fewer optical pulses.

In order to elucidate the operational mechanism of the synaptic nano-device, Fig. 5 presents the schematic equivalent circuit model and band diagram of the metal‒semiconductor‒metal (M₁‒S‒M₂) configuration. Within the synaptic device, light absorption and carrier generation are primarily facilitated by the Ga₂O₃ nanowires. Consequently, the predominant source of the resulting photocurrent is the nanowires interfacing with the designated electrodes. During operation, M₁ and M₂ are subjected to negative and positive biases, respectively (Fig. 5(a)). The device's functionality, characterized by a Schottky contact in Fig. 5(b), is predominantly influenced by the reverse Schottky barrier (M₁‒S, Φ_d1) with specific emphasis on the M₁‒S interface. Furthermore, the observed decrease in resistance pre- and post-UV light exposure, along with the reduction of Φ⁰_d1 and R_NW, are identified as critical factors contributing to the augmented photogenerated current.

A Schottky electroshock (Φ⁰_d1) is established at the M₁‒S junction, with the presence of defects within the nanowire illustrated in Fig. 5(c). Upon stimulation by a light pulse, the synaptic nano-device undergoes carrier separation, with the reduction of Φ⁰_d1 promoting carrier transport (Φ¹_d1 in Fig. 5(d)). Notably, some photogenerated holes are captured by oxygen vacancies (V_o) to form a relatively stable complex (V₀²⁺), as represented by the subsequent equation:

5VO+2h+=VO2+.

This mechanism hinders the recombination of electron-hole pairs, causing certain photogenerated electrons to remain in the conduction band^[46]. Consequently, the photocurrent persists for a duration following the cessation of light pulse excitation (Fig. 5(e)). Upon the initiation of a light pulse in this scenario, a greater number of photogenerated electrons and holes are generated (Fig. 5(f)), resulting in an augmented photocurrent relative to the previous state, thereby improving synaptic characteristics. The order of the Φ_d1 values mentioned above is Φ⁰_d1 > Φ²_d1 > Φ¹_d1 > Φ³_d1.

To demonstrate its application potential, we utilize the synaptic Ga₂O₃ nano-device to emulate neural networks for training and testing the MNIST dataset. The MNIST dataset comprises hand-written digit images of 28 × 28 pixels. We utilize supervised learning techniques, employing backpropagation and stochastic gradient descent algorithms to train the neural network model. Each MNIST image corresponds to a 28 × 28 pixel input. The entire neural network system consists of three components: the input layer, the hidden layer and the output layer. Specifically, the input layer consists of 784 input neurons, the hidden layer consists of 100 neurons. The output layer consists of 10 neurons, corresponding to digits 0‒9 (Fig. 6(a)). Throughout the training process (Fig. 6(b)), synaptic weights are preserved, which are extracted from experiments involving different enhancement and inhibition conductance. The formulas are as the follows^[47]:

6GP=B(1−e(−PA))+Gmin,

7GD=−B(1−e(P−PmaxA))+Gmax,

8B=Gmax−Gmin1−e−PmaxA,

9Gnorm=Gn−GminGmax−Gmin.

G_P and G_D represent the conductance values related to enhancement and inhibition, respectively. G_max, G_min, and P_max represent the maximum conductivity, minimum conductance, and the corresponding number of pulses, respectively. A is a nonlinear parameter and the value of B is a function of A. They can be used to calculate synaptic weight distribution, which are shown in Table 1. In terms of data processing, the input current should first be converted into conductance values, then all conductance values should be normalized. The data should be fitted and entered into the MNIST database to simulate the neural network. As illustrated in Fig. 6(c), 10 000 handwritten digital pictures should be recognized. After 12 training sessions, the accuracy of digital recognition exceeds 90% (Fig. 6(d)). From Fig. 6(e), 20 groups of recognition results are randomly selected from them. Hence, the research results indicate that the system based on Ga₂O₃ synaptic nano-device has the strong digital image recognition ability.

In this work, we have successfully developed a synaptic nano-device based on Ga₂O₃ nanowires, possessing key features such as simulating STDP and optical dependency similar to biological synapses. It has the stability in long-term stimulation and storage. By modulating the light stimulus conditions, the transition of synaptic device can be facilitated from STM to LTM. Similar to the human brain, it is found that the excitatory postsynaptic current can not only recall the memory level under the same stimulation in the relearning stage, but also achieve the higher level. In addition, the synaptic weight can rise to about 32 after being stimulated by light pulses, which reflects the excellent learning performance of the device. The frequency and power of optical pulse stimulation are also demonstrated to have the ability affecting the synaptic weight of synaptic device. With a shorter duration of optical pulse stimulation and fewer optical pulses, the previous cognitive level can be reached during the relearning process. This synaptic nano-device consumes only 2.39 × 10^‒11 J of energy per individual synaptic event. The application of this synaptic device has also been successfully demonstrated in neural morphological computation. Therefore, this work is expected to drive the development of ultra-low-power artificial intelligence systems and hardware-based neuromorphic computing.