Traditional computing relies on the Von Neumann architecture, where processing and memory are separate entities. In contrast, neuromorphic computing aims to overcome the limitations inherent in the Von Neumann architecture by replicating the structure and functionality of the human brain to combine processing and memory^[1−3]. The human nervous system consists of approximately 10¹¹ neurons and 10¹⁵ synapses each of which consumes energy of about 1−100 fJ during information processing^[4−6]. Synapses play a vital role in the human brain, acting as dynamic connectors that enable neurons to communicate efficiently. They modulate the strength and timing of the signals transmitted between neurons, adapting in response to activity, which is fundamental for learning and memory. Therefore, artificial synaptic devices are foundational elements in this computing paradigm^{[7, 8]}. Recent advancements have shown a variety of synaptic devices offering unique properties^{[9, 10]}. Traditionally, the modulation of synaptic weight in these devices relied on electrical stimulation. However, this approach often encounters bandwidth constraints, which can impede processing speed. To address these challenges, optical excitation presents a more effective alternative. It offers low crosstalk and high bandwidth capabilities, surpassing the limitations of electrical excitation. Photoelectronic synaptic devices, which harness the synergistic benefits of both optical and electronic modulation, have emerged as a promising pathway^{[11, 12]}.

Currently, a variety of materials have been applied to photoelectronic synaptic devices, including two-dimensional (2D) materials^[13−15], oxides^{[16, 17]}, organic compounds^{[18, 19]}, and phase change materials^{[20, 21]}. These devices have demonstrated capabilities in emulating fundamental synaptic functions, such as photonic potentiation and electronic depression, which are vital for the processes of learning and memory in artificial neural networks. Oxide-based devices, in particular, have attracted significant attention due to their straightforward fabrication process and superior light-responsive abilities, making them ideal candidates for the construction of large-scale, highly integrated neuromorphic systems^{[22, 23]}.

Indium gallium zinc oxide (IGZO), a transparent high-mobility oxide semiconductor, is renowned for its applications in thin-film transistors, especially in liquid crystal displays (LCDs), due to its transparency and stability^[24−26]. Beyond its established applications, the potential of IGZO in neuromorphic computing is an area of emerging interest. The unique properties of IGZO, such as its high electron mobility and stability, make it a compelling choice for developing more efficient and scalable neuromorphic devices. In particular, the material's ability to operate at low voltages while maintaining high performance is beneficial for reducing power consumption, a critical factor in large-scale implementations. Although its application in neuromorphic transistors has been explored^{[27, 28]}, the use of IGZO in two-terminal neuromorphic devices is still in the early stages. Two-terminal devices are known for their simplistic fabrication, high integration capability, and seamless compatibility with CMOS circuits^{[29, 30]}. Developing IGZO-based two-terminal devices could lead to significant advancements in neuromorphic computing, potentially leading to advancements in device efficiency, integration density, and overall performance.

In this study, we have successfully developed an IGZO-based two-terminal artificial synaptic device that exhibits effective simulation of key synaptic functions, including excitatory post-synaptic current (EPSC), paired-pulse facilitation (PPF), long-term potentiation (LTP), long-term depression (LTD) and the learning-forgetting-relearning mechanism of the human brain. A model has been introduced to elucidate the underlying mechanisms of these behaviors. The device's synaptic weight, derived from light pulse stimulation, has been integrated into a three-layer artificial neural network (ANN). This integration has been effective, as the network successfully recognized handwritten digits from the MNIST database^[31] with an accuracy of 93.4%. These results highlight the potential of our device in the advancement of neuromorphic computing (NC) systems, not only highlight the efficiency and effectiveness of our IGZO-based synaptic device. The implementation and outcomes of this study pave the way for future research in neuromorphic computing, indicating the potential for the development of more advanced and efficient artificial intelligence systems.

Firstly, the Si substrate was cleaned sequentially in acetone, ethanol, and deionized water using ultrasonic agitation for 5 min each, followed by drying in a nitrogen environment. A 10 nm titanium (Ti) film and a 70 nm nickel (Ni) film were then deposited on the substrate using physical vapor deposition (PVD), serving as the adhesion layer and the bottom electrode, respectively. Subsequently, a 60 nm IGZO film was deposited to form the dielectric layer using plasma-enhanced atomic layer deposition (ALD) at a growth temperature of 200 °C. Finally, the top electrode was defined lithographically into a rectangle measuring 80 μm, and a 100 nm thick layer of indium tin oxide (ITO) was deposited over it using the PVD method, completing the top electrode structure.

The composition of the devices was determined by X-ray photoelectron spectroscopy (XPS). The electrical properties of the devices were measured by an Agilent B1500 semiconductor device analyzer.

The structure and fabrication process flow of our artificial synaptic device is illustrated in Fig. 1(a). This design is instrumental in simulating the functionalities of biological synapses which are critical junctions for efficient information transmission between neurons. As depicted in Fig. 1(b), in biological synapses, neurotransmitters are released from the presynaptic neuron and cross the synaptic cleft to bind to receptors on the postsynaptic neuron, facilitating the propagation of nerve impulses. Our device features a dielectric layer composed of an IGZO film, and the composition of this layer is crucial for the device's performance. Fig. 1(c) presents the X-ray photoelectron spectroscopy (XPS) characterization of the dielectric layer, confirming the elemental of composition the IGZO film.

The conductance modulation of the device through consecutive voltage scans is illustrated in Figs. 2(a) and 2(b). A negative voltage scan ranging from 0 to −2 V resulted in a gradual decrease in the device's conductance, while a positive voltage scan from 0 to 2 V led to a gradual increase in conductance. Furthermore, it is noteworthy that the device exhibits pronounced rectification characteristics, with the current under negative voltage being over four orders of magnitude higher than under positive voltage. To emulate long-term potentiation (LTP) and long-term depression (LTD), which are crucial for neural network adaptability, we applied 50 positive pulses (1 V, 200 ms) followed by 50 negative pulses (−1 V, 200 ms) consecutively, with a 1-s time interval. As shown in Fig. 2(c), this resulted in an initial conductance increase followed by a decrease, indicative of LTP and LTD, respectively.

Fig. 3 illustrates the rectification and conductance modulation mechanisms of the device. Ni is a metal with a high work function, and therefore, the current model at the Ni−IGZO interface can be explained using the thermionic emission model. As shown in Fig. 3(a), there is a Schottky barrier at the Ni−IGZO junction. When a high potential is applied to the nickel electrode (i.e., a negative voltage is applied to the device), the Schottky junction is forward-biased. Conversely, when a low potential is applied to the nickel electrode (i.e., a positive voltage is applied to the device), the Schottky junction is reverse-biased, resulting in a modulated current. Fig. 3(b) depicts the conductance modulation mechanism of the device. The device conducts electricity based on oxygen vacancy conduction due to the relatively thick dielectric layer of IGZO (60 nm). The top electrode, ITO, acts as an oxygen capture layer, creating an oxygen-rich region near the top electrode and an oxygen-deficient region near the bottom electrode. Most oxygen vacancies are concentrated on the side closer to the bottom electrode. When a negative voltage is applied to the device, oxygen vacancies migrate toward the side closer to the top electrode and are captured and neutralized by ITO, resulting in a decrease in device conductance. On the other hand, when a positive voltage is applied, the oxygen capture layer releases oxygen vacancies, leading to an increase in oxygen vacancy concentration and an increase in device conductance.

After verifying the device's electrical pulse conductance modulation characteristics, we proceeded to investigate its photoelectric pulse conductance modulation properties. EPSC of synapses refers to the transient increase in electrical activity in a neuron's postsynaptic membrane, which can lead to the generation of action potentials. As depicted in Fig. 4(a), EPSC was recorded in response to a series of optical pulses with varying pulse widths. The light source had a wavelength of 320 nm, the power density of the optical pulse we used was 4.51 mW/cm², and a small read voltage of 10 mV was applied. It's evident that as the pulse duration increased, the response current of the synaptic device exhibited nearly linear growth. In Fig. 4(b), a pair of optical pulses with a 1-s time interval was applied to the device. The EPSC generated by the first pulse was denoted as A1, while the EPSC produced by the second pulse was labeled as A2. Due to the persistence of photogenerated carriers after the first pulse, A2 was observed to be greater than A1. This successful simulation replicates the PPF observed in biological synaptic responses. PPF is a mechanism that contributes to the dynamic regulation of synaptic transmission and neuronal communication, contributing to more effective information processing. Furthermore, as illustrated in Fig. 4(c), it became evident that the memory feature of the device could be enhanced with an increase in the frequency of light stimulation. Over a span of 60 s, the synaptic weight transitioned from short-term plasticity (STP) to long-term plasticity (LTP) as the pulse frequency increased. Fig. 2(d) showcases the LTP achieved by applying 50 consecutive light pulses with a wavelength of 320 nm, a pulse width of 1 s, and a frequency of 0.5 Hz. LTP could facilitate learning and memory processes within the organism. The plot exhibits the nearly linear weight modulation of the device under light pulses, establishing a robust basis for its utilization in the subsequent focus on neural network applications. In addition to these findings, we simulated the process of synaptic learning, forgetting, and relearning. Initially, we administered 50 optical pulses to the device to facilitate the learning process. Subsequently, there was a spontaneous decline in conductance over 200 s, followed by the application of additional optical pulses for relearning. As illustrated, only 16 pulses were required during the relearning process to achieve the current level of conductance. This demonstrates the potential of our synaptic device in future applications related to biomimetic neural systems, emphasizing its adaptability and utility.

To assess the prospective application of our device in neural networks, we established an artificial neuromorphic network (ANN) for handwritten digit recognition, utilizing our artificial synaptic device and drawing from the modified National Institute of Standards and technology (MNIST) database. The ANN architecture comprised three layers: an input layer with 256 neurons, a hidden layer with 64 neurons, and an output layer with 10 neurons, as depicted in Fig. 5(a). We employed synaptic weights derived from the conductance of long-term potentiation (LTP), as illustrated in Fig. 2(d), to establish the linear weight layers within the network. Fig. 5(b) demonstrates the progression of recognition accuracy throughout 1000 epochs of training with the backpropagation (BP) algorithm. After 1000 epochs, the network achieved an impressive recognition rate of 93.4%. The average confusion matrix, presented in Fig. 5(c), illustrates the outcomes of digit recognition training. The horizontal axis represents the actual labels of the images in the test set, while the vertical axis represents the labels assigned by the neural network. Following 1000 training epochs, the prominently green diagonal elements signify successful recognition of the digit images. These findings underscore the device's potential for image recognition and its contribution to advancing neuromorphic computing.

In summary, this study has provided compelling evidence for the remarkable potential of IGZO in the development of two-terminal artificial synaptic devices, while elucidating the underlying mechanisms. The device effectively emulated a diverse range of synaptic characteristics, including EPSC, PPF, LTP, LTD, and the learning-forgetting-relearning mechanism. Furthermore, when integrated into an artificial neural network, it achieved an impressive recognition accuracy of 93.4% in identifying handwritten digits from the MNIST database. These results collectively highlight the promise of IGZO-based memristors as a transformative avenue for advancing the field of neuromorphic computing systems. This research paves the way for future innovations in artificial intelligence and cognitive computing.