Realize low-power artificial photonic synapse based on (Al,Ga)N nanowire/graphene heterojunction for neuromorphic computing

As the Internet of Things (IoT) and artificial intelligence (AI) develop rapidly, the von Neumann bottleneck becomes increasingly pronounced during the processing of complex information that involves interactions with a dynamic environment and various uncertainties.^1,2 In contrast, the human brain exhibits excellent efficiency in solving complex and unstructured problems.³ The human brain is an interconnected network that is composed of organic biological microfibers (neurons and synapses), which can be used for information transport.⁴ Drawing inspiration from the neural architecture of the human brain, an artificial neural network (ANN) that emulates the functions of biological neural networks has garnered significant attention from the scientific community in recent years.^3,5,6 Synaptic devices capable of synchronously storing and processing information are the essential components of an artificial neural network.⁷ Hence, the development of optoelectronic devices that can mimic synaptic behavior is quite crucial and necessary.³ 

Due to the advantage of fast and broad-band information transmission with low-power consumption, photonic synaptic devices are considered the ideal device to simulate the biological synapse, which could confer significant advantages for neuromorphic computing.^8–11 In general, light-stimulated synaptic devices contain active materials that have the ability to stimulate history-dependently, such as persistent photoconductivity (PPC).^9,12,13 The varied conductivity or current can be utilized to emulate the functions of biological synapses under stimuli.^9,13,14 Furthermore, (Al,Ga)N-based materials exhibit a stable PPC effect at room temperature, providing a promising route for the development of light-stimulated synaptic devices.^15,16 In addition to their tunable bandgap and high stability, (Al,Ga)N-based materials are one of the most promising candidate materials for synapses devices.¹⁵ Moreover, combined with the excellent characteristics of a large surface-to-volume ratio and efficient charge carrier transport, these nanowires (NWs) can improve the memory duration of the photocurrent.^17–22 

On the other hand, the heterojunction-based synapses device shows exceptionally sensitive photoconductance properties, and the photo-response spectral range can be easily controlled by using the absorber with an appropriate bandgap.²³ Graphene has good conductivity and ultrahigh transmissivity in the UV range, which makes it suitable to be used in connecting NWs and conducting current.^24,25 In our previous studies, we fabricated the GaN-based NWs and utilized them in photodetectors successfully.^26–28 However, to date, very few works have been reported about utilizing (Al,Ga)N NW/graphene heterojunction in light-stimulated artificial synapses, not to mention studying the performance of power consumption and mechanisms.

In this work, we have demonstrated a light-stimulated synaptic device based on (Al,Ga)N NW/graphene heterojunction. This artificial synaptic device is used to emulate multiple functionalities of biological synapses, including spike-timing-dependence plasticity (STDP), spike-number-dependent plasticity (SNDP), learning behavior, and the transition from short-term memory (STM) to long-term memory (LTM). Apart from the experiments, we have also investigated the underlying mechanisms of this synaptic device. To evaluate the learning capabilities of the synaptic devices, an artificial neural network (ANN) for mimicking bio-vision has also been established and demonstrated.

By molecular beam epitaxy (MBE, Vecco G20), the (Al,Ga)N NWs [Fig. 1(a)] were grown on the n-type Si(111) substrates. The N atoms were supplied by the N plasma cell, while the Al and Ga atoms were generated by the Al and Ga effusion cells. To eliminate native oxides in the growth chamber before epitaxial growth, the Si substrates were heated to 900 °C for 15 min. First, a thin AlN buffer segment deposited on the Si surface forms islands by the Stranski–Krastanow growth mode, which acts as a material collector and a seed for the NW growth.^29,30 After that, a GaN section was grown with a nominal Ga flux of 2.0 × 10⁻⁸ Torr for 100 min. Then a (Al,Ga)N segment was grown with a nominal Al/Ga flux ratio of 1.5 for 30 min. To characterize the growth rate of (Al,Ga)N NWs, two AlN segments were inserted into the (Al,Ga)N section with a nominal Al flux of 1.5 × 10⁻⁸ Torr for 4 min. During the epitaxial process, the nitrogen flow rate and plasma power were kept at 4.8 sccm and 450 W, respectively.

In order to fabricate artificial synaptic devices, a thick layer of SiO₂ was deposited on the surface of (Al,Ga)N NWs, followed by the selective removal of some portions of the SiO₂ film to expose a certain square area of (Al,Ga)N NWs [Fig. 1(b)]. Subsequently, a backside electrode of Ti/Al (50/200 nm) was deposited onto the Si substrate by electron beam evaporation [Fig. 1(c)]. Then the sample was annealed in an N₂ atmosphere at 400 °C for 30 min to form an ohmic contact. After that, graphene was transferred onto the NW surface by wet transfer [Fig. 1(c)]. After that, the artificial synaptic device was fabricated successfully by depositing another electrode consisting of Ti/Pt/Au (25/55/300 nm) onto the graphene surface within a certain area [Fig. 1(d)]. To study the mechanism and stability, another synaptic device (Device-HNO₃) was prepared for comparison, whose NWs were treated with dilute nitric acid (HNO₃) for 4 h.

Based on the experimental results of device parameters, such as nonlinear values, a multilayer perceptron (MLP) neural network model consisting of 784 input neurons, 100 hidden neurons, and 10 output neurons was established. In the network, each synaptic device was treated as a neuron. 28 × 28 image pixel information corresponds to the 784 input neurons. The Modified National Institute of Standards and Technology (MNIST) dataset was used for input. The sigmoid function was used as the activation function in this model. Furthermore, the output of ten classes of digits (0–9) was transformed into one pot vector for recognition, such as 2 => [0, 0, 1, 0, 0, 0, 0]. To optimize the weight value, a random gradient descent method was employed for continuous iteration. Moreover, the back-propagation algorithm was used to continually adjust the threshold values and weight of the network, updating weight values and minimizing the total square error in real time. Finally, the MNIST handwritten dataset was able to achieve accurate number recognition through the accurate evaluation and prediction of these numerical values.

The morphology and element distribution of (Al,Ga)N NWs were characterized in this study through scanning electron microscopy (SEM), spherical aberration corrected scanning transmission electron microscopy (AC-STEM), and high-resolution energy-dispersive X-ray (EDX) mapping. X-ray photoelectron spectroscopy (XPS) was used to characterize surface band bending (BB) and chemical states. All XPS spectra were analyzed by MultiPak software. Both core level and valence band maximum (VBM) were calibrated against C1s (284.8 eV). The measurement of current–time (I–T) characteristics was conducted using an Agilent B1505A semiconductor parameter analyzer. To determine the quality of the grown graphene layers, Raman spectroscopy was carried out using a laser excitation of 532 nm. To characterize the optical properties, a photoluminescence (PL) system with a 213 nm laser was utilized. In the response measurements, a function generator (FY6900-20M) was employed to control the 310 nm LED, which served as the light source.

In the biological nervous system, the synapse is an important structure to complete the signal transmission between two adjacent neurons and is composed of the pre-synapse, post-synapse, and synaptic cleft.³¹ The schematic images of biological neurons and synapses are illustrated in Figs. 1(e) and 1(f). When the biological pre-synapse is stimulated, the pre-synapse membrane will trigger the rapid release of neurotransmitters. Then the receptor on the post-synapse membrane opens the Ca²⁺ channel to realize signal transmission.³² In this work, the structure of synaptic devices based on (Al,Ga)N NWs is designed to emulate the functionalities of biological synapses [Figs. 1(d)–1(f)]. The generation and transport of carriers in the artificial synaptic device under light stimulation are used to emulate the transport of neurotransmitters in the biological synapse.

In the biological nervous system, a brief and slight increase in synaptic weight will lead to the formation of STM, which is located in the hippocampus for only a few minutes at most. If repeated stimulation is applied to the synapse, the weight of the synapse will continue to increase. After that, STM will transfer from the hippocampus to the cerebral cortex and become LTM, where it can be long-lasting or even permanent.⁴⁹ Inspired by the biological synapses, the STDP and SNDP of the (Al,Ga)N NW synaptic device are explored and shown in Fig. 3. As clearly shown in Fig. 3(a), a small EPSC triggered by a short t_on (0.05 s) decays rapidly, while a larger EPSC triggered by a longer t_on (0.5 s) requires more time to reach the level of the black curve, indicating a transition from STM to LTM. To reach the same baseline level (orange dashed line) after undergoing five light pulses [Fig. 3(b)], the synaptic device exhibits a higher EPSC when the t_off is shorter (0.1 vs 0.9 s). Similar phenomena can be observed under different light power densities [Figs. S2(a)–S2(d)]. The spike numbers in Figs. 3(c) and 3(d) are one and five, respectively, leading to different EPSC results. When t_on and t_off are both 0.1 s, the EPSC results of the device triggered by five light pulses [Fig. 3(d)] are higher than those triggered by only one light pulse [Fig. 3(c)]. Therefore, the synaptic device exhibits SNDP performance. Under simulation with the different light power densities, the SNDP characteristics can also be observed in Figs. S2(e) and S2(f). For the synaptic device, the ability to transfer STM to LTM depends on the light-pulse time of duration, frequency, pulse number, and light intensity, reflecting the synaptic plasticity of the device.

In the synaptic device, (Al,Ga)N NWs are used as light absorbers. The heterojunction between graphene and (Al,Ga)N NWs can generate a built-in electric field, which facilitates the separation of photogenerated carriers. When the device is stimulated by light, photogenerated carriers will be generated and separated under the built-in electric field [Fig. 4(c)]. Subsequently, some of the photogenerated carriers are trapped by capture centers (e.g., Al vacancies and graphene defects), while others transport to the electrodes, contributing to the current signal. When the light is off, these photogenerated carriers captured could slow down the speed of current recovery [Fig. 4(d)]. The larger BB in Device-HNO₃ can make the transport of carriers more difficult [Fig. 4(e)], leading to a lower EPSC [Figs. 4(a) and 4(b)]. Furthermore, due to the increased Al vacancies in Device-HNO₃ (Fig. S4), more carriers will be captured. When the light stimulus is removed, the carriers in the trapped states need more time to transport [Fig. 4(f)], resulting in a longer decay time [Fig. 4(b)]. Therefore, graphene defects, surface band bending, and vacancies can be the key reasons contributing to the synaptic characteristics. In such a synaptic device, the incident light, the light-generated carriers, and the electrodes are used to play the roles of the action potential, neurotransmitters, and pre-synaptic/post-synaptic membranes in the biological synapse, respectively.

In this work, an artificial synaptic device based on (Al,Ga)N NW/graphene heterostructure has been successfully fabricated. Such a photonic synapse has been demonstrated to have the characteristics of STDP, SNDP, and light-intensity dependence. The energy consumption of the artificial synapse can be only 2.6 µJ for a single synaptic event. Furthermore, the PPF index of the photonic synapse can reach 202%. Thanks to the dynamic “learning–forgetting” performance, human learning behavior can also be emulated in light-stimulated synaptic devices. It is found that the reservoir effect of the graphene, the surface band bending, and the Al vacancies on the nanowire surfaces can contribute to the synaptic characteristics. Based on the experimental conductance for long-term potentiation and long-term depression, the recognition accuracy of simulating the neural network can reach 92%. Hence, such artificial synapses based on NW/graphene heterostructures could provide a pathway for applications requiring low-power consumption, including bio-realistic artificial intelligence, neuromorphic computing systems, etc.

See the supplementary material for additional details on the data analysis.

The authors are grateful for the Key Research Program of Frontier Sciences, CAS (Grant No. ZDBS-LY-JSC034), the Research Program of Scientific Instrument and Equipment of CAS (Grant No. YJKYYQ20200073), and the National Natural Science Foundation of China (Grant No. 62174172). The authors are thankful for the technical support from Vacuum Interconnected Nanotech Workstation (Nano-X, Grant No. F2309), Platform for Characterization and Test of SINANO, CAS.

The authors have no conflicts to disclose.

M.Z. completed all experiments in device fabrication, device measurements, and the corresponding data collections and analyses. Y.Z. conceived the idea and guided the work. Y.Z. and X.G. completed the SEM measurements. Q.Z. and M.Z. completed the numerical simulations. M.Z., Y.Z., and J.Z. completed the mechanism study. M.Z., Y.Z., and S.L. wrote the original draft of this work. S.L. and Y.Z. carried out the funding acquisition and project administration. Y.Z. carried out all MBE experiments. M.Z., Y.Z., J.Z., and M.J. carried out the methodology and visualization of this work. M.Z., J.Z., and M.J. performed the investigation. All authors reviewed this manuscript.

Min Zhou: Conceptualization (equal); Data curation (lead); Formal analysis (equal); Investigation (equal); Methodology (lead); Writing – original draft (equal); Writing – review & editing (equal). Yukun Zhao: Conceptualization (equal); Funding acquisition (equal); Project administration (equal); Supervision (lead); Writing – original draft (equal); Writing – review & editing (equal). Xiushuo Gu: Formal analysis (equal); Investigation (equal); Methodology (equal); Writing – review & editing (equal). Qianyi Zhang: Investigation (equal); Software (lead); Writing – review & editing (equal). Jianya Zhang: Formal analysis (equal); Validation (equal); Writing – review & editing (equal). Min Jiang: Methodology (supporting); Validation (equal); Writing – review & editing (equal). Shulong Lu: Funding acquisition (equal); Supervision (equal); Writing – review & editing (equal).

The data that support the findings of this study are available from the corresponding authors upon reasonable request.