Inspired by the biological nervous system, the neuromorphic hardware implementation based on compute-in-memory (CIM) architecture and highly adaptive computing mode are promising to significantly improve computer efficiency and performance. At present, it is still a great challenge to integrate system-level neuromorphic hardwares, and neuromorphic computing with high recognition accuracy can be realized by adopting synaptic devices in combination with neural network algorithms. Therefore, an optical synaptic device based on back-to-back Schottky junction (B-B SJ) is presented, and then some common synaptic plasticities are emulated, such as post-excitatory synaptic currents (EPSCs), short-to-long-term plasticity, interval-dependent paired-pulse facilitation (PPF), and learning-forgetting-relearning process. Additionally, the memristor-based convolutional neural network (M-CNN) is constructed by mapping the device conductance change to the weight change of the convolutional neural network, and its applications in image recognition are evaluated. Meanwhile, the experimental results show that the recognition accuracy can reach 95.12% and demonstrate the potential applications of devices in neuromorphic computing.

Optical synaptic devices have been proposed based on B-B SJ. The device conductance is modulated by light-induced Schottky barrier modulation, with simultaneous non-volatile conductance state regulation achieved via the silicon dioxide interface trapping effect. Synaptic behavior such as EPSC, PPF, short-to-longterm plasticity, and PPF has been successfully emulated by adopting B-B SJ devices. Furthermore, by extracting the device conductance and mapping the conductance range into the M-CNN algorithm, image information recognition has been accomplished. The results indicate that by adopting the M-CNN algorithm, this neuromorphic device demonstrates outstanding performance in image recognition tasks, with an accuracy of up to 95.12%.

An M-CNN is constructed to test its image recognition performance. The conductance values of the devices are mapped as weight values for image recognition. Figure 4(a) presents a schematic diagram of this process, while Fig. 4(b) shows the feature maps of the convolutional layer. Different numbers of pulses are applied to the device, which results in three distinct conductance ranges as illustrated in Fig. 4(c). The corresponding changes in conductance values for these three scenarios are depicted in Fig. 4(d), indicating that an increase in the pulse number enhances the conductance range of the device. The confusion matrices and accuracy distribution plots for the M-CNN recognition rates corresponding to different pulse numbers are shown in Fig. 5. A comparison with the results in Table 1 reveals that the image recognition network constructed by the B-B SJ artificial synapse device performs well in image recognition tasks, demonstrating high accuracy and further validating the effectiveness of the device for neuromorphic computing.

A B-B SJ artificial synapse device is fabricated to simulate various plasticity behaviors of biological synapses, including EPSC, short-to-long-term plasticity, interval-dependent PPF, and learning-forgetting-relearning process. Additionally, the corresponding conductance parameters are extracted from this artificial synapse device, and a three-layer CNN based on this device is constructed. This network achieves a recognition accuracy of 95.12% in tests on the MNIST handwritten digit dataset. These results demonstrate the device’s advantages in image information processing and confirm its potential as an artificial synapse device for neuromorphic computing applications.