An electronic synapse device based on aluminum nitride memristor for neuromorphic computing application

Brain-inspired computing is believed to have a better performance compared with the conventional von Neumann computing. The synaptic electronic device is the most important component of a neuromorphic circuit. In this study, we present an aluminum nitride (AlN) based memristor as the synaptic weight element in a functional neural network for handwritten digit recognition. Reliable and stable resistive switching behaviors were successfully demonstrated in the AlN based memristor. Moreover, it also possesses excellent features for neuromorphic applications such as long retention (>10(4) s), and multi-level storage. Continuous and smooth gradual set and reset switching transition can be modulated by applying appropriate compliance current limits and reset stop voltages. We particularly examined long-term potentiation and long-term depression and improved the linearity by optimizing pulse response conditions. Finally, the symmetric and linear synaptic behaviors which can be utilized in a neural network simulation are obtained. Simulations using the MNIST handwritten recognition data set prove that the AlN based memristor can operate with an online learning accuracy of 95%. Our work suggests AlN based memristor has potential for using as an electronic synapse in future neuromorphic systems.