A Study on Ion-based Synaptic Transistors for Neuromorphic Applications

Abstract

With increasing demand for processing unstructured data, neuromorphic computing that simulates the human brain has been intensively studied. Efficient data processing and low-power operation of neuromorphic systems enable the development of artificial intelligence, such as deep learning and pattern recognizing. For the realization of neuromorphic hardware with a high recognition accuracy, it is essential to develop a synaptic device with ideal characteristics. Two-terminal (2-T) synaptic devices such as resistive random-access memory and phase-change memory have been actively studied based on their simple structure. However, due to the nature of 2-T devices that control conductance through localized regions of a few nm, there are many limitations such as non-ideal synaptic characteristics, stochasticity, and read-disturbance issue. To solve these limitations, an ion-based synaptic transistor, also called electrochemical random-access memory (ECRAM), has been proposed. In ECRAM devices, the conductivity is determined by the ion concentration in the channel. The bulk ion concentration can be precisely tuned by the gate pulse, resulting in ideal weight-update linearity. Based on near-ideal synaptic characteristics, ECRAM devices are attracting attention as promising synaptic devices. Nevertheless, the ECRAM device has a disadvantage in that the programming speed is slow because the synaptic weight is changed by the electrochemical reaction. Furthermore, in the case of proton- and lithium-ion-based ECRAM devices, there are inevitable drawbacks such as open circuit potential and poor retention, which degrade training and inference characteristic. Moreover, the three-terminal structure based ECRAM device has a large cell size (8F²) compared to the 2-T synapse devices (4F²). This dissertation focused on developing an ideal synaptic device with improved disadvantages such as slow programming speed and low area density. By understanding the working principle and key parameters of ECRAM, I designed near-ideal proton-, lithium-, and oxygen-ion-based ECRAM devices with near-ideal synaptic characteristics. In this process, I presented device design guidelines based on physical parameters affecting synaptic properties. I also improved the program speed through high temperature training by developing a micro-heater integrated ECRAM device. Finally, by using an ion-permeable graphene electrode, I developed a high-density (4F²) vertical sensing ECRAM device with improved synaptic characteristics and reliability.