Neuron devices using ferroelectric-based FET and its neuromorphic applications

Abstract

This dissertation aims to suggest a device structure that can improve the area efficiency of ferroelectric device-based neuron devices and to find a suitable neuromorphic application. Low power consumption during polarization switching makes FeFET neurons advantageous for low-power operation. Furthermore, Leaky Integrate-and-Fire (LIF) neuron behavior can be implemented in a Leaky FeFET without additional components such as a capacitor, comparator, and amplifier. However, the simple structure of neuron devices has been nullified with the additional components to implement inhibitory connection in many practical applications. In this study, a Leaky FeFET with a Split-Gate structure was proposed as a neuron device. It was confirmed that both excitatory and inhibitory connections were implemented without introducing additional devices. At the same time, to reduce standby power, a compact structure was proposed to combine FeFET and an additional nFET, which is only controlled by excitatory gate input. The device can save standby power in the off-state by suppressing the tail current, which is a chronic problem in FET-based neuron devices. The fabricated device shows both LIF and inhibitory operation of neurons. Furthermore, a 3-layer SNN simulation based on the Split-Gate Leaky FeFET neuron shows an accuracy of over 90% for pattern recognition examples, thereby verifying its validity as a neuron device. While previous works considered the Leaky FeFET as a general LIF neuron for conventional SNN applications, it was confirmed that the Leaky FeFET-based neurons are specialized for Leaky & Integrate (LI) behavior. In this study, Leaky FeFET-based neurons were applied to reservoir computing, which is specialized to process sequential data. Time-delayed feedback of LI-enhanced neurons is suitable for the physical reservoir that requires temporary memory effects and non-linearity between input and output. In the proposed reservoir computing scheme with the Leaky-FeFET reservoir, the memory capacities (MCs) were 2.08 and 1.71 for short-term memory and parity check tasks, respectively. As an aside, device simulation for antiferroelectric FET was also conducted to check whether retention degradation in antiferroelectric FET can contribute to reservoir computing. Improvement in operation speed and energy efficiency is expected while it shows comparable reservoir computing performance with Leaky FeFET neurons. This study can contribute to the implementation of bio-inspired low-power artificial neural networks by applying the unique characteristics of FeFET-based neuron devices.