ReRAM-Based Neuromorphic Devices with Signal Enhancement Properties

Synaptic mimicking devices not only have the potential to replace existing memory-based storage media, but also to integrate with CMOS technology or independently mimic the functions of brain cells. With the advent of various AI models like Chat-GPT, the hardware requirements for conventional von Neumann computing systems have significantly increased. To address this, parallel processing systems based on graphics cards have been utilized, and recently, application-specific integrated circuits (ASICs) designed specifically for AI systems have been employed. However, substantial power consumption and large-scale infrastructure are still required. As AI technology becomes increasingly integrated into daily life, the development of low-power, compact intelligent semiconductors for ICT convergence devices have become essential, accelerating research beyond the von Neumann architecture. The human brain, composed of 100 billion neurons and 100 trillion synapses, is more energy-efficient in data analysis and processing than any existing computing system. Neuromorphic computing, which mimics the functions of the brain, is expected to be a major direction for future AI development.<br/>ReRAM, a typical synaptic device made mostly of oxides, operates through the movement of oxygen ions or charge trapping through vacancies. Unlike conventional memory systems, this allows various resistance levels to be used, thus enabling synaptic learning. However, this mode of operation, which is controlled by voltage, presents a challenge: as the resistance decreases as the learning progresses, higher drive voltages are required to maintain the linearity of the learning.<br/>We have fabricated a signal-amplifying layer using the Mott insulator VO₂ and created a bi-layer structure with an extremely thin Al₂O₃ layer on top to implement resistance change characteristics for a synaptic device. The Mott insulator VO₂ exhibits an insulator-metal transition where resistance decreases as temperature increases. As the resistance of the synaptic device decreases, the high current generated can create Joule heating, which can induce an additional reversible resistance reduction characteristic. This allows for a more linear reduction in resistance change at the same voltage.<br/>The resistance change characteristic is implemented through the movement of oxygen ions in the insulator layer, where a highly resistive Al₂O₃ is placed. When oxygen ions move towards the metal side at the interface, the barrier becomes extremely thin, enabling tunneling and forming a low resistance state. Conversely, when oxygen ions move towards the bulk region, the barrier thickness increases, forming a high resistance state. This characteristic allows for resistance change implementation in an extremely thin local area. Additionally, this asymmetric structure can provide the advantage of self-rectification.<br/>Through this work, we introduced the concept of signal amplification by incorporating Mott insulators and successfully implemented a ReRAM-based synapse mimetic device with improved linearity.