Keonhee Kim1,2,Jae Gwang Lim1,2,Yeonjoo Jeong1,Jaewook Kim1,Suyoun Lee1,Joon Young Kwak1,Jongkil Park1,Gyu Weon Hwang1,Kyeong-Seok Lee1,Seongsik Park1,Hyun Jae Jang1,Byeong-Kwon Ju2,Jong Keuk Park1,Inho Kim1
Korea Institute of Science and Technology1,Korea University2
Keonhee Kim1,2,Jae Gwang Lim1,2,Yeonjoo Jeong1,Jaewook Kim1,Suyoun Lee1,Joon Young Kwak1,Jongkil Park1,Gyu Weon Hwang1,Kyeong-Seok Lee1,Seongsik Park1,Hyun Jae Jang1,Byeong-Kwon Ju2,Jong Keuk Park1,Inho Kim1
Korea Institute of Science and Technology1,Korea University2
The rapid advancement of artificial intelligence (AI) has created a demand for computing architectures capable of meeting its complex processing requirements. Traditional von Neumann architectures are facing challenges in meeting these demands, prompting the exploration of alternative technologies. Neuromorphic computing has emerged as a promising candidate to address this challenge. Memristive devices have gained significant attention as potential artificial synapses for neuromorphic computing hardware. While numerous studies have been conducted, further improvements are needed across various aspects of memristive devices.<br/>This study emphasizes the significance of analog switching with highly linear updates in synapse weights as a critical parameter. Additionally, the high self-rectifying characteristics of memristive devices were explored to suppress crosstalk in memristive neural networks, thereby enhancing scalability. We propose a memristive device consisting of Cu/TaO<sub>x</sub>/IGZO/Pt with dual Schottky barriers at the interfaces TaO<sub>x</sub>/IGZO and IGZO/Pt, respectively. TaO<sub>x</sub> serves as a switching layer, and IGZO as a current limiting layer. The energy barrier at the interface of IGZO/Pt induces a self-limiting effect in a positive bias. Limiting the current prevents excessive formation of Cu conductive filaments, thereby leading to analog synaptic switching. Another energy barrier at the interface of TaO<sub>x</sub>/IGZO tunes the overall filament size by limiting the current, providing self-rectifying characteristics as well. Oxygen stoichiometry in both oxide layers was found to be a key factor to modulation of the dual Schottky barriers. The memristive devices with optimally controlled dual barriers offer high linearity in synapse weight updates, long retention time as well as high self-rectifying ratio. We discussed the origin of the energy barrier modulations induced by the control of oxygen stoichiometry. Lastly, we implemented a small-scale crossbar array and demonstrated vector matrix multiplication, showing its potential as a next-generation application.