Analog Synaptic Characteristic in Bilayered Gd-Doped Ceria and CeO₂ Memristors for Neuromorphic Computing

Neuromorphic computing, inspired by the human brain that consumes less than 20 W of power, is emerging as an alternative to the existing von Neumann computing system suffering from its limits in device scalability and delayed data transfer between memory and processing unit [1]. In order to develop neuromorphic computing systems, significant efforts have been made to develop artificial electronic synaptic devices that mimic biological synaptic functions. Among various synaptic devices, analog resistance switching-based oxide memristive devices have promising opportunities due to their unique properties such as high density, good scalability, and high energy efficiency. In this study, we have demonstrated linear, symmetric, and analog synaptic characteristics in the bilayer structures of Pt/Gd-doped ceria (GDC)/CeO₂/Pt and Pt/CeO₂/Gd-doped ceria (GDC)/Pt memristors with different oxide stacking order. In particular, Gd-doped ceria, being one of the oxides having high ionic conductivity among mixed oxygen ionic-electronic conductors. Also, Gd-doped ceria has been reported to form a high concentration of oxygen vacancies [2]. As the concentration and mobility of oxygen-based defects play a significant role in oxide-based memristors, the bilayered CeO₂ and GDC memristors were fabricated and their synaptic properties were investigated. The use of bilayer memristors by stacking CeO₂ with more oxygen-deficient GDC turned out to have improved linearity and symmetry in potentiation and depression behaviors thanks to redistribution of oxygen vacancies between CeO₂ and GDC. The Pt/GDC/CeO₂/Pt memristors showed the potentiation behavior with decreasing resistance upon applying positive voltage at top Pt electrode while the application of negative voltage caused the depression with increasing resistance. The Pt/CeO₂/GDC/Pt memristors with reverse stacking order of GDC and CeO₂ exhibited opposite dependence of resistance change on the applied voltage polarity, which is associated with the resistance switching mechanism of energy barrier modulation at GDC/CeO₂ interface and Schottky barrier modulation at the interface with electrodes. Pattern recognition simulations using the modified MNIST dataset showed accuracy levels of up to 88% in bilayer memristors, confirming the potential of these bilayer memristors as artificial synapses for neuromorphic computing.<br/><br/>References<br/>[1] C. Li, M. Hu, Y. Li, H. Jiang, et al. Nat. Electron. 1, 52 (2018)<br/>[2] R. Schmitt, J. Spring, R. Korobko, et al. ACS NANO, 11, 8881-8891 (2017)