Non-Volatile Analog Synapse Characteristics in Bi-Layered Nb₂O₅/CeO₂ Memristors for Artificial Neural Network in Neuromorphic Computing Systems

Data-centric applications such as artificial intelligence demand novel energy-efficient computing systems including processing-in-memory (PIM) and neuromorphic computing. Recently, non-filamentary valence change memory (VCM) type memristors have been widely researched as one of the promising candidates for PIM and neuromorphic systems because of their superior uniformity and stablity of analog resistive switching. The analog memristor crossbar array is an adequate platform for vector-matrix multiplication operations in artificial neural network. However, inferior retention properties of interfacial type VCM memristors due to the volatile redistribution of oxygen ions have not been fully resolved yet.<br/>In this study, bi-layered memristors with cerium oxide (CeO₂) and niobium oxide (Nb₂O₅), i.e., Nb₂O₅/CeO₂, were investigated, where CeO₂ and Nb₂O₅ are suitable material for interfacial VCM owing to their tunable valance states (Ce³⁺, Ce⁴⁺ and Nb⁵⁺, Nb⁴⁺, Nb²⁺) and high oxygen ion diffusivity. ^{[1], [2]}<br/>Active inter-migration of oxygen ions between CeO₂ and Nb₂O₅ enables analog resistance change as synaptic weight update with long-term stability. The bi-layered Nb₂O₅/CeO₂ memristors displayed polarity-dependent analog resistance switching properties. Upon positive pulsing, oxygen ions move from CeO₂ to Nb₂O₅; thereby increasing conductance of the device (potentiation). Then, they reversibly turn back upon negative pulsing, decreasing the device conductance (depression). It was also confirmed that the Nb₂O₅/CeO₂ memristors had enhanced dynamic switching range (&gt;10⁵), analog linearity, symmetry, and retention properties, as compared to single-CeO₂ device. Multi-level conductance states were found to retain non-volatile property even after repeated potentiation and depression cycles. The analog resistance change was originated from the changes in oxygen vacancy concentration in CeO₂ layer, which caused difference in Schottky barrier height between CeO₂ and bottom electrode. And Nb₂O₅ layer worked as oxygen ion reservoir supplying oxygen vacancies to CeO₂ and holding oxygen ions for non-volatile retention stability. TEM and XPS analyses revealed the presence of abundant oxygen vacancies in CeO₂ and Nb₂O₅ layer for the enhanced oxygen ions redistribution. Furthermore, the Nb₂O₅/CeO₂ memristor maintained highly selective synaptic weight update properties in array structures without additional selector. Thanks to non-linear current-voltage characteristics of Nb₂O₅/CeO₂ memristor, sneak leakage current and unintended weight update were efficiently prevented in the array architecture. Using the obtained weight update characteristics, the 94.6 % of pattern recognition accuracy was achieved in simulation of MNIST handwritten patterns using NeuroSim program.<br/>In conclusion, these enhanced synapse characteristics of bi-layered Nb₂O₅/CeO₂ memristor crossbar array demonstrated the potential of proposed device to be applicable to integrated neuromorphic system that implements training and inference operations for neuromorphic computing.