Gd-Doped CeO₂/CeO₂ Bi-Layer Memristors-Based Reservoir Computing—Study on Optimal Parameters of Memristors for Pattern Recognition

Reservoir computing (RC) is a new powerful framework for effectively processing complex data, including pattern recognition, speech recognition, time series analysis, and natural language processing. Unlike DNNs, requiring the entire weight network to be trained by the gradient method, RCs have an advantage in increased learning speed with reduced cost because training is performed only at the readout while maintaining the rest of the network fixed. Among several device candidates for hardware implementation of RC systems, memristor is a proper and efficient device for RC framework because it has nonlinear functions and exhibits memory characteristics [1]. Analyzing the time-dependent weight dynamics of memristors is important to extract various spatial and temporal characteristics of the reservoir in performing spatiotemporal operations in RC. These features include high dimensionality of reservoir, nonlinearity with respect to weight update, nonlinear short-term plasticity (STP), and class separation property. In particular, the weight update is performed consecutively to obtain time-varying reservoir state, where a paired-pulse facilitation (PPF) occurs that impacts on the nonlinearity characteristics of the weight updates. In addition, the reservoir state must have short-term decay characteristics to process spatiotemporal information via effectively separating complex sequential information without superimposing reservoir states.<br/>In this study, we experimentally demonstrate RC system with Pt/Gd-doped CeO₂ (GDC)/CeO₂/Pt memristors exhibiting time-dependent weight update and decay characteristics, which benefit to realize RC systems. The bi-layered oxide memristors based on the same CeO₂ matrix, i.e., GDC/CeO₂, are characterized to have more stable and controllable redistribution of oxygen vacancies between CeO₂and GDC layers under the electric field. CeO₂has a high oxygen ion conductivity and variable valence states of Ce cations, which are favorable properties as a resistive switching material. GDC was also reported to have increased oxygen ion conductivity via increased oxygen vacancy concentration as a result of Gd doping in CeO₂. The proposed GDC/CeO₂ bi-layered oxide memristor has been demonstrated to have analogous synaptic weight update, PPF, STP characteristics and their adjustment could be improved compared to the single-layered CeO₂ or GDC memristors [2]. This GDC/CeO₂ memristors were employed to perform RC by storing reservoir states with controllable weight update and decay characteristics that effectively separate complex sequential information without superimposing the states. The RC performance of the memristor was evaluated by a 4-bit pattern verification as applying various pulse signals to the memristor using Modified National Institute of Standards and Technology (MNIST) handwriting database to train and test the RC system. Pattern recognition simulations showed accuracy levels of up to 90 % in bilayer memristors, confirming the potential of these bilayer memristors as artificial synapses for neuromorphic computing. In addition to the experimental study, the short-term decay characteristics were fitted using an exponential decay curve to obtain the value of �� (time constant). Then, we propose optimal parameter values of the memristor for RC system by tuning �� and PPF index to trace the conditions that can distinguish between reservoir states. Therefore, the dynamics of memristors in complex spatiotemporal tasks could be investigated to derive optimal parameters, which can then be implemented in RC systems to achieve high-performance recognition and prediction.<br/><br/><b>References</b><br/>[1] ZHANG, Guohua, et al., <i>Adv. Funct. Mater., 33(42)</i>, 2302929, 2023<br/>[2] Moon, Sola, et al.,<i> J. Alloys Compd., 963, </i>171211, 2023