Homogeneous Memristors with Tunable Decay Dynamics for Self-Adaptive Reservoir Computing
Yongfei Li, Wei Tang, Zhiyuan Li, Weihao Niu, Shuangshaung Han, Xiangshui Miao, Rui Yang
Abstract
Leveraging inherent nonlinear dynamics, memristors have demonstrated superior performance in reservoir computing (RC). However, the use of different materials for reservoir nodes and readout layers poses significant challenges to integration. Moreover, the reported RC systems generally employ fixed reservoir nodes with limited temporal dynamics, which severely restricts the processing of sequences with complex temporal features in practical applications. Here, a homogeneous RC system based on the CMOS-compatible TiO x and AlO y thin films has been demonstrated. The reservoir nodes, which require nonlinear temporal dynamics, are prepared based on the TiO x /AlO y dynamic memristors. Furthermore, the readout layers are implemented using the nonvolatile memristors with AlO y /TiO x stacked structures. Furthermore, an effective approach to modulate the time constants of dynamic memristors is proposed by controlling the reading V bias, thereby expanding the precise prediction temporal scale to 10 in the Hénon map benchmark task. Additionally, a self-adaptive RC system is simulated, and the recognition accuracy improves from 82.0% to 93.9% by controlling the time constants of reservoir nodes for dynamic gesture recognition involving complex temporal features. This work provides a promising demonstration of enhanced performance in complex temporal sequence processing by developing a homogeneous self-adaptive RC system based on CMOS-compatible oxides.