Enhancing Computing Capacity via Reconfigurable MoS₂-Based Artificial Synapse with Dual Feature Strategy for Wide Reservoir Computing

Reservoir computing (RC) has garnered considerable interest owing to its uncomplicated network structure and minimal training costs. Nevertheless, the computing capacity of RC systems is limited by the material-dependent physical dynamics of reservoir devices. In this study, an efficient neuromorphic reservoir device with adjustable reservoir states, achieved through the development of an electrically tunable three-terminal charge trap memory, is introduced. This device utilizes molybdenum disulfide (MoS₂) as the channel material and a perhydropolysilazane-based charge trap layer. Notably, the absence of a tunneling layer in the device structure enables dynamic resistive switching, characterized by outstanding endurance and an excellent memory window. Furthermore, by implementing a simple input decay and refresh scheme, a reconfigurable neuromorphic device capable of multiple feature extraction and functioning as an artificial synapse is developed. The device's efficacy is validated through device-to-system-level simulations within a hardware-based wide RC (WRC) system, resulting in an improved recognition rate in the MNIST hand-written digit recognition task from 87.6% to 91.0%, a testament to the enhanced computing capacity. This strategic approach advances the development of hardware-based WRC systems, marking a significant step toward energy-efficient reservoir computing.