Dynamic memristor array with multiple reservoir states for training efficient neuromorphic computing

Abstract

In this study, we evaluated the performance of a Pt/Al/TiO_y/TiO_x/Al₂O₃/Pt RRAM array device in synaptic and reservoir computing applications. The device exhibited excellent switching characteristics and consistent set processes, along with verifying 100 cycles of DC endurance and cell-to-cell properties. Furthermore, over 10⁴ retention time, the device displayed gradual current decay leading back to its initial high-resistance state, revealing the presence of short-term memory characteristics. Additionally, by leveraging potentiation and depression, paired-pulse facilitation, spike-number-dependent plasticity, spike-amplitude-dependent plasticity, spike-rate-dependent plasticity, and Pavlovian conditioning, we replicated the mechanisms of the biological brain in terms of both short- and long-term memory within our memristor array technology. We also implemented a 4-bit reservoir computing system by leveraging the nonlinear dynamics of the device, adding to its computer-favorable applications. Finally, through analyzing the temporal changes based on a stimulus frequency in a 5 × 5 synaptic arr ay image training process, we concluded that the Pt/Al/TiO_y/TiO_x/Al₂O₃/Pt device is suitable for application in neuromorphic systems.