Physical Reservoir Computing Of Single-walled Carbon Nanotubes Network Complexed With Polyoxometalate Acid Molecules

With the rapid development of 5G,AI,and Io T,the demand for high-performance realtime computing continues to grow.As a result,the power consumption of silicon-based hardware increases exponentially,arousing more and more attention to neuromorphic hardware without silicon.Reservoir computing(RC)is a kind of computational framework with low training cost and fast learning.Studies have shown that the single-walled carbon nanotube(SWCNT)network can be used as a physical reservoir in RC.In this paper,the physical reservoir computing of the SWCNT network complexed with polyoxometalate(POM)was systematically studied especially for its implementation and capability by using the lab-built experimental platform.First,an experimental platform based on the NI data acquisition board and MATLAB was built to configure and evaluate the physical reservoir with its status information extracted and provided support for the RC research.Next,by using PBMA dielectric layer,the POM/SWCNT network was fabricated on the PCB substrate with a 10×10 microelectrode array.Then,a physical RC framework to implement NARMA10 task was set by the experimental platform with an input channel and 80 readout nodes using the ridge regression algorithm.The experimental results showed that the normalized root-mean-square error(NRMSE)of offline training was reduced from 0.21 to 0.019,and the short-term memory capacity(MC)increased from 6 to 20.5.By a three-dimensional behavioral space consisting of kernel rank(KR),generalization rank(GR)and memory capacity,the quality of SWCNT network was evaluated with coverage of behavioral space significantly increased by the decoration of POM.Finally,a cellular automata(CA)model of the POM/SWCNT network was used to simulate spiking and RC performance.The results showed that the RC performance of the POM/SWCNT network can be improved by increasing the input signal gain and the number of sampling microelectrode in the substrate.In addition,a method to modulate the internal dynamic and RC performance of the POM/SWCNT network by redox gas molecules was proposed.This paper showed that the RC performance of the SWCNT network was improved by POM decoration,providing fundamentals for the further application of the POM/SWCNT network in physical reservoir computing.