Polarity-Resolved and Nonlinear HfZrO x Memcapacitive Dynamics for Mask-Free Reservoir Computing

Abstract Physical reservoir computing (RC) seeks to harness intrinsic nonequilibrium device dynamics, but practical implementations often still rely on external feature extraction and random masking. Here, we exploit the polarity-resolved and nonlinear voltage-state dynamics of a solution-processed HfZrOx (HZO) memcapacitor as distinct computational functions. The polarity-dependent response maps positive and negative temporal variations of electromyography signals into separated reservoir states, enabling bipolar differential RC without conventional time- or frequency-domain feature extraction and achieving 91.67% gesture-recognition accuracy. The experimentally calibrated nonlinear voltage-state response is further reused layer by layer to diversify reservoir states without externally generated random masks, increasing the accuracy from 94.17% to 98.75%. Because both input and output are represented in the voltage domain, the device is also compatible with direct interlayer signal transfer without current-to-voltage conversion. These results show how intrinsic memcapacitive dynamics can replace externally imposed processing steps in physical RC.