Enhanced Ferroelectricity of Hf-Based Memcapacitors by Adopting Ti Insert-Layer and C–V Measurement for Constructing Energy-Efficient Reservoir Computing Network

Abstract

Hf-based ferroelectric memcapacitors only consume dynamic power with the merits of reliable nonvolatile storage and Si-process compatibility, which is an outstanding artificial synapse for constructing energy-efficient neuromorphic computing networks. In this paper, the ferroelectricity of Hf_0.5Zr_0.5O₂ (HZO) memcapacitor is improved by the co-optimization of process design and electrical measurement with various thicknesses of the Ti insertion layer and conditions of Capacitor–Voltage (C–V) tests. Material characterization indicates the Ti insertion layer reduces the m-phase and increases the ratio of the o-phase in HZO film. The wake-up-free behaviors are achieved in the Ti insertion layer memcapacitors with an endurance of ≈10⁹ cycles. Furthermore, ferroelectric properties are further enhanced after C–V measurement with the 1nm-thick Ti insertion layer showing the largest remanent polarization (2Pr≈41.02 µC cm⁻²). Subsequently, a full hardware-implemented hierarchical parallel reservoir computing (RC) network is constructed using 34 HZO memcapacitive synapses. The proposed network achieves high recognition accuracy (≈96.10%) and low dynamic power consumption (≈0.15 fJ per input) with the MNIST dataset. These findings indicate the feasibility of developing a highly energy-efficient, fully hardware-implemented, hierarchical parallel RC neural network.