High-performance in domain matching epitaxial La:HfO2 film memristor for spiking neural network system application

Abstract

Next-generation synaptic devices with multiple non-volatile states, high endurance and high-temperature operation are highly desired in the era of big data. Here, high-performance memristors are fabricated using La: HfO₂(HLO)/La_2/3Sr_1/3MnO₃(LSMO) heterostructures on Si substrate, with domain matching epitaxial structure using SrTiO₃(STO) as buffer layer. The devices possess high reliability, nonvolatility, low fluctuation rate (<2.5 %) and the highest number of states per cell (32 states or 5 bits) among the reported Hf-based ferroelectric memories at room temperature (25 °C) and high temperature (85 °C). Moreover, the device exhibits high endurance of 10⁹ cycles and excellent uniformity at the room and high temperatures. The functionality of long-term plasticity in the synaptic device is obtained with high precision (128 states), reproducibility (cycle-to-cycle variation, ∼4.7 %) and linearity. Then, we simulate one system using the stable performance at high temperature that detects the speed of moving targets, which achieves high accuracy of 98 % and 99 % on Human Motion and MNIST datasets, respectively. Furthermore, we have built a hardware circuit to realize a spiking neural network (SNN) system for digital pattern online learning, which demonstrates the capability of the device in brain-like computing applications.