SrTiO3-based memristor with metal/oxide bilayer electrode for high recognition accuracy neuromorphic computing

Abstract

Memristors can modulate conductance and have multiple levels of storage, which has received great attention in the field of artificial synapses. Due to the disadvantages of low switching ratio, low stability and so on with single metal electrode memristor. Metal/oxide bilayer electrode memristor of Pt/La0.7Sr0.3MnO3/SrTiO3/Nb: SrTiO3 device (Pt/LSMO/STO/NSTO) by inserting a transition metal oxide electrode LSMO between metal electrode Pt and resistance material STO. The oxygen vacancies in LSMO layer can reduce the barrier height (Φ) and the barrier width (Wd) of STO/NSTO interface, resulting in larger on/off ratio (1.2 × 105), smaller Vset (0.58 V) and higher stability (0.124/0.18) than metal single electrode memristor without LSMO (on/off ratio = 9 × 103, Vset = 0.9 V, σ/µ = 0.23/0.25). In addition, it effectively simulates the features of artificial synapses and accomplishes the function of both D-latch and decimal logic neuron computation. In particular, the convolutional neural network based on metal/oxide bilayer memristor realizes the high-precision recognition of traffic signals, demonstrating high recognition rates of 95.4% for the traffic dataset, and the recognition accuracy still remains above 80% even in 10% Gaussian noise.