Self-Selective Crossbar Synapse Array with n-ZnO/p-NiO_x/n-ZnO Structure for Neuromorphic Computing

Artificial synapse devices are essential elements for highly energy-efficient neuromorphic computing, which can overcome the current bottle neck issue of the von Neuman computing system.^1,2 They are implemented as crossbar array architecture, where highly selective synaptic weight update for training and sneak leakage-free inference operations are required. In this study, self-selective bipolar artificial synapse device was proposed with n-ZnO/p-NiO_x/n-ZnO heterojunction and its analog synapse operation with high selectivity was demonstrated in crossbar array architecture without the aid of selector devices.<br/><br/>In unit device, the dynamic range of weight update reached 17.6 times of initial value and the updated memory retained about 34.6 % after 30 min, exhibiting both short-term plasticity and long-term plasticity. The nonlinearity factor in half-voltage operation scheme, which is the ratio of current at full voltage to half voltage, was 19.4 and the dynamic range of the full-biased cell was 27 times wider than that of the half-biased cell.<br/><br/>In 32×32 crossbar array structure, with 10 µm line width, the device showed analog weight update with dynamic range of more than 100 times of the initial state. The direct measurements of weight update in the crossbar array exhibited a high conductance change reaching about 14.4 times in the full-biased selected cell, while it was less than 1.8 times in the unselected cells (non-biased and half-biased). The pattern recognition accuracy simulated with the obtained multiple cycles of synaptic weight update profile using CrossSim ver. 2.0 Training exhibited about 89.0 % of accuracy within 20 epochs of training.<br/><br/>The high nonlinearity of the device was from the additional current flow by Zener tunneling effect. In the low field regime, the diffusion current overcoming the p-n junction barrier dominated. However, when strong field was applied, Zener tunneling current flowed additionally from the valence band of the p-NiO_x to the conduction band of the n-ZnO due to the bending of the band in the reverse bias depletion region, which provided highly nonlinear current-voltage characteristics at both voltage polarities for self-selecting function.<br/><br/>Analog synaptic weight update of the device originated from voltage-driven redistribution of oxygen ions inside n-p-n oxide structure. Application of positive bias at top electrode attracted oxygen ions to the upper region of ZnO and NiO_x and bent the energy band accordingly. It lowered the potential barrier height at the bottom p-NiO_x/n-ZnO junction, which increased the current in the low field regime by increasing the diffusion current. In addition, the increased band offset at the upper n-ZnO/p-NiO_x junction, as a result of oxygen redistribution, increased the current in the high field by lowering the Zener breakdown turn-on voltage. Therefore, the increased conductance as potentiation behavior could be achieved by applying positive bias at top electrode. Inversely, application of negative bias at top electrode repelled oxygen ions to the bottom region of the n-ZnO and p-NiO_x lowering the conductance back to the initial state as depression behavior.<br/><br/><b>References</b><br/>[1] N.K. Upadhyay et al., <i>Advanced Materials Technology</i>, <b>4</b>(4), 1800589, 2019<br/>[2] X. Duan et al., <i>Advanced Materials</i>, <b>36</b>(14), 2310704, 2024