A Flexible Complementary Memristive Synapse for Convolutional Neural Network Applications

Abstract The memristive artificial synapse, which typically operates on the basis of various filamentary conductance-update mechanisms, has been extensively utilized to implement hardware neural networks for brain-inspired computing. However, the stochastic conductance-switching characteristics that arise from uncontrollable filamentary evolution within the two-terminal device have long hindered hardware neural networks from achieving training and recognition performance comparable to their software neural network. Here, we propose a flexible complementary memristive synapse that achieves highly linear conductance-update behavior while preserving a large dynamic range. This synapse consists of a homogeneous bilayer of two amorphous indium–gallium-tin oxide (a-IGTO) films deposited under different oxygen partial pressures: an oxygen-deficient, VO-rich layer formed without O2 supply and an oxygen-supplied, VO-poor layer formed with O2 supply, exhibiting incremental and decremental conductance-update behaviors, respectively. The distinct oxygen-vacancy-related states and filament-evolution characteristics of the two layers suggest vacancy-mediated layer-to-layer redistribution, which can mutually compensate the nonideal conductive filament evolution of the two different layers, thereby facilitating more gradual modulation of the net conductance. The proposed flexible synapse exhibits robust mechanical stability and durability after 12,000 bending–release cycles, as confirmed by electrical characterization while maintained in the bent state at a bending radius of 2 mm. Finally, the applicability of the proposed flexible synapses to HW-NNs is evaluated through a VGG-8-based CIFAR-10 image recognition simulation using experimentally measured synaptic characteristics for weight updates in the two fully connected layers.