Memristors with nonstoichiometric tungsten oxide (WOx) as an active layer, derived from the oxidation of atomically thin two-dimensional tungsten diselenide (WSe2), enable the creation of the monolithic layered structure of WOx/WSe2. These devices are promising candidates for emulating various biological synaptic functions in the human brain. In this study, we fabricate monolithic few-layer WOx/WSe2 memristors with precisely controlled WOx thickness by UV-ozone treatment from 1L to 9L, depending on chuck temperature. The postsynaptic responses of the topmost single-layer (1L) oxidized WSe2 and fully (9L) oxidized WSe2 memristors exhibit sharply contrasting behaviors, which can be applied to mimic the heterosynaptic plasticity in the CA1 region of the hippocampus. Beyond the significance of emulating the biological synaptic characteristics, we explore the feasibility of using each oxidation-layer-controlled memristor as a hardware accelerator. Their performances are assessed through application in a CIFAR-10 pattern recognition task using a convolutional neural network. Pattern recognition rates of 84% and 71% are obtained for the 1L and 9L WOx-based devices, respectively. We also examine the applicability of a synaptic cell composed of devices with oppositely switched characteristics. Consequently, the synaptic weight—defined as the difference in conductance between two synaptic devices—can be either increased (potentiated) or decreased (depressed) by simultaneously updating both devices with the same voltage signal. This weight update concept achieves a moderate recognition rate of 85.94% when using an MNIST pattern-based recognition task, simplifying the complex weight-adjustment process.
Read full abstract