Superlow Power Consumption Artificial Synapses Based on WSe2 Quantum Dots Memristor for Neuromorphic Computing

Abstract

As the emerging member of zero-dimension transition metal dichalcogenide, WSe₂ quantum dots (QDs) have been applied to memristors and exhibited better resistance switching characteristics and miniaturization size. However, low power consumption and high reliability are still challenges for WSe₂ QDs-based memristors as synaptic devices. Here, we demonstrate a high-performance, superlow power consumption memristor device with the structure of Ag/WSe₂ QDs/La_0.3Sr_0.7MnO₃/SrTiO₃. The device displays excellent resistive switching memory behavior with a R _OFF/R _ON ratio of ~5 × 10³, power consumption per switching as low as 0.16 nW, very low set, and reset voltage of ~0.52 V and~ -0.19 V with excellent cycling stability, good reproducibility, and decent data retention capability. The superlow power consumption characteristic of the device is further proved by the method of density functional theory calculation. In addition, the influence of pulse amplitude, duration, and interval was studied to gradually modulating the conductance of the device. The memristor has also been demonstrated to simulate different functions of artificial synapses, such as excitatory postsynaptic current, spike timing-dependent plasticity, long-term potentiation, long-term depression, and paired-pulse facilitation. Importantly, digit recognition ability based on the WSe₂ QDs device is evaluated through a three-layer artificial neural network, and the digit recognition accuracy after 40 times of training can reach up to 94.05%. This study paves a new way for the development of memristor devices with advanced significance for future low power neuromorphic computing.

Figure 1

(a) HR-TEM image of WSe₂ QDs. (b) and (c) are XPS analysis results of WSe₂ QDs: (b) W 4f core spectra and (c) Se 3d core spectra. (d) I-V curves of Ag/WSe₂ QDs/LSMO/STO memristor clearly display resistive switching characteristics. (e) The logarithm form of (d). (f) Comparison of the power consumption of the device with the values of other QDs-based memristors.

Figure 2

(a, b) Distribution histogram and Gaussian fitted curves of set and reset voltage. (c) Statistics of high and low resistance over 100 cycles. The read voltage is 0.2 V. (d) The ratios of R_OFF/R_ON of Ag/WSe₂ QDs/LSMO/STO device. (e) Retention data at HRS and LRS of the device in the room temperature. The read voltage is 0.5 V. (f) The cumulative probability plot of the HRS and LRS. (g) and (h) are the linear fitted curves of LRS and HRS by ln(J) ∝ 1/E, demonstrating the trap-assisted tunneling (TAT) conduction mechanism. (i) and (j) are the fitted curves of LRS and HRS by formulas (1)–(3). (k) The density of states of pristine WSe₂ and the five defect models. The corresponding crystal structures are also shown. The dark green, light green, and gray balls represent upper layer Se atoms, lower layer Se atoms, and W atoms, respectively; “d” represents defect sites.

Figure 3

The conductance of the device for 30 pulse cycles was measured under different pulse (a) amplitudes, (b) durations, and (c) intervals. The device conductance was measured under a train of positive pulses: (d) the pulse duration and interval are both 50 μs and different pulse amplitudes. (e) The pulse amplitude and interval are 4 V and 50 μs, respectively, and different pulse durations. (f) The pulse amplitude and duration are 4 V and 50 μs, respectively, and different pulse intervals.

Figure 4

(a)–(c) Comparison of EPSC measurement response (orange curves) and fitted curves (green curves) under the condition of different pulse numbers. (d)–(f) Comparison of EPSC measurement response (orange curves) and fitted curves (green curves) under different square wave amplitudes. (g)–(i) Comparison of EPSC measurement response (orange curves) and fitted curves (green curves) under different pulse durations. (j)–(l) Comparison of EPSC measurement response (orange curves) and fitted curves (green curves) at different intervals.

Figure 5

Simulation of the characteristics of STDP and PPF in biological synapses. (a) Schematic illustration of the structure of biological synapses. (b) The relationship between the pulse number and the resistance of the device. Applying a negative/positive pulse to the device will cause a decrease/increase in resistance, which represents the modulation of synaptic weight owing to enhancing or suppressing the pulses. (c) Schematic diagram of the pulse waveforms applied to the device for STDP simulation. (d) Measured STDP characteristics of Ag/WSe₂ QDs/LSMO/STO device, the green lines are the curves fitted by Equation (5). (e, f) Measured PPF characteristics of Ag/WSe₂ QDs/LSMO/STO device, (e) and (f) are the test results after applying positive and negative voltage pulses, respectively.

Figure 6

Simulation of neural network based on WSe₂ QDs device. (a) A three-layer neural network structure is shown. It contains input layer, hidden layer and output layer. (b) Image representation of some handwritten digits in the MNIST dataset. (c) The comparison between the recognition accuracy of the optical recognition dataset of handwritten digits in the neural network simulation and the ideal case. After 40 training sessions, the ideal case recognition accuracy reaches 96.71%, and the device-based recognition accuracy reaches 91.59%. (d) Regarding the comparison of the recognition accuracy of the MNIST dataset in the neural network simulation with the ideal case, after 40 training sessions, the ideal case recognition accuracy reaches 98.19%, and the device-based recognition accuracy reaches 94.05%.