Figure 2 shows the RS behaviors of ten cycles in a single IGZO and bi-layer IGZO/ZnO memristors. This result shows the typical bipolar RS characteristics with the bottom electrode grounded in electrical measurements. The memristor devices are initially in HRS state. With a compliance current of 1 mA, a voltage sweep is applied to the top electrode from 0 to 2 V (meaning arrow “1” in the figure), and the RS behaviors of a single IGZO and bi-layer IGZO/ZnO memristors change from the HRS to the LRS. The compliance current is set to prevent the memristor devices from being permanently damaged by a sudden increase in current levels. Once again, the voltage sweep is applied from 2 to 0 V (meaning arrow “2” in the figure) to measure the current and the RS behaviors of a single IGZO and bi-layer IGZO/ZnO memristors remain in the LRS and show a high current level. The arrows “1” and “2” in the figure are called “SET process,” which means that the RS behaviors of a single IGZO and bi-layer IGZO/ZnO memristors change from HRS to LRS. Contrary to the SET process, the process of changing from arrow “3” to “4” in the figure is called “the RESET process,” which means that the RS behaviors of a single IGZO and bi-layer IGZO/ZnO memristor changes from LRS (arrow “3”) to HRS (arrow “4”). These results can benefit the binarized neural networks (BNNs) because the conductance values (inverse of resistance value) are used as synaptic weights. The set voltages of a single IGZO and bi-layer IGZO/ZnO memristors are 1 V and 0.9 V, respectively. In contrast, the reset voltages of a single IGZO and bi-layer IGZO/ZnO memristors are about -1.8 V and -0.7 V, respectively. Therefore, in the case of IGZO/ZnO structure, the memristor device can be used as a low-power application in terms of operating voltage, suggesting that power consumption can be dramatically reduced.
X-ray photoelectron spectroscopy (XPS) measurements were performed to investigate the chemical composition of a single IGZO and bi-layer IGZO/ZnO memristors, to verify the proportions of the oxygen vacancy. Figure 3 shows the XPS analysis result of the O1s spectra in the surface following the deposition of HRS and LRS layers for a single IGZO and bi-layer IGZO/ZnO memristors, respectively, with Gaussian peak fitting. For a single IGZO and bi-layer IGZO/ZnO memristors, the proportions of the oxygen vacancy peak (O1) of HRS layer are 45.2% and 38.2%, respectively, that of LRS layer for both memristors is about 43.4%. As the ratio of Ga and Zn determines the oxygen concentration in HRS and LRS layers during IGZO sputter-deposition, the increase in the ratio of Ga and Zn increase the number of non-oxygen vacancies in the memristor devices, resulting in a lower conductivity layer. The oxygen vacancies modulate HRS layer (oxygen-rich) into LRS layer (oxygen-deficient). The VRESET of bi-layer IGZO/ZnO memristors- is lower than that of a single IGZO, which implied that oxygen vacancy filaments could be easily ruptured due to lower oxygen vacancy peak HRS layer.
To verify the mechanism of the RS behaviors for a single IGZO and bi-layer IGZO/ZnO memristors, the corresponding I–V characteristics of the SET and RESET processes are plotted in Fig. 4. For a single IGZO and bi-layer IGZO/ZnO memristors, a linear fitting slope based on the experimental data is close to 1, which means the linear relationship between current and applied voltage33. The charges originating from the metal electrode interface are thought to be trapped by the empty trap sites of IGZO and ZnO in the HRS layer. As the electric field across memristor devices increases, the steep current for a single IGZO is followed by a quadratic term (I∝V2) with the increase of the injected charges when the conductive filaments form between two electrodes as shown Fig. 4. When the empty trap sites are gradually occupied fully, the slope of the fitting line decreases about 2, indicating that the conduction enters the trap-free space charge limited current (SCLC). It implies that the SCLC is dominant because most injected electrons contribute to the current component34–40.
However, the slope of the fitting line at the high electric field for bi-layer IGZO/ZnO memristor is found to be 4.0, which means that the Schottky mechanism is dominant. The Schottky mechanism in the high electric field may be due to oxygen vacancies close to the metal/metal-oxide interface. The I–V characteristics in the RESET process are dominated by the Schottky and ohmic mechanism, as shown in Fig. 4. After a single IGZO and bi-layer IGZO/ZnO memristors were changed from LRS to HRS; the switching behaviors are controlled by the interface properties due to the Schottky mechanism. It is worth noting that the SCLC is the main conduction mechanism in the SET process for a single IGZO and bi-layer IGZO/ZnO memristors. We conclude the Schottky mechanism observed in a single IGZO and bi-layer IGZO/ZnO memristors at the RESET process is attributed to the interface barrier.
Figure 5 shows the long-term potentiation/depression (LTP/LTD) characteristics with applied positive/negative pulses for an amplitude of 2 V in the two memristors. The positive (2.0 V, 400 ns) or negative (−2 V, 400 ns) voltage pulses with the interval time (4.5 μs) are applied on the memristor devices, and then, the current is measured by a reading voltage pulse (0.2 V, 1 μs) after each pulse. The LTP and LTD characteristics exhibit gradual potentiation and depression characteristics in synaptic weight depending on the input spiking signal, which can be used to evaluate whether memristors can learn or not. When a potentiating input signal train consisting of positive pulses with an amplitude of 2 V is applied on the top Ti metal of the memristor synapse (pre-neuron), the synaptic weight is changed progressively in the increase of the current, which means that the oxygen vacancies are injected into the RS layer, and then this process is formed between TE and BE for potentiation. It can emulate the potential of oxygen vacancies for neuromorphic computing, which enhances the synapse weight by releasing neurotransmitters. In contrast, when a depressing input signal train that consists of negative pulses with an amplitude of -2 V is applied on the top metal, the synaptic weight is depressed progressively, and the conductive path formed by the oxygen vacancies move away from the bottom metal, resulting in the decrease of the current. The nonlinearity for the memristor devices is quantitatively given by equation (1)
where G is the change in the conductance of memristive devices (equivalently, synaptic weight), and GLinear is the linear change in conductance (determining training accuracy41).
Figure 5 shows that the nonlinearity of LTP and LTD characteristics in the bi-layer IGZO/ZnO memristor is 6.77% and 11.49%, respectively, while these for in a single IGZO memristor is 20.03% and 51.1%, respectively42. The linearity and symmetric LTP and LTD characteristics for the bi-layer IGZO/ZnO memristor show more improved linearity than for a single IGZO memristor. Therefore, the high electron conductivity of the ZnO layer in the bi-layer IGZO/ZnO memristor plays an important role in charge carriers to be injected easily under a small set voltage and a reset voltage switching behavior to form the conductive path.