Synthesis of composite films for ZnO-based memristors with superior stability

Memristors have unique non-volatile characteristics that potentially can emulate biological synapses for applications in neural computing systems. However, the random formation of conductive filaments in these devices can cause various unreliability problems. In this work, films of a composite of ZnO nanoparticles and carbon nanotubes were prepared as functional layers for memristors by an in-situ growing strategy (ZnO@CNT-IS) using a straightforward high-temperature annealing treatment. This approach allowed for the formation of a high-quality films with uniform loading of ZnO nanoparticles on the carbon nanotubes, which contributed to a lower formation energy for oxygen vacancies and increased electron transfer rate. As a result, the memristors exhibited faster switching response speed, lower power consumption, and a stabilised switching ratio even after 2000 switching cycles. Based on the analog switching behaviour, the ZnO@CNT-IS-based devices showed significant biological synapse functions and plasticity, indicating their potential for high-density storage and neuromorphic computing.


Introduction
Memristors are passive devices with non-linear voltage-current relationships and a resistance that varies with its recent history of applied voltage.They have potential applications for data storage and in neuromorphic computing [1][2][3].Typically they are constructed as simple metal-insulating-metal sandwiches, with small size and low power consumption [4,5].
ZnO films can be prepared by traditional methods such as magnetron sputtering [24], vacuum thermal evaporation [25], sol-gel method [26], chemical vapour deposition [27], pulsed laser deposition [28] and spin coating [29].The latter method is a cost-effective and user-friendly option, offering a broad range of film options.However, the random distribution of ZnO nanoparticles (NPs) can result in unstable switching [30], whilst agglomeration of ZnO NPs can cause large surface roughness which will change the electric field distribution and reduce the uniformity and stability of the switch [31].Furthermore, the zinc interstitial and oxygen vacancy defects in ZnO films can damage switching performance and reduce the reliability of devices [32].
Considerable efforts have been devoted to addressing these particular issues and enhancing performance.For example, Zhao et al devised ZnO nanowires to create one-dimensional conductive paths, which improve carrier transport and result in better switching resistance states [33].Wang et al developed a ZnO NPs/CuO nanowire heterojunction RS layer that demonstrated write-once-read-many-times performance [34].The novel two-dimensional material MXene has also been employed as a carrier for ZnO as an RS layer, enabling excellent switching characteristics that are adjustable in multiple configurations [35].In recent years, carbon nanomaterials, with their impressive excellent electrical, chemical and solution-processable properties, have also been considered for potential applications within ZnO memristors [36][37][38].In table 1 we list some examples of work on ZnO NPs-based memristors from recent years, including device structure, preparation method, and figures for the device performance.
Although the use of carbon nanotubes as a functional layer for memristors has been shown to be a viable option [39][40][41], whether carbon nanotubes can serve as carriers for ZnO NPs to enhance the dominant mechanism for conductivity depends on their ability to create oxygen vacancy conducting path.Therefore their true viability remains unclear.In this study we attempt to address this issue.Specifically we use a special filmforming process to create new memrister devices, and we compare the advantages of this new approach to more traditional spin-coating methods.
In the next section (section 2) we describe the experimental methods, which include the preparation of memristors based on an intermediate layer of ZnO with CNTs prepared using the new in-situ method, together with devices made, for comparison, of pure ZnO and from solutions of ZnO and CNTs (sections 2.2 and 2.3).This is followed by the Results section, which includes characterisation of the device surface (section 3.1) and medium (section 3.2), and studies of the electrical properties (section 3.3), electrical performance (Section 3.4) and analogue switching behaviour (section 3.5) of the memristors.Some supporting results are given in the electronic supporting information (ESI).We argue in the Conclusions (section 4) that the new ZnO@CNT medium prepared by the new in-situ method offers significant advantages in terms of both performance and durability.

Preparation of ZnO@CNT-IS precursor solution
The pristine ZnO precursor solution was prepared by dissolving Zn(NO 3 ) 2 • 6H 2 O (1.08 g) into 18 ml DMF solvent.For the ZnO@CNT-IS precursor solution, 10 ml CNT dispersions were added into the ZnO precursor solution and stirred at room temperature for at least 4 h.For comparison, a ZnO solution was prepared by ultrasonically dispersing ZnO nanoparticles directly into the DMF solution, and a ZnO@CNT solution was prepared by ultrasonically dispersing ZnO nanoparticles directly into the CNT dispersion.

Fabrication of the devices
The device was fabricated with a structure of Al/ZnO@CNT-IS/ITO, where Al was used as the anode, ZnO@CNT-IS as the functional layer, and ITO as the cathode.The ITO-glass substrates were cleaned with acetone, ethanol, and deionised water by ultra-sonication each for 10 min, and then dried at 100 °C for 20 min.The ZnO@CNT-IS precursor solution was spin-coated on the ITO substrate with a low speed of 500 rpm for 6 s, and a high speed of 1500 rpm for 30 s, followed by heating at 50 °C to remove excess organic solvent and to obtain the composited Zn(NO 3 ) 2 • 6H 2 O@CNT film.A subsequent thermal annealing treatment was carried out under dry N 2 atmosphere at 400 °C for 4 h, which allowed the Zn(NO 3 ) 2 to decompose into ZnO through the reaction [50] 2Zn [Finally, an Al electrode with an area of dimensions 200 × 200 μm was deposited via thermal evaporation with shadow masks.The full process is illustrated schematically in figure 1.For comparison, Al/ZnO/ITO and Al/ ZnO@CNT/ITO devices were fabricated by varying the functional layers, which were directly spin-coated using corresponding ZnO and ZnO@CNT solutions respectively.

Thin-film, surface and device characterisation
The structure of the sample was characterised by x-ray diffraction (XRD, XSAM 800) with scattering angles from 10-90°.The surface morphologies of the films were observed by atomic force microscopy (AFM, Park XE7) and scanning electron microscopy (SEM, thermo scientific Apreo 2C).The surface chemical composition and valence bond distribution of the composites were analysed by x-ray photoelectron spectroscopy (XPS, Thermo Fischer, ESCALAB Xi+, USA) with Al Kα x-rays as the excitation source.The electrical performance and synaptic behaviour of Al/ZnO@CNT-IS/ITO devices were measured by a Keithley 2400 source meter.

Characterisation of the device surfaces
Figure 2 shows images of the surfaces of the ZnO@CNT/ITO and ZnO@CNT-IS/ITO devices; corresponding results for the Al/ZnO device are given in the ESI (figure S1).Both films show a degree of roughness, but greatly reduced in the ZnO@CNT-IS/ITO.The AFM analysis gives roughness values measured as 9.25nm in the ZnO@CNT-IS film, and the higher value of 14.35 nm in the ZnO@CNT film.In comparison, the measurements gave a much higher value of 35.40 nm for the roughness of the pure ZnO film.
The surface roughness is likely to arise from ZnO nanoparticles failing to dissolve properly in the DMF, which meant that the ZnO nanoparticles were not uniformly dispersed during the spin coating process.This results in a thin film of lower quality and hence an RS of reduced stability and durability.On the other hand, the ZnO@CNT-IS film, prepared by the in-situ synthesis, gave a highly-soluble precursor solution.This resulted in the formation of a high-quality film, with the ZnO NPs uniformly loaded onto the carbon nanotubes.The ZnO@CNT-IS film rarely showed the same aggregation of the ZnO NPs as found on the ZnO@CNT films, as illustrated in the contrast between figures 2(b) and (d).Overall, the XRD and XPS data strongly support the conclusion that the ZnO was successfully loaded onto the CNT, and that the CNTs can provide oxygen vacancy transport paths which may enhance switching response and provide more oxygen vacancy transport paths.The I-V curve of the Al/ZnO@CNT-IS/ITO memristor was measured by applying a direct current (DC) sweeping voltage of 0 V → +5 V → 0 V → −5 V → 0 V with a 1 mA compliance current and a scan rate of 0.1 V s −1 .A sample result is shown in figure 4(b).During process 1 in the SET stage, the device was in the highresistance state (HRS).As the positive scan voltage increased from 0 V to +5 V, the conductive filament was formed, and the device switched into the low resistance state (LRS).In process 2, the positive voltage was then decreased from +5 V to 0 V, with the device remaining in the LRS.In process 3 during the RESET stage, as the negative voltage changed from 0 V to −5 V, the conductive filament was broken, and the device switched into the HRS.During process 4, as the negative voltage was reduced from −5 V to 0 V, the device remained in the HRS.The device shows the bipolar resistive switching behaviour, and a gradual change in resistance was observed during both the SET (voltage change 0 V → +5 V) and RESET (voltage change 0 V → −5 V) stages.This indicates that the conductance of the device can be continuously regulated at either positive or negative voltages, and that it exhibits good non-volatility.

Electrical properties of the Al/ZnO@CNT-IS/ITO memristor
To better understand the RS mechanism of the Al/ZnO@CNT-IS/ITO memristor, we analysed the carrier transport mechanism in both the OFF and ON states by plotting I log -V log curves across the entire sweep region.Figure 4(c) shows the logarithmic I -V curve of the Al/ZnO@CNT-IS/ITO device in the positive sweeping region with some values of the gradient k At a low positive voltage, the size of the slope is approximately 1 (e.g.1.10, 1.20), indicating that I ∝ V, characteristic to the normal Ohmic behaviour.As the voltage increases, the size of the slope increases approximately to 2 (e.g.2.06, 2.25), passing though the regime I ∝ V 2 as described by the Mott-Gurney law [52], and then to higher gradients (2.76, approximately 3), I ∝ V n , as described by the Mark-Helfrich equation [53].This shows that the carrier transport in the HRS follows the space charge limited conduction (SCLC) models.Additionally, as the positive voltage decreased, the slope decreased again, indicating that the conduction in the LRS changes back towards Ohmic behaviour.The same conduction mechanism is observed throughout the negative sweeping region in figure 4(d), indicating a consistent trend.
Figure 5 presents our interpretation of the mechanism for the formation of the conductive filaments for Al/ ZnO@CNT-IS/ITO devices.Upon applying a positive voltage, we observed the accumulation of oxygen vacancies at the negative electrode.With increasing voltage, the conductance filaments quickly emerged through adsorption of oxygen atoms on the ZnO@CNT surface.Carbon vacancies play a key role in capturing oxygen atoms, thereby reducing the formation energy of oxygen vacancies, promoting uniformity in the conduction path, and increasing the number of conductance filaments [54,55].As the conductance filaments are formed, the device transforms from HRS to LRS.Conversely, a negative voltage leads to the depletion of oxygen vacancies and the breakdown of the conductance filaments, returning the device to the HRS.The presence of the CNTs was found to enhance the uniformity and stability of the conductance filaments, resulting in bipolar resistive switching behaviour and a reversible conductive path.

Electrical performance of the Al/ZnO@CNT-IS/ITO memristor device
Here we analyse the effect of our new in-situ synthesis on the electrical performance of devices that can be made with our memristor.We subjected the devices to consecutive voltage pulses of ±5 V, switching between the ON  and OFF states as shown in figure 4. Figures 6(a)-(c) show the cumulative probability of the three different devices when in both states, namely for the Al/ZnO@CNT-IS/ITO, Al/ZnO@CNT/ITO and Al/ZnO/ITO devices respectively, after subjected to many repeat cycles.In the case of Al/ZnO@CNT-IS/ITO there appears be no overlap of the distributions of resistance values in the HRS (OFF) and LRS (ON), but in the other two cases there is overlap between the distributions for these two states.This separation of the resistance of the two states shows one clear advantage of the Al/ZnO@CNT-IS/ITO device.
In figures 6(d)-(f) we show the cumulative probability of the three different devices when in the ON state for different read voltages.Again, we see no overlap of the LRS state in the Al/ZnO@CNT-IS/ITO device under different read voltages (figure 6(d)), but in the other two devices, we do see the overlap as in figures 6(e) and (f).
Our results show that the Al/ZnO@CNT-IS/ITO devices show enhanced behaviour compared to the Al/ ZnO@CNT/ITO and Al/ZnO/ITO devices.Our explanation of this is that in the in-situ synthesis has avoided an uneven dispersion of the ultrasonically-dispersed ZnO nanoparticles.In the Al/ZnO@CNT/ITO and Al/ ZnO/ITO devices the uneven dispersion of ZnO nanoparticles led to fewer conductance filaments with a more random distribution.In Al/ZnO@CNT/ITO we observed an overlap of HRS and LRS, resistance dispersion, and an unstable distribution of the resistance state, due to the agglomeration of ZnO as a result of non-uniform loading onto the CNT.On the other hand, the introduction of CNT through the in-situ synthesis strategy in the Al/ZnO@CNT-IS/ITO device resulted in stable HRS and LRS throughout 2000 cycles, showing a concentrated distribution of HRS and LRS and superior endurance properties.Additionally, the Al/ZnO@CNT-IS/ITO device exhibited a stable non-volatile nature, with no significant fluctuation in the time retention property over 8000 s.In contrast, the HRS and LRS of Al/ZnO/ITO and Al/ZnO@CNT/ITO devices gradually began to fluctuate at around 100 s and broke down at roughly 300 s, indicating poor time retention property.
Figures 6(g)-(i) show the results of a test of the long-term resilience of the three devices.We switched the devices into ON and OFF states, and we used a reading voltage of 0.1 V to record the resistance for 8000 seconds.The Al/ZnO@CNT-IS/ITO device exhibited a stable non-volatile nature, with no significant fluctuation in the time retention property over the period of the test.On the other hand, the Al/ZnO/ITO and Al/ZnO@CNT/ Figure 6.Top row, cumulative probability of (a) Al/ZnO@CNT-IS/ITO, (b) Al/ZnO@CNT/ITO and (c) Al/ZnO/ITO devices obtained by measuring with a constant cycling of voltage pulses of ± 5 V, showing distributions in both the ON (low-resistance state) and OFF (high-resistance state) states.Middle row, cumulative probability of (d) Al/ZnO@CNT-IS/ITO, (e) Al/ZnO@CNT/ITO and (f) Al/ZnO/ITO devices obtained by measuring with different voltage levels in the ON state.Bottom row, The variation of the resistance of the two states of (g) Al/ZnO@CNT-IS/ITO, (h) Al/ZnO@CNT/ITO, and (i) Al/ZnO/ITO devices with time.
ITO devices gradually began to fluctuate at around 100 s and broke down at roughly 300 s, indicating poor time retention property.
One important aspect of device reliability is the delay time for pulse response.To measure the delay time for Al/ZnO/ITO, Al/ZnO@CNT/ITO, and Al/ZnO@CNT-IS/ITO devices, a series of pulse mode tests were conducted.A pulse voltage of 0.1 V and a width of 3 s was applied to the device, and the time for the current response to switch from a HRS to a LRS was measured.The Al/ZnO/ITO device exhibited a longer delay time of 10.2 ms shown as in figure 7(a), whereas the Al/ZnO@CNT/ITO (Figure 7(b)) and Al/ZnO@CNT-IS/ITO (figure 7(c)) devices have the lower response times of 2.3 ms and 2.1 ms, respectively.A reduced response time will lead to a reduced power consumption.These improved response times were attributed to the presence of πelectrons in the CNT, which enhance both the electron transfer rate and oxygen vacancy formation.Furthermore, the response time is on the millisecond time scale, which is somewhat different from the switching behaviour that usually occurs on the time scales of picosecond and nanosecond [56].This may be related to parasitic capacitance effects, since larger parasitic capacitance can lead to longer response times [57].Comparing the response times of the Al/ZnO@CNT/ITO and Al/ZnO@CNT-IS/ITO devices it can be seen that the quality of film formation in the Al/ZnO@CNT-IS/ITO device also contributes to the speed of response.
We attribute the better long-term reliability to the introduction of CNT through in-situ synthesis method can capture more oxygen atoms, and improve uniformity in the path and the number of conductance filaments, which results in a stable resistance state and long-time endurance properties.Furthermore, the Al/ZnO@CNT-IS/ITO device maintained superior stability at different compliance currents compared to Al/ZnO@CNT/ITO and Al/ZnO/ITO devices.

Analogue switching behaviour of Al/ZnO@CNT-IS/ITO device
The analog switching behaviour of Al/ZnO@CNT-IS/ITO devices is demonstrated in figures 8(a) and (b) for various cycles.The deviceʼs conductance gradually increases or decreases with consecutive positive or negative voltage changes, respectively.
Subjecting a device to 10 consecutive +1 V/ −1 V pulses, figure 8(c), resulted in a gradual increase or decrease of its normalised conductance with successive positive or negative pulses.Figure S5 shows the data before normalisation, where the ON/OFF ratio, obtained from minimum and maximum conductance values, was approximately 8. When continuously modulated by a stream of pulses, the device exhibits nonlinear transmission properties, indicating its potential to simulate synaptic behaviours, and this is what we investigate in this final section.
Short-term synaptic plasticity was simulated using paired-pulse facilitation (PPF) [58,59].The principle of the PPF test is that for a forward pulse voltage, two consecutive voltage pulses are applied.The current response of the second one, I 2 , will be higher than the first one, I 1 , and the effect will be weakened with increasing time between the two pulses.For negative pulse voltage, when applying two consecutive pulses, the current response of the second pulse will be lower than the first one, and the effect will be weakened as the time between the two pulses increases.We used a voltage amplitude of 1 V with a time separation of 100 ms.The excitatory postsynaptic current (EPSC) produced by the second pulse is significantly greater than that of the first pulse [60], and the enhancement effect gradually decreases with increasing interval time between the two pulses.
The PPF is defined in the measurements as assume that it can be fitted by a function of form where A 1 and A 2 are the conductance values after the first and second pulses, respectively, and τ 1 and τ 2 are the fast and slow decay times.As shown in figures 8(d) and (e), as the time interval Δt increased, it became evident that the PPF exhibited an exponential-like decay, culminating in the PPF index reaching near-zero at an interval of approximately 800 ms.This observation strongly indicates that the device possesses the ability to simulate short-term memory.
Based on the analog switching behaviour of the artificial synapse using on the Al/ZnO@CNT-IS/ITO device, we performed measurements designed to simulate long-term synaptic plasticity [61].Figure 8(f) shows the results from the simulation of LTP and LTD [62,63].Initially, the device was subjected to 300 consecutive positive pulses with an amplitude of 0.1 V and a width of 10 ms.This was followed by 300 consecutive negative pulses with the same amplitude and width.As the number of positive pulses increased, the deviceʼs conductance also increased gradually until it reached a constant level, as shown in figure 8(f).Conversely, as more negative pulses were received, the conductance gradually decreased.This progressive change in conductance shows that the device is capable of producing both long-term potentiation (LTP) and long-term depression behaviour (LTD), and analog long-term synaptic plasticity.

Conclusions
In this work, a ZnO@CNT-IS-based memristor was successfully prepared by a simple in-situ synthesis method.Compared with the Al/ZnO/ITO and Al/ZnO@CNT/ITO devices, the Al/ZnO@CNT-IS/ITO device exhibits a more concentrated threshold voltage distribution, faster switching speed (2.1 ms), longer retention time (8000 s), and higher switching endurance (2000 cycles).This is because the introduction of the CNT increases the number of carbon vacancies, which can trap oxygen atoms and provide more oxygen vacancies, thereby increasing the number of conductance filaments formed and improving the uniformity of the conductance filaments.At the same time, the π-electrons in the carbon nanoparticles accelerate the transformation of the electrons, thus increasing the formation rate of the oxygen vacancies.Since the ZnO@CNT-IS single-device structure exhibits obvious non-volatile emulating characteristics, it can emulate biological synapses such as EPSC/IPSC, PPF, and LTP/LTD, which provides a basis for application in a number of research fields such as neural networks and artificial intelligence.

3. 2 .
Characterisation of the ZnO@CNT medium X-ray diffraction patterns of a ZnO@CNT-IS film is shown in figure3(a).It appears that the standard zinc blende crystalline phase of ZnO is formed within the film after the thermal decomposition of Zn(NO 3 ) 2 • 6H 2 O.The diffraction pattern in figure 3(a) is compared with that given by the standard reference phase of zine blende from the International Centre for Diffraction Data, ICDD#01-089-7102.The XPS data, figures 3(b)-(d), show that the ZnO was loaded onto the carbon nanotubes, and give information concerning the elemental composition and the chemical valence state of ZnO@CNT-IS.The survey spectrum shown in figure S4 of the ESI clearly shows the presence of C, O, and Zn elements.The peaks at 1022.3 and 1045.4 eV are associated with the Zn 2p3/2 and Zn 2p1/2 states respectively (figure 3(b)), indicating that zinc is present in the form of Zn 2+ [51].The C 1s XPS spectrum in figure 3(c) shows not only the peaks for the C=C bond (284.7 eV) and C-C bond (285.5 eV), but also for C=O bonds (286.6 eV) and O-C=O bonds (288.7 eV), indicating that carbon has the potential to capture oxygen vacancies from the ZnO.Additionally, the O 1s XPS spectrum displays peaks for the Zn-O bond (531.3 eV), C-O bond (532.3 eV) and C=O bond (532.7 eV).The characteristic peak at 531.3eV corresponds to the oxygen ion, O 2− , within the ZnO matrix.

Figure 4 (
a) presents a schematic diagram of the measurements of the memristor.The device has Al electrodes on the top surface, an intermediate ZnO@CNT-IS composite layer, and a bottom electrode made of ITO.A Keithley-2400 source meter was used to measure the electrical performance of the device.

Figure 2 .
Figure 2. (a) , (b) AFM (a) and SEM (b) images of the ZnO@CNT film.(c) , (d) AFM (c) and SEM (d) images of the ZnO@CNT-IS film.The AFM images shows a lower degree of surface roughness in the ZnO@CNT-IS film.

Figure 4 .
Figure 4. (a) Schematic diagram of the structure of the Al/ZnO@CNT-IS/ITO device; (b) I-V curves of the Al/ZnO@CNT-IS/ITO device obtained with a scan rate of 0.1 V s −1 ; (c,d) I V log log | | | | -[ curves of positive (c) and negative (d) applied voltages.The numbers at different points on the curves indicate the values of the gradient k I V dlog dlog | | | | = .

Figure 5 .
Figure 5. Schematic diagram of the SET/RESET processes.

Table 1 .
Examples of previous work.