Analog HfxZr1‐xO2 Memristors with Tunable Linearity for Implementation in a Self‐Organizing Map Neural Network

Doped‐metal oxide‐based memristors, with the potential for improved switching performance and capability for multi‐bit information storage, are attractive candidates in the implementation of artificial neural network (ANN) hardware systems. However, performance and process considerations such as switching behavior and complementary‐metal‐oxide‐semiconductor (CMOS) process compatibility remain a challenge. This study shows that amorphous Zr‐doped HfO2 (HZO) memristors fabricated via a co‐sputtering approach improve the switching performance by providing a controllable knob to modulate defects in the switching layer. At the same time, it satisfies the CMOS process compatibility requirements for industry adoption. HZO memristors with optimized stoichiometry exhibit 30% reduced switching voltages and 50% faster switching as compared to control HfO2 memristors. Concurrently, this study shows that high linearity analog states tuning is achievable via a programming scheme that utilizes voltage pulses with increasing amplitudes. This study further shows via simulation evaluation that HZO memristors implemented in a self‐organizing‐map (SOM) network for Fashion MNIST database classification, achieve an accuracy of 92% with short training cycles. The results thus pave a potential pathway for further development of CMOS process compatible HZO memristors for use in future storage and computing applications.


Introduction
In order to facilitate more extensive implementation of next generation computing-in-memory systems, oxidebased memristors are widely considered to be one of the most promising building blocks.They satisfy the key performance requirements including the ability for analog switching behavior, fast switching, non-volatility, and low power consumption.These are on top of having a simple device structure and low processing temperature to realize high-density memory array integration via stacking multiple memory layers at the complementarymetal-oxide-semiconductor back-endof-line (CMOS BEOL) process. [1]ono-oxides memristors have been widely reported with characteristics suitable for implementation as analog memory cells in artificial neural networks (ANN).Despite the promises, state of the art mono-oxide memristors still falls short in areas such as poor linearity modulation and high operating voltages. [2]A potential avenue for further performance optimization can be via a controlled approach in introducing new elements to form binary or higher-order oxide memristors.In recent years, the growth of binary or multi-element metal oxides can easily be achieved in many advanced manufacturing technologies.Through process including chemical doping or controlled defects concentration (in the form of oxygen vacancies) introduction, the resistive switching (RS) layer of memristors, with clear indications of performance improvement has been demonstrated.Several works on doped-memristors, including Al/HfO 2 /Cu, [3] Au doped HfO 2 , [4] Pt-dispersed SiO 2 , [5] have been reported with performance.However, the main challenge to these efforts is that the materials used are not CMOS process compatible.The search for compatible doping elements while providing obvious performance advantage remains an area of research interest.In this aspect, the doping of Zr in a conventionally used HfO 2 switching layer in the form of Hf x Zr 1-x O 2 (HZO) ternary oxide could provide a feasible solution.First, both HfO 2 and ZrO 2 are high-k materials that play an important role in the size scaling of CMOS technology, and thus highly CMOS process compatible.Second, HZO thin-film is known to exhibit different memory properties such as either ferroelectric or memristive when presented in either crystallized or amorphous phase.1c,6] Process engineered HZO memristors with tunable linearity suitable for ANN implementation, enabled by both material and operation co-optimization are reported in this work.There have been several reports on HZO memristors with some improved performances, but still lack systematic studies and evaluation of potentials for hardware implementation.Furthermore, the reported memristors fabrication process involves hightemperature conditions, thus not satisfying the requirements for CMOS BEOL integration. [7]To circumvent that, the HZO memristors studied in this work are prepared by co-sputtering ZrO 2 and HfO 2 targets at room temperature.This provides the capability to both control the thickness and modulate the Zr composition in the HZO film.At the same time, the entire fabrication process is extremely friendly for CMOS BEOL integration.In our study, the best performing HZO memristor is obtained with Zr at% = 63%, with low operating voltages below 0.85 V, good uniformity with the coefficient of variation of 0.605/0.398 in high resistance state/low resistance state (HRS/LRS), retention exceeding 5 × 10 4 s under 85 °C, fast switching speed down to 40 ns, and low power consumption of 60 pJ.As an analog memristor, the HZO memristor also exhibits excellent analog state tunability using a programming scheme that utilizes increasing voltage pulse amplitude.The potential of the HZO memristors to be used as computation memory cells in a self-organizing map (SOM) ANN for Fashion MNIST classification is evaluated via simulations, achieving the accuracy of 92% after 20 epochs of training cycles.These results provide a potential pathway for further development of HZO memristors for use in future storage and computing applications.

Results and Discussion
The HZO and control memristors were fabricated with typical cross-point structures with the process flow shown in Figure 1a.The high-resolution transmission electron microscopic (HRTEM) image (Figure 1b) shows the amorphous nature of the HZO layer, while together with the high-angle angular dark field-scanning transmission electron microscopic (HAADF STEM) image (Figure S1, Supporting Information) further confirms the thickness of the deposited layers in the memristors stack.Scanning electron microscope (SEM) image of the fabricated devices with zoom-in image of the memristor in Figure S2 (Supporting Information) show the cross-point structure of memristors from the top view.The memristors have both the bottom electrode (BE) and top electrode (TE) comprised of Ti/Pt metals with the thicknesses fixed at 5/20 nm, which were deposited via the e-beam evaporation process.The resistive switching (RS) layers for the mono-oxide memristor were obtained via a standard single-sputtering approach, while those of the HZO memristors were obtained via a co-sputtering approach using stoichiometric HfO 2 and ZrO 2 targets.The different stoichiometries of the HZO RS layers were obtained by varying the individual sputter powers of the HfO 2 and ZrO 2 targets.Three different stoichiometries of Hf x Zr 1-x O 2 thin films with x value of 0.37 (HZO-1)/ 0.66 (HZO-2)/1 (HfO 2 ) were fabricated, with those atomic ratios confirmed via X-Ray photoelectron spectroscopy (XPS) in Figure S3 (Supporting Information).The respective deposition conditions used for are indicated in supplementary Table S1 (Supporting Information).All the switching layers were controlled to a nominal thickness of ≈5 nm, confirmed using an ellipsometer.Details of the fabrication process are provided in the Experimental Section.
Material characterizations were first performed to elucidate the physical properties of the RS layer.It is well-known that polycrystalline HZO crystallized in the orthorhombic phase exhibits ferroelectric properties.However, in our experiments, no crystallization process step was performed on the HZO RS layer, and hence HZO RS layer should remain in the amorphous state.From the observation of zoom-in image of Figure 1b, together with the corresponding FFT of the selected region (Figure S4, Supporting Information) and the XRD spectra [8] (Figure S5, Supporting Information) obtained from the HZO films, no signal related to HZO crystallinity was obtained.The results confirm that the fabricated HZO films in our work have an amorphous phase.It was also confirmed via a capacitance versus voltage (C-V) measurement under different frequencies of the HZO-1 memristor as shown in Figure S6 (Supporting Information), where it shows no typical hysteresis loop attributed to ferroelectricity.In addition, typical RS behavior was observed in the freshly fabricated device during polarization measurement (Figure S7, Supporting Information), suggesting the preservation of the RS properties of the HZO layer.Thereafter, atomic force microscope (AFM) images with a scanning range of 1 × 1 μm 2 were collected from the deposited RS layers of all three thin films.Despite slight differences, all the thin films show a surface root mean square roughness (R q ) of about 0.3 nm (Figure 1c).The switching performance of memristors is known to be highly dependent on the oxygenrelated defects in the RS layer. [7,9]In order to quantify for that, XPS spectra of the three different RS layers -HfO 2 , HZO-1/2 were obtained and fitted by using Shirley background subtraction and Gaussian functions, as shown in Figure 2a-c.
As referenced in previous literature, the O1s spectra of XPS can be deconvoluted into three components including lattice oxygen (LO), non-lattice oxygen (NLO), and absorbed oxygen, based on the position of the referenced characteristics binding energies. [9,10]The same approach was used to analyze all the RS layers, where the same components are deconvoluted.For the HfO 2 thin film, the lowest binding energy peak located at ≈530.6 eV (blue) corresponds to the LO or metal-oxide (M-O), while the rest located at ≈531.7 eV (red) and ≈532.8 eV (green) correspond to the NLO in the HfO 2 film and the absorbed oxygen (e.g., hydroxyl group, -OH) respectively, as shown in Figure 2a.Similarly, for the HZO-1/2 thin films, the lowest binding energy peaks located at ≈530.0 and ≈530.4 eV (blue) correspond to the LO resulting from a stoichiometric metal-oxide bond; the medium binding energy peaks at 531.6 and 531.8 eV correspond to the NLO, which can be indirectly interpreted to be one type of oxygen defects in the form of oxygen vacancies (V o ) in the oxygendeficient region [10] ; the highest binding energy peaks centered at ≈532.7 eV (green) correspond to the absorbed oxygen, as shown in Figure 2b,c.From the observation of O1s spectra of the three thin films, the largest amount of NLO in HZO-1 is observed, indicating the largest concentration of V o in the HZO-1 thin film.This can be further verified by the ratio of the intensity of LO and NLO (O L /O NL ); changes in the relative intensity of O L /O NL can be correlated to the relative concentration of V o in the thin film. [10]Figure 2d shows the plot of O NLO /O LO as a function of x value of Hf x Zr 1-x O 2 , indicating that the relative V o concentration as a function of Zr% intensity (inversely related to the x value).The V o concentration in the RS layer is observed to follow an increasing trend as the x value of Hf x Zr 1-x O 2 decreases while the Zr% intensity increases; this suggests a lower formation energy for V o with increasing Zr concentration. [11]The increase in the V o concentration is beneficial to the I-V characteristics with reduced switching voltage, as well as reliability parameters such as uniformity and endurance as discussed in the later part of the manuscript. [12]C I-V electrical characteristics of the HfO 2 and HZO-1/2 memristors were obtained with a compliance current (CC) of 1 mA to avoid premature breakdown of the devices.The I-V behavior of the memristors averaged over 20 DC sweep cycles are plotted out in Figure 3a, where all the memristors regardless of Zr% exhibit bipolar resistive switching characteristics.It can be observed that among the three different stoichiometry memristors, the HZO-1 memristor shows the lowest switching voltage, i.e., set voltage (V set ) ≈0.85 V and reset voltage (V reset ) ≈−0.75 V; the HZO memristors are further discussed.Figure 3b further shows the DC switching characteristics of the HZO-1 memristor over 150 consecutive switching cycles (50 th , 100 th , and 150 th cycles shown).Same measurement was performed on HZO-2 memristor shown in Figure S8 (Supporting Information), while all the DC I-V switching cycles of three types of memristors are shown in Figure S9 (Supporting Information).A high ON/OFF ratio of two orders and excellent cycle-to-cycle (C2C) variations are obtained despite the harsh conditions in switching in DC.Narrowing down to both HZO-1/2 memristors, their LRS/HRS resistance cumulative distribution function (CDF) curve over150 cycles each were extracted and plotted out in Figure 3c.Here, we define the coefficient of variation (C•V) as a normalized measure of the degree of dispersion of a probability distribution, defined as the ratio of standard deviation to mean (C•V = /μ).By applying this parameter to the measured devices, the HZO-1 memristor exhibits a C•V of 0.398, as compared to that of HZO-2 with a C•V of 0.709.In Figure 3d, the distributions of the V set /V reset of the memristors extracted over 150 switching cycles are shown.Despite that both types of HZO memristors have a relatively tight distribution, the distribution of reset voltage of HZO-1 memristor is 34.4% smaller than the HZO-2 memristor, which is indicated by the C•V of V set /V reset for HZO-1 and HZO-2 memristor (0.253/0.082 vs. 0.229/0.125).Combined with the comparison of switching voltages among three devices, it can be concluded that HZO-1 memristor possesses a lower operating voltage, as well as better uniformity among the two, and is further investigated.
In addition to the better DC electrical characteristics, the HZO-1 memristor is further characterized using AC pulse measurements, showing lower switching voltages, fasterswitching speed, and concurrently a lower switching energy as compared to the control HfO 2 memristor.Figure 4a,b show the switching comparison of the HfO 2 and HZO-1 memristors.The switching speeds extracted in this work are defined as the latency between the full width at half maxima (FWHM) of applied write pulse and response current of the memristors.As shown in Figure 4a, the time difference between the FWHM of the applied write pulse and response current are 3.06 μs and 3.12 μs, respectively for the set process, translating to a switching speed of 60 ns for the HfO 2 memristors.Using the same methodology, the switching speeds of the set process of the HZO-1 memristor are extracted to be 40 ns, as shown in Figure 4b.Consequently, the switching energy of the HZO-1 memristor defined as the product of applied voltage, response current, and switching time, is extracted to be 60 pJ, which is 77.8% smaller than 270 pJ of the control HfO 2 memristor.The switching speed of the reset process is then extracted using the same methodology as well, as shown in Figure S10 (Supporting Information).We can see that the reset switching speed of the HZO-1 memristor is also faster than that of the HfO 2 memristor with 100 ns versus 120 ns, respectively.
The DC endurance of the HZO-1 memristor is shown in Figure 5a.The HZO-1 memristor was further switched over 550 DC sweep cycles, maintaining its operating voltages and displaying almost stable LRS/HRS resistances.The resistance of each cycle is extracted at 0.2 V read voltage (V read ).The device could switch between HRS and LRS with ON/OFF ratio ≈15.Similar to the HZO-1 memristor, the HZO-2 memristor was also tested with over 180 DC sweep cycles in Figure 5b.Although HZO-2 memristor exhibited a higher ON/OFF ratio value than HZO-1, devices fabricated in that stoichiometry suffered from a lower endurance and broke down permanently within 180 DC cycles.Although the HZO-1 memristor has a smaller ON/OFF ratio, the larger amount of V o in the film is beneficial for reduced switching voltages together with improved switching speed.The AC endurance of the HZO-1 memristor was further characterized, as shown in Figure 5c.In order to perform a complete write/erase cycle, a positive pulse (1.5 V/ 1 μs) was applied to the device followed by a negative pulse (−1.8 V/ 1 μs); the currents were read at 0.2 V.It can be seen that more than 10 5 continuous write/erase cycles, under the same set/reset pulse with stable ON/OFF ratio of ≈10 can be achieved.We have shown earlier that the retention of HZO-1 memristor can exceed 10 4 s under room temperature condition. [13] The HZO-1 memristor exhibits several performance advantages including switching voltages and the largest DC endurance.In addition, further electrical characterizations are conducted in this work including high AC endurance exceeding 10 5 cycles, fast set speed of 40 ns, stable multi-state retention, and stable high temperature (85 °C) retention, which was not reported in the similar memristors works compared.
To further understand the RS mechanism of HZO memristor, the transport mechanism was analyzed from both set and reset process.It is widely agreed that the transport of memristors include that of ohmic conduction, space-charge limited conduction (SCLC), Flower-Nordheim (FN) tunneling, Poole-Frenkel (PF) emission and Schottky emission. [9]For our HZO memristor in the high-field region of the HRS and low-field region of the LRS states, respectively (Figure 6a,b), Schottky emission and ohmic conduction are at play.The details of the analysis are shown in Figure 6c-f and as discussed.In the LRS of the set process, the ln(I) is linearly proportional to ln(V) with a slope of 1.082 (Figure 6c), indicating that the ohmic conduction mechanism is due to the formation of conductive filament (CF). [15]In the HRS of the set process, it can be seen that the linear fit of ln(I) versus (V 1/2 ) corresponds to Schottky emission as shown in Figure 6d.This confirms that the transport mechanism of HRS at high-field region is dominated by the Schottky emission described by Equation (1): [16] where T is the absolute temperature, k is the Boltzmann constant, q is the charge of electron and  B is the Schottky barrier height.Similarly, in the reset process, the memristor shows an ohmic behavior with a slope of ln(I) versus ln(V) of 1.089 in the LRS (Figure 6e) when the CF is formed.From the curve fitting result of ln(I) versus (V 1/2 ) shown in Figure 6f, it suggests that Schottky emission is the dominating transport mechanism when the CF ruptures with a larger applied V reset .Based on the above analysis, we conclude that the RS mechanism of the HZO memristor is dominated by both ohmic conduction and Schottky emission, explained as follows.During the RS process, the formation of the CF, which consists of V o , results in an ohmic conduction in the LRS while, in HRS, the conduction mechanism transits into Schottky emission because of the rupture of the CF and a larger number of defects accumulating between the Pt electrode and the CF. [9,15]This phenomenon is consistent with the results of XPS, where the V o concentration is relatively high in the HZO RS layer as compared to the control HfO 2 RS layer.The use of memristor as analog weights is popular because of the ability to continuously modulate the device conductance.Considering the overall performance of both memristors, HZO-1 with improved endurance, stable retention, multiple analog states, fast switching speed, and lower switching voltages was selected to be implemented in the SOM ANN.In the following, the analog states tunability of the HZO-1 memristors are characterized.The potentiation (P) and depression (D) are described as an increase or decrease in conductance modulation via a series of programming and erasing scheme.The SET voltage pulses can be used to realize potentiation, whereas the RESET voltage pulses enable depression.However, if the width or amplitude of the pulse signal is not suitable, it can be completely switched to the maximum HRS and minimum LRS.To avoid this, different pulse conditions are experimented, and the appropriate pulse scheme is chosen.Two different programming schemes were selected, namely, voltage pulses with increasing amplitude (P: 0.8-1.29 V, 1 μs; D: −0.9∼−1.39V, 1 μs), and voltage pulses with constant amplitude and pulse width (P: 0.45 V, 1 μs; D: −0.6 V, 1 μs) for comparison.Figure 7a,b show the P/D cycle of the memristor under 50 pulses, and the fitting curve under five consecutive P/D cycles.Figure 7a is the result of using a constant voltage amplitude pulse scheme (scheme 1).However, when the RESET pulse is applied, the conductance of the device drops abruptly, and many missing states appear, which is undesirable for use in an ANN.On the other hand, when voltage pulses with increasing amplitude scheme (scheme-2) were chosen, as shown in Figure 7b, significantly better linearity and cycle-to-cycle variation of the memristors can be achieved.The HZO-1 memristor is further shown to retain its different resistance states (eight of them) over 1000 s each (Figure S11, Supporting Information), demonstrating longterm behavior and multi-bit storage, suitable for use as weights in an ANN.
To evaluate the linearity of the potentiation and depression, the change in conductance with the number of pulses can be described as: where G p and G d are the conductance of potentiation and depression, respectively.N and N max are the number and the maximum number of programming pulses, respectively.G max and G min are the maximum and minimum values of conductance.A p and A d represent the linearity of potentiation and depression, respectively.The fitting curves of the conductance potentiation and depression in programming scheme-1 and scheme-2 are shown in Figure 7a,b.In order to enable direct and effective comparison, the data in our experiments were all normalized to a range of 0-1.The fitting linearities of the potentiation and depression in scheme-2 are extracted as A p = 0.350 and A d = 0.405, which are higher than that in scheme-1 (A p = 0.179 and A d = 0.056).
The comparison of the conductance states and linearity of the devices studied in this work with others reported in the literature are shown in Table S2 (Supporting Information).
Based on the potential of analog characteristics of the device, an unsupervised SOM network that can map complex and nonlinear high-dimensional data to a low-dimensional space with simple geometric structure and mutual relationships [17] is constructed based on the HZO-1 memristor for image classification, exhibiting strong adaptability to non-ideality compared with conventional ANN. [18]The SOM simulation in this work achieves a high image recognition accuracy despite a non-linear weight update.This is due to the intrinsic topology-preserving nature and competitive learning features in the SOM network, as well as the online training and the write-verify device programming scheme to ensure that the correct resistance state has been achieved.The detailed process of the analog weights programming in the SOM network is described and shown in Figure S12 (Supporting Information).The SOM neural network is constituted by an input layer with 784 neurons and a competitive layer with 100 competitive neurons arranged regularly in a 2D topological space, [17a] as shown in Figure 7c.The weights that fully connect the input and competitive neurons are implemented with memristors. [19]he adopted dataset contains 6000 images with an image size of 28 × 28 (784 pixels), including four types of clothing images adapted from the Fashion MNIST dataset [20] (0:'t shirt ', 1:' trouser ', 2:' sandal ', 3:' bag') (see Methods section for details).During the unsupervised network training, the weights that connect to the winning neuron and its neighboring neurons within the influence radius are updated, with which the network is trained to identify the common features among the input data. [21]The weight updating during the training is based on the conductance updating of the memristors considering cycleto-cycle variation.
The network's classification ability was evaluated after testing with 4000 images from four selected categories.The curves in Figure 7d compare the classification accuracies of the SOM network implemented with scheme-1 and scheme-2, respectively.The image classification accuracy of SOM implemented with the scheme-2 reached 92% (as shown in Table S3, Supporting Information) and remained stable after 20 epochs of network training, exhibiting strong tolerance to device variation.In contrast, the accuracy of the SOM network implemented with scheme-1 reached an average accuracy of 75% with high fluctuations due to the undesirable loss of conductance states during LTD and higher weight updating nonlinearity and variation during the weight updating.
Figure 7e,f depicts the weight maps during the training process for SOM networks implemented with two schemes.Figure 7f shows the weight map updating from the initial random state to four clustered categories during the training, [22] demonstrating that the SOM network based on the scheme-2 effectively learned the input features and stored the learned information in weights. [23]In contrast, the change of weight maps from the initial random state to vague contours of the clothing in Figure 7e suggests that the SOM network implemented with the scheme-1 only learned and stored partial input features.The learning ability difference between two SOM networks arises from device nonideality-caused discrepancy between the ideal weight update values and the actual update values of the device conductance during network training.This discrepancy will lead to the deviation of network learned features during each weight update, resulting in low recognition accuracy and high fluctuations.

Conclusion
In summary, we report on the enhanced performance of HZO-based memristor over mono-oxide HfO 2 memristor.The Zr alloying approach introduces larger V o concentration, resulting in lower operating voltage (forming/set/reset voltages ≈1.8 V/0.85 V/−0.75 V) in the optimized HZO memristor (HZO-1) in this work, while retaining decent DC endurance, retention time, switching speed, and energy.We also evaluated two different programming schemes, namely the constant-amplitude voltage pulses scheme and increasing-amplitude voltage pulses scheme.Excellent linearity can be obtained via the latter scheme, beneficial for weights implementation in an ANN.In addition, when evaluated in a SOM ANN using simulations, our memristors perform well in Fashion MNIST classification, achieving an accuracy of 92% with short training cycles.With further optimization of the HZO RS layer process, we expect the overall uniformity and variability of the HZO memristor to further improve, thus paving a potential pathway for large-scale implementation of alloyed memristors in storage and computing applications.

Experimental Section
Device Fabrication: The bottom electrode composed of Ti (5 nm)/Pt (20 nm) was deposited on a 285 nm-SiO 2 -capped Si substrate using an electron-beam evaporator.Next, three different stoichiometry of Hf x Zr 1-x O 2 thin-films with x of 0.37/ 0.66/1 were fabricated (HZO-1/HZO-2/HfO 2 , respectively) via a sputtering approach using stoichiometric HfO 2 and ZrO 2 targets, with all the thickness of 5 nm.In this work, HfO 2 was obtained via a standard single sputtering approach.A HfO 2 target was used to deposit the HfO 2 thin-film on to the BE at room temperature using RF sputtering.The working pressure of Ar was 3 mT, while the RF power was 60 W, as shown in Table S1 (Supporting Information).Then the HZO thin-films were obtained via a co-sputtering approach using stoichiometric HfO 2 and ZrO 2 targets.The individual sputter powers of the HfO 2 and ZrO 2 targets were varied to deposit the different stoichiometries of HZO thin-films.For HZO-1 thin-film, the sputter powers of HfO 2 and ZrO 2 targets were 53 and 125 W, respectively, while that of HZO-2 thin-film were 80 and 80 W. At last, the top electrodes were obtained with the same process as the bottom electrodes, composed of Ti (5 nm)/Pt (20 nm).
Characterizations: Voltage bias was applied to the TE with BE grounded.XPS analysis was performed by EscaLab 250Xi from Thermo Fisher Scientific.All the measured thin-films with the thickness of ≈5 nm were deposited on the pure SiO 2 -on-Si wafer under the same conditions as the corresponding device.All binding energy data of spectra were calibrated by the C 1s signal from surface adsorbed hydrocarbon at 284.8 eV.Electrical characteristics were measured using a Keysight B1500A semiconductor parameter analyzer and Everbeing C-4 probe station.
Training and Test Datasets: The dataset used in this article is the Fashion-MNIST dataset, which was an open-source image library available on GitHub.It contains ten categories of images that represent a concrete human necessity: wears.This model still used its 28*28 resolution for training and testing, and selects 6000 images from each of the four categories: (0:'t shirt ', 1:' trouser ', 2:' sandy ', 3:' bag') for model training.For the same category, using the test set in the Fashion-MNIST dataset, 1000 pictures in each of the four categories were used to verify the classification and recognition accuracy of the network.

Figure 2 .
Figure 2. XPS O1s core-level spectra of a) HfO 2 , b) HZO-1 and c) HZO-2 thin-film, and d) plot of the O NL /O L as a function of x value of Hf x Zr 1-x O 2 .

Figure 3 .
Figure 3. DC switching characteristics.a) The average I-V curves of the three different memristors fabricated in this work over 20 DC switching cycles.b) The 50 th , 100 th , and 150 th DC switching cycles of the HZO-1 memristor, cumulative distribution of the c) LRS/HRS and d) operation voltages measured over 150 cycles for both the HZO-1/2 memristors.

Figure 4 .
Figure 4. a) HfO 2 and (b) HZO-1 memristor switched using AC pulse.A set switching time down to 40 ns was measured for the HZO-1 memristor as compared to 60 ns for the control HfO 2 memristor.

Figure 5 .
Figure 5. DC endurance characteristics of a) the HZO-1 memristor over 550 switching cycles, and b) HZO-2 memristors over 180 switching cycles with a V read of 0.2 V. c) AC endurance characteristics of HZO-1 memristor, and d) Retention characteristics of HZO-1 memristor at 85 °C, with currents read at 0.2 V.

Figure 6 .
Figure 6.Transport mechanism analysis of the HZO-1 memristor in (a) reset and (b) set process, respectively.Curve fitting of (c,d) set process and (e,f) reset process, both of which indicate ohmic transport mechanism in LRS and Schottky emission in HRS.

Figure 7 .
Figure 7. Fashion MNIST image classification based on HZO-1 memristor implemented SOM neural network.Standard deviation and fitting curve of five conductance modulation cycles with pulses with a) constant amplitude and b) increasing amplitude.c) The structure diagram of the SOM neural network.The input layer is composed of 784 neurons corresponding to the number of pixels in the input image, while the output layer is composed of 100 competitive neurons.The analog weight map of the winning neuron displays the weight values between the input layer and a single output neuron after the training.d) Image classification results of the SOM neural network implemented with scheme-1 and scheme-2.The weight maps of all the competitive neurons before training, and after 5, 20, 35, 50 epochs training with e) scheme-1 and f) scheme-2, respectively.

Table 1 .
Comparison of the memristors studied in this work with alloyed-memristors reported in literature.