Multifunctional Ag–In–Zn–S/Cs3Cu2Cl5‐Based Memristors with Coexistence of Non‐Volatile Memory and Volatile Threshold Switching Behaviors for Neuroinspired Computing

Lead‐free all‐inorganic Cs3Cu2Cl5 perovskites, a member of the metal‐metal halide material family, have attracted pronounced attention owing to their low toxicity, facile fabrication strategies, considerable ambient stability, and intriguing photoelectric properties. However, the application of environmentally friendly copper‐based Cs3Cu2Cl5 in memristors has been rarely reported to the authors’ knowledge. Herein, multifunctional memristors with the coexistence of non‐volatile memory (MS) and volatile threshold switching (TS) behaviors are introduced based on an innovative Ag–In–Zn–S/Cs3Cu2Cl5 heterostructure. The inserted Ag–In–Zn–S quantum dots layer may provide an effective method for guiding the formation of the dominant metallic Ag filaments, resulting in considerably stable and controllable multiple switching behaviors. Additionally, the heterostructure memristor is capable of imitating some essential biological synaptic functions, including long‐term potentiation (LTP), long‐term depression (LTD), and the short‐term memory (STM) to long‐term memory (LTM) transition. Furthermore, the famous conditioning Pavlov's dog experiment corresponding to associative learning is electronically simulated by the studied device. Moreover, utilizing the devices’ LTP and LTD properties, relatively high recognition accuracies for small and large digits datasets are achieved through a three‐layer artificial neural network, revealing the feasibility of implementing neuromorphic computation using heterostructure memristors.


Introduction
The efficient and highly parallel neuromorphic computing inspired by biological brains, a promising alternative to existing conventional computing, has been intensively studied to overcome the key computing bottlenecks of traditional complementary metal-oxide-semiconductor (CMOS) hardware platform based on von Neumann architecture, especially in the face of ever-increasing computing efficiency demands. [1][2][3] The investigations and fabrications of electronic biomimetic devices that can simulate biological learning and memory behaviors are considered to be the first step toward the hardware implementation of neuromorphic computing systems. [4] Artificial electrical synapses, one of the essential elements for building effective brain-like neuromorphic computing systems, have been substantially emulated by well-investigated nonvolatile memristors with similar working principles of biological synapses. [5,6] Nonvolatile memristors, playing a pivotal dual role in data storage and computing, are highly efficient for computing in-memory Lead-free all-inorganic Cs 3 Cu 2 Cl 5 perovskites, a member of the metalmetal halide material family, have attracted pronounced attention owing to their low toxicity, facile fabrication strategies, considerable ambient stability, and intriguing photoelectric properties. However, the application of environmentally friendly copper-based Cs 3 Cu 2 Cl 5

in memristors has been rarely reported to the authors' knowledge. Herein, multifunctional memristors with the coexistence of non-volatile memory (MS) and volatile threshold switching (TS) behaviors are introduced based on an innovative
Ag-In-Zn-S/Cs 3 Cu 2 Cl 5 heterostructure. The inserted Ag-In-Zn-S quantum dots layer may provide an effective method for guiding the formation of the dominant metallic Ag filaments, resulting in considerably stable and controllable multiple switching behaviors. Additionally, the heterostructure memristor is capable of imitating some essential biological synaptic functions, including long-term potentiation (LTP), long-term depression (LTD), and the short-term memory (STM) to long-term memory (LTM) transition. Furthermore, the famous conditioning Pavlov's dog experiment corresponding to associative learning is electronically simulated by the studied device. Moreover, utilizing the devices' LTP and LTD properties, relatively high recognition accuracies for small and large digits datasets are achieved through a three-layer artificial neural network, revealing the feasibility of implementing neuromorphic computation using heterostructure memristors.
applications. [7] In addition, but equally important, volatile memristors can be functioned as promising selector devices for their capability of solving the leakage current in a crossbar array structure, thus enhancing the stability of neuromorphic computing systems. [8] There is no doubt that choosing fairish medium switching materials to design innovative device configurations is one efficient way to fabricate functional memristors.
Recently, the low-dimensional lead-based halide perovskites have attracted stupendous attention owing to their relatively remarkable physical characteristics, including tunable bandgap, easy synthetic procedures, long carrier diffusion lengths, fast carrier mobilities, and so forth. [9][10][11] Taking advantage of those properties, combined with the reported presence of shallow defects, the lead-based halide perovskites have been explored as promising resistive switching (RS) materials for ion migration-related memristors. [12,13] However, the environmental unfriendliness and instability of toxic lead-based halide perovskites have been considered as tough problems, which seriously hinder their further practical application. Therefore, there is strong demand for exploiting lead-free halide perovskites with non-toxicity and air-stability for the development of advanced memristors.
Very recently, Cu-based perovskite Cs 3 Cu 2 Cl 5 , one kind of all-inorganic enthralling materials, has been numerously investigated as one of the most competitive alternatives in the fields of UV photodetectors, light-emitting diodes, and X-ray detection, due to their low toxicity and highly stable crystal structures in the ambient environment. [14][15][16] Nonetheless, there are scarce investigations on the adoption of Cs 3 Cu 2 Cl 5 to construct memristors until now. Additionally, it has been reported that the multiple functions memristors can coexist with nonvolatile memory (MS) and volatile threshold switching (TS) phenomena in a single device, which is very beneficial for high-density data storage due to the highly compact integration manner. [17] Nevertheless, the multifunctional RS properties are hardly been implemented simultaneously in a single perovskite-based memristor, let alone for Cu-based perovskite Cs 3 Cu 2 Cl 5 . From these viewpoints, it is worth exploring the RS behaviors of lead-free perovskite Cs 3 Cu 2 Cl 5 and developing its application in multifunctional memristors with an engineered device structure. On the other hand, it has been experimentally reported that the introduction of quaternary Ag−In−Zn−S (AIZS) quantum dots (QDs) can improve the memristive device performance due to their excellent stability and high carrier mobility. [18] Therefore, the utilization of AIZS QDs serving as an insertion layer is conducive to producing advanced memristors with high stability and uniformity, which are suitable for constructing robust neuromorphic computing systems.
In this present work, we have provided a simple method to obtain stable and controllable coexistence of MS and TS behaviors endowed through adjusting the compliance current (CC) in a designed heterogeneous memristor. The device is composed of AIZS/Cs 3 Cu 2 Cl 5 bilayer materials with distinct silver (Ag) ion mobilities. In addition, the RS behaviors may be dominated by metallic Ag ions diffusion along the location of AIZS QDs acting as seeds, allowing rapid penetration and kinetically stable formation of conductive filaments (CFs). [18] Furthermore, we used the heterostructure memristors to emulate various biological behaviors, such as the short-term memory (STM) to long-term memory (LTM) transition, long-term potentiation (LTP), and long-term depression (LTD). Moreover, the successful electrical simulations of Pavlov's dog experiment reveal the associative learning ability of our memristors. Finally, the resulting high classification accuracies for small and large digits datasets have been achieved via a three-layer neural network training based on the experimentally measured conductance states. All the results suggest that the as-investigated multifunctional heterostructure memristors have great application potential in environment-friendly neuromorphic devices.

Results and Discussion
Our as-fabricated memristors made of AIZS/Cs 3 Cu 2 Cl 5 functional heterostructure were sandwiched between Ag and tungsten (W) electrodes to form crossbar array structures. The simple structure diagram of the as-synthesized AIZS QDs is displayed in Figure 1a, where the QDs with uniform size distribution were prepared analogously according to the detailed process flow reported previously. [18] Besides, the insertion of AIZS QDs in the memristor structure, suppressing the redistribution of ions, has huge positive effects on enhancing the stability of RS performance, which has been experimentally confirmed in our previous research. [18] The crystalline structure of the AIZS QDs is characterized by X-ray diffraction (XRD). An XRD spectrum of the AIZS QDs sample is exhibited in Figure 1b. Three broad prominent peaks, disclosing the nanoscale dimension of QDs, are approximately corresponding to the (002), (110), and (112) indices at 2θ positions of 27.54°, 45.74°, and 53.66°, respectively. [19] The photoluminescence (PL) characteristics of AIZS QDs at an excitation wavelength of 370 nm are shown in Figure 1c. It can be obviously seen that upon photoexcitation, the PL peak emission wavelength of AIZS QDs in toluene is around 619 nm, demonstrating a large Stokes shift of 249 nm. Figure 1d depicts the schematic illustration of the studied vertical device Ag/AIZS/Cs 3 Cu 2 Cl 5 /W with crossbar array construction. More details of device fabrication have been described in the Experimental Section, and the manufacturing process flow has also been clearly visualized in Figure S1, Supporting Information.
The top view of a 4 × 4 crossbar array characterized by an optical microscope is clearly demonstrated in Figure 1e, where metal pads (300 × 300 µm) are marked with black dashed lines, and the core intersection representing the fabricated heterostructure device (100 × 100 µm) is illustrated by red dashed lines. The cross-sectional view scanning electron microscopy (SEM) image of a heterostructure device with multiple layers is schematically displayed in Figure 1f, where the thickness of the AIZS QDs layer is about 240 nm, as well as Cs 3 Cu 2 Cl 5 layer, is around 80 nm. The crystal structure of the Cs 3 Cu 2 Cl 5 perovskites particles can be viewed from different coordinate perspectives as shown in Figure 1g and Figure S2a, Supporting Information, in which Cs 3 Cu 2 Cl 5 employs an orthorhombic crystal system identified by the powder XRD pattern ( Figure S2b, Supporting Information). [20,21] Cs, Cu, and Cl atoms are represented by red, blue, and green balls, respectively. It can be vividly observed that the framework of Cs 3 Cu 2 Cl 5 perovskites particles consists of [Cs 3 Cl 5 ] 3− anionic chains that www.advelectronicmat.de are spatially isolated by Cs + cations. [14] Such XRD data demonstrates that the crystallinity of our Cs 3 Cu 2 Cl 5 perovskite synthesized by using the hot-injection method is quite good. The corresponding SEM image shows granule-like shaped particles of Cs 3 Cu 2 Cl 5 that have an average diameter of ≈400 nm, as illustrated in Figure 1h. Additionally, a large Stokes shift of 216 nm between the PL (peak at 527 nm) and PLE spectra (peak at 311 nm) can be calculated from Figure 1i, thereby revealing good optical characteristics.
Upon preparing the materials and fabricating the studied devices, the electrical RS properties of the heterostructure memristors have been experimentally investigated in detail, as shown in Figure 2. Figure 2a depicts the typically volatile TS behaviors with multiple direct current (DC) current-voltages (I-V) curves recorded at a comparatively low CC of 100 µA. Accordingly, the device can be unidirectionally switched from high-to low-resistance states (HRS to LRS) corresponding to the set process, and then automatically return to its original state (i.e., HRS) without applying a bias voltage, which may be originating from the spontaneous dissipation of thin and unstable nanoscale CFs. [22,23] The switching voltages of volatile RS are extracted from consequent sixty-five cycles, as replotted with a histogram in Figure 2b. It is found that the ranging of switching voltages is from 0.5 to 0.72 V, which can be well fitted by Gaussian functions. Figure 2c demonstrates the respective I-V curves of positive bias responding to the varying CC of 0.1, 1, 10, and 100 µA, showing the definite volatile RS behavior under relatively low CC. Intriguingly, it has been reported that www.advelectronicmat.de the RS behavior can be probably tuned from volatile to nonvolatile switching by increasing the value of CC. [24,25] Therefore, further research on whether the non-volatile characteristics can be obtained in our studied device by increasing CC is worthwhile. Surprisingly, Figure 2d exhibits the device conductance as a function of CC, consequently revealing that increasing CC can induce the transformation from volatile TS to non-volatile MS. It should be pointed out that there is an increase in the low conductance state (LCS) when the CC is increased (<200 µA). Such obvious decay/increase in HRS/LCS under higher CC may be related to the high conductivity of more residual Ag CFs in the Cs 3 Cu 2 Cl 5 layer, as clearly illustrated in Figure S3, Supporting Information. [26,27] Once the value of CC is larger than 200 µA, the device can maintain the high conductance state (HCS, or LRS), that is, a reverse voltage is required to switch the device back to HRS. To further evaluate the memory performance of non-volatile MS behaviors of the heterostructure memristors, systematic analyses of electrical RS characteristics have been performed, as shown in Figure 2e-i. Figure 2e illustrates the repeating 100 semilogarithmic I-V dual sweep curves of the studied device under CC of 300 µA. The device can be set from HRS to LRS at positive voltage and reset in reverse at negative voltage respectively, showing a typical bipolar RS behavior. As can be seen from Figure 2e, the gradual decrease of current in the process "3" marked in the reset process could be explained by the joule heating-assisted electric field effect that governs the slow oxidation of Ag filament, which leads to the gradual thinning of the Ag filament. [28,29] In addition, the consecutive 400 cyclic measurements have been carried out at room temperature to demonstrate the cycle-to-cycle variation of the heterostructure memristors, as displayed in Figure 2f. Notably, both HRS (i.e., OFF-State) and LRS (i.e., ON-State) are respectively distributed in a sufficiently narrow region, verifying the uniformly stable formation and fracture of dominant Ag CFs in a bettercontrolled manner. On the other hand, the endurance characterization using another pulsed voltage stresses method will be investigated for future work. [30,31] Furthermore, the cumulative distributions of V set and V reset have been processed with the coefficients of variation of 7.03% and 22.52% respectively, as www.advelectronicmat.de noted in Figure 2g. Specifically, the range of V set is from 1.32 to 1.75 V while the V reset ranges from −0.75 to −0.31 V, exhibiting relatively low variability in operation voltages. Moreover, data retention, a key metric in memory applications, has been non-destructively tested over 2000 s with a switching ratio (i.e., HRS/LRS or memory window) ≈10 2 using a non-disruptive reading voltage of 0.1 V under open-air conditions, as presented in Figure 2h. It is reasonably hypothesized according to the steady trend that longer lifetimes can be expected from the retention results of HRS and LRS. Finally, the device-to-device variability has been investigated as plotted with a box plot in Figure 2i, where the values of both HRS and LRS of twelve devices under the same testing conditions are distributed in a small range respectively.
Overall, the above electrical results analyzed in detail indicate the relatively good stability and high reproducibility of our multifunctional memristors even compared to other reported perovskite-based devices, which have been listed in Table 1.
In addition, it can be also concluded that our Cs 3 Cu 2 Cl 5 -based devices have other distinct advantages, such as relatively low operation voltages and large memory windows, multifunctional RS behaviors, and various synaptic simulations. Except for the comparison limited to the perovskite material system, we have supplemented and summarized device parameters corresponding to recently reported state-of-the-art memristors with the coexistence of MS and TS behaviors, as listed in Table S1, Supporting Information, suggesting that the investigated heterostructure devices have relatively excellent performance. A method of employing a sacrificial layer (e.g., a parylene layer with ultrahigh stability and ultralow permeability) can be tried to further develop Cs 3 Cu 2 Cl 5 perovskites-based memristors compatible with CMOS manufacturing processes in the near future. [32] In terms of the time-dependent switching stability, the corresponding I−V curves and the resulting distributions of resistance states are demonstrated in Figure S4, Supporting Information, disclosing that the switching characteristics of our as-investigated memristor have only slight degradation after 100 days when compared with the initial value. The predictably excellent temperature stability of the memristors will be confirmed by testing the uniformity of temperature-dependent I−V curves in the following research due to that both AIZS and Cs 3 Cu 2 Cl 5 materials have quite good environmental stability within the appropriate temperature. [21,33] It is of interest to gain insight into the switching mechanisms of our studied heterostructure memristors by means of curve-fitting and electrode-cutting operations. With regard to the positive and negative voltage sweepings of non-volatile RS, the I-V curves have been re-plotted in double-log scale, as displayed in Figure 3a,b respectively. The linear relationships between logV and logI with the fitting slopes of ≈1 (1.01 and 0.99, respectively) show that the LRS across the entire range of applied bias obeys the Ohmic conduction related to the filamentary mechanism. Meanwhile, the multiple randomly selected LRS curves of volatile RS have also been well fitted in Figure S5, Supporting Information, showing that the Ohmic behavior is consistent with the LRS mechanism of non-volatile RS due to the slopes of ≈1 (ranging from 1.01 to 1.11). But for HRS, the fitted slopes have three distinct values with an increasing trend, corresponding to Ohmic region (I ∝ V), Child's square law region (I ∝ V 2 ), and current increase region (I ∝ V n , n > 2) respectively, which indicates that the RS mechanism of HRS highly matches to the trap-controlled space charge limited current (SCLC) conduction. [34][35] It has been demonstrated that the SCLC is closely associated with the injected carriers trapping or de-trapping by the inherent pre-existing defects in the switching layer. [36] The increase in slopes may be ascribed to the fact that the traps are gradually filled with the charge carriers. Coincidentally, the presence of defects/traps in Cs 3 Cu 2 Cl 5 perovskite particles has been validated from strong evidence, that is, timeresolved PL decay result discussed in Figure S6, Supporting Information. The inherent trap states may be induced by the comparatively lower ionic radius of Cl − . [37] By fitting the data To experimentally validate the localized nucleation property of CFs, as schematically illustrated in Figure 3c, the Ag TE was first cleaved into two parts (TE1 and TE2) by the tungsten probe tip after the device was set to the LRS, and then the TE1 was again cut into TE1-1 and TE1-2 using the same approach. The resistance states involving the same W bottom electrode (BE) and varying cut TEs were read individually at 0.1 V, as shown in Figure 3d. The TE2 and TE1-2 exhibit the HRS, disclosing that there are no forming CFs beneath the TE2 and TE1-2. As expected, the resistance of TE1-1 has a similar value as compared to the initial uncut Ag TE, experimentally confirming the localized nature of CFs once again. In addition, Figure 3e shows the device area dependence of the LRS and HRS, which suggests that the value of LRS is independent of the device size while that of HRS is inversely proportional to the device size, further supporting the filamentary RS conduction. [38] Furthermore, the sudden increase in current during the set process in Figure 2a,e is a sign of electrochemical metallization filament-type switching behavior. [39] Under these cases, combined with the generally accepted redox reactions of active Ag TE, we speculate that the RS  Figure 3f. It should be noted that the Ag element in AIZS QDs is less active compared to the metallic form as it can form covalent bonds with other elements. In addition, the surface of QDs was covalently capped by organic ligands, which further prevents the migration of elements in the current circumstance, revealing that the Ag element in the AIZS layer cannot migrate and participate in the RS process under the presence of an electric field. As consequence, the AIZS layer is not subject to the change in stoichiometric ratio, which has been verified by the stable and reproducible experimental data in Figure 2f. To be more specific, the ionized Ag + is first generated from the oxidation of Ag TE under the excitation of an external positive bias. More importantly, Ag ions then migrate into the heterostructure through the guidance of AIZS QDs as well as the assistance of defects in the Cs 3 Cu 2 Cl 5 layer. It should be pointed out that the thickness of the Cs 3 Cu 2 Cl 5 layer is small enough, enabling enough electric fields to benefit the migration of positively-charged Ag ions through the whole Cs 3 Cu 2 Cl 5 layer. Or contrariwise, the Ag atoms are produced from the reduction of Ag ions by the negatively-charged electrons and finally accumulated to establish the complete penetration of the CFs between the two electrodes, thereby leading to the transition from HRS to LRS. It is worth mentioning that the higher the CC value, the stronger and more stable the CFs, and vice versa. [40,41] Such is why our single heterostructure memristor has versatile functions, that is, the coexistence of non-volatile MS and volatile TS behaviors. Reversely, the device can be reset back to HRS when applying the negative bias voltage, since the shaped filaments underwent chemical dissolution resulting in the local dissolution of CFs. [42] The resultant transformation from volatile TS to non-volatile MS behaviors may have potentially strongly correlated effects on the modulation of device conductance. Neurobiologically, memory responses in the human brain can be generally classified into two categories, that is, short-and long-term plasticity. [42] Psychologically, human memory was akin to describe as STM and LTM, corresponding to a rapid decline (<1 min) and slow decay process respectively. [43] Biologically, both plasticity or memory types, enabling the achievements of the acute computational and memorization functions, are closely related to the connection strength (i.e., synaptic weight) change. [44,45] Additionally, it is widely accepted that the transition from STM to LTM can be mimicked by providing proper rehearsals as well. [46,47] What is more, the active electrode-based memristors have been extensively studied owing to the fact that field-driven Ag or Cu ion diffusion kinetics share interesting similarities with potential-induced dynamics of the biological Ca 2+ , which is genetically suitable for mimicking bio-synaptic behaviors through weight (conductance) modulation. [48] As shown in Figure 4a, the device current increases roughly in a certain trend with 10 consecutive stimulation voltage pulses (amplitude of 3 V and width of 10 ms [3 V, 10 ms]), followed by readout voltage pulses (0.1 V, 10 ms). Once removing the electrical voltage pulses, the increasing current has not been upheld and quickly recovered back to the initial low current after 60 s, experimentally demonstrating the realization of the STM behavior of the AIZS/Cs 3 Cu 2 Cl 5 device. At this stage, the forming Ag CFs seem to be weak and unstable due to insufficient pulse stimulation. In contrast, by increasing the number Figure 4. Demonstration of non-associative learning behavior. The current response to a) 10 and b) 30 presynaptic pulses, implementing STM and LTM behaviors respectively. c) The schematic presenting the transition from STM to LTM through the rehearsal process. d) Device conductance stimulated by different pulse numbers was recorded at the initial stimulus, final stimulus, and 60 s after removing stimulation respectively. www.advelectronicmat.de of stimulations from 10 to 30 without changing other pulse parameters, the device current eventually increases to a higher level and tends to saturate as exhibited in Figure 4b, stemming from the formation of strong and stable CFs. Surprisingly, the decay in current or conductance is just slight when comparing the conductance for 60 s before and after stimulation, implying that the device conductance can be maintained for a longer period. It can be clearly observed from Figure 4d that the conversion from STM to LTM has been realized by our AIZS/ Cs 3 Cu 2 Cl 5 device through repeating rehearsal illustrated in Figure 4c, resembling the volatile TS to non-volatile MS transition, which can be speculatively interpreted by the formation of CFs from weak to strong under more pulses applied on the device.
Non-associative learning behavior has been discussed and analyzed in Figure 4. Apart from this regard, associative learning, another type of sophisticated learning mechanism generally occurring in biological living, enables the brain to establish new associations similar to a connate unconditional response after repeating training of conditional stimulus. [49,50] More interestingly, the classical conditioning Pavlov's dog experiment, a well-established and famous example of associative learning, has been successfully emulated by employing Ag/ AIZS/Cs 3 Cu 2 Cl 5 /W devices, as demonstrated in Figure 5. [51][52][53] During the mimicking process, a group of 10 pulses with an amplitude of 1 V (low)/3.2 V (high) and width of 10 ms (1 V/3.2 V, 10 ms) has been utilized to imitate the ringing bell (conditioned stimulus)/bone food (unconditioned stimulus) respectively. It can be seen from Figure 5a that the bone food stimulus alone can obviously induce salivation of the dog as analogous to HCS (>1 mS) of the device, while the followed one-off ringing bell stimulus has no capability to trigger any salivation (<1 mS). By adjusting the stimuli, the experimental imitation results show similar conclusions compared to Figure 5a as displayed in Figure 5b, suggesting that twice-off training cannot establish an effective connection between the bone food and the ringing bell. Figure 5c depicts a 27-repetition training process, in which, the connection (conductance) is gradually built (increased) by simultaneously providing two excitatory stimulations, eventually leading to the salivation of the dog (HCS of the device). Upon receiving such a sequence of training, as a result, the dog (device) is still activating salivation (maintaining HCS) with ringing bell (low-amplitude pulses) stimulation only, that is, the dog has learned to associate the ringing bell with bone food, therefore confirming the excellent associative learning ability of our heterostructure memristor, as demonstrated in Figure 5d. It should be referred to that the training-established conditioning association (training-modulated HCS) can be extinct (reduced) to the initial state when applying the ringing bell alone after a long time (not shown here), thus revealing the time-related forgetting behavior of the brain. The multi-input synaptic behaviors of the AIZS/ Cs 3 Cu 2 Cl 5 device are quietly meaningful for the development of complex brain-like computing systems.
To further investigate the neuromorphic computational simulation capabilities of customized AIZS/Cs 3 Cu 2 Cl 5 heterostructure memristors, a simple three-layer artificial neural network consisting of one hidden layer has been developed to perform image recognition by employing CrossSim simulator, as depicted diagrammatically in Figure 6a. The stimulation is based on the experimentally measured conductance states with continuously tuned LTP and LTD. As displayed in Figure 6b, the LTP process can be mimicked by applying consecutive positive voltage pulses (2 V, 15 ms) while the LTD process can be driven by positive voltage pulses (−2 V, 15 ms), Where the dynamic range (i.e., conductance ON/OFF ratio) is calculated about 64.95. Such a relatively large dynamic range is conducive to providing many distinct conductance states. The power consumption per pulse has been calculated as low as 3.9 × 10 −7 J, which could be further reduced by employing narrower and/or lower voltage pulses in the future. The performance of an artificial neural network using back-propagation algorithm has been evaluated by training small (8 × 8 pixels) and large (28 × 28 pixels) images of handwritten digits from the "optical recognition of handwritten digits" and Modified National Institute of Standards and Technology datasets respectively. [54][55][56] On the basis of 10 output neurons, the 784 (64) pre-neurons and 300 (36) hidden neurons correspond to the simulation of the large (small) image of handwritten digits.
To maximize the usage of the device's dynamic range, the lookup table update model has been adopted to choose the appropriate conductance range. [57] In addition, nonlinearities (NLs), a critical non-ideality affecting pattern recognition accuracy, have been extracted from the above full range and clipped range of LTP/LTD curves to quantitatively evaluate device performance, as shown in Figure S7, Supporting Information. The corresponding conductance NLs for clipped LTP and LTD are approximately 0.46 and 4.91, while those for LTP and LTD in Figure 6b are about 0.53 and 5.54, respectively. Such reduction results in smaller NLs, which is essential for improving the accuracy of image recognition, as compared in Figure S7c,f, Supporting Information. As the simulation results show in Figure 6c,d, the maximum network classification accuracy rates for small and large digits datasets are 92.43% and 91.36% Figure 5. Electronic implementation of Pavlov's dog experiment by using four processes to demonstrate associative learning behavior. Dog, bone food, ringing bell, and dog's response correspond to the device, six highamplitude pulses (3.2 V with a width of 10 ms), six low-amplitude pulses (1 V with a width of 10 ms), and device conductance respectively. a-b) Before repeated training: Separated stimulation with bone food (salivation) and ringing bell (no salivation). c) Repeated training with simultaneous bone food and ringing bell stimuli. d) Stimulation with ringing bell only (salivation).

www.advelectronicmat.de
respectively after 20 training epochs, indicating comparable high accuracies to the ideal numerical results. It is worth pursuing that more in-depth research is warranted to improve the recognition accuracy of the devices based on AIZS/Cs 3 Cu 2 Cl 5 heterogeneous structure, such as improving the linearity of the conductance modulation through advanced device engineering. Meanwhile, the LTP and LTD processes can be mimicked continuously for six cycles by applying 140 positive (1 V, 15 ms) and negative (−1 V, 15 ms) pulses per cycle, clearly reflecting the reproducible conductance tuning behavior of the studied device, as exhibited in Figure S8, Supporting Information.

Conclusion
In summary, Cs 3 Cu 2 Cl 5 perovskite particles have been prepared and then utilized to fabricate multifunctional memristors. To help in improving the reliability, an AIZS QDs layer was interposed between the Ag TE and Cs 3 Cu 2 Cl 5 layers to construct a heterogeneous structure AIZS/Cs 3 Cu 2 Cl 5 as a functional layer. In addition, the versatile devices exhibit excellent RS properties, including a controllable CC-induced volatile TS to non-volatile MS transition, good device-to-device and cycle-to-cycle variations, and concentrated operating voltage distributions. Furthermore, the RS mechanisms have been investigated based on curve-fitting, electrode-cutting, and area-changing operations. More importantly, some important synaptic functions have been mimicked by our studied heterostructure memristors, for example, LTP, LTD, and STM to LTM transition. Moreover, the simulation of the famous Pavlov's dog experiment has been electrically carried out, successfully demonstrating an associative learning behavior. Finally, a simple three-layer neural network has been built based on the experimentally measured device properties to perform image recognition, and the resulting classification accuracies approach 91.36% and 92.43% corresponding to the large and small digits datasets respectively. These elaborated results shed light on that our designed heterostructure memristors with stable and multifunctional performance exhibit significant potential in developing efficient neuromorphic computing systems.

Experimental Section
Cs 3 Cu 2 Cl 5 Material Preparation: In terms of the colloidal synthesis of halide perovskite particles, cesium oleate (CsOL) is considered the most commonly used Cs precursor. First, Cs 2 CO 3 (0.324 g), 1-octadecene (ODE, 10 mL), and oleic acid (OA, 2 mL) were put into a 50 mL 3-neck flask. After that, the mixture was heated to 120 °C in an inert environment and the temperature was kept at 120 °C for half an hour, which was then heated at 100 °C under continuous stirring and nitrogen flow in an oil bath for preparing the subsequent synthetic Cs 3 Cu 2 Cl 5 materials. After preparing the CsOL precursor, the synthesis of cesium copper chloride Cs 3 Cu 2 Cl 5 perovskite particles has been carried out by a hot-injection method. CuCl (4 mmol), ODE (20 mL), OA (2 mL), and oleylamine (OAm, 2 mL) were loaded into another 50 mL 3-neck flask, followed by heating at 120 °C for 30 min. After completely dissolving CuCl powders, the reaction temperature was changed to 150 °C, and then 12 mL CsOL precursor was quickly injected into the solution for subsequent reaction. Afterward, the as-obtained solution was cooled in an ice-water bath www.advelectronicmat.de for 10 s. The cooled crude solution was centrifuged at 9000 rpm for 7 min to discard the supernatant. Next, the remained precipitation was centrifuged at 7000 rpm for 5 min and cleaned with n-hexane several times. Consequently, the reactant was dried in a vacuum oven at 60 °C for 12 h, leading to the successful synthesis of Cs 3 Cu 2 Cl 5 materials, which can be used to fabricate the studied memristors.
Material and Device Characterizations: The characterizations of as-received Cs 3 Cu 2 Cl 5 perovskite particles and AIZS QDs were carried out by photophysical and morphological measurements. The crystal phases of powder AIZS QDs were confirmed from the XRD pattern obtained by an advanced diffractometer system with a Cu-Kα X-ray tube (D/max 2500/PC, Rigaku, Japan). In addition, PL characteristics including steady-state and time-resolved fluorescence spectroscopy of both materials were performed on a PL system with Photoluminescence Spectrometer (Edinburg FLS1000). Moreover, the size and morphology of Cs 3 Cu 2 Cl 5 perovskite particles were characterized by using a field emission SEM (SUPRA55, Zeiss). Finally, the 4 × 4 crossbar array structures were also characterized by SEM and optical microscope, respectively.
Device Fabrication and Measurements: Heterostructure memristors with stacked structure Ag/AIZS/Cs 3 Cu 2 Cl 5 /W had been fabricated in 4 × 4 crossbar arrays on a two-inch wafer using a bottom-up approach. At first, the initial SiO 2 /Si substrate was cleaned with ethanol and water. After drying, 100-nm-thick W BE was deposited by the DC magnetron sputtering method with the aid of a hard mask. Next, the as-prepared Cs 3 Cu 2 Cl 5 perovskite particles dissolved in toluene solution (30 mg mL −1 ) were spin-coated with a spin rate of 500 rpm followed by an annealing process at 110 °C. After that, AIZS QDs well dissolved in toluene solution were transferred onto the Cs 3 Cu 2 Cl 5 film via the same spin-coating approach and annealing treatment to form the final AIZS/Cs 3 Cu 2 Cl 5 heterostructure (≈250/80 nm). Consequently, 120-nm-thick Ag TE was deposited by magnetron sputtering through another hard mask covering the heterostructure. The side lengths of the TE, BE, and intersection device were all about 100 µm. All the electrical characterizations of the studied heterostructure memristors were performed by Keithley 4200A-SCS semiconductor device parameter analyzer equipped with Cascade MPS150 manual probe station under open-air conditions at room temperature. The W BE was always grounded while the Ag TE was applied with the positive or negative bias determined by the switching operation.

Supporting Information
Supporting Information is available from the Wiley Online Library or from the author.