Memristive Synapse Based on Single‐Crystalline LiNbO3 Thin Film with Bioinspired Microstructure for Experience‐Based Dynamic Image Mask Generation

One of the key steps toward constructing neuromorphic systems is to develop reliable bio‐realistic synaptic devices. Here, memristors based on single‐crystalline LiNbO3 (SC‐LNO) thin film are fabricated as artificial synapses. A reservoir of oxygen vacancies is induced by Ar+ irradiation to resemble synaptic vesicles containing neurotransmitters. Phenomena of saturation and adaptivity, short‐term plasticity, paired‐pulse facilitation, paired‐pulse depression, and long‐term potentiation are successfully mimicked. The dynamic transition from sensory memory to short‐term memory, and further to long‐term memory, is also successfully emulated for multipattern memorization. In addition, first, taking advantage of short‐ and long‐term synaptic plasticity is proposed, to realize experience‐based image mask generation with different stimuli schemes. During the experience‐based generation process, memristive multi‐value masks (MMVMs) are generated with different numbers of stimuli applied to the memristor at each pixel, which corresponds to the times the region occurred in the history image set. The experience‐based memristive multi‐value mask successfully extracts multiple regions of interest with different priorities. This work demonstrates that the memristor based on Ar+‐irradiated SC‐LNO thin film with bioinspired microstructure shows great potential in future neuromorphic systems for experience‐based intelligent image processing.

of these mentioned above indicate the significant role synaptic vesicles play in realizing diverse synaptic functionalities.
In order to develop desired synaptic device for neuromorphic systems, various memristors based on different resistive switching layers have been developed to emulate the synaptic characteristics, such as TaO x /Ta 2 O 5 -based memristor, [19] HfO 2based memristor, [20] Hf-AlO x -based memristor, [21] LiNbO 3 -based memristor, [22,23] and Nb 2 O 5 /NbO x -based memristor. [24,25] These memristor can mimic essential analog synaptic functionalities with the characteristics of resistance plasticity, and also show the advantages of low power consumption, high switching speed, reliability, and high scalability. Especially, the memristors based on drift and diffusion of mobile species (such as Ag + [13,26,27] or oxygen vacancies (V o s) [28][29][30] ) that exhibit biorealistic dynamics similar to synaptic dynamics have been investigated. However, the memristive synapse with a specific structure to mimic the structure of biological synapse with synaptic vesicle is rarely reported, although wide studies have indicated that the state of vesicles (e.g., number, release probability, depletion, quantal variability, etc.) in biological synapses is an important contributor to diverse synaptic plasticity. [16,31] Memristor with synaptic plasticity has been used in different neuromorphic systems for different purposes, such as image processing. [32][33][34] In image processing, for processing largescale high-dimensional image data, mask operation is crucial to remove noise and shrink the dimensionality of the data representation while retaining the information we are interested in. [35] For example, mask operation can also be used to extract structural features and exclude other easily-confused information of the image, which has been widely applied in remote sensing [36,37] and medical monitoring. [38] If the mask can be generated based on the experience (i.e., the history image the intelligent system received), the intelligent systems can be able to adapt to unforeseen situations in the complexity of the world without being pre-defined with a set of static trained offline, which will make artificial intelligence (AI) move toward next step. [39,40] In order to realize experience-based image mask generation, the mask needs to be generated according to the external input in the history, in which the synaptic plasticity of memristor under external stimuli can be fully used of. It also put forward higher requirements for the memristive device, such as reliable nonvolatile characteristic, multilevel stability, and ultra-low device variation (for generating large-scale image masks).
In this work, we demonstrated a reliable single-crystalline LiNbO 3 (SC-LNO)-based synaptic device with bioinspired microstructure for mimicking the structure and functions of biological synapses. A reservoir of oxygen vacancies (OVR) layer with a high concentration of oxygen vacancies (V o s) is induced by Ar + irradiation to imitate synaptic vesicles containing neurotransmitters. Similar to vesicles, under stimulus, it is from the OVR that V o s move in the memristive layer to change the resistance of memristor (corresponding to the synaptic plasticity). The presented device exhibits uniform and reliable short-and long-term synaptic plasticity. The dynamic transition between sensory memory (SM), short-term memory (STM), and longterm memory (LTM) is successfully emulated. By using a 5 × 5 memristive synapse array, the experimental implementations of multi-pattern memorization and multi-value dynamic image masking have been realized with reliable synaptic plasticity of the memristive synapses. For extracting the multiple regions of interest (ROIs) with different priorities that occurred in the history image set, the desired memristive multi-value mask (MMVM) was automatically generated by applying different pulse stimulus schemes with different numbers of stimuli corresponding to the times of region occurred in the history, validating the feasibility of the MMVMs for practical applications in experience-based intelligent image processing. Figure 1a shows the schematic diagram of a memristor based on Ar + -irradiated SC-LNO thin film to mimic a typical axondendrite synapse in the cortex neural network. In this type of biological synapse, action potentials activated vesicles docked at the axon terminal (i.e., presynaptic membrane) to release neurotransmitters that transmit signals to the connected dendrite terminal. [41] Analogously, the top electrode (TE) and bottom electrode (BE) of the sandwich-like memristive device can be regarded as the axon terminal and dendrite terminal, respectively. Likewise, the functional layer (SC-LNO layer) can be equivalent to the substrate for neurotransmitter diffusion in the synaptic cleft ( Figure 1b). Significantly, as interpreted in Figure 1c, for resembling the vesicles docked at the axon terminal, a reservoir of the oxygen vacancies (OVR) was introduced on the surface of SC-LNO thin film by Ar + irradiation (see Figure S1, Supporting Information, and Experimental Section). [42,44] The drift process of V o s can simulate the release of neurotransmitters in synaptic vesicles. Based on X-ray photoelectron spectroscopy analysis, the occurrence of the peak associated with V o s [45] and a pair of new peaks corresponding to Nb 4+ (3d 3/2 ) and Nb 4+ (3d 5/2 ) [42,46] verified the formation of V o s through Ar + irradiation ( Figure S2, Supporting Information). In previous works, it has been proven that Ar + irradiation is an effective method to regulate the resistive switching behavior of oxide-based memristor by modulating the V o s in the oxide thin film. [47][48][49] Before mimicking synaptic functions, the electrical characteristics of the memristor are evaluated first. As presented in Figure 1d, during electroforming, the V o s in OVR drift toward the BE when the device current jumps suddenly, indicating the formation of V o channels (OVCs). [50][51][52] After electroforming, the device shows analog resistive switching behavior at dual sweep voltage. Figure 1e and Figure S3a (Supporting Information) display the reproducibility of device resistance (measured at a read voltage of 0.5 V) over 100 switching cycles. The results show very low cycle-to-cycle variation, only 3.32% for HRS and 1.47% for LRS. Besides, the device exhibits excellent uniformity with low device-to-device variation (down to 4.91%, Figure S3b, Supporting Information), which can be attributed to the high-quality SC-LNO thin film. The area dependence of the device's HRS and LRS was further investigated. As shown in Figure 1f, the resistances are independent on the electrode area, indicating the local filamentary conduction. [53][54][55] In addition, we validated that the resistance state of the presented device can be tuned continuously with 20 identical voltage pulses. As plotted in www.advelectronicmat.de Figure 1g, the current gradually increases after each pulse stimulus. This process is similar to the accumulation of Ca 2+ ions caused by repeated stimuli in biological synapses, which in turn leads to the potentiation of synaptic strength. These dynamic characteristics of the Ar + -irradiated SC-LNO memristor can be utilized to simulate Ca 2+ dynamics and further mimic the synaptic functionalities. [56,57] In the biological synapse, the postsynaptic potential generated by Ca 2+ influx refers to excitatory postsynaptic current (EPSC). [15] Its magnitude or intensity can be modulated by action potentials through altering the permeability of Ca 2+ influx in the presynaptic membrane, which is a significant synaptic behavior. [17] For evaluating the plasticity characteristics of the memristive synapse, a single voltage pulse (with pulse amplitude of 5 V and duration of 50 ms), which is equivalent to the action potential in biological synapse, is applied to the TE. Meanwhile, the postsynaptic current (PSC) is recorded by a sequence of read voltage pulses with an amplitude of 0.5 V. As shown in Figure 2a, the PSC rises rapidly from the initial state (≈1.65 µA) to the peak value, i.e., EPSC, and then enters a spontaneous decay process until it is relatively steady. The voltage pulse causes the clustering of V o s to form OVCs, which switches the device to a relatively low resistance state. [56] For the spontaneous decay process, it can be fitted by an exponential decay function:

Results and Discussion
where I(t) is the PSC value at the time of t and I 0 is the stabilized current. A is the pre-factor, and τ is decay time constant. Notably, the τ corresponds to the relaxation time of the decay process in biological synapse. [58] In this work, the τ of the memristive synapse is 7.43 ms. In addition, in order to analyze PSC clearly, the memory level is defined here as the magnitude of the relatively steady-state current, and it implies the residual OVCs in the Ar + -irradiated SC-LNO layer. [59] Based on the above results, we further investigated the synaptic plasticity of the memristive synapse. Figure 2b and Figure S4 (Supporting Information) display the EPSC, τ, and memory level under the stimulus of different single voltage pulses with different amplitudes. The pulse duration is fixed at 50 ms. Higher positive pulse amplitude causes an enhanced EPSC and a higher memory level, as well as a large τ. That is the spike-amplitude-dependent plasticity. In addition, the pulse duration could also tune the synaptic characteristics. Within a time scale of tens of milliseconds, EPSC, τ, and memory level www.advelectronicmat.de are rapidly enhanced with the increase of pulse duration. While the pulse duration exceeds hundreds of milliseconds, EPSC, τ, and memory levels are gradually saturated (Figure 2c and Figure S5, Supporting Information). This behavior may be due to the relative saturation of OVC size at a large pulse duration (>200 ms), which is similar to the postsynaptic receptor saturation of bio-synapses that contributes to STP. [31] Interestingly, the τ of the device is on the same order with the counterpart of bio-synapses, that is, tens of milliseconds. [18] These indicate that the memristive synapse shows abundant plasticity by tuning the pulse parameter, manifesting the potential implementation of learning strategies in neuromorphic systems. [60] STP reflects the dynamic activity for adaption and synaptic computation that take place on time scales of tens to hundreds of milliseconds. [58] In STP categories, two notable forms are affected by the interval (Δt) between paired pulses, i.e., pairedpulse facilitation (PPF) and paired-pulse depression (PPD). They refer that, when two pulse stimuli are applied successively, the second stimulus will generate a larger or a smaller EPSC than the first pulse, which plays a key role in decoding temporal information in auditory and visual signals. [18] To simulate STP in the memristive synapse, two groups of paired pulses with intervals of 10 and 25 ms are applied to the TE, respectively. Here, the pulse amplitude is set to 5 V, and the pulse duration is set to 10 ms. As shown in Figure 3a, the EPSC of the second pulse is enhanced more significantly when the interval between paired pulses is shorter (10 ms). It indicates that the V o s in the OVC triggered by the first pulse are unable to completely diffuse, and a shorter spontaneous decay process leaves thicker residual OVC. When the second pulse arrives, the thicker residual OVCs are involved in the reconstruction process and facilitated a larger EPSC. This behavior is analogous to the dynamics of residue presynaptic Ca 2+ concentration induced by previous stimulus in biological synapses. [14] In the same fashion, the change of synaptic weight (Δw), corresponding to the increment of EPSC, is also found to decrease with the increase of pulse interval (Figure 3b).
In addition, successive pulses with shorter intervals can significantly increase the Δw. Conversely, a long interval leads to a decrease in increment or even a decrease in EPSC (Figure 3c). It is worth mentioning that the memristive synapse shows excellent uniformity of Δw modulation with low device variation (randomly measured from five memristive synapses, Figure S6, Supporting Information), which are much lower than the counterpart of synaptic devices based on the migration of Ag + . [26] By using this memristive synapse, we demonstrated PPF and PPD by applying a pulse sequence with frequency of 5 Hz, 20 Hz, and 100 Hz. As shown in Figure 3d, the EPSC increases following high-frequency stimuli (100 Hz) (facilitation), while it experiences a decrease once the stimuli frequency drops to 5 Hz (depression). Moreover, the effect of a pulse sequence with the same frequency of 20 Hz on EPSC is converted from facilitation to depression after the experience of stimuli with frequency of 100 Hz, which exhibits the intriguing adaptivity of the memristive synapse. This important feature of biological synapses, i.e., adaptation of sensory information processing, [18] which has not been clearly demonstrated in previously memristive synaptic devices and can promote the realization of adaptation of robots. [61,62] Unlike the temporary modifications of synaptic weight in STP, long-term potentiation (LTP) plays its role in learning and memorizing by changing the synaptic weight more persistently. LTP can be realized with strong or high-frequency stimuli, and the retention of synaptic weight lasts from hundreds of milliseconds to days or even longer. [63] In our memristive synapse, the LTP property was explored by applying identical repeated stimuli and recorded the dynamic PSC after www.advelectronicmat.de N stimuli. As shown in Figure 3e, the dynamic evolution of normalized memory level represents the phenomenon of LTP. After undergoing high-frequency stimuli, the PSC enters a permanent potentiation process, and the normalized memory level increases with the increase of the number of stimuli, indicating the OVCs in SC-LNO layer are thickened by repeated voltage pulses. [28] This result is analogous to the synaptic behavior in which the synaptic strength can be enhanced by increasing the number of repeated stimuli. [64] In addition, these characteristics of STP and LTP are well correlated to the human memory in the brain, that is from sensory memory (SM) to short-term memory (STM) and eventually to long-term memory (LTM), which are widely accepted models in psychology (Figure 3f). [65,66] In this model, information from the environment is first encoded in SM. If the information is perceived, it is processed further into STM. Moreover, if the information in STM needs to be www.advelectronicmat.de preserved, it can be transformed into LTM by repeated rehearsal or training. Analogously, the dynamic process of the simplified memory model is simulated in our memristive synapse. As illustrated in Figure 3g, when the EPSC triggered by new information is higher than the perceived threshold (2.0 µA), information is stored in STM, whereas repeated rehearsal events result in LTM. The relaxation time constant, τ, and memory level increase during repeated rehearsal, indicating the memory consolidation. Through this bionic feature, the memristive synapse can be used to develop artificial adaptive sensory receptors that filter out innocuous information, when the EPSC triggered by environmental stimulus does not reach the perceived threshold, and encode critical stimulus, which would significantly improve the energy efficiency for information memorizing and processing. [62,67] To verify the memory capability concretely, multi-pattern memorization comprising "E", "S", and "D" is experimentally realized in 5 × 5 memristive synapse arrays (Figure 4). Before applying stimuli, all memristive synapses in the array are set to a uniform initial state after electro-forming, and show a 100% device yield during programming and erasure operations ( Figure S7, Supporting Information). Furthermore, under diverse device operations, the PSC values of all memristive synapses possess excellent device-to-device uniformity, i.e., ultralow variations (Δ/μ), with 0.67% for initialization, 1.44% for programming, and 1.47% for erasure ( Figure S7b, Supporting Information).
The pattern of "E", which is programmed by applying two groups of 10 stimuli with different pulse durations (10 and 500 ms) and a fixed pulse interval of 10 ms (Figure 4a). First, the pattern of "E" is temporarily stored in the synapse array soon after applying the group of 10 repeated voltage pulses with a duration of 10 ms, but it is dramatically degraded after 60 s exhibiting the STM behavior. However, after applying the second group of ten successive voltage pulses with a larger duration of 500 ms, the image is still preserved firmly over 60 s, demonstrating the dynamic transition to LTM (Figure 4a). The pattern of "S", which is programmed by applying two groups of 10 stimuli with different pulse intervals (40 and 2.5 ms) and a fixed pulse duration of 50 ms (Figure 4b). Similar transition of STM-to-LTM observed in the memorization process of pattern "E" is also demonstrated in the pattern of "S" by applying two groups of 10 pulses with different intervals (i.e., 40 and 2.5 ms) (see Figure 4b).
As for the pattern of "D", as presented in Figure 4c, the 5 × 5 synapse array successfully demonstrates the transition from SM to STM, and finally to LTM by synchronously increasing the pulse duration and decreasing the pulse interval. Meanwhile, the dynamic PSCs of the three memory modes are extracted and presented in Figure 4d (the memristive synapses marked  Figure 4c with dotted boxes). In SM mode, the PSC to the weak stimulus (with a duration of 10 ms and an interval of 40 ms) never exceeds the perceived threshold (2.0 µA, as shown in Figure 4d (left graph)), hence the image ("D") is almost invisible. With the enhancement of stimulus, the PSC rises above the threshold and enters into STM, but as time goes by (after 60 s) it gradually decreases to about 2.0 µA, indicating the forgetting process in STM (Figure 4c,d (in-between graph)). When the stimulus is adequate for LTM, the PSC remains at a high level, even with spontaneous decay (Figure 4d (right graph)). In this mode, the image ("D") is stored persistently (Figure 4c). Above demonstrations of memorization model exhibit that the SC-LNO synapse possesses extensive memorizing and learning capabilities.
Furthermore, based on the reliable short-and long-term plasticity of SC-LNO-based synapse, we proposed a dynamic image masking strategy to demonstrate its application in image processing. The synaptic plasticity of memristive synapses can be used to generate a mask according to the different external stimuli, while PSCs are mapped to the matrix values of mask (M(i,j )). Then by multiplying the image pixel matrix (I(i,j )) with the mask matrix, i.e., G(i,j ) = I(i,j ) × M(i,j ), regions of interest (ROIs) can be extracted, while other incurious regions (ROB) can be filtered out, as shown in Figure 5a. Here, the value of I(i,j ) is the gray value at each pixel of the image. In addition, to extract regions with different interest levels, it can generate memristive masks with different mask values. A memristive multi-value mask (MMVM) based on the three memory modes (as described in Figure 4c,d) was first demonstrated. As shown in Figure 5b, the PSC of the memristive synapse can be modified to expected status by different stimulus schemes, due to its inherent non-volatile dynamic characteristics. As presented in Figure 5c, the obtained PSCs of the three memory modes are normalized to [0,1] (normalized generating process is described in Figure S8, Supporting Information), and the 5 × 5 multi-value MMVMs, i.e., MASK-SM, MASK-STM, and MASK-LTM, are generated. Notably, the memristive synapses without stimuli are at the initial state (about 1.65 µA), corresponding to a value of "0" in the mask. Furthermore, we would like to highlight that the reliable synaptic plasticity of SC-LNO memristive Schematic diagram of stimulus scheme and the corresponding mask for multiple ROIs with different priorities (right column). The stimulus schemes are: P 1 (Duration = 500 ms, Δt = 2.5 ms, number = 10), P 2 (Duration = 500 ms, Δt = 2.5 ms, number = 5), and P 3 (Duration = 500 ms, Δt = 2.5 ms, number = 1). g) Input image included multiple ROIs with different priorities and the masked images after processing by the generated MMVM. The priority of three ROIs (ROI #i, i = 1, 2, 3) is P 1 > P 2 > P 3 .

www.advelectronicmat.de
synapse and excellent device-to-device uniformity are guarantee for mask generation.
Based on aforementioned strategy, a simulated MMVM generated by a 256 × 256 SC-LNO synaptic array is used for extracting ROI in a grayscale image (256 × 256 pixels, Figure 5d). The values in the mask matrix according to the pixel of ROI are set as 0, 0.3, and 1, respectively, while the values of other regions (i.e., ROB) were set as "0" (see Figure S9, Supporting Information). As shown in Figure 5e, the ROI is effectively extracted with the masks, while ROB is filtered out. More importantly, MMVM can not only effectively extract ROI, but also reflect the interest level of the region according to the gray level of the masked image, which is of great significance for the image processing tasks. The multi-valued mask generated based on different pulse stimulus schemes provides feasibility for processing multiple ROIs.
Practically, in image processing tasks, there are often multiple ROIs with different priorities in an image, which also represents the relative importance of each ROI in the image. In order to realize experience-based mask generation based on the images received in history, the mask generated by applying different number of pulse stimuli which is corresponding to the times of region occurred in a history image set. Specifically, there is a history image set with 10 grayscale images for the learning process (see Figure S10, Supporting Information). In the image set, the times of the three regions occurred are 10, 5, and 1, respectively. For generating the desired MMVM, the pulse stimuli which matches with the number of occurrences, are applied to the memristor in SC-LNO synaptic array at the corresponding pixel. Thus, the PSCs after different numbers of stimuli can be mapped to the mask matrix (Figure 5f ). By normalizing from experimental results, the matrix values in the mask are specified as follows, 1, ROI #1, stimulus number is 10 0.9, ROI # 2, stimulus number is 5 0.8, ROI #3, stimulus number is 1 0, ROB, stimulus number is 0 As shown in Figure 5f (right column), the experiencebased MMVM for extracting the three ROIs (labeled as ROI #i, i = 1, 2, 3) with different priorities (P 1 > P 2 > P 3 ) is automatically generated in the simulation based on the test data of practical synaptic plasticity. After mask operation, three ROIs in the input image are effectively extracted according to their priorities (Figure 5g). In addition, if the times of the three regions occur in the history image set are same, the desired mask can be obtained by applying the same pulse stimuli scheme to the memristive synapses corresponding to the pixel of three regions, and then the multiple ROIs with the same priority in the image (P 1 = P 2 = P 3 ) can be effectively extracted (see Figure S11, Supporting Information). These results indicate that experience-based image mask generation can be realized with our memristor with reliable synaptic plasticity, showing great application potential of our memristive device in experience-based intelligent image processing.

Conclusion
We have demonstrated an SC-LNO-based memristor as a memristive synapse. A reservoir of oxygen vacancies (OVR) is induced on the surface of SC-LNO layer by Ar + irradiation to imitate the synaptic vesicles containing neurotransmitters. This memristive synapse with bioinspired microstructure successfully mimicked synaptic functionalities, such as the phenomenon of saturation and adaptivity, STP, PPF, PPD, and LTP. These functions are realized by the modulation of OVCs in the SC-LNO layer, which resembles the dynamical behaviors in biological synapses. By using a 5 × 5 memristive synapse array, the dynamic transition from SM to STM and further to LTM, is also successfully emulated for multipattern memorization. The synaptic characteristics in the memristive synapse are quite stable with tiny fluctuations and low device variation. Furthermore, an effective multi-value dynamic image masking strategy based on the SC-LNO synapse is proposed. By taking advantage of reliable synaptic plasticity, the experience-based image mask generation based on the images in history has been successfully realized. The generated MMVM successfully extracts multiple ROIs with different priorities. These results demonstrated that the reliable SC-LNO synapse has a great potential to be used in future neuromorphic systems for image processing and experience-based learning.

Experimental Section
Device Fabrication: The high-quality SC-LNO thin films were developed by the Crystal-ion-slicing technique ( Figure S1, Supporting Information). i) A 500 µm-thick Z-cut SC-LNO wafer was used for substrate material. ii) A buried damaged layer (about 600 nm) was created by high-energy He + deep-ion implantation with energy of 300 keV. The fluence of He + ions was 8 × 10 16 cm −2 . iii) After implantation, a Pt (120 nm) layer, as the BE, was deposited by sputtering on the implanted wafer. Then, SiO 2 bonding/insulating layers were deposited on the implanted wafer and another pristine SC-LNO wafer by plasma-enhanced chemical vapor deposition. iv) The two SC-LNO wafers were directly bonded together at room temperature via hydrophilic bonding. v) An SC-LNO thin layer was lifted off under thermal annealing treatment at 250 °C. vi) After that, a typical thickness (300 nm) of the SC-LNO thin film was fabricated after fine chemical mechanical polishing. vii) Then, the interfacial engineering was conducted on the as-prepared SC-LNO thin film via low-energy Ar + irradiation. Here, the acceleration voltage was 70 eV. After irradiation for 16 mins, the thickness of SC-LNO thin film was reduced from 300 to 120 nm. Finally, the Au circular electrodes (200 nm in thickness) with a diameter of 200 µm were deposited on top of the vias by a magnetron sputtering method. viii) Therefore, the final configuration of the presented Ar + -irradiated Au/SC-LNO/Pt memristive device was realized.
Characterization and Methods: The morphology and cross-section structures of the developed devices were investigated by scanning electron microscopy (Gemini) and transmission electron microscopy (Jeol JEM-2100). The current-voltage properties (I-V) of the device were measured by using a Keithley 2400 SourceMeter. Besides, the synaptic characteristics of the memristive synapse were characterized in dark ambient conditions by a semiconductor analyzer (Keithley 4200) with the aid of a probe station.

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