Flexible Solution‐Processable Black‐Phosphorus‐Based Optoelectronic Memristive Synapses for Neuromorphic Computing and Artificial Visual Perception Applications

Being renowned for operating with visible‐light pulses and electrical signals, optoelectronic memristive synaptic devices have excellent potential for neuromorphic computing systems and artificial visual information processing. Here, a flexible back‐end‐of‐line‐compatible optoelectronic memristor based on a solution‐processable black phosphorus/HfOx bilayer with excellent synaptic features, toward biomimetic retinas is presented. The device shows highly stable synaptic features such as long‐term potentiation (LTP) and long‐term depression (LTD) for repetitive 1000 epochs, having 400 conductance pulses, each. The device presents advanced synaptic features in terms of long‐term memory (LTM)/short term memory (STM), as well as learning–forgetting–relearning when visible light is induced on it. These advanced synaptic features can improve the information processing abilities for neuromorphic applications. Interestingly, the STM can be converted into LTM by adjusting the intensity of light and illumination time. Using the light‐induced characteristics of the device, a 6 × 6 synaptic array is developed to exhibit possible use in artificial visual perception. Moreover, the devices are flexed using a silicon back‐etching process. The resulting flexible devices demonstrate stable synaptic features when bent down to 1 cm radius. These multifunctional features in a single memristive cell make it highly suitable for optoelectronic memory storage, neuromorphic computing, and artificial visual perception applications.


Introduction
The development of neural networks in the visual cortex of the human brain supports the memory-based nature of human DOI: 10.1002/adma.202300446 vision. [1,2] The advancement of artificial vision systems for autonomous control by recognizing and processing real-time visual information has encouraged this ecological visual perception. Similar to the biological visual system, the artificial vision system is made up of photodetectors to gather visual data, a memory unit to store the imaging data, and a processor to process images, carry out neuromorphic computations, and recognize any object. In order to deploy effective brain-like vision systems, it is essential to realize a single imaging unit with a combination of built-in memory and signal processing capability. Due to their innate optical sensitivity and memory properties, optoelectronic memories and synaptic devices have been explored as viable devices for artificial visual systems. [3,4] These devices have shown the ability to perform image pre-processing and neuromorphic computation for machine vision in more recent inventive breakthroughs. [5,6] The conventional complementary metal-oxidesemiconductor (CMOS) artificial vision systems have significant challenges such as high-power consumption, delayed data access, and hardware termination due to the separation of its functional components including sensory terminals, computing, and memory units, [7][8][9][10][11][12] which pose significant obstacles to their ability to adapt to future neuromorphic systems. Motivated by the human vision architecture, optically functional neuromorphic memristive devices have been developed for their usage in neuromorphic vision sensors. [5,[13][14][15][16][17] Those devices have shown great attention for the neuromorphic vision sensors because of the combination of optical sensing, memory, and processing functions in the same optical memristive device. [18,19] These excellent features in the photonic/optical memristor reduce the computational latency and power consumption. [20,21] Among many interesting and promising materials, lately, layered black phosphorus (BP) has aroused great attention from the optoelectronic device community due to its exciting properties such as high carrier mobility and a strong in-plane anisotropy in its electronic properties. As a result, several optical and electrical synaptic devices based on black phosphorus have been reported which could lead to many exciting advances in bio-inspired electronics. Nevertheless, most of these devices use various exfoliation techniques to transfer the 2D material onto the substrate. Although such approaches do not require a high thermal budget, their main drawback is that large sized crystals of the 2D material are needed for obtaining nanosheets with large dimensions. On the other hand, solution processable techniques such as spin coating and drop casting allow for a large area coverage while being low cost with a low thermal budget, which is necessary when dealing with highly sensitive optical components on a chip, as well as being critical for allowing compatibility with back-end-of-line devices. In this work, we report a flexible bilayer HfO x /solution processable BP based optoelectronic memristive device that can display both electrical and optical synaptic features in a single memristive cell. The electrical synaptic features such as long-term potentiation (LTP), long-term depression (LTD), short-term plasticity (STP), and paired pulse facilitation (PPF) are effectively accomplished by using a series of voltage pulses. More analyses display that, by tuning the interval time of visible-light pulses, the light illuminated STP can be accomplished. Importantly, the advanced neuromorphic features like photo-synaptic-current (PSC), photonic PPF, short-term memory (STM), long term memory (LTM), and learning-forgettingrelearning process are achieved successfully. Using the illumination of 405 nm wavelength light, the photonic memristor array can well recognize the mimicry of human visual perception and visual memory by tuning intensities and exposure time of light. These results confirm that the photonic memristor displays a promising ability to build up human visual systems. Table 1 shows a comparison between the resistive switching characteristics of the demonstrated Cu/HfO x /BP/Pt optoelectronic synaptic device in this work and previously reported 2D-materials-based synaptic devices.

Results and Discussion
The schematic structure of the biological human vision system as well as the neuromorphic vision system of optoelectronic memristive devices that are capable of sensing and processing data is illustrated in Figure 1a. The visual cortex and retina, both, are the leading components of the human visual system and are responsible for a variety of visual functions. The eyeball's retina captures the information of the image and transforms it into electrical signals. These electrical signals then get transmitted to the visual cortex via an optic nerve. And finally, the visual cortex interprets and processes the visual information which is received from the eyes. [22][23][24] The proposed device in this work, which can emulate the functions of the retina and visual cortex in terms of detecting optical signals/images and processing them, is based on a memristive structure with a bilayer of HfO x /solution processable BP flakes. The fabrication process flow is explained in detail in the experimental section. Moreover, extensive structural characterization is performed to confirm the composition of the fabricated devices. More specifically, the HR-TEM with EDS spectrum was used to confirm the cross-sectional images (scales: 200 and 100 nm) with layer-by-layer thickness and elemental profile of the Cu/HfO x /BP/Pt device, as shown Figure 1b-d respectively. The EDS line profile validates the presence of copper (Cu), phosphorus (P), oxygen (O), hafnium (Hf) and platinum (Pt) in the device. Various Cu, P, Hf, O, and Pt element mapping validate the presence of a multilayer structure, as shown in Figure 1e. From the XRD spectra in Figure 1f, the diffraction peaks at various angles confirm the presence of BP. [25,26] The XPS spectrum also validates the surface chemical composition and chemical valence of BP, as shown in Figure 1g. In fact, the phosphorus P 2p photoelectron spectrum exhibits the spin-orbital splitting doublet located at 129.5 (2p 3/2 and 131.2 (2p 1/2 ) eV, which are characteristic of black phosphorus. [27,28] The various electrical characterizations were done to determine whether the introduction of the BP flakes in the HfO x /BP bilayer device shows superior performance as compared to the HfO x single layer device. Figure 2a-f illustrates the I-V, DC endurance, and voltage distribution for HfO x based and HfO x /BP devices respectively. The other electrical features for both HfO x and HfO x /BP devices are shown in (Note S2, Supporting Information). The electrical characteristics confirm that the HfO x /BP bilayer device has superior characteristics than HfO x single layer one. The current compliance of 1 mA was used during the SET operation in both devices. More specifically, the threshold voltage distribution of the HfO x /BP ( Figure 2f) device is smaller and more concentrated than the HfO x one ( Figure 2c). Moreover, the threshold voltages of HfO x (inset of Figure 2c) and HfO x /BP (inset of Figure 2f) of 100 cycles were extracted and the distribution histogram of the threshold voltage was obtained (to extract V set and V reset from the HfO x and HfO x /BP devices, the first 100 continuous I-V switching cycles were taken from both devices and noted their corresponding V set or threshold voltages and V reset ). In HfO x device (Figure 2c), the wider dispersions in the V set and V reset voltages are attributed to the stochastic growth and breach of conductive filament (CF). Due to the wider distribution in Vset and Vreset, the HfO x device shows poor endurance (100 cycles only) with wider fluctuations in both LRS and HRS (Figure 2b). [29,30] The effective RS region is restricted in the HfO x /BP device so that the HfO x /BP device possesses sharp distributions of V set and V reset , leading to the stable RS characteristics ( Figure 2f). [30][31][32] Therefore, the HfOx/BP device depicts good endurance of at least 1000 cycles without any degradation with highly stable LRS and HRS. The result enables the precise control of the V set and V reset programming voltages for neuromorphic computing applications. Due to the excellent characteristics of HfO x /BP device, further characterization of the HfO x /BP device is conducted as will be explained below.
The multistate feature of the memristive device was measured using the reset stop voltages from −1. The non-volatility of data storage capability was confirmed by a retention test in the memristive device as shown in Figure 2h. The multilevel resistance values in both LRS and HRS are stable with the on/off ratios of around two orders in magnitude, and no significant degradation is observed for more than 10 4 s and are predictable up to 10 years of lifetime. [33,34] The high-speed dynamic pulse induced RS properties by using a pulse height of 0.9 V for the SET condition and -0.8 V for the RESET condition under the pulse width of 100 ns. Figure 2i displays the AC durability of the device which can well maintain its both LRS and HRS states for more than 10 6 cycles without any degradation. The read voltage of 0.2 V (can use 0.1 and 0.3 V also) was used to measure the multilevel DC endurance, AC endurance as well as for the multilevel retention test. Figure 3 schematizes the switching mechanism for Cu/HfO x /BP/Pt memristive device. Metal filaments develop from the top electrode (Cu) to a bottom electrode (Pt) during the forming, which involves the Cu/HfO x /BP/Pt being exposed to a positive voltage for the first time [29,30] As a result, the Cu filament is created in the HfO x layer by the Cu oxidation process, Cu ion transport, nucleation, and Cu filament growth. Identifying the variation in Cu ion transport rates between HfO x and BP layers assist in predicting the subsequent filament growth process in BP. The HfO x /BP of the solid electrolyte will enable the formation of an hourglass-shaped Cu filament. [31,32] The HfO x layer significantly reduces the Cu ion transport, which results in a low ion transport rate in the oxidized layer, as shown in Figure 3b (forming 1-1). For the Cu filament to grow continuously, the nucleation and transit of Cu ions must be in balance. After the filament spreads over the HfO x and BP interface and the Cu filament diameter steadily grows (Figure 3c, forming 1-2), the Cu ion transmission rate significantly increases in the BP layer due to its high thermal conductivity [35] than HfO x . [36] The growth of a complete Cu filament causes the HfO x /BP device to enter a low resistance state (LRS) after the filament reaches the bottom Pt electrode. Additionally, the interface of HfO x and BP is probably the weakest part of the whole filament due to differing ion-mobility characteristics. [37,38] The filament ruptures at its weakest point when a proper negative voltage is used because of the Joule heat produced by Cu diffusion or oxidation, which causes the switching of the device from LRS to high resistance state (HRS), which is depicted in Figure 3d. Moreover, thermal conductivity is the primary factor to decide the location of the CF rupture during reset process. In the Cu/HfO x /BP/Pt device, there is an additional interface between HfO x and BP films. Thus, the thermal conductivity of the BP film in the device will also affect the reset process during negative bias. The thermal conductivity of BP and HfO x are 12 and 0.68 W m −1 K −1 , respectively. [35,36] So, the heat transfer from the filament region to the BP layer will be higher than HfO x layer because of that the thermal conductivity of BP is higher than HfO x . During the reset, the higher thermal conductivity of the BP layer increases the thermal dissipation through the bottom electrode interface, which pushes the maximum temperature point of CFs away from the bottom electrode and toward the HfO x /BP interface of the device. Therefore, the rupture of CF takes place easily near the center of the CF through the assistance of the local Joule heating, leading to the partial rupture of CF with the gap formation in the device (shown in Figure 3d). The HRS becomes the LRS when positive voltage is applied to the Cu top electrode, which causes the Cu ions to accumulate at the bilayer interface and form a full CF, shown in Figure 3e. When the HfO x /BP device is cyclically converted between the positive and negative voltages, this displays a typical bipolar switching. After the Forming process, the Cu filament nucleation/growth zone is condensed and largely fixed, which significantly reduces the variation of CF in the HfO x /BP device. The SET and RESET operations of HfO x /BP device take place at HfO x /BP interface; as a result, the switching voltage of the device is comparatively decreased, and reliability rises. [30,32,37,38] Therefore, the HfO x /BP has excellent characteristics as compared to the HfO x based device.
The resistive switching mechanism of the Cu/HfO x /Pt device is also explained in Figure 4. Figure 4a shows the schematic structure of the fresh device. During the forming process, when a positive bias is applied to the Cu top electrode, the Cu atoms are oxidized and converted into the Cu ions. The generated electric field by the applied positive voltage between top and bottom electrodes drifts the ionized Cu atoms from the top electrode toward the bottom electrode, resulting in the formation of Cu filament, as shown in Figure 4b. During the reset process, when a negative voltage is applied to the Cu top electrode, a high current flows through the CF, and this results in Joule heating effects, causing the complete rupture of the CF from the bottom electrode. [39,40] Moreover, when the negative voltage is applied on the Cu top electrode, it also causes an electrochemical reaction where Cu ions drift back to the Cu top electrode, changing the low resistance state (LRS) of the device to the high resistance state (HRS), as displayed in Figure 4c. Hence, the Cu/HfO x /Pt device has the higher resistance value of the HRS, which is shown in Figure 2a,b. However, the complete rupture of the CF would re-sult in the position change of the subsequent CF growth during switching cycles, which is shown in Figure 4b,d, causing unstable resistive switching properties and wider fluctuations in V set and V reset . [30,39,40] The wider voltage fluctuations in the V set and V reset degrades the device performance. As a result, the Cu/HfO x /Pt device shows poorer characteristics.
To simulate the bionic synaptic plasticity, voltage pulses were applied to the memristive device, as shown in Figure 5a. The consistent LTP and LTD were accomplished using the repeated 400 identical positive voltage pulses (0.9 V/10 μs) and subsequent 400 identical negative voltage pulses (−0.8 V/10 μs). The result confirms that the synaptic weight (conductance) of the device can be enhanced or reduced by applying positive and negative voltage pulses. The voltage pulse schemes for emulating synaptic functions are mentioned in Note S4 (Supporting Information). We also measured our device using various identical positive/negative voltage pulses, as shown in Figure 5b. As shown in Figure 5c, the linear conductance response of the device with ON/OFF conductance ratio of more than 100, which is suitable enough for accurate training in MNIST pattern recognition. [41][42][43] The conductance of the device was increased with increasing the voltage pulses. The PPF is one significant part of synaptic plasticity that plays a crucial role in the accomplishment of superior learning and memory, as shown in Figure 5d. Consecutive electrical pulses are used to achieve PPF index by calculating (A2−A1/A1) × 100, which shows stable and gradual decrease in PPF index mimicking mammalian synaptic function. After applying presynaptic voltage, the change in excitatory current is termed as A1. The increase in current for second pulse as postsynaptic pulse in neuron is termed as excitatory postsynaptic current (EPSC). [44,45] The increase in current after applying postsynaptic spike is shown as A2. To confirm the feasibility of the memristive synapse, the LTP and LTD is achieved in form of conductance (synaptic weight) states. Figure 5e,f shows the first 10 and last 10 LTP/LTD cycles out 1000 cycles. All 1000 epochs data is shown in the Note S5 (Supporting Information). Figure 5g demonstrates the artificial neural network (ANN) implemented with electrical device synapse for the image classification task. The ANN architecture is formed by three layers with 28 × 28 input neurons, 100 hidden neurons, and ten output neurons respectively. The ANN model is trained and tested on the Modified National Institute of Standard and Technology (MNIST) database, where input image has the same size of 28 × 28 and the pattern recognition results (0-9) are presented through the ten output neurons. The experiment result is shown in Figure 5h. After 150 training epochs, the accuracy becomes stable at ≈90.16%, which is slightly lower than the software default (97.17%). Figure 5i presents the confusion matrix for initial state and after training 20 and 40 epochs respectively. The simulation details of ANN are mentioned in (Note S6, Supporting Information). The matrix transforms from chaos to standard, each line represents the decrease of recognition mistake and knowledge gradually learnt by electrical device synapse. The experimental result highlights the learning ability of our electrical device synapse and its future application for image recognition in neuromorphic computing.
Besides the electrical synaptic features, we proposed that our device also shows the change in PSC when light is illuminated on it. Here, we study the optoelectronic synaptic features of the device using a series of optical light pulses with various intensities and time durations, as schematically demonstrated in Figure 6a. In the human brain, pre-synaptic neuron relates to the post-synaptic neuron. Our two-terminal optoelectronic memristive synapse is like a biological synapse in the human brain. Figure 6b shows the light illuminated PSC response of the device when optical light (405 nm) with 68 mW cm −2 was induced for 5 s. The PSC of the device was increased to 57 μA with an obvious PSC change of ≈9 μA. This obvious increment in the PSC is ascribed to the photoconductive effect of HfO x and BP within the device. After removing the optical light, the device shows the gradual decay in the photocurrent instead of recovering initial current quickly, that could be attributed to the persistent photoconductivity [46] This PSC decay could be ruptured by applying the negative voltage pulses. As depicted in Figure 6b, the decay current of the device drops suddenly from the 57 to 49 μA (initial state) when −1 V voltage pulse with 1 ms is induced at 50 s, confirming the photonic potentiation and electrical depression (elimination). We further examine the light irradiated synaptic plasticity (PPF) of the device, as shown in Figure 6c. The photonic PPF was studied when the time interval (ΔT) between two identical optical pulses is varied. The PSC (A2) of the device after the second illumination with the intensity of 68 mW cm -2 (time duration: 5 s, interval: 10 s) is higher than that the PSC (A1) after the first illumination with the same light intensity, which is closely related to ΔT. The photonic PPF can also be calculated using the same equation as electrical PPF, as presented in Figure 6c. The curve displays superb mimicking of synaptic photonic PPF feature that is valuable synaptic function when it comes to neuromorphic vision systems. [43,47] The PSC of the memristive device by applying the different levels of intensities (dark, 32, 48, 57, and 68 mW cm −2 ), as illustrated in Figure 6d. The higher intensity indicates larger current conductance, and its decay also depends on the intensity of light.
The light with smaller intensity shows slight enhancement in the current conduction and decaying to its initial state after remov-ing the power, which is like the STM in human brain. Once the light is removed at the higher level of intensity, the PSC is decayed gradually and maintains at a high level above the initial state, which behaves like an LTM in the human brain. The decay rate of the LTM is demonstrated in Figure S6 (Supporting Information), which confirms that the high-level PSC of the device can be retained more than 350 s. Furthermore, we also measured the transformation from STM to LTM by modifying the illumination time (2, 3, 4, and 5 s) of light, which is displayed in Figure 6e. The higher illumination time of light indicates high PSC and slower decay rate. In short, the PSC response of the optoelectronic memristive device could be converted from STM to LTM by adjusting either intensity of light or light illumination time, suggesting the strong ability in mimicking superior memory functions for visible light. The opto-electronic synaptic features of Cu/HfO x /Pt single layer device are shown in (Note S7, Supporting Information). The working mechanism of the Cu/HfO x /BP/Pt optoelectronic memristive device mainly relies in providing charge trapping centers for photogenerated charge carriers (electrons/holes), which store the photogenerated electrons/holes even after the removal of optical light stimuli. In this case, the optical information can be stored in the Cu/HfO x /BP/Pt memristive synaptic device because of that the light intensity defines the amount of photogenerated charge carriers, which is generally related to the possibility of charge trapping at the HfO x /BP interface. [48][49][50][51][52] Therefore, tuning the light intensity provides an approach to achieve multilevel resistance states in the Cu/HfO x /BP/Pt device. Furthermore, light stimulation with various intensities correspond to different optical excitation energies and different amount of light absorption within the HfO x /BP interface, which may also induce to different charge trapping/detrapping and multiple resistance states. The excellent optoelectronic synaptic characteristics of the Cu/HfO x /BP/Pt device can be attributed to the interface barrier between HfO x and BP, which prevents the recombination of photoexcited electron-hole pairs to increase the carrier lifetime and realizes the nonvolatile high current state after removing the light stimulus. Atkinson et.al., presented a multistore model for human memory. This model is still a highly established model in psychology. [53,54] This model states that new information is stored for a very short time in the sensory register ((sensory memory (SM)), and then that information is transmitted from short-term storage (STM) to a permanent long-term storage (LTM)), as shown in Figure 6f. It is believed that STM could be converted to the LTM due to the repetition of rehearsal process. We also exhibit the memorization level, which is depicted in Figure 6g. The memorization level in SM is increased slightly with initial rehearsals, and then, in STM the memorization level is increased temporarily before decaying. Following repeated rehearsals are observed to result in LTM, leading to permanent memorization. We have also shown the learning forgetting process for our optoelectronic memristor by applying the continuous turning ON/OFF the light, which is depicted in Figure 6h. Here, turning ON the light represents the learning and relearning behavior in the device while turning OFF the light symbolizes the forgetting behavior. The photocurrent of the device is increased with light exposure and then degrades to its intermediary level after a given period, indicating that the learned information is slightly forgotten with time. After continuous repetition of the learning or relearning process, the PSC conductance of the device is enhanced marginally (A7 > A6 > A5 > A4 > A3 > A2 > A1), representing that the earlier learned information can considerably boost the memory ability. The PSC conductance of the device reaches its maximum level (A7) after seven continuous times of the learning and relearning process, indicating the transformation from STM to LTM is achieved in the optoelectronic memristive device. This excellent repeatability in PSC response can ease the simulating of the superior synaptic function in synaptic plasticity, which is represented by the "learning-forgetting-relearning" process.
In visual memory, it is noticeable that memory mode can be strengthened by increasing the number of cycles (ON/OFF) or giving stimulus over time, which is also named memory recall. [55] Similar to the human visual system, our device has shown the excellent property of sensing optical vision stimulus with the corresponding conductance change, which can be viewed as a vision memory. Based on the device learning and forgetting process, the vision memory function is demonstrated in a 6 × 6 optoelectronic memristor array, as illustrated in Figure 7a. The training process is to use 405 nm light as the stimulus to remember the special pattern. As shown in Figure 7b, the densities of regions I, II, III, IV, and V are stimuli with different light intensities at 68, 57, 48, 32 mW cm −2 and dark, respectively. To be mentioned, in region VI, optical signals are designed to discretely give in different areas to introduce variations among different light intensities. Figure 7c illustrates the  (2,3,4, and 5 s), followed by photocurrent decay when light is turned off. f) Multi-storage psychological paradigm of the human memory system, which consists of three kinds of memories: sensory memory (SM), short-term memory (STM), and long-term memory (LTM). This is also a learning-forgetting-relearning procedure. g) Memory strength model in the device, which is inspired by the multi-storage model. First, the new information is stored in SM and then the information is stored in STM for a short time, after repeated rehearsals the STM converts into the LTM. At the lower repeated rehearsals, the STM could not convert into the LTM, shown in red line, whereas higher rehearsal repetition forms STM to LTM, depicted in blue line. h) The learning-forgetting-relearning procedure for seven cycles.  conductance state change with potentiation and depression state according to the different optical stimulus period. The results indicate that memory strength happens with the increasing optical stimulus time, which brings information storage among the device (potentiation). On the contrast, the conductance decrease happens after removing the light source, also named as the depression state. By tuning the light stimulus, it can modulate the device conductance efficiently through the "learning-forgettingrelearning" rules without any additional voltage pulse, which effectively emulates the human visual perception property. The inherent light sensing ability makes it suitable to operate and sense different light intensities along with duration time, which makes it a candidate for future artificial visual perception.
To further empower the optoelectronic synapses and allow them to take different shapes and configurations for various applications including wearables, the device must perform good stability under bending conditions. It is worth to note that all the active layers in the device are considered to be biocompatible, except for the substrate silicon which requires special packaging in this case. Figure 8a shows a photograph of the flexible device during measurements. The flexible device shows stable LRS/HRS values when bending radius was changed from flat to 1 cm, which is depicted in Figure 8b. Figure 8c depicts the DC endurance of the flexible device for 1 cm bending. The device shows stable performance for 100 cycles without any degradation. The synaptic feature (LTP/LTD) was also measured for the flexible device when bent with 1 cm bending radius. The device depicts 25 epochs using 400 potentiation pulses and 400 depreciation pulses (Figure 8d). These results confirm that the flexing process did not degrade the performance of the synaptic devices.

Conclusion
The optoelectrical synaptic behavior of the HfO x /BP could be the hardware implementation of neuromorphic computing sys-tem with superior advantage of low power consumption and processing speed. The device has shown highly stable synaptic features such as LTP and LTD for repetitive 1000 epochs having 400 conductance pulses each. The device showed advanced synaptic features in terms of LTM, STM, STP, and learning-forgettingrelearning when visible light was induced on it. To highlight the advantages of optical learning-forgetting-relearning process, a 6 × 6 optoelectronic synapse array consisting of HfO x /BP was eventually constructed and used to mimic the human visual perception and visual memory functions. The integrated functions of optical sensing and synaptic learning behavior make it a candidate for future neuromorphic computing systems as well as optoelectronic memory storage for artificial visual and wearable applications.

Experimental Section
Device Fabrication: The proposed device was fabricated on a Si wafer. Before fabrication, the Si wafer was cleaned with isopropyl alcohol (IPA) and deionized (DI) water and then dried it with N 2 gas. First, a 200 nm SiO 2 was grown on Si wafer by plasma-enhanced chemical vapor deposition (PECVD) at 400°C. Then, 30 nm Ti as an adhesion layer and 100 nm Pt as a bottom electrode were deposited by sputtering at room temperature. After Pt, a solution processable black phosphorus (BP) was deposited by drop casting with the thickness of ≈40 nm. Next, 20 nm HfO x as a switching layer was grown by atomic layer deposition (ALD) at 250°C. Subsequently, a 100 nm Cu top electrode was deposited using shadow mask (100 μm in dia) by RF sputtering with pure Ar ambiance at 10 mTorr pressure to prepare new structure-based Cu/HfO x /BP/Pt memristive device. The schematic illustration of the fabrication process flow of the memristor is shown in Note S1 (Supporting Information). Also, the Cu/HfO x /Pt memristive device with the thickness of HfO x (20 nm) was also fabricated for comparison.
Device Flexing: To thin down the devices for physical compliance, a deep reactive ion etching (DRIE) tool was used. A thick photoresist (≈10 μm) was spin-coated on the sample to protect the active devices Adv. Mater. 2023, 35,2300446 from the top surface. Next, the sample was turned upside-down on a carrier wafer for etching the backside bulk silicon. The whole back-etching process was divided into multiple steps to make sure that the required thickness was achieved without over-etching. In the first four steps, the sample thickness was reduced from 525 to 100 μm, and, the final four steps were used to achieve ≈40 μm with a reduced etch rate. The etch process was carried out at a temperature of -20°C, 1500 W ICP power and 60 W RF power, 20 mTorr pressure, and 80 sccm SF6 flow. In between each etching step, the actual sample thickness was measured using a physical profilometer in order to confirm the thickness of the substrate. Once the expected thickness was achieved, the photoresist was stripped using the acetone/IPA, and the sample was placed on a semi-circular support for further characterization.
Characterization and Measurement: The cross-sectional structure and layer by layer materials composition were analyzed using the highresolution transmission electron microscope (HR-TEM, JEOL JEM-2010F). The crystal structure of the spin coated and drop casted MoS 2 layers was characterized using Bruker D8 Advance X-ray diffraction (XRD) system with a Cu K ( = 1.5405 Å) source at 40 kV. The X-ray photoelectron spectroscopy (XPS) was performed on BP sample in a high vacuum using a Kratos Amicus XPS system equipped with a monochromatic Al K X-ray source operating at 10 kV. The electrical characteristics of the device measured by an Agilent B1500A semiconductor device parameter analyzer. A visible blue light-emitting diode source (405 nm) was used for the photoinduced measurements (SDL-405-LM-100T).

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