A Bioinspired Ultra Flexible Artificial van der Waals 2D‐MoS2 Channel/LiSiOx Solid Electrolyte Synapse Arrays via Laser‐Lift Off Process for Wearable Adaptive Neuromorphic Computing

Wearable electronic devices with next‐generation biocompatible, mechanical, ultraflexible, and portable sensors are a fast‐growing technology. Hardware systems enabling artificial neural networks while consuming low power and processing massive in situ personal data are essential for adaptive wearable neuromorphic edging computing. Herein, the development of an ultraflexible artificial‐synaptic array device with concrete‐mechanical cyclic endurance consisting of a novel heterostructure with an all‐solid‐state 2D MoS2 channel and LiSiOx (lithium silicate) is demonstrated. Enabled by the sequential fabrication process of all layers, by excluding the transfer process, artificial van der Waals devices combined with the 2D‐MoS2 channel and LiSiOx solid electrolyte exhibit excellent neuromorphic synaptic characteristics with a nonlinearity of 0.55 and asymmetry ratio of 0.22. Based on the excellent flexibility of colorless polyimide substrates and thin‐layered structures, the fabricated flexible neuromorphic synaptic devices exhibit superior long‐term potentiation and long‐term depression cyclic endurance performance, even when bent over 700 times or on curved surfaces with a diameter of 10 mm. Thus, a high classification accuracy of 95% is achieved without any noticeable performance degradation in the Modified National Institute of Standards and Technology. These results are promising for the development of personalized wearable artificial neural systems in the future.


Introduction
The 4th industrial age, characterized by the Internet of Things and artificial intelligence (AI), has led to a substantial increase in data generation across electronic devices with a projected 170 (zettabyte) to be generated in 2025. [1] In addition, there has been a huge demand for the collection and processing of the massive data generated in daily life due to various personalized portable devices. These personalized wearable and portable sensors have the advantage of receiving tactile, pressure, rotational, and optical signals from the human body. Thus, enabling them to provide immediate feedback to the electronic devices. [2][3][4][5] Despite the advances in novel materials and device architectures that could improve the sensitivity and resolution, there are still some challenges, including high computing power, high latency of data transmission, and physical separation of processing and memory units. [6,7] The conventional approach for processing the sensing data is to transmit them to external servers and subsequently compute them with the off-chip devices. [8] This off-chip computing method causes an information bottleneck in which the sensing and processing units for temporal sensing signal data are physically separated. Furthermore, AI functions are becoming highly important, meaning that interpreting sensing data to deliver meaningful real-time information on wearable devices is needed. [9,10] Thus, hardware units with small sizes and low power consumption to process the signals are required.
To date, many efforts have been made to develop flexible neuromorphic devices optimized to an artificial neural network (ANN) for memory processing combined with functional nanomaterials and devices. [11][12][13][14] Oh et al. reported a sensory-neuromorphic system for sign language translation based on a flexible artificial synapse with a silicon-indium-zinc-oxide (SIZO)/ion-gel hybrid structure. [15] Lee et al. fabricated an organic optoelectronic sensorimotor synapse using a stretchable organic nanowire synaptic transistor. The light signals converted to presynaptic signals triggered postsynaptic potentiation to actuate the artificial muscle. [16] Kim et al. reported an artificial afferent nerve to detect the movement of an object by converting pressure information to an action potential. [17] Moreover, Choi et al. implemented a synaptic transistor with triboelectric rotation sensors to mimic an energyscavenging artificial nervous system. [18] Various stretchable and flexible devices have been developed, but most of their electrolytes are organic polymers or ion gel-based devices. Therefore, efficient processing of a large amount of information using highly integrated neuromorphic circuit devices is limited.
More recently, flexible 2D artificial neuromorphic synapse devices based on inorganic materials, including graphene, transition metal dichalcogenides (TMDs), and boron nitrides [19] have been reported with highly integrated array functionality. [20] Notably, these atomically thin 2D-layer materials, which differ from traditionally bulky materials, exhibit exceptional properties such as low-power switching capability, [21] electrostatic gate tunability, [22] and mechanical flexibility. [23,24] Furthermore, novel neuromorphic synaptic devices built by van der Waals (vdW) heterostructures with ideal surface characteristics have been successfully demonstrated. [25][26][27][28][29][30][31][32][33] For instance, Choi et al. reported a curved neuromorphic image sensor array using a 2D-MoS 2organic heterostructure inspired by the human vision recognition system. [5] In addition, Hu et al. reported an ultralow power optical synapses based on 2D-MoS 2 layers by indium-induced surface charge doping for biomimetic-eyes applications. [4] Despite the remarkable progress in flexible artificial synapse array devices based on 2D materials, there are still some challenges in scaling up the production of these 2D materials and achieving large-area uniformity because of the inherent film transfer process. Thus, a back-end of line-compatible process (< 450°C) and the availability of an innovative flexible substrate [34] are necessary for 2D material-based flexible synaptic devices and neuromorphic edge systems.
Here, we demonstrate an excellent linear and symmetric neuromorphic performance with an inorganic flexible neuromorphic device using Li ions, including a solid-state electrolyte and a semiconducting 2D-MoS 2 channel. In addition, we present a method for fabricating a neuromorphic device with a MoS 2 channel and a LiSiO x (LSO) electrolyte on a flexible substrate without the mate-rial transfer process. For the substrate, a flexible polymer, colorless polyimide (cPI), was deposited on a solid glass substrate. The laser lift-off (LLO) method was used to facilitate the attachment and detachment of the device on the cPI. MoS 2 was synthesized directly on the cPI using plasma-enhanced chemical vapor deposition (PECVD) below the glass transition temperature (T g ) of the polymer to prevent polymer degradation. Without a transfer step, the occurrence of defects during transfer can be suppressed and the processing time can be shortened. This method may be applied to large-capacity and large-area semiconductor processing lines in the future.

Device Fabrication Process
An illustration of the flexible all-solid-state artificial synapse array fabricated via the LLO process is shown in Figure 1a. As shown in the photograph and optical microscopy image in Figure 1b and inset, the neuromorphic device is an ultraflexible three-terminal synaptic transistor. The few-layer MoS 2 is a semiconducting channel, where the drain current (I D ) flows in the channel through the source and drain electrodes. The LiSiO x layer used as an electrolyte exists between the gate electrode and the channel. When a gate voltage (V G ) is applied, the mobile Li ions in the electrolyte diffuse and affect the conductivity of the channel, resulting in neuromorphic synaptic characteristics.
The fabrication process of the ultraflexible neuromorphic synaptic device, shown in Figure 1c, is described here (see Fig-ures S1 and S2, Supporting Information). A Si layer, which was used as a sacrificial layer during the LLO process, was deposited on a cleaned glass substrate via PECVD. An IPI-C solution was spin-coated on Si/glass and heated at 400°C for 3 h to synthesize the cPI substrate. Using a laser with a power of 18 kW and wavelength of 532 nm, the Si layer under the cPI was removed, except for the edge of the Si layer, for the cPI substrate to adhere to the edge of the glass. A 30 nm thin film of Al 2 O 3 was deposited using atomic layer deposition (ALD) to smoothen the surface of the cPI. The roughness (R q ) of the cPI surface is low, ranging from 1.237 to 0.326 nm ( Figure S3, Supporting Information). A twostep PECVD was carried out to directly grow MoS 2 on the cPI substrate under low temperatures. First, a MoO 3 film was deposited on the substrate using thermal evaporation, which reacted with H 2 S in the H 2 atmosphere at 200°C forming the 2D-MoS 2 layer. A photoresist (PR) mask was developed and etched using reactive ion etching (RIE) to lithograph the MoS 2 layer in the form of a channel. Pd was deposited by e-beam evaporation and lifted off by developing a PR mask for the source/drain electrodes on the MoS 2 channel. The shape of the gate PR mask was placed between the source and the drain to deposit the LSO, which was used as a solid electrolyte dielectric by cosputtering Li 2 O and SiO 2 targets. The gate metal was then deposited on the LSO film immediately by an e-beam, followed by the lifting-off process. To protect the LSO film and synaptic transistor device from external environmental effects, an Al 2 O 3 passivation layer, 30 nm thick, was deposited on the device using an ALD process. When used as a flexible synaptic device, a free-standing thin-film device can be easily obtained by cutting the edge fixed with the Si layer.

Microstructure Analysis of Flexible Synaptic Devices
A qualitative analysis was performed to determine the heterostructural elements in the synthesized ultraflexible neuromorphic synaptic device. Figure 2a shows the cross-section of the flexible solid-state electrolyte synaptic transistor device. The 2D-MoS 2 channel is stacked on the buffer layer along with Al 2 O 3 on the cPI flexible substrate. The LSO electrolyte and gate metal are stacked on the channel, with the Al 2 O 3 passivation layer on top. Figure 2b shows the tunneling electron microscopy (TEM) image of the cross-sectional sample of the device prepared using a focused ion beam (FIB). Each layer is clearly distinguished in the order described in Figure 2a. Figure 2c shows the distribution of atoms using energy-dispersive X-ray spectroscopy (EDX) mapping, which is sharply divided for each layer. Mo and S are detected in the channel, while Si and O are detected in the LSO electrolyte layer. Pd, which was used as an electrode, and Al and O, used in the passivation and buffer layers, are detected in each region. The EDX line profile also shows a clearly distinguishable atomic region ( Figure S4, Supporting Information).
Furthurmore, structural analysis was carried out to clarify the layer-to-layer arrangement of two dimensional MoS 2 channel. A high-resolution tunneling electron microscopy (HRTEM) image of the MoS 2 layer. A layered structure with a lattice distance of ≈6.2 Å consistent with the van der Waals layer MoS 2 is observed, as shown in Figure 2b,d. The fast Fourier transform (FFT) pattern derived from the HRTEM image, where the pattern for the (002) plane of layered MoS 2 is observed, as shown in Figure 2e. [35] Raman spectroscopy of MoS 2 deposited on SiO 2 substrate. The peak positions of the Raman spectra are located at 382 and 405 cm −1 , corresponding to the in-plane vibration of Mo and S atoms (E 1 2g ) and the out-of-plane (A 1g ) vibration of S atoms, respectively, as shown in Figure 2f. The difference between the two main peaks is ≈23 cm −1 , indicating that MoS 2 is an average of four to five layers, which agrees with the TEM cross-sectional image in Figure 2d. [36,37] The X-ray photoelectron spectroscopy (XPS) analysis k-m) The XPS spectrum core-level spectrum of Li 1s, Si 2p, and O 1s. The spectrum is obtained by depth-XPS at the gate region between S/D electrode. As the measured area is larger than the width of LSO electrolyte, the passivation layer is detected at the same time.
of MoS 2 was conducted at the core-level of Mo 3d and S 2p. The Mo 4+ peaks of MoS 2 are observed at 229.1 and 232.2 eV. The peak at 226.3 eV corresponds to the MoS 2 peak at the core-level of S 2s. In S 2p, the S 2− peaks of MoS 2 were observed at 161.8 and 163 eV, [38] indicating that the MoS 2 layer grows evenly on the Al 2 O 3 /cPI substrate with an average of five to six layers without oxide residue.
The LSO electrolyte was analyzed to detect the Li ions. An electron energy loss spectrum (EELS) measured in the LSO region exhibited peaks corresponding to Li-K and Si-L 2,3 , confirming the presence of Li and Si in LSO, as shown in Figure 2i. [39,40] The LSO deposited on the SiO 2 substrate was analyzed using Fourier transform infrared (FT-IR) spectroscopy, as shown in Figure 2j  In the O 1s spectrum, peaks appeared at 532.7, 531.9, and 531.7 eV [42,43] indicating that LSO consists of Li 2 O, LiSiO x , and SiO 2 , respectively. The Al 2 O 3 peak at 531 eV is observed because the measurement depth of XPS is larger than the width of the electrolyte, detecting the passivation layer simultaneously. The depth of the XPS spectrum for the gate electrolyte as a function of the etching time and the XPS spectrum for cPI are displayed in Figures S5 and S6 in the Supporting Information. It is confirmed that the layered MoS 2 channel and LSO electrolyte layer are well deposited on the cPI layer.

Electrical Analysis and Operation Mechanism
Neuromorphic synaptic devices can mimic real human synaptic behaviors. In the human synapse, chemical stimulation is provided from presynaptic neurons to postsynaptic neurons for the transduction of biosignals, which has been applied to electrolyte transistor-based neuromorphic devices. The drain current (I D ) flows from the source to the drain through the channel, applying a presynaptic stimulus through the gate (V G ). Eventually, the mobile ions in the electrolyte affect the conductivity of the channel and amplify the postsynaptic signal. [44] The basic operation between the human brain and the synaptic device is illustrated in Figure 3a. The dual-sweep transfer curve (I D -V G ) of a three-terminal Li-based electrolyte synaptic transistor is shown in Figure 3b. The I D gradually increases when V G is swept from 0 to 5 V. The I D exhibits an anticlockwise behavior without any decrease in I D when the V G is swept from 5 to 0 V. This anticlockwise hysteresis of our synaptic devices is ascribed to the mechanism of the electrochemical doping process. The mobile cations (e.g., Li + , Mg 2+ , and Cu 2+ ) in the electrolyte layer approach the channel by an external electric field and intercalate the channel as a dopant to enhance I D . [45][46][47] Conversely, the mobile cations doped in the channel are deintercalated in the negative sweep region (from 0 to −5 V and from −5 to 0 V), causing I D to gradually www.advancedsciencenews.com www.small-methods.com diminish. Figure 3c shows the sequential transfer curves (I D -V G ) obtained by applying a gate voltage in succession. As V G sweeps in the positive voltage direction (from 0 to 5 V and from 5 to 0 V), I D rises step-by-step by ≈0.4 nA. Conversely, I D gradually decreases when V G sweeps (from 0 to −4 V and from −4 to 0 V) in the negative voltage direction. It should be noted that the mobile Li ions in the LSO electrolyte layer can precisely control the channel conductivity by successive electrical gate pulses. [48,49] To investigate the operation mechanism, the capacitance as a function of the operating frequency (C-f) is presented in Figure 3d. The capacitance is less than 0.3 pF at high frequencies, between 10 4 and 10 5 Hz, which is related to the bulk electrolyte, which of capacitance has a fast charging speed. The capacitance increases between frequencies of 10 3 and 10 4 Hz because of the electrochemical doping, which is affected by the electric double layer (EDL) at low frequencies below 10 3 Hz. The existence of the electrochemical doping process can be explained by the relation between frequency and capacitance, as shown in Equation (1) where is a parameter related to the Warburg impedance. The slope increase below 10 3 Hz and between 10 3 and 10 4 Hz indicates that the neuromorphic behavior mechanism is driven by electrochemical doping, as described in the transfer curve graph. [48,50] Figure 3e illustrates the operating mechanism of the synaptic transistor using external gate voltages. Li cations exist in the electrolyte at 0 gate voltage. In the potentiation state (V G > 0), Li cations move from the electrolyte to the channel surface. The cations temporarily form an EDL and dope the channel by intercalation into the MoS 2 channel layer. In the depression state (V G < 0), the cation returns to the electrolyte and the channel is deintercalated. [49] The electrical pulse and read scheme for V G and V D are shown in Figure 3h, along with the long-term potentiation (LTP) and long-term depression (LTD). To identify the synaptic behavior of the spike, the following equations are defined [20,25,51] G max∕min = G max ∕G min (2) where G max and G min are the maximum and minimum conductance in each spike, respectively; G p and G d are the conductance of potentiation and depression, respectively; and n is the pulse number. The nonlinearity of the potentiation and depression are represented as p and d , respectively. AR is short for the asymmetric ratio. Figure 3i shows the cycle-to-cycle endurance, repeated 20 times in a row. The synaptic characteristics of the potentiation and depression are maintained within 400 gate pulses without any deterioration. This means that the reversible reaction for intercalation and deintercalation works well with the novel MoS 2 and LSO heterostructures. Figure 3j shows the LTP and LTD characteristic curves according to the pulse number by magnifying one spike. As shown in Figure S7 in the Supporting Information, the energy consumption of a single spike is 0.55 nJ. In Figure S8 in the Supporting Information, current decay characteristics of excitatory/inhibitory postsynaptic current in our device are also depicted. The maximum to minimum conductance ratio (G max/min ) is ≈1.05, the nonlinearity for potentiation and depression ( p and d ) is 0.55 and −1.00, respectively, and the AR, showing the symmetricity between potentiation and depression, is 0.22. Our all-solid-state electrolyte synaptic transistor with highly integrated functionality on a flexible cPI film shows superior linear and symmetric performance compared with those reported in the literature (see Table 1). Figure 3k,l shows the statistical distributions of each state obtained from Figure 3i. All the conductance states are distinctively separated over 400 pulses. Thus, the minimal variation induced by mobile Li-ions indicates that the weight updates can be linear and symmetric. As a result, linear potentiation and depression can significantly improve the accuracy of the pattern recognition ability, as explained later in the simulation section. Figure 3m,n is histograms that show the nonlinearity of each potentiation and depression. They are distributed with an average of 0.3 in potentiation and −2.8 in depression, indicating that the shape of each spike is similar. The retention test of five different conductance states are also displayed in Figure S9 in the Supporting Information. The statistical distributions in G max /G min ratio of 45 random LSO synaptic devices and retention performance of five different conductance states on cPI flexible substrate are shown in Figure S10 in the Supporting Information.

Synaptic Characteristics on Mechanical Stress
The mechanical flexibility of the all-solid-state 2D MoS 2 /LiSiO x synaptic array was characterized for wearable edge-computing device applications. Contact between the electrodes was enabled by fixing the flexible device on the semi-cylindrical pedestal, as shown in Figure 4a,b. The mechanical flexibility of our synaptic device was measured with diameters of 30, 20, and 10 mm. The strain applied to the device with each diameter was calculated as follows where is the strain, t is the thickness of the layer, and ∅ is the diameter of the device. [52] Figure 4c shows the normalized endurance graph for LTP and LTD characteristics of a flexible device in a flat state and a device bent by 30, 20, and 10 mm (the maximum conductance is 1 and the minimum is 0). The stable and repeatable endurance characteristics are maintained after 400 cycles, even when the device is folded with a diameter of 10 mm ( = 0.25%  c) The normalized conductance plot of bended device with strain (Ɛ) of 0%, 0.08%, 0.13%, and 0.25% (gray, red, orange, and blue). d) The photo of a bending machine which helps to bend and flatten the flexible synaptic device. e-g) The plot of on/off ratio (G min /G max ), nonlinearity of potentiation ( p ), depression ( d ), and asymmetric ratio (AR) as bending cycle.
in the synaptic characteristics under endurance cyclic tests with mechanical strain stress were well maintained without noticeable performance degradation ( Figure S10, Supporting Information).
The mechanical durability of the flexible synaptic transistor was also measured against the number of bending cycles. Figure 4d shows the bending machine used for the continuous bending test of the device. In the repeatable bending-cycle test, the device is bent from a flat state (top) to a curved surface with a di-ameter of 5 mm (bottom). Figure 4e,f,g shows the statistical distributions of the synaptic device for the G max /G min , p , d , and AR parameters with respect to the number of bending cycles (up to 700 cycles). The relative ratio of G max /G min is maintained within the range of 1.18-1.22, even after 700 bending cycles. The nonlinearity of the potentiation is maintained at ≈2. In the case of depression, nonlinearity tends to decrease from −4 to −8, but the asymmetry rarely changes from 0.6 to 0.7. The durability of the flexible device is maintained even after 700 bending cycles, exhibiting an artificial synapse performance ( Figure S11, Supporting Information).
To evaluate the neuromorphic performance of all-solid-state flexible 2D MoS 2 /LiSiO x synaptic transistors, an artificial neural network was simulated based on the parameters extracted from the flat and bent states of the flexible neuromorphic devices. Figure 5a depicts a three-layer neural network structure for recognizing handwritten digits with 28 × 28 pixels using the Modified National Institute of Standards and Technology (MNIST) dataset. The three-layer network included an input layer (784 neurons), a hidden layer (300 neurons), and an output layer (ten neurons). The circuits of the synaptic transistor-based crossbar array for matrix operations are illustrated in Figure 5b. The recogni-tion accuracy of ANN is subject to G max /G min , nonlinearity, and the asymmetric ratio of devices. The mapping of the synaptic weights of the synaptic devices in the flat state as a training epoch is shown in Figure 5c, and the recognition accuracy as a function of the training epoch is shown in Figure 5d. As the training epoch increases, the recognition accuracy approaches 94.5%, and the "3" digit shape in the image mapping becomes clearer, whereas the accuracy of ideal numeric is 98.1%. The accuracy of the various bend states reaches 95.2% (flat), 94.3% (30 mm), 92.4% (20 mm) as radius, 93.2% (0 cycle), 91.6% (100 cycles), 90.1% (200 cycles), and 90.0% (300 cycles) as bending cycles after 30 epochs (Figure 5e,f). These results indicate that our flexible artificial synapse has a high recognition accuracy in any bend state.

Conclusion
We showed that flexible artificial synapse arrays could be fabricated using the LLO and PECVD methods without a transfer process. In the flat state, this device exhibited high step-specific differentiation with a nonlinearity of ≈0.55 and −1.00, an asymmetry ratio of ≈0.22, and maintaining its performance even after 20 cycles. It was also shown that bending was performed up to a maximum curved surface of 10 mm and that the neuromorphic performance was maintained even after 700 cycles. The MNIST simulation was carried out using these parameters, and an accuracy of up to 90.0% or more, even in any bending state, was confirmed. Our sample synthesis method has the potential to grow into a photoreactive wearable device in future studies, as the fabrication method is highly flexible. In addition, it suggests a novel method for synthesizing neuromorphic devices with excellent performance, which can facilitate massive data processing using artificial neural networks.

Experimental Section
Fabrication of cPI Film and Laser-Lift-Off: A fabricated cPI film with an approximate thickness of 20 μm was deposited on amorphous silicon (a-Si) sacrificial layer upon cleaned glass with 50 × 50 × 1.8 mm 3 . The glass was cleaned with sonication for 5 min soaking in isopropyl alcohol, acetone, and diluted water. After drying in oven at 100°C, the glass was loaded onto the PECVD chamber (SNTEK, Korea) heated to 250°C for deposition of 100 nm thick Si layer. A plasma was generated under SiH 4 condition with H 2 at 400 mTorr with radio frequency (RF) power and then Si layer was deposited. The IPI-C solution (5000 cps and a solid content of 18.5 wt%, IPITECH, Korea) mixed with 3-glycidoxypropyltrimethoxysilane (98%, Sigma-Aldrich, USA) spin-coated onto the a-Si sacrificial layer substrate at 400 rpm using a spin-coater. The solution-coated glass was annealed at 350°C for 4 h in a box furnace (Thermal System and Technology, Korea). The Si-layer sandwiched between cPI film and glass substrate was lifted-off by a laser with 532 nm of wavelength. The laser power was 196.5 mW and was controlled using external attenuator, frequency, and diode current (Kortherm Science, Korea). The 30 nm thick Al 2 O 3 passivation layer was deposited on LLO treated cPI film using ALD (Lucida M100, NCD, Korea) with trimethylaluminum ((CH 3 ) 3 Al) precursor and H 2 O at 200°C.
Growth of MoS 2 Channel: 2D-MoS 2 channel layer was grown at lowtemperature (< 250°C) by two-step PECVD as referred to the previous study. [53] A 3 nm thick MoO 3 film was deposited on cPI substrate by thermal evaporator (REP5004, SNTEK, Korea) at rate of 0.2 Å s −1 . The MoO 3 deposited cPI was loaded on a chamber of atmosphere pressure (AP) PECVD system (A TECH System, Korea) with a cylindrical rotary electrode which was heated to 200°C. The gas mixture of H 2 and 0.1 mol% of H 2 S/He was injected into the chamber until a working pressure of 100 Torr and then plasma generated to sulfurize MoO 3 film by rotating the cylindrical electrode with very high frequency power of 400 W. As-grown MoS 2 layer was patterned using RIE under Ar plasma with PR (AZ5214, Micro-Chemicals, Germany).
Fabrication of MoS 2 /LSO Synaptic Transistor Device: The source/drain electrode of Pd (50 nm) was lithographically patterned on MoS 2 channel using thermal evaporation and wet lift-off by spraying acetone. The LSO electrolyte was deposited using cosputter (MIBE-6000, Korea Vacuum, Korea) with Li 2 O (99.90%, iTASCO, Korea) and SiO 2 target (99.65%, iTASCO, Korea) through negatively patterned PR mask. The Ar plasma was generated at the RF power of 50 and 100 W under pressure at 20 mTorr until the LSO was deposited to thickness of 100 nm. The gate electrode of Pd was immediately deposited onto LSO by thermal evaporation as same condition of source/drain electrode. After the LSO/gate was patterned with wet lift-off, the passivation layer of Al 2 O 3 was deposited by same process above. As-fabricated MoS 2 /LSO neuromorphic array was detached from glass substrate by cutting cPI edge to be used as flexible device.
Characterization: The cross-sectional sample of device was fabricated using FIB (Helios 5 UX, Thermo Fisher Scientific, USA) and transferred on TEM grid. The atomic structure of each layer was analyzed by Cs-Corrected field emission (FE)-TEM (JEM-ARM200F, JEOL, Japan) with HR-TEM image, EDX (with Mn K ), and EELS. The MoS 2 film was measured by Raman spectroscopy (XperRAM-CS, NANOBASE, Korea) with a laser of 532 nm. The composition of device was analyzed by XPS (NEXSA, Thermo Fisher Scientific, USA). The etching of depth XPS was operated with sputter energy of 2 kV and the size of sputtering was 2 × 2 mm. The measured XPS spot size was 100 μm. The atomic composition of LSO was confirmed by attenuated total reflection mode of FT-IR (Nicolet iS10, Thermo Fisher Scientific, USA). The surface roughness of cPI was confirmed using atomic force microscopy (AFM) (NX10, Park System, Korea). Electrical neuromorphic performance was analyzed by Keithley 4200A semiconductor analyzer under ambient temperature and pressure. The capacitance of device was measured using LCR meter (E4980A, KEYSIGHT, USA).
Simulation of Handwritten Pattern Recognition: The simulation was performed on CrossSim platform in Python environment. A data set of 28 × 28 pixels handwritten images from Modified National Institute of Standard and Technology (MNIST) database are used for training and testing. A three-layer network used for backpropagation consisted of 784 input neurons, 300 hidden neurons, and 10 output neurons. The input neurons and output neurons corresponded to 28 × 28 pixels and 0 to 9 respectively. An entire 60 000 images were used for training and 10 000 images were used for assessing the accuracy.

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