Near‐Infrared‐Sensing Flexible Organic Synaptic Transistors with Water‐Processable Charge‐Trapping Polymers for Potential Neuromorphic Computing/Skin Applications

Neuromorphic devices, which can mimic the human body's neural system, are rising as an essential technology for artificial intelligence. Here, two types of organic synaptic transistors (OSTRs), OSTR‐A and OSTR‐B, are fabricated on either glass or polymer film using water‐processable charge‐trapping gate‐insulating layers that are prepared by reacting ethylenediamine (EDA) and poly(2‐acrylamido‐2‐methyl‐1‐propanesulfonic acid) (PAMPSA). OSTR‐A is designed to function as a basic artificial synapse by gate pulse stimulation only, while OSTR‐B has additional near‐infrared (NIR)‐absorbing conjugated polymer layers for further sensing of NIR light upon gate voltage stimulations. The PAMPSA:EDA films are found to contain permanent charge bridges (ion pairs of –SO3− +NH3‐) that play a charge‐trapping role in OSTRs. Both devices with the PAMPSA:EDA layers exhibit clear postsynaptic current (PSC) signals upon gate voltage pulses, leading to long‐term potentiation/depression characteristics. The flexible OSTR‐B devices can sense the NIR light (905 nm) upon gate pulse stimulation and their PSC signals are well maintained even after bending (>5000 times). Artificial neural network simulations disclose that the flexible OSTR‐B devices can stably perform synaptic operations under the NIR light with high accuracy (>90%) even after repeated bending (5000 times), indicative of potential use in artificial neuromorphic skin applications.


Introduction
[3] As a tremendous amount of data is supplied from surrounding environments, AI systems need to effectively process such huge data with high accuracy. [4,5]n this regard, a parallel computing method has been suggested and actively studied in terms of software coding aspects, but it has practically a limitation when it comes to conventional hardware systems that are rooted in a von Neumann architecture of serial processing device components. [6,7]10][11] As biological synapses operate by sending and receiving neurotransmitting molecules, the speed of a single synapse component is restricted by the molecular diffusion regime but can be efficient because the whole synaptic processes occur in a parallel mode in the actual neural network systems. [12,13]In principle, the synaptic signals between presynapse and postsynapse components show a spike-type increasing and decreasing shape due to the diffusion of neurotransmitting molecules. [14,15]Therefore, most studies have so far been conducted to make hysteresis signals by introducing inorganic Neuromorphic devices, which can mimic the human body's neural system, are rising as an essential technology for artificial intelligence.Here, two types of organic synaptic transistors (OSTRs), OSTR-A and OSTR-B, are fabricated on either glass or polymer film using water-processable charge-trapping gateinsulating layers that are prepared by reacting ethylenediamine (EDA) and poly (2-acrylamido-2-methyl-1-propanesulfonic acid) (PAMPSA).OSTR-A is designed to function as a basic artificial synapse by gate pulse stimulation only, while OSTR-B has additional near-infrared (NIR)-absorbing conjugated polymer layers for further sensing of NIR light upon gate voltage stimulations.The PAMPSA: EDA films are found to contain permanent charge bridges (ion pairs of -SO 3 À þ NH 3 -) that play a charge-trapping role in OSTRs.Both devices with the PAMPSA:EDA layers exhibit clear postsynaptic current (PSC) signals upon gate voltage pulses, leading to long-term potentiation/depression characteristics.The flexible OSTR-B devices can sense the NIR light (905 nm) upon gate pulse stimulation and their PSC signals are well maintained even after bending (>5000 times).Artificial neural network simulations disclose that the flexible OSTR-B devices can stably perform synaptic operations under the NIR light with high accuracy (>90%) even after repeated bending (5000 times), indicative of potential use in artificial neuromorphic skin applications.[34][35][36][37][38] To date, a couple of OMTR-based artificial synapse devices have been reported utilizing organic materials such as poly(vinylidene fluoride-co-trifluoroethylene) (P(VDF-TrFE)), poly (3,6- [16,[38][39][40][41][42][43][44] However, research has been reported on organic neuromorphic transistors that can sense near-infrared (NIR) light, even though NIR-sensing organic phototransistors have been extensively studied as a detector that can be used in the light distance and ranging (LiDAR) systems for autonomous vehicles, drones, humanoid robots, etc. [45][46][47][48][49][50][51] In this work, we tried to fabricate OSTRs by introducing waterprocessable charge-trapping gate-insulating layers (CTGILs) that were prepared by water-based acid-base reactions of ethylenediamine (EDA) and poly(2-acrylamido-2-methyl-1-propanesulfonic acid) (PAMPSA) at various EDA molar ratios.Two types of OSTRs, OSTR-A and OSTR-B, were designed to act as a basic artificial synapse (for neuromorphic computing application) and a NIR-sensing artificial synapse (for neuromorphic skin application), respectively.OSTR-A has a structure of substrate/ gate electrode/CTGIL/source-drain electrode, while OSTR-B has a structure of substrate/gate electrode/CTGIL/source-drain electrode/interfacial dielectric layer (IDL)/NIR-sensing layer (SL).Both indium tin oxide (ITO)-coated glasses and poly(ethylene naphthalate) (PEN) films were employed as a substrate for OSTR-B.Results showed that both OSTR-A and OSTR-B exhibited an effective postsynaptic current (PSC) signal, leading to clear short-term and long-term potentiation (STP)/depression (STD) characteristics, upon stimulation with gate voltage pulses.The flexible OSTR-B could deliver clear PSC signals with an increased intensity under the NIR light condition (λ = 905 nm) even after bending 5000 times.The artificial neural network (ANN) simulations disclosed that the OSTR-A and flexible OSTR-B showed >90% accuracy even after 5000 times bending.

Results and Discussion
As illustrated in Figure 1a, signal transmission in the human body occurs via the movement of neurotransmitter molecules, between presynapse and postsynapse in neurons, which are known to be activated by a presynaptic nerve impulse. [52]In the case of organic synaptic transistors (OSTRs), the activation role can be played by the pulses of gate voltage, which gradually induce and accumulate charges due to the CTGILs. [17]In this study, as a CTGIL, PAMPSA:EDA was prepared by acid-base reactions of PAMPSA and EDA in water (see Figure 1b).The PAMPSA:EDA layers, which contain permanent charge bridges (twin charge sites), were employed as a CTGIL for the two types of OSTRs (see Figure 1c): OSTR-A and OSTR-B were designed to act as a basic artificial synapse (activation by gate voltage pulse only) and an NIR-sensing artificial synapse (activation by gate voltage pulse and NIR light), respectively.Poly(3-hexylthiophene) (P3HT) and poly[{2,5-bis-(2-octyldodecyl)-3,6-bis-(thien-2-yl)pyrrolo [3,4-c]pyrrole-1,4-diyl}-co-{2,2 0 -(2,1,3-benzothiadiazole)-5,5 0 -diyl}] (PODTPPD-BT) were introduced as a channel layer (CL) and a NIR-SL, respectively, while poly(methyl methacrylate) (PMMA) was employed as an IDL between the P3HT and PODTPPD-BT layer in order to protect the bottom P3HT layer upon the spin-coating process of the top PODTPPD-BT layer (note that the P3HT/PMMA/PODTPPD-BT multilayer is hereafter called the channel/dielectric/sensing (CDS) structure).OSTR-A can be applied to a neuromorphic computing system that mimics the human brain, while OSTR-B can be used as a neuromorphic skin that can particularly detect NIR light (LiDAR application).
To understand the influence of EDA content on the performance of OSTRs, the molar ratio of EDA to PAMPSA was varied up to 50 mol%.As shown in Figure 1d, the pH of aqueous PAPMSA solutions slowly increased from 1.33 (PAMPSA only) to 1.57 (EDA = 30 mol%) and then jumped up to 6.13 (EDA = 50 mol%).The ionization percentage of sulfonic acid groups gradually decreased with the EDA molar ratio, indicative of the reduced number of protons (H þ ) that were dissociated from sulfonic acid groups.This result supports that PAMPSA and EDA underwent acid-base reactions in the aqueous solutions.The PAMPSA:EDA films, which were coated on quartz substrates, were colorless and almost transparent in the visible light range (see Figure 1e).Note that the PAMPSA:EDA solution (EDA = 50 mol%) showed a light yellow but the color of its film could not be clearly observed.After spin-coating processes for device fabrication (OSTR-A and then OSTR-B), each layer was subjected to step-by-step inspection using an optical absorption spectrometer and an optical microscope.As displayed in Figure 1f, the major characteristic peaks of each layer were clearly measured for the polymeric multilayer (P3HT/PMMA/ PODTPPD-BT) for OSTR-B (see the spectrum of CDS structure).Note that a flexible type of OSTR-B was also fabricated using poly(ethylene naphthalate) (PEN) film substrates (see the inset photo in Figure 1f ).
The basic transistor performance of devices was measured to understand whether the PAMPSA:EDA layers act as a chargetrapping layer leading to the hysteresis required for synaptic transistors.As shown in Figure 2a, the hysteresis of transfer curves during forward and backward sweeps of gate voltages (V G = AE5 V) gradually increased with the EDA content in the presence of improved off-current level, even though the maximum drain current (I D ) level was slightly reduced. [53,54]owever, at EDA = 50 mol%, the maximum drain current level became considerably low, leading to a poor on/off ratio (see transistor parameters in Table S1, Supporting Information, and output curves in Figure S1, Supporting Information).Hence the EDA content of 30 mol% was chosen for the initial assessment of the synaptic transistor.
To understand the origin of hysteresis, the PAMPSA:EDA films were investigated using X-ray photoelectron spectroscopy (XPS).As shown in Figure 2b (top panel), the N1s peak at around 401.5 eV, which corresponds to the C-NH 3 þ group, appeared in the PAMPSA:EDA films, and its intensity gradually increased with the EDA content.Note that no N1s peak of the C-NH 3 þ group was measured in the case of pristine PAMPSA film.However, as the EDA content increased, the O1s peak at around 531.5 eV, which corresponded to the sulfonic acid (SO 3 H) group, was gradually shifted to a lower binding energy direction (≈531.2eV corresponded to SO 3 À group) (Figure 2b (middle panel)). [55]As shown in Figure 2b (bottom panel), the S2p peak was also shifted to a lower binding energy direction, indicating a gradual change in the sulfur atom environment with the EDA content.These XPS results basically informed that the permanent charge units (-SO 3 Àþ NH 3 -) were formed in the PAMPSA:EDA layers, which can be further supported by the Fourier transform-infrared spectrometer (FT-IR) spectra (see the analysis result in Figure S2, Supporting Information).As shown in Figure 2c, the grazing-incidence wide-angle X-ray scattering (GIWAXS) measurement uncovered that the characteristic out-of-plane (OOP) peak of the pristine PAMPSA film at q xy = 0.595 Å À1 , which corresponds to a d-spacing of 1.2 nm, gradually decreasing with the EDA content (see corresponding 2D-GIWAXS images in Figure S3, Supporting Information).As shown in Figure 2d, the atomic force microscopy (AFM) measurement delivered a featureless and random surface morphology for the PAMPSA:EDA films irrespective of the EDA content (see all AFM images in Figure S4, Supporting Information).Here, the reducing trend of chain stacking (OOP peak) was quite in good agreement with that of surface roughness (R g ) (see Figure 2e).These results can be explained by the reduced chain stacking that was caused during the formation of permanent charge bridges (i.e., by the existence of EDA molecule) between the PAMPSA chains.Based on the above analysis results, a mechanism can be briefly proposed for the hysteresis phenomenon in the present transistor devices (see Figure 2f ): 1) before applying gate voltage (V G = 0 V); 2) applying negative gate voltage (V G < 0 V), leading to the formation of field-effect charges; and 3) removing gate voltage (V G = 0 V), leading to the trapping of field-effect charges by the ion pair (-SO 3 À þ NH 3 -) bridges in the PAMPSA:EDA layers.Here it is noted that the movement of ions in the PAMPSA:EDA films was not detected by the impedance measurement (see the Nyquist plot in Figure S5, Supporting Information).Based on the above performance and hysteresis characteristics of devices, the PAMPSA:EDA (EDA = 30 mol%) films were chosen to examine the synaptic behavior of OSTR-A devices (glass/ ITO(G)/PAMPSA:EDA/P3HT/Ag(S/D)).First, the single-gate pulse test was performed to check the retention capability of OSTR-A devices.As shown in Figure 3a, the drain current, that is, a PSC, quickly increased upon applying a single pulse of gate voltage (duration = 500 ms).When no further gate bias was applied after the single-gate pulse (open-circuit condition), the PSC slowly decayed with time and still stayed at a higher PSC level.However, the PSC decay was relatively quick in the case of applying V G = 0 V (short-circuit condition) because of the built-in potential caused by the trapped charges in the PAMPSA:EDA layers close to the P3HT layer.This trend was obviously measured for all the gate voltage pulses (from À1 to À5 V) even though the peak PSC values almost gradually increased with the gate voltage (see Figure S6, Supporting Information).As shown in Figure 3b, applying the repeated gate voltage pulses (pulse number = 10, 20, 40) resulted in further improved retention of PSC level in the case of no gate bias condition.This behavior implies that the charge trapping in the PAMPSA:EDA layers could be improved by the repeated pulse action of gale voltage.These results evidence that the present OSTR-A devices deliver good retention characteristics suitable for neuromorphic device applications.
Considering both single and repeated gate pulse test results (Figure 3a,b), the PSC retention part can be controlled by adjusting the level of gate voltage.In this study, however, the synaptic stimulus-response characteristics of OSTR-A devices were investigated focusing on the influence of gate voltage pulse frequency ( f P = 0.4-1.67Hz) and interval (0.1-2.0 s) in the longterm potentiation (LTP) part by fixing a simple short-circuit condition (V G = 0 V) for the part of long-term depression (LTD).As shown in Figure 3c, 50 excitatory and inhibitory PSC signals were acquired under gate voltage pulses at V G = À2 V (duration period = 500 ms), followed by constant gate bias of V G = 0 V (short-circuit condition) (see each condition on top of Figure 3c).Upon applying gate voltage pulses, STP/STD signals were certainly measured irrespective of the frequency of gate voltage (see the enlarged graphs in Figure S7, Supporting Information).However, as the V G frequency increased, the PSC intensity (both excitatory and inhibitory spike levels) increased together.The higher the frequency of gate voltage, the higher the PSC intensity.Consequently, the 40 STP/STD processes delivered obvious LTP/LTD characteristics which were noticeably affected by the frequency of gate voltage.Here it is noted that the inhibitory PSC (IPSC) in the LTD stage was gradually improved with the V G frequency (see Figure 3d), even though some portions of trapped charges were inevitably released by the action of built-in potentials.As plotted in Figure 3e, the excitatory PSC (EPSC) between the 1st and 40th pulses was gradually changed with the frequency of gate voltage (from EPSC = 38 nA at f P = 0.4 Hz to 117 nA at f P = 1.67 Hz).Here it is noted that the change of PSC in synaptic devices can be directly related to the change of conductance. [56]ence, the paired-pulse facilitation (PPF) characteristics between two STP/STD pulses were analyzed according to the PPF equation (PPF index = A 2 /A 1 Â 100% where A 1 and A 2 are the maximum (peak) PSC values of the first and second STP/STD curves).As plotted in Figure 3f, the PPF value was considerably influenced by the pulse interval of gate voltage (PPF = 142% and 108% at the pulse interval of 0.1 and 3.5 s, respectively).This result implies that the stimulus-response characteristics of OSTR-A can be effectively adjusted and optimized by varying the pulse condition of gate voltage.The nonlinearity (NL) of LTP and LTD signals, as its concept is shown in Figure 3g, was calculated using the equation of conductance (G) that is converted from the PSC values (see Figure S8a, Supporting Information).As shown in Figure 3h (top panel), the NL ratio (NR LTD/LTP ) of LTD-to-LTP curves was slightly higher than 1.0 for all the V G frequency ranges, indicating that the NL of LTD curves was relatively higher than that of LTP curves.This asymmetrical NL can be attributed to the relatively easy release of trapped charges in the LTD stage (due to the built-in electric field under the present short-circuit condition) compared to the charge-trapping process in the LTP stage (note that this can be improved by achieving an optimal V G condition in the LTD stage, as supported from Figure 3b).However, the NR LTD/LTP value showed a decreasing trend with the V G frequency ( f P ), informing that the reduced NL of LTD can be achieved at higher f P , which is in good agreement with the symmetricity trend (see Figure 3h bottom panel and Figure S8b, Supporting Information).
By applying the optimized process conditions of glass-based OSTR-A devices, flexible OSTR-A devices were fabricated using PEN film substrates.On top of the flexible OSTR-A devices, the PMMA and PODTPPD-BT layers were consecutively formed leading to flexible OSTR-B devices with a structure of PEN/ ITO(G)/PAMPSA:EDA(30 mol%)/P3HT/PMMA/PODTPPD-BT/ Ag(S/D).As illustrated in Figure 4a, a tensile mode bending test (bending radius = 7.5 mm, angle = 91°) was carried out for the flexible OSTR-B devices inside an argon-filled glovebox system.As plotted in Figure 4b, irrespective of bending number, the PSC signals in the LTP and LTD regions were clearly measured upon stimulations of gate voltage pulses (V G = À3 V, V D = À2 V, f P = 1.0 Hz) in the dark.The PSC intensity in LTP/LTD signals was slightly reduced with the bending number but its reduction was almost saturated after 1500 times (see the PSC change in Figure 4c).This result reflects that the flexible OSTR-B can stably operate with effective PSC (STP/STD) signals, leading to LTP/ LTD behavior even after bending up to 5000 times.The evidence of good STP/STD signal generation can be supported by the trend of PPF index that is always higher than 100% even after bending 5000 times (see the PPF values in Figure 4c).In terms of NL, as shown in Figure 4d   the same NIR light condition, the increased STP/STD signals compared to those in the dark were measured with keeping the signal shape.This result indicates that the photocurrent caused by the PODTPPD-BT layer obviously affected the STP/ STD signals.However, as shown in Figure 5a (bottom panel), after bending 5000 times, the intensity of STP/STD signals was slightly reduced in the dark but a similar increase in the PSC intensity was measured upon illumination with the NIR light.In particular, the increased PSC intensity of STP/STD signals under the NIR light illumination was still measured even after 5000 times bending (see the data for 500 and 1500 times-bending cases in Figure S9, Supporting Information).In more detail, as compared in Figure 5b, the PSC intensity of LTP and LTD curves was slightly reduced with the bending number but noticeably stabilized after bending more than 500 times (see the PSC change in Figure S10, Supporting Information).Under NIR illumination, as plotted in Figure 5c (top panel), the NL of LTP curves showed an increasing trend (from 1.65 to 3.65) with the bending number but the NL of LTD curves was not much affected by the bending number (3.16-3.39).This trend was almost similar to that in the dark as discussed in Figure 4d.As a result, interestingly, the NL ratio of light-to-dark conditions (NR LIGHT/DARK ) fell into the range between 0.95 and 1.27, implying that the NIR light-sensing of PODTPPD-BT layer in the flexible OSTRA-B devices did not affect the stimulation action of gate voltage pulses for the generation of effective STP/STD signals, leading to LTP/LTD behavior.
Based on the above results, a mechanism is proposed for the operation of flexible OSTR-B devices (see Figure 6).First, the electric field-induced charge carriers are generated upon gate voltage pulses in the dark (V G < 0 V and V D < 0 V).Here, it is noted that some of the electric field-induced charges are still trapped at the interfaces between the PAMPSA:EDA and P3HT layers at V G = 0 V during pulsing operations.Second, upon NIR illumination, photoinduced excitons are generated in the PODTPPD-BT SL and some excitons in the bottom part of the PODTPPD-BT layer dissociate into individual charges (holes and electrons) under the electric field between the source and drain electrodes.Of the individual charges separated from the excitons, some electrons can induce dipoles in the thin PMMA interlayer, which can generate additional holes in the P3HT CL.Therefore, both electric field-induced and NIRinduced charges (holes) result in higher drain current signals.Then, by setting V G = 0 V (short-circuit condition) in the pulse operation, both trapped charges (from electric field-induced charges) and NIR-induced charges lead to an increasing level of PSC signals.
Finally, the PSC data from the OSTR-A and flexible OSTR-B devices (EDA = 30 mol%) were used for the recognitionsupervised learning simulations, which were conducted using the CrossSim platform based on the multilayer ANN and the Modified National Institute of Standards and Technology (MNIST)-handwritten digit image database. [57]As illustrated in Figure 7a, the input layer size of the MNIST image was 28 Â 28 pixels (784 neurons), leading to 300 neurons (hidden layer) and 10 neurons (output layer).The synaptic weight of ANN here was calculated and updated in real time using a backpropagation algorithm with the PSC (drain current) signals between presynapse (source) and postsynapse (drain).As a result, the accuracy of OSTR-A reached >90% after 40 epochs irrespective of the frequency of gate voltage (see Figure 7b).Note that the ANN simulations were successfully performed with a similar quality of PSC shapes in previous studies. [17,58]This result indicates that OSTR-A can be applied as an artificial synapse device with frequency-adjustable accuracy characteristics.As displayed in Figure 7c, the ANN simulations proved that the OSTR-B devices after 5000 times bending could undergo neural processes with quite a high accuracy of >90% (40 epochs) in the dark and under the NIR light.Considering the high accuracy of >90% even after bending, the present flexible OSTR-B devices have great potential as NIR-sensing synaptic devices for artificial (neuromorphic) skin applications in the future.In terms of practical applications, as suggested in Figure 1c, the LiDAR I-based flexible OSTR-B devices can be mounted as an electronic skin on humanoid robots or as a near-field sensing unit on human arms.to act as a basic artificial synapse and NIR-sensing artificial synapse, respectively.The XPS and FT-IR measurements disclosed that permanent charge bridges (ion pairs of -SO 3 À þ NH 3 -) were formed in the PAMPSA:EDA films, while the AFM and GIWAXS examinations informed the gradual change in the surface and bulk nanostructures of PAMPSA:EDA films.The best hysteresis characteristics in the dark were obtained for OSTR-A at the EDA content of 30 mol%, which could deliver good retention characteristics leading to effective PSC signals upon stimulations with gate voltage pulses.The EPSC/IPSC signal of OSTR-A was intensified at the higher frequency ( f P ) of gate voltage pulses, while the PPF value became higher at a lower interval of gate voltage pulses.Although the solution-processed PMMA/PODTPPD-BT layers were coated on top of OSTR-A, the flexible OSTR-B devices could generate good STP/STD (LTD/ LTP) signals.Even after bending 5000 times, the flexible OSTR-B devices delivered a well-maintained PSC signal shape in the presence of slightly reduced PSC and PPF.In particular, upon illumination with the NIR light (905 nm), the PSC signal of OSTR-B devices was noticeably increased, leading to a successful NIR sensing in the middle of STP/STD operations.The ANN simulations showed that both OSTR-A and OSTR-B devices can be operated with an accuracy of >90%.Furthermore, even after 5000 times bending, the flexible OSTR-B devices could excellently perform both synaptic actions and NIR light sensing at the same time with >90% accuracy (40 epochs).

Experimental Section
Materials and Solutions: EDA and PAMPSA (15 wt% in H 2 O, weightaverage molecular weight = 2000 kDa) were purchased from Sigma-Aldrich.The aqueous solutions of PAMPSA and EDA were prepared by varying the molar ratio of EDA (0, 10, 30, 50 mol%) to PAMPSA, followed by acid-base reactions at 25 °C for 24 h.The P3HT polymer (weight-average molecular weight = 70 kDa, regioregularity = 97%), which was purchased from Rieke Metals, was dissolved in toluene for the preparation of P3HT solutions (solid concentration: 13 mg mL À1 ).The PMMA polymer (weight-average molecular weight = 120 kDa), which was purchased from Sigma-Aldrich, was dissolved in 2-butanone solvent for the preparation of PMMA solutions (solid concentration = 10 mg mL À1 ).The PODTPPD-BT polymer (weight-average molecular weight = 8.7 kDa, polydispersity (PDI) = 1.37) was synthesized via a Stille coupling reaction according to the same procedure in our previous works. [59,60]The PODTPPD-BT solutions were prepared using toluene as a solvent (solid concentration = 15 mg mL À1 ).
Device Fabrication: Indium tin oxide (ITO)-coated glasses were subjected to typical photolithography and etching processes leading to the patterned ITO (1 mm Â 12 mm stripe)-coated glasses.The same patterning process was used for the ITO-coated PEN films (thickness = 0.13 mm).The patterned ITO substrates were cleaned with acetone and isopropyl alcohol using ultrasonication baths for 30 min each.The wetcleaned ITO-patterned substrates were dried with a nitrogen flow, followed by UV-ozone treatment for 20 min (UV intensity = 28 mW cm À2 ).Next, for the fabrication of OSTR-A devices, the reacted PAMPSA:EDA solutions were spun on the ITO-patterned substrates at a spin speed of 2000 rpm for 60 s and soft baked at 85 °C for 1 h.Then, the P3HT CLs were spin coated on the PAMPSA:EDA layers at 1500 rpm for 30 s and soft baked at 85 °C for 15 min.These samples were moved into a vacuum chamber, which was installed inside an argon-charged glovebox system, and loaded on a metal shadow mask.After evacuating the chamber and reaching a base pressure of <1 Â 10 À6 Torr, the 60 nm-thick silver (Ag) source (S) and drain (D) electrodes were deposited on the P3HT layers via thermal evaporation in a vacuum.The deposited S/D electrodes defined a channel length and width of 70 μm and 2 mm, respectively.For the fabrication of OSTR-B devices, OSTR-A devices were moved to a nitrogen-filled glovebox system and the 50 nm-thick PMMA layers were spin coated on the S/D electrode-coated faces of OSTR-A at 2000 rpm for 60 s.After soft baking at 85 °C for 30 min, the PODTPPD-BT SLs were spin coated on the PMMA layers at 1500 rpm for 30 s and soft baked again at 85 °C for 15 min.Note that no damage was found on both the P3HT and PMMA layers during the spin-coating process for the fabrication of OSTR devices (see Figure S11, Supporting Information).All the fabricated devices were stored inside the argon-charged glovebox system to prevent any deterioration by oxygen and/or water molecules before measurement.
Measurement and Characterization: The pH of PAMPSA:EDA solutions was measured using a pH meter (Model AB15, Fisher Scientific).A UV-vis-NIR spectrometer (Lambda 750, PerkinElmer) was used to measure the optical absorption spectra of solutions and films.XPS measurement was conducted using an XPS system (monochromatic Al-Kα X-ray source: 1486.6 eV, Theta Probe AR-XPS system, Thermo Fisher Scientific).The nanostructure of PAMPSA:EDA films was measured using a synchrotron-radiation GIWAXS (wavelength = 1.192165Å) equipped at the 3C beamline in the Pohang Accelerator Laboratory (PAL).The analysis of GIWAXS data was carried out according to the previously reported method. [61]The surface morphology of PAMPSA:EDA films was characterized using AFM (nanoscope V multimode 8, Bruker).FT-IR (Frontier, PerkinElmer) with an attenuated total reflection mode was used for the measurement of functional groups in the PAMPSA:EDA films.The capacitance and impedance spectra of PAMPSA:EDA films were measured using an impedance analyzer (VersaSTAT 4, Ametek).The performances and stimuli-response characteristics of synapse transistors (output, transfer, gate voltage pulse mode, etc.) were measured using a semiconductor parameter analyzer (Keithley 2636B) and a function generator (Keithley AFG1022), which were operated using specialized software.NIR laser diode (wavelength: 905 nm, VD9030V, Delos Laser) was used to illuminate the channel part of OSTR-B.ANN simulation of handwritten digits was performed using a Python-based program that loads the MNIST database on the basis of the CrossSim platform. [62]The neural network consisting of the three-layer system (784 Â 300 Â 10) had 784 input neurons connected to a 28 Â 28 pixel binarized MNIST input image, which went through 300 hidden neurons, and finally, 10 output neurons corresponding to each number from '0' to '9'.The present simulation used the synaptic weight that was calculated based on the difference in conductance (G) of S/D electrodes in synaptic devices (OSTR-A and OSTR-B).65]

Figure 1 .
Figure 1.a) Schematic diagram of a biological synapse system and neurons.b) Scheme for the acid-base reaction between PAMPSA and EDA leading to PAMPSA:EDA (ionic bridge structure).c) Illustration of device structures and synaptic behaviors of two types of organic synaptic transistors: OSTR-A and OSTR-B are illustratively connected for possible applications of AI in a humanoid robot (right).d) pH (left) and ionization percentage (right) for the PAMPSA:EDA solutions according to EDA molar ratio (mol%) (inset: photographs of solutions).e) Optical absorption spectra (normalized) of PAMPSA: EDA films coated on quartz substrates (soft-baking at 85 °C for 1 h) according to the EDA content (inset: photographs of films).f ) Optical absorption spectra of P3HT (CL), PODTPPD-BT (NIR-SL), and CDS layers (substrate: quartz) (inset: photograph of flexible OSTR-B).

Figure 2 .
Figure 2. a) Transfer curves (dark condition) of OSTR-A devices according to the EDA content (0, 10, 30, 50 mol%): 'F' and 'B' denote forward and backward sweeps, respectively.b) XPS spectra (N1s, O1s, S2p) of the PAMPSA:EDA films.c) 1D GIWAXS profiles (OOP direction) of the PAMPSA:EDA films (see the d-spacing value (d)).d) AFM image (height mode) of the pristine PAMPSA and PAMPSA:EDA (EDA = 50 mol%) films.e) Correlation between RMS roughness (R g ) and OOP peak intensity (d = 1.2 nm) as a function of the EDA content.f ) Graphical illustration of a possible mechanism for hysteresis characteristics due to charge-trapping effects by the formation of permanent charge bridges (between -SO 3 À and NH 3 þ ions) in the PAMPSA: EDA CTGIL of in OSTR devices.

Figure 3 .
Figure 3. a) PSC signals of OSTR-A devices upon single pulse of gate voltage (V G = À0.5-À5.0V, pulse width = 500 ms, V D = À2 V) in the case of two different decay modes after the gate pulse: (red lines) no V G bias (open-circuit condition), (blue lines) V G = 0 V bias (short-circuit condition).b) PSC signals of OSTR-A devices upon gate voltage pulses (pulse number = 10, 20, 40) and no V G bias (open-circuit condition) after V G pulses.c) PSC signals of OSTR-A devices in the dark (short-circuit condition) by varying the frequency ( f P ) of V G pulses (inset: the enlarged PSC signals at the initial stage): (LTP stage) V G,pulse = À2 V (potentiation)/0 V (depression), pulse width (W P ) = 500 ms, pulse number = 40, V D = À2 V; (LTD stage) short-circuit operation (V G = 0 V).d) PSC signals in the LTD stage (short-circuit condition) as a function of pulse number according to the frequency of V G pulses ( f P = 0.4-1.67Hz).e) EPSC peak signals as a function of the frequency ( f P ) of V G pulses at the 1 st and 40 th potentiation pulses.f ) PPF index according to the pulse interval at V G, pulse = À2 V and V D = À2 V. g) Example of NL analysis for LTP/LTD curves.h) (top panel) NL ratio of LTD to LTP (NR LTD/LTP ) and (bottom panel) symmetricity as a function of f P .

Figure 4 .
Figure 4. a) Illustration for the bent state for the flexible OSTR-B device (see photographs of real OSTR-B devices on the right).b) PSC signals of flexible OSTR-B devices as a function of pulse number in the dark (see the bending number on each graph): (LTP stage) V G,pulse = À3 V (potentiation)/0 V (depression), V D = À2 V, f P = 1 Hz; (LTD stage) short-circuit operation (V G = 0 V).c) PSC change (%) and PPF (%) as a function of bending number (0-5000 times).d) NL (top) and NL ratio (NR LTD/LTP , bottom) as a function of bending number.
(top panel), the tensile bending of flexible OSTR-B devices increased the NL of LTP curves from 1.3 to 3.6.However, the NL of LTD curves (3.3-3.5) was almost not much affected by the tensile bending of flexible OSTR-B devices.As a result, the NL ratio of LTD to LTP (NR LTD/LTP ) was lowered with the bending number (see Figure 4d bottom panel).Next, the influence of tensile bending was investigated on the NIR-sensing synaptic behavior of flexible OSTR-B.As shown in Figure 5a (top panel), before device bending, the flexible OSTR-B devices delivered good STP/STD signal shapes in the dark.After long-term depression in the dark, the base (PSC) signal was slightly increased upon the NIR light illumination (905 nm), which can be attributed to the contribution of photogenerated charges in the PODTPPD-BT layer of flexible OSTR-B.When the gate voltage pulses were applied to the flexible OSTR-B under

Figure 5 .
Figure 5. a) PSC signals of flexible OSTR-B devices before and after bending 5000 times in the dark and upon NIR light illumination (λ = 905 nm, P IN = 3.8 mW cm À2 ): (LTP stage) V G,pulse = À3 V (potentiation)/0 V (depression), f P = 1 Hz, V D = À2 V; (LTD stage) short-circuit operation.b) PSC signals of flexible OSTR-B according to the pulse number in the dark and upon NIR light illumination (see the bending number on each graph).c) NL (top) and NL ratio of dark to light (NR LIHGT/DARK ) as a function of bending number.

Figure 6 .
Figure 6.Illustration of the charging mechanism under NIR light illumination in the flexible OSTR-B devices: a) off-state (no bias) in the dark, b) charge generation in the P3HT CL upon applying gate voltage pulse (V G < 0 V) in the dark, c) exciton generation by the illumination of NIR light, d) exciton dissociation (a bottom part of the PODTPPD-BT layer) into individual charges under the electric field between the source and drain electrodes, e) formation of dipoles in the PMMA layer by the charges (electrons) separated from the excitons in the PODTPPD-BT layer and generation of new charges (holes) in the P3HT layer, and f ) charges (electric field-induced and NIR-induced) remained in the P3HT CL after removing the gate voltage pulse.

Figure 7 .
Figure 7. a) Schematic diagram of multilayer ANN simulations for the recognition of MNIST handwritten digit images and synaptic weights as the difference of conductance of two equivalent devices: The NIR (905 nm) light stimulation is marked on the circuit diagram (left bottom).b) Recognition accuracy of OSTR-A for the handwritten digit images as a function of training epoch according to f P (0.4-1.67 Hz) (inset: accuracy as a function of f P ).c) Recognition accuracy of 5000 times-bent flexible OSTR-B for the handwritten digit images as a function of training epoch in the dark and under NIR light illumination (905 nm).