One-dimensional organic artificial multi-synapses enabling electronic textile neural network for wearable neuromorphic applications

See allHide authors and affiliations

Science Advances  10 Jul 2020:
Vol. 6, no. 28, eaba1178
DOI: 10.1126/sciadv.aba1178


One-dimensional (1D) devices are becoming the most desirable format for wearable electronic technology because they can be easily woven into electronic (e-) textile(s) with versatile functional units while maintaining their inherent features under mechanical stress. In this study, we designed 1D fiber-shaped multi-synapses comprising ferroelectric organic transistors fabricated on a 100-μm Ag wire and used them as multisynaptic channels in an e-textile neural network for wearable neuromorphic applications. The device mimics diverse synaptic functions with excellent reliability even under 6000 repeated input stimuli and mechanical bending stress. Various NOR-type textile arrays are formed simply by cross-pointing 1D synapses with Ag wires, where each output from individual synapse can be integrated and propagated without undesired leakage. Notably, the 1D multi-synapses achieved up to ~90 and ~70% recognition accuracy for MNIST and electrocardiogram patterns, respectively, even in a single-layer neural network, and almost maintained regardless of the bending conditions.


Wearable electronics have attracted considerable attention as a human-friendly emerging platform that closely interacts with our environment and surroundings using adaptable devices attached to or placed in the human body (13). Thus, the development of wearable devices with various electronic functionalities and suitability for attachment to the human body is necessary (15).

As a unit for the wearable device, a one-dimensional (1D) fiber based on organic materials is promising for smart electronic textiles (e-textiles) owing to its unique characteristics; e.g., it is soft, light, and flexible and has a simple fabrication process (68). In particular, a thread-like fiber device unit is comfortable to bend and can be integrated directly into fabrics and maintain its inherent functions on soft and curved skin (68). Because of these unique characteristics, the 1D fiber unit has potential for complex and area-scalable wearable electronic applications based on weaving technology (6, 7). In this regard, various 1D fiber-based devices have been extensively investigated and demonstrated, including energy harvesters (9, 10), sensors (11, 12), actuators (13), light-emitting diodes (8, 14), transistors (15), logic circuits (16), and memories (17, 18). However, the conventional computing platform consisting of these 1D fiber-based device constituents (e.g., transistor-based logic circuits) may not be suitable to immediately process and classify the numerous unstructured sensing data continuously received from the human body or surroundings. This is because (i) the current wearable battery technology has a limited lifetime and (ii) the classical computing hardware architecture requires massive computing resources and power even for simple classification tasks owing to the von Neumann bottleneck between the processor and memory (19).

A brain-inspired computing architecture might be suitable for an e-textile computing platform because of the potential to efficiently process the large amount of unstructured sensing data, including diverse and complex signals from the human body or the surrounding environment (2022). Furthermore, crossbar-type e-textiles based on weaving technology using a 1D fiber device are highly suitable for implementing artificial neural networks (ANNs) capable of efficient parallel signal processing and recognizing timely signal features. In this regard, the first step for realizing wearable neuromorphic computing applications is to implement a 1D fiber artificial synaptic platform that is capable of performing diverse essential synaptic functions and has mechanical and electronic stability. However, the 1D fiber artificial synapse and its applicability have rarely been studied despite the aforementioned necessity of constructing a wearable computing system.

Recently, Park et al. (23) fabricated a two-terminal memristor synapse by crossing two yarns coated with reduced graphene oxide (rGO). Their results are noteworthy in that they indicate the potential for yarn-based artificial synapses. However, an integrated array structure using only rGO-coated yarns can suffer from undesired neural signals passing through neighboring nodes, which leads to severe degradation of the learning accuracy (24). For a high-level learning capability, a 1D fiber artificial synaptic platform should be capable of independently addressing and transmitting the neural signal without unwanted interference between the synapses (24). In addition, it must have the ability to perform high-level synaptic activities equipped with a large number of multistates, sustaining the switching linearity and durability under repetitive bending stress.

In this study, we fabricated several ferroelectric organic transistors on a 1D Ag wire using poly(vinylidene fluoride-trifluoroethylene) [P(VDF-TrFE)] and used them as a 1D fiber-shaped multisynaptic device platform for an e-textile neural network. The 1D fiber-shaped device can be easily coiled or bended on tubes of various forms and operated well under the bending conditions without severe switching degradation. They can exhibit the well-defined essential synaptic functions through meticulous modulation of the degree of polarization of their ferroelectric domain according to the electrical stimulus. In addition, their synaptic characteristics can have excellent reliability under repeated pulse cycles and mechanical bending stress. Notably, a NOR-type synaptic array is simply fabricated via cross-connecting 1D multi-synapses with Ag wires, where the postsynaptic responses received by individually addressing the synaptic cells can be integrated. With the conventional stochastic gradient ascent/descent learning algorithm in the single-layer neural network, recognition accuracies are up to ~90 and ~70% for Modified National Institute of Standards and Technology (MNIST) and electrocardiogram (ECG) patterns, respectively. In particular, the initial accuracy for the ECG pattern was maintained regardless of the mechanical bending stress.


Fabrication of 1D artificial multi-synapses

Figure 1A shows the fabrication process for 1D artificial multi-synapses implemented by substrate-free ferroelectric organic transistors on a thin Ag wire (diameter of 100 μm). An organic ferroelectric P(VDF-TrFE) film, which functioned as a common gating dielectric layer, was uniformly and directly coated onto the whole surface of the Ag wire via the dip-coating method using a capillary tube equipped with a printing speed controller (15, 25). After annealing at 140°C for 1 hour in an air convection oven to crystallize the coated film, 50-nm-thick pentacene as an active semiconductor channel was thermally deposited on the upper semicircle of the P(VDF-TrFE)/Ag wire. Then, Au metal was patterned on top of the pentacene channel through a shadow mask with intervals of 15 μm. The Au was used as source and drain electrodes. Since the channel width (W) and length (L) of the 1D device are determined by the semicircle of the 1D fiber and the intervals of the source and the drain, respectively, in the case of our device, W = 157 μm and L = 15 μm. The average dielectric capacitance of the Au/P(VDF-TrFE)/Ag junction was measured as 4.75 ± 0.3 nF/cm2 at different positions at a frequency of 10 kHz, indicating that the ferroelectric film was coated uniformly on the Ag wire without the formation of severe pinholes (fig. S1). Figure 1B shows magnified cross-sectional and top-surface scanning electron microscopy (SEM) images of a completed 1D organic multisynaptic device. Note that the thickness of the P(VDF-TrFE) film is about ~608 nm. Although this thickness is enough to prevent an undesired gate leakage current (fig. S2) and the electrical short, it would lead to a high operating voltage at the same time (25). Therefore, it is important to find optimal conditions between the operating voltage [or the thickness of P(VDF-TrFE)] and device performance/stability depending on the target applications. Each constituent layer of the device was clearly distinguished, and the patterning produced a gap of ~500 μm between synaptic units, as shown in Fig. 1B. In this result, a 30-mm-long completed 1D device contained ~24 individual synapses, as shown at the right of Fig. 1C. Owing to the fiber-shaped device form and the inherent flexibility of the organic components, the 1D organic multi-synapses can be easily coiled or bent on tubes of various forms and kept at least eight times ON-OFF ratio under tangled conditions without critical device failure under coiling radius of 1 mm (left of Fig. 1C and fig. S3) (2629), exhibiting a potential for wearable and flexible neuromorphic electronics.

Fig. 1 Fabrication process of the 1D organic multi-synapses and neural signal transmission.

(A) Fabrication process and schematics of the 1D organic artificial multi-synapses. (B) Cross-sectional and top-surface (bottom left inset) SEM images of the completed 1D synaptic device. (C) Optical photograph of the completed 1D synaptic device (right) and the coiled form on a glass tube with a diameter of 1.2 mm (left). (D) Schematic of the biological multipolar neurons. The strength of the postsynaptic response passing through each synapse is determined by the degree of w. Photo credit: Tae-Wook Kim, Jeonbuk National University.

Signal transmission through 1D artificial multi-synapses

The flexible fiber-shaped device containing multisynaptic channels can imitate the signal transmission of a biological multipolar neuron that propagates divergent signals to several connected neighboring neurons in a high spatial dimension (30, 31). It can propagate a signal from a multipolar preneuron to many other postneurons through connected synapses. In general, an axon of one preneuron is connected parallel to the dendrites of many other postneurons through multiple synaptic clefts where the chemical-to-electrical signal is converted (Fig. 1D) (30, 31). Whenever the presynaptic pulse from an axon of a preneuron stimulates the connected synapses, the output signal at each dendrite of the postneuron [i.e., the postsynaptic response or current (IPSC)] may be fired individually (30, 31). Each output signal is determined by the connection strength (i.e., synaptic weight, denoted w) between the two neurons. For example, when w1 > w2 in Fig. 1D, the dendrite of the first postneuron can generate a higher postsynaptic response than that of the second postneuron, even when the same presynaptic pulse is applied. The dendrite can change its size and shape in response to the presynaptic pulse; therefore, the w is capable of gradual modification. This ability of w to change over the input pulse and time is defined as synaptic plasticity, which is widely recognized to be the fundamental principle underlying memory and learning processes in the human brain (30, 31). In the architecture of our device, the Ag wire acts as the axon of a preneuron and transfers the presynaptic pulse as an electrostatic gating voltage (VG). The pentacene conductance (G) generated at each synapse, which represents the w, determines the IPSC that propagates to the postneuron.

Synaptic characteristics and switching mechanism

The 1D organic multi-synapses can implement two essential synaptic functions distinguished by a temporal or persistent change in the synaptic weight: (i) short-term plasticity (STP) and (ii) long-term plasticity (31, 32). Figure 2A shows the IPSC changes over time generated by different single presynaptic pulses of VG = −30 V with pulse widths (PW) of 100 ms (blue line) and 500 ms (red line) at VDS (the drain-source voltage) = −10 V. When the shorter PW (VG = −30 V for 100 ms) was applied to the preneuron, the IPSC was temporally changed; i.e., its value returned to the original one. This IPSC behavior is analogous to the STP of the biological synapse, which is excited for only a short period. In contrast, when the longer PW (VG = −30 V for 500 ms) was applied to the preneuron, the IPSC increased slightly, from 1.88 × 10−11 to 3.05 × 10−11 A, and then retained this value for a longer period. This IPSC behavior is analogous to the long-term potentiation (LTP) that represents memory consolidation in the brain (30, 32). The VG pulse for increasing IPSC is called the potentiating pulse.

Fig. 2 Synaptic characteristics and switching mechanism of the 1D organic multi-synapses.

(A) IPSC responses triggered by different potentiating pulses of VG = −30 V for PW = 100 and 500 ms at VDS = −10 V. The right schematic presents the cross section of the 1D synaptic device and the postsynaptic responses generated by different degrees of downward P in the P(VDF-TrFE) layer according to the width of VG. (B) IPSC responses triggered by 20 repeated potentiating pulses of VG = −30 V for 300 ms with Δt = 0.5, 1.5, and 8 s, respectively (left). Retention tests after the 20 potentiating pulses with different Δt (middle). The right schematic presents the postsynaptic responses generated by different degrees of downward P in the P(VDF-TrFE) layer for Δt = 0.5 s and 8 s, respectively. (C) LTP and LTD of IPSC with different pulse amplitudes of VG ranging from ±20 to ±40 V. The number of potentiating pulses and depressing pulses is 80. VDS and PW are set as −10 V and 500 ms, respectively. The right schematic presents the postsynaptic responses generated by different degrees of P in the P(VDF-TrFE) layer for different pulse amplitudes (VG = ±20 and ± 40 V) at a fixed PW (500 ms). (D) Plot of the NL and dynamic range (inset) with respect to VG (ranging from ±20 to ±40 V). (E) Repetitive transitions of the LTP/LTD of IPSC for the 1D artificial multi-synapses during 100 cycles. (F) Red and blue circles represent the first and last three cycles, respectively, for the LTP/LTD of IPSC in (E) (marked as red and blue boxes, respectively). (G) Schematic of one bending cycle with R = 2.5 mm. (H to J) Repetitive transitions of the LTP/LTD of IPSC at different fixed bending radii (R = ∞, 5, and 2.5 mm, respectively) after 100 bending cycles. Photo credit: Seonggil Ham, Korea University.

As the switching principle, these synaptic characteristics originated from the change in the polarization (P) of the ferroelectric domains in the P(VDF-TrFE) layer according to the magnitude of the applied VG (33). As shown in the right in Fig. 2A, a longer VG (−30 V for 500 ms) can sufficiently downward-polarize the domain in the P(VDF-TrFE) to the direction of the Ag gate compared with a shorter VG (−30 V for 100 ms). Thus, several hole carriers can be accumulated at the interface between the pentacene channel and the P(VDF-TrFE) layer, increasing the IPSC. The LTP can be mimicked because the accumulated holes at the interface are retained to some extent owing to the remnant downward P. Considering the similarity of the operation to a biological synaptic cleft, the accumulated holes and the downward P degree can be regarded as the quantity of the neurotransmitters released and the concentration of the Ca+ influx into the axon terminal of the preneuron, respectively (30).

In addition, the w can be controlled by changing the number of voltage pulses applied in a given time; that is, the degree of the IPSC increases can be affected by the time interval (Δt) between the potentiating pulses. Figure 2B shows the IPSC changes for 20 repeated pulses (VG = −30 V for 300 ms) with respect to Δt (0.5, 1.5, and 8 s). A shorter Δt (0.5 s) led to a larger and more rapid increase in the IPSC. This is because the potentiating pulses with a shorter Δt further induced the downward domain of P before the postsynaptic response to the previous pulse was completely attenuated. In contrast, the longer Δt (8 s) could fully attenuate the response; hence, the final IPSC hardly changed (right schemes of Fig. 2B). This Δt dependency on the IPSC is similar to the phenomenon of temporal summation of the signal propagation that occurred when the input graded potentials from one biological preneuron were close together in a given time. This is regarded as one of the main features of the spike rate–dependent plasticity (SRDP) (34). Our device also exhibits symmetric spike timing–dependent plasticity (STDP) (35) according to the Δt of the two input spikes between the preneuron and the postneuron (fig. S4) (36). As shown in the middle of Fig. 2B, the different degrees in the decay of IPSC according to the Δt were observed. While a shorter Δt leads to a larger and more rapid increase in the IPSC, the relaxation degree of initial IPSC increases. However, in the case of Δt = 8 s, the IPSC value returns to its original value after ~200 s (like the STP). In the cases of Δt = 0.5 and 1.5 s, the IPSC values are still higher than its original value even after 200 s despite their relaxation (like the LTP). After enough time, however, the IPSC is expected to return to its original values. In addition, we performed the retention test for the completed LTP and long-term depression (LTD) state for 104 s and for different LTP states for 103 s as a function of different pulse numbers from 10 to 30 (fig. S5). We observed commonly that the IPSC corresponding to the LTP state slightly relaxed as the time is increased, which can potentially be exploited for the LTP. This phenomenon might have originated from the polarization relaxation of the ferroelectric domain (3738). The polarization loss can be accelerated by polarization shielding effects of injected and trapped mobile charges and by the depolarization field inevitably remaining in the ferroelectric film during retention interval (3940). However, we should note that the IPSC relaxation could substantially affect the learning accuracy if enough time has passed, which is one of the challenges of the ferroelectric-based synapse (fig. S6). Figure 2C shows the gradual LTP and LTD of IPSC depending on the continuous potentiating and depressing input pulse trains. When the polarity of the presynaptic pulse is changed from negative to positive, the direction of P is changed to the opposite (upward polarization of ferroelectric domain), which can repel the accumulated holes and deplete the interface region. Consequently, the channel conductance (or IPSC) gradually decreases during the positive depressing input pulse train. This gradual IPSC reduction is analogous to the LTD of a biological synapse (30). Because a higher input pulse can further change the direction of P, the dynamic range of IPSC can be increased as the VG increases from ±20 to ±40 V (figs. S7 and S8) (41). For both LTP and LTD, the number of available states is 80.

Simultaneously, the NL, which reflects the nonlinearity in the w update for the LTP and LTD response, is increased at a higher VG because of the rapid saturation of P (Fig. 2D). It is well known that a larger dynamic range and lower NL are beneficial for a higher cognition ability in ANNs (41). Therefore, it is important to identify the optimal operation condition for VG in which LTP and LTD can function. For our 1D synaptic device, VG = ± 30 V for 500 ms was selected as the optimal presynaptic pulse because of the high ratio of the dynamic range to the NL compared with other VG conditions (fig. S7). However, we also found that the conductance levels before and after applying 80 LTP/LTD pulses are not matched well (Fig. 2C), which might cause an accuracy loss during the learning process. This phenomenon might be associated with the deep trapping of the charge carriers at the interface of P(VDF-TrFE) and semiconductor when a high (or repeated) gate electric field was applied (42). In addition, space charges can be injected from electrodes into ferroelectrics and trapped at the boundary of crystallites or/and by defects (43). The interaction between trapped charges and electric dipoles in ferroelectrics can inhibit ferroelectric switching and result in the occurrence of polarization fatigue. For this reason, it needs to constrain the total number of input pulses (up to 30) for each LTP/LTD function to prevent the loss of the learning ability.

Electrical and mechanical stability of synaptic functions

For practical and robust wearable neuromorphic device applications, stable synaptic functions under repeated pulse cycles and mechanically deformed conditions are essential. Figure 2E shows the repetitive transitions between the LTP and LTD functions over 6000 continuous input pulses (corresponding to 100 cycles). The number of presynaptic pulses for one cycle was 60, each consisting of 30 potentiating pulses and 30 depressing pulses (VG = ±30 V and PW = 500 ms). As shown in Fig. 2E, stable LTP/LTD functions under 6000 input pulses were demonstrated. Furthermore, the IPSC values for the first and last three cycles (indicated by the red and blue boxes, respectively, in Fig. 2E) were almost identical (red and blue circles correspond to cycles 0 to 3 and 97 to 100, respectively, in Fig 2F). This indicates the reproducibility and robust control of the synaptic weight. We also performed a mechanical stability test for the LTP and LTD functions, as shown in Fig. 2 (G to J). To facilitate this, the 1D device was placed onto poly(ethylene terephthalate), and both ends of the device were secured with tape. Here, one bending cycle was defined as the device being bent with a certain bending radius (R = 2.5 mm) and then returned to the flat condition (Fig. 2G). After 100 bending cycles, the LTP and LTD functions were repeatedly measured at different fixed bending radii (R = ∞, 5, and 2.5 mm) (Fig. 2, H to J). For R = ∞, 5, and 2.5 mm, the bending strain (ε) was estimated as 0, 3, and 6%, respectively, considering the bending radius and the device thickness (fig. S9) (44). This experiment confirmed that the 1D device can exhibit mechanically stable synaptic functions. In addition to this, we should note here that an additional protecting or shielding layer is required to eliminate expected risks related to their contact with the human body while keeping the electrical function of the devices against various environmental conditions (45).

NOR-type textile array structure consisting of 1D multi-synapses

Owing to its excellent flexibility and multisynaptic channels, our 1D device can be easily extended into a wearable textile array structure, e.g., a NOR-type synaptic array as the ANNs (bottom circuit diagram of Fig. 3A) (36). This textile array resembles a biological neural network (top schematic of Fig. 3A) because the IPSC determined by the w on each channel can be transmitted and integrated. In addition, this array can independently address the channels and control the w, preventing undesired neural signal transmission through other neighboring channels. This undesired leakage could result in the misreading of the IPSC flowing through the designated postneuron, and the learning and memory capability of the array would be severely degraded (24). Although the 1D device provides a direct implementation for an array without a complex fabrication and deposition process, it presents challenges owing to the time-consuming process and the interconnection between the wires (7). Nevertheless, demonstrating this proof of concept provides an adequate means to evaluate the potential of the textile synaptic array, where the postsynaptic response can be properly integrated and propagated when several presynaptic pulses are received with Δt from multiple preneurons (bottom circuit diagram of Fig. 3A). Figure 3B shows the real-time change of the IPSC integrated from the 2 by 2 array (two 1D devices and two Ag wires) when potentiating VG1 and VG2 pulses (−30 V for 500 ms) with Δt = 5 s were alternately applied at each preneuron. Figure 3C shows the change of the IPSC integrated from the 3 by 2 array (three 1D devices and two Ag wires) as a function of the learning phases (#). Each learning phase is determined by a combination of the completed LTP and LTD states (by the VG sweep) of the cells in the 3 by 2 array (top table of Fig. 3C). When the number of learning phases is increased, the integrated IPSC is larger owing to the increased number of the completed LTP states. This demonstrates that the output can be differently changed according to the weight update of each cell in the array. As shown in Fig. 3D, we extended and reconstituted our device into a 10 by 12 array structure on a glass substrate containing 60 synaptic cells using 10 1D devices and 12 Ag wires. Before interconnecting common drain and source lines, we investigated IDS-VG characteristics for individual 60 synaptic cells and analyzed the switching variation (Fig. 3, E and F, and fig. S10). The completed LTP state, the completed LTD state, and the dynamic range of the cells were found to be 1.5 ± 0.1 × 10−8 A, 9.5 ± 1.1 × 10−11 A, and 314 ± 43, respectively, exhibiting distinguishable switching states. After interconnecting with Ag wires in the form of the cross-point array, we encoded the pattern of “k” in the 10 by 12 array matrix, which can generate six IPSC values integrated from “1” to “6” rows in the array (Fig. 3G). Note that the blue and light blue regions in the array represent the completed LTP and LTD states, respectively. Then, we investigated whether each output can integrate the current information of individual synaptic cells. As shown in the right of Fig. 3G, each output signal transmits in the form of a summed current in a common drain line and is dependent on the number of the completed LTP states that are similarly observed in Fig. 3C. These results indicate that the 1D multi-synapses can be extended to the array architecture for the ANNs, where it could allow the individual integration of multiple postsynaptic responses without the undesired leakage. We should note that, however, there are many challenges and issues for realizing the circuit level of ANNs if the 1D synaptic cell having quite low IPSC (~nA) is used. Details are discussed in the Supplementary Materials.

Fig. 3 NOR-type textile array consisting of 1D multi-synapses.

(A) Schematics of the signal transmission in a biological neural network (top) and the circuit diagram of NOR-type 2 by 2 array (bottom), representing the integrated postsynaptic response generated by two presynaptic pulses with different timings. (B) Plot of the integrated IPSC in the 2 by 2 array with alternately applied potentiating VG1 and VG2 pulses (−30 V for 500 ms) with Δt = 5 s at each preneuron. The inset is a photograph of the 2 by 2 array. (C) Plot of the integrated IPSC in the NOR-type 3 by 2 array with respect to the learning phases from 1 to 8. The top table shows the learning phases that are determined by a combination of the completed LTP and LTD states (by the VG sweep) of the cells (A, B, and C cells). The inset is a photograph of the 3 by 2 array. Note that each integrated IPSC was statistically obtained from three different 3 by 2 arrays. (D) A photograph for the 10 by 12 array on a glass substrate where it is placed inside the probe station (top). The inset shows a magnified photograph for 6 by 2 array. Schematic of the 10 by 12 array structure on a glass substrate containing 60 synaptic cells using 10 1D devices and 12 Ag wires (bottom). (E) Overlay plots of IDS-VG curves of the 60 synaptic cells at VDS = −10 V. (F) Statistical histograms of IDS for the completed LTP (ON) and LTD (OFF) states and the dynamic range for the 60 cells. (G) Encoded pattern of k in the 10 by 12 array matrix (left). The blue and light blue regions in the array represent the completed LTP and LTD states, respectively. The right graph shows six IPSC values integrated from 1 to 6 rows in the array. Photo credits: (B and C) Seonggil Ham, Korea University; (D) Tae-Wook Kim, Jeonbuk National University.

Recognition simulation for MNIST and ECG patterns

To evaluate the learning capability of the textile array comprising 1D multi-synapses, we simulated MNIST pattern recognition based on the fitting results of the LTP and LTD at the selected VG = ±30 V for 500 ms (figs. S11 and S12 and table S1) (41, 4649). After only 10 learning epochs, recognition accuracy of ~90% was achieved, which is slightly higher than those of other synaptic device platforms with a single-layer neural network (Fig. 4A) (41, 5052). Considering the cell-to-cell variation of the LTP/LTD functions, however, the recognition accuracy is decreased up to ~82% at 10 epochs (figs. S13 and S14). This is because the change of consequential vector calculation between the input signal and programmed conductance configuration would result in some inversion for output neurons. Here, one epoch corresponded to 60,000 learning phases. The left inset of Fig. 4A shows reshaped 28 by 28 contour images from the w before and after 10 epochs. The final contour image after 10 epochs had a similar pattern to the number “3,” indicating a well-recognized output. The confusion matrix in the right inset of Fig. 4A exhibits the classification result between the output and targeted MNIST patterns. We used 60,000 and 10,000 different MNIST patterns for the learning and the inference test, respectively. The diagonal tiles in the matrices are mostly saturated in red, indicating that each digit pattern was correctly inferred. The details of the single-layer neural network and learning algorithm were described in previous work and are also shown in figs. S15 and S16 (24, 53). According to the accuracy result, the e-textile neural network based on our 1D multi-synapses could perform the learning and recognizing tasks properly.

Fig. 4 Recognition simulation for MNIST and ECG patterns.

(A) Recognition accuracy for MNIST patterns with respect to the number of learning epochs. The left inset shows a reshaped 28 by 28 contour image of the digit 3 from w before and after 10 epochs, and the right inset shows the confusion matrix for a classification test involving 10,000 MNIST images after 10 epochs. (B) Schematics of the five classes of ECG waveforms: N, S, V, F and Q. (C) Constituents of a single-layer neural network for the N ECG pattern recognition process in which 187 input neurons and 5 output neurons are fully connected by 935 artificial synapses. (D) Changes in the w values of 187 artificial synapses connected to the output neuron corresponding to the N class during 3200 learning phases. (E) Changed w values from all ECG classes after 50 learning epochs. (F) Recognition accuracy for the ECG patterns during 50 learning epochs in the initial state and after 100 bending cycles with different fixed bending radii (R = ∞, 5, and 2.5 mm). (G) Confusion matrices for a classification test of the 800 ECG patterns in the initial state and after 100 bending cycles with different fixed bending radii (R = ∞, 5, and 2.5 mm).

Beyond simple digit recognition, to apply a wearable and intelligent health care device to the human body, the ability to immediately and accurately classify diverse biometric sensing information even under the bending condition is necessary. As a proof of concept, we selected a heartbeat dataset comprising five different arrhythmia ECGs categorized by different waveforms and morphologies (54, 55). ECGs can be reliably and effectively used for measuring the functionality of the cardiovascular system. Each ECG signal can be represented as “N”, “S”, “V”, “F”, or “Q” class according to normal and abnormal heartbeats that can be used to diagnose cardiovascular diseases (Fig. 4B) (55). Details regarding the classification criteria for the ECG are presented in table S2 (5557). However, in most cases, the waveforms of ECG signals appear very similar to each other and are difficult to classify, making the real-time diagnosis of the relevant diseases prone to errors. The resemblance between the ECG waveforms makes it difficult to interpret the biometric states of the heartbeat without expert medical knowledge, which might limit the development of a smart ubiquitous health care monitoring system.

To interpret the ECG signals using our e-textile neural network, we used the shape-related features of each signal preprocessed with a sampling frequency of 125 Hz extracted from the dataset used by Kachuee et al. (55). Via this process, we separated each ECG waveform into 187 sequential electrical potentials with an interval of 8 ms and then normalized each potential in the range of 0 to 1. Figure 4C shows an example of a single-layer neural network for the N class recognition process. This network consists of 187 inputs (corresponding to the number of electrical potentials) and five output neurons (corresponding to five ECG classes) connected by 935 artificial synapses. When each electrical potential in the ECG waveform is fed to each corresponding input in the neural network, the corresponding w is repeatedly updated through the stochastic gradient ascent/descent learning algorithm, which is the same method that was used for the MNIST pattern recognition (figs. S15 and S16). Figure 4D shows the changes in all the w values (187 weights) connected to the output neuron corresponding to N during the learning phases (equal to one epoch in the case of the ECG). As shown, they became a representative feature of the N class. After 50 epochs, the w for each class had distinctive and representative waveforms, implying that the learning process for the ECG signals was complete (Fig. 4E). However, the accuracy was ~70% even after 50 epochs, which was lower than that for the MNIST patterns. This might be associated with the limited ECG dataset, which yielded a nonoptimal w in the neural network. We used 3200 and 800 different ECG patterns for the learning and the inference test, respectively; thus, the dataset was substantially smaller than that of the MNIST patterns. In such a small dataset, it could be more complicated to extract a common feature of the ECG patterns (fig. S17), and, we excluded all the synaptic characteristics of the device and used only arithmetic operation that can be considered as an ideal recognition process in the present dataset (fig. S18). As shown in fig. S18, the maximum accuracies are ~91 and ~71% for MNIST and ECG patterns, respectively, which are slightly higher than the simulation result reflecting the device information. Hence, we think that the recognition accuracy can be improved by training the network with a larger amount of ECG signals and not by the improvement of device performance and the change of the recognition process.

To confirm the mechanically stable ECG recognition capability, we estimated the accuracies according to the fitting results of the LTP and LTD functions of the 1D multi-synapse at different fixed bending radii (R = ∞, 5, and 2.5 mm) after 100 bending cycles (Fig. 4F). Because each LTP and LTD function changed very little even under the bending conditions (Fig. 2, G to J, and table S3), the ECG recognition accuracy was almost maintained within ~2% degradation. This result can be concretely visualized by the confusion matrix (5 by 5) for inferred and targeted ECG signals (Fig. 4G). A similar degree of color saturation of the individual tiles in the confusion matrices was observed regardless of the bending conditions, and the diagonal tiles exhibited the greatest saturation, as expected. All the results are consistent with Fig. 4F.


In summary, 1D fiber-shaped artificial multi-synapse–based ferroelectric organic transistors were successfully fabricated on a thin Ag wire. Diverse and well-defined synaptic functions, such as STP, LTP, LTD, SRDP, and STDP, were demonstrated. In addition, the transitions between the LTP and LTD functions were stable under repeated pulse cycles, even after 100 bending cycles at a fixed bending radius (R = ∞, 5, and 2.5 mm). Our 1D multi-synapses can be used as the basic component for an e-textile neural network because they can be easily extended in the form of a NOR-type synaptic array. Various textile synaptic arrays such as 2 by 2, 3 by 2, and 10 by 12 were fabricated and investigated as a proof of concept, and the propagation of the integrated output signal through each synaptic cell without undesired neural signals was demonstrated. Notably, the 1D multi-synapses achieved up to ~90 and ~70% recognition accuracy for MNIST and ECG patterns, respectively. In particular, the initial accuracy for the ECG pattern was maintained within a margin of error of ~2% regardless of the mechanical bending stress. The results indicate that our 1D synaptic architecture represents an important step toward the development of an e-textile neural network for wearable neuromorphic electronic applications.


Fabrication process of 1D multi-synapses

Silver wires (φ = 0.1 mm, 99.99%; Nilaco Corp.), as the gate electrode, were used to fabricate fiber-shaped multisynaptic device with a bottom-gate/top-contact structure. The wires were washed with isopropyl alcohol and dried by air blowing. P(VDF-TrFE) (70:30 mole percent copolymer) purchased from Solvay was dissolved in N,N-dimethylformamide (99.8% purity; 28 weight %; Sigma-Aldrich) and deposited on the metal wires by dip-coating (30 mm/min) using a capillary tube and a printing speed controller, as shown in Fig. 1A. The P(VDF-TrFE) film was annealed at 140°C for 1 hour in a convection oven. Then, a 50-nm-thick layer of pentacene (99.9% purity; Sigma-Aldrich) was evaporated at a pressure of approximately 5 × 10−7 torr. Last, Au source/drain electrodes (50 nm thick) were deposited on the pentacene through a shadow mask using thermal evaporation. The channel length (L) and width (W) of the 1D multi-synapses were 15 and 157 μm, respectively. Fiber-shaped capacitors were fabricated on an Ag wire with Au/P(VDF-TrFE)/Ag layers, and the dielectric capacitance was evaluated at different positions along the fiber (fig. S1).

Thin film and device characterization

Field-emission SEM (FEI Nova NanoSEM 200) was used to measure P(VDF-TrFE) film morphology and thicknesses of each constituent layer in the 1D multi-synapse. The electrical characteristics of the 1D fiber-shaped multisynaptic devices were measured using a semiconductor parameter analyzer (4155C, Keysight) equipped with a pulse generator (81104A, Keysight) and a Keithley 4200 semiconductor characterization system.

Single layer–based simulation for pattern recognition

Neural network simulations with single layer based on 1D multi-synapses were performed using the MNIST handwritten dataset and ECG Heartbeat Categorization Dataset, which was preprocessed and segmented by M. Kachuee from the MIT-BIH Arrhythmia Dataset. The further details of configuration of the network and dataset are described in figs. S11, S12, S15, and S16 and table S2.


Supplementary material for this article is available at

This is an open-access article distributed under the terms of the Creative Commons Attribution-NonCommercial license, which permits use, distribution, and reproduction in any medium, so long as the resultant use is not for commercial advantage and provided the original work is properly cited.


Acknowledgments: We appreciate J. H. Park and I. Yeo for discussion and comments on simulation results. Funding: This work was supported by the National Research Foundation of Korea (NRF-2019R1A2C2003704, 2019R1A6A3A01095700, and 2020R1A2C2010163), the KU-KIST Research Fund, a Korea University Grant, and Samsung Electronics. Author contributions: T.-W.K. and G.W. conceived and designed the experiments. S.H. and M.K. performed the experiments and analyses. S.J., J.J., and S.C. performed the pattern recognition simulation. All authors discussed the results and contributed to the manuscript. G.W. oversaw the project, revised the manuscript, and led the effort to completion. Competing interests: The authors declare that they have no competing interests. Data and materials availability: All data needed to evaluate the conclusions in the paper are present in the paper and/or the Supplementary Materials. Additional data related to this paper may be requested from the authors.
View Abstract

Stay Connected to Science Advances

Navigate This Article