Harnessing glycol–alkyl copolymerization to realize nonvolatile and biologically relevant synaptic behaviors

Abstract

Organic electrochemical synaptic transistors (OESTs) are attracting growing attention for neuromorphic computing, yet their long-term stability remains constrained by uncontrolled ion dynamics. Previous studies have incorporated glycol side chains to facilitate ionic transport, but a systematic understanding of how copolymerization with hydrophobic alkyl units governs ion doping and retention is still lacking. Here, we establish a rational backbone–side chain copolymer design strategy that precisely regulates ionic interactions, crystallinity, and charge transport. We also reveal clear correlations between copolymer structure, ion dedoping dynamics, and nonvolatile retention. These structural advantages enable the faithful emulation of key biological behaviors including paired-pulse facilitation, spike-timing dependent plasticity, and long-term potentiation/depression (LTP/D) with high linearity and stability. Based on these properties, the device achieved a high accuracy of 94.1% in ANN-based recognition simulations for MNIST handwritten digits. This work demonstrates that systematic glycol–alkyl copolymer engineering provides a robust and predictive design principle for high-performance neuromorphic synapses, moving beyond empirical side-chain modifications.

---

New concepts

OESTs are essential synaptic devices for implementing neuromorphic computing because they can mimic neurotransmitter release through ion diffusion within the polymer channel layer from electrolytes. Therefore, research is underway to introduce glycol side chains into organic materials to enhance device performance. However, systematic studies on the effects of copolymerization between hydrophobic alkyl units and hydrophilic glycol units on device characteristics remain insufficient. This study clearly establishes the correlation between structural changes induced by copolymerization of glycol and alkyl chains and the resulting property changes, including ion doping mechanisms. Copolymerization of glycol chains effectively controls ion transport dynamics and enhances nonvolatile performance, enabling the realization of key biological synaptic functions such as PPF, STDP, high-pass filtering, and long-term potentiation/depression (LTP/D). These findings establish systematic glycol–alkyl copolymerization as a robust molecular design for high-performance artificial synaptic devices and provide new insights for realizing next-generation neuromorphic computing.

Introduction

The advent of the big data era has led to the rapid growth of artificial intelligence, which has dramatically increased the amount of data that needs to be stored and processed.^1–7 However, the traditional von Neumann architecture, in which memory and computation units are separated, and operations are performed sequentially, suffers from severe bottlenecks in data transfer and consumes large amounts of energy.^8–10 These intrinsic limitations have spurred the search for new computing methods capable of processing vast amounts of complex data both quickly and efficiently.^11–13 The human brain operates at low power, possesses high fault tolerance, and can rapidly process data storage and processing in parallel.^14–17 Accordingly, neuromorphic computing, which mimics human brain-like information processing, has emerged as a next-generation computing paradigm and has stimulated extensive research efforts to realize practical hardware platforms.^18–21

To implement neuromorphic computing, artificial synapse devices that reproduce the function of biological synapses, the elementary units of the neural network, are indispensable.²² Previous approaches focused on memristors, two-terminal devices that exhibit resistance-switching and nonvolatile memory characteristics under applied bias.^23–27 Although memristors can emulate synaptic functions, their structural simplicity often results in poor selectivity and stability.^28–30 To address these limitations, three-terminal transistors have been proposed, introducing an additional gate electrode to provide improved control over external stimuli.^31–33 Inorganic semiconductors-based transistors offer fast switching speeds and high durability, but they require high-temperature processes and complex fabrication procedures.^34,35 In contrast, Organic semiconductors-based transistors have attracted widespread interest because of their compatibility with low-cost fabrication, rich synthetic tunability, and efficient ion–electron coupling.^36–40 Among these devices, organic electrochemical synaptic transistors (OESTs) have shown promise by mimicking neurotransmitter release in biological synapses through ion diffusion into the polymer channel layer. This mechanism provides rapid switching and controllable conductance modulation, making OESTs a powerful platform for studying neuromorphic function.^41–44

To realize high-performance artificial synapse devices, it is essential to achieve effective control of synaptic weights together with robust nonvolatile characteristics. One promising strategy has been the incorporation of polar glycol side chains into the polymer channel layer.^45–48 Glycol groups not only enhance the solubility of conjugated polymers but also facilitate smooth ionic transport, thereby improving doping efficiency and synaptic response.^49–51 Several studies have investigated blending alkylated polymers with glycolated polymers or tuning the length of glycol side chains to optimize performance.^52,53 These studies show that glycol incorporation is a rational approach for enhancing synaptic properties. Nevertheless, most of these efforts have been limited to empirical modification, and systematic investigations of copolymer structures that combine hydrophilic glycol side chains with hydrophobic alkyl side chains remain scarce. A deeper understanding of how such copolymerization governs ion doping, back diffusion, and retention characteristics is still required.

In this study, we demonstrate that copolymerization of alkyl and glycol side chains offers an effective strategy to control ion transport dynamics and improve nonvolatile performance in OESTs. By monitoring current changes under applied bias, we show that glycol incorporation suppresses ion back diffusion and strengthens ion–polymer interactions. Complementary ultraviolet-visible spectroscopy and grazing incidence wide-angle X-ray scattering (GIWAXS) analyses further reveal how glycol copolymerization enhances doping efficiency and modifies crystalline ordering within the polymer matrix. These structural advantages enable the fabricated devices to sustain current for over ∼10² seconds after pulse application and faithfully reproduce key biological synaptic behaviors, including paired-pulse facilitation (PPF), spike-timing dependent plasticity (STDP), high-pass filtering, and long-term potentiation/depression (LTP/D). Building upon these properties, the copolymer-based OESTs achieve 94.1% accuracy in artificial neural network simulations for handwritten digit recognition. Collectively, our results establish systematic glycol–alkyl copolymerization as a robust molecular design principle for advancing high-performance artificial synapses and provide new insight into the structure–transport–function relationships that underpin neuromorphic computing.

Results and discussion

Fig. 1a shows the signal transmission process in the human nervous system, which consists of presynaptic and postsynaptic neurons and neurotransmitters. When a presynaptic neuron receives an input spike, it releases neurotransmitters toward the postsynaptic neuron. The receptors on the cell membrane of the postsynaptic neuron then absorb the neurotransmitters and generate an excitatory postsynaptic potential (EPSP). At this point, the synaptic weight can be adjusted depending on the intensity and interval of the stimulus. Based on this basic principle, biological behaviors, including synaptic weight updates, are implemented. To mimic biological synapses, our artificial synapses adopt the structure of organic electrochemical synaptic transistors (OESTs). This structure contains electrolytes that function as neurotransmitters, with the gate and source/drain electrodes performing the roles of presynaptic and postsynaptic neurons, respectively. In the OESTs, an electric field is formed in the electrolyte by the input pulse applied to the gate electrode, causing mobile ions to interact with the polymer film, which acts as a receptor, resulting in the generation of excitatory postsynaptic currents (EPSCs) and mimicking biological behavior (Fig. 1b). To investigate the effect of glycol groups on the copolymer synaptic properties, we compared films containing only alkyl groups with polymer films copolymerized with glycol groups as the channel layer. Fig. 1c shows the polymer structure. The polymer containing only alkyl chains (OD) is poly[3-(5-(3,6-dibutoxy-5-methylthieno[3,2-b]thiophen-2-yl)thiophen-2-yl)-6-(5-methylthiophen-2-yl)-2,5-bis(2-octyldodecyl)-2,5-dihydropyrrolo[3,4-c]pyrrole-1,4-dione] (ODTT), and the polymer copolymerized with alkyl chains (OD) and glycol chains (EG) in a 1 : 1 ratio is poly[3-(5″-(2,5-bis(2-(2-(2-methoxyethoxy)ethoxy)ethyl)-4-(5′-methyl-[2,2′-bithiophen]-5-yl)-3,6-dioxo-2,3,5,6-tetrahydropyrrolo[3,4-c]pyrrol-1-yl)-[2,2′ : 5′,2″-terthiophen]-5-yl)-6-(5-methylthiophen-2-yl)-2,5-bis(2-octyldodecyl)-2,5-dihydropyrrolo[3,4-c]pyrrole-1,4-dione] (EGODTT) (Fig. 1d and Fig. S1–S6, SI). To fabricate artificial synapses, ODTT and EGODTT were spin-coated onto a Si/SiO₂ substrate to form the channel layer, and Au electrodes were deposited using a thermal evaporator (for details, see Experimental section in SI).

Fig. 1 Synapse transistors based on ODTT and EGODTT organic semiconductors. (a) Schematic diagram of the biological synapse in the human brain, consisting of presynaptic neurons, postsynaptic neurons, and neurotransmitters. (b) Structure of OESTs fabricated by simulating biological synapses. (c) Chemical structures of OD (left) and EG (right). (d) Mixing ratio of ODTT and EGODTT.

We measured the transfer curve to analyze the electrical characteristics of OESTs. Fig. 2a and b shows the transfer curves of ODTT- and EGODTT-based OESTs, respectively. The properties of the devices were obtained by sweeping the gate voltage from +3 to −3 V under a drain voltage (V_D) of −0.2 V. To quantitatively compare the performance of each OESTs, we extracted the hysteresis window (ΔV_hys), threshold voltage (V_th), maximum current (I_D,max), and transconductance (g_m) from the transfer characteristics (Fig. 2c). The hysteresis window was defined as the voltage difference between the forward sweep and reverse sweep that results in half of the maximum current. The transconductance was defined as the output current with respect to voltage applied and calculated using the following equation: g_m = ∂I_D/∂V_G. The measurement results showed that the EGODTT-based devices exhibited a wider hysteresis window than the ODTT-based devices. The enhanced performance is attributed to glycol copolymerization, which facilitates efficient ion doping and increases channel polarity, resulting in slower ion back diffusion within the electrolyte. Additionally, EGODTT-based devices were found to have a relatively low V_th, large I_D,max, and g_m. These results suggest that copolymerizing a monomer containing a glycol chain with a monomer having alkyl groups enables efficient charge transport and achieves stronger bonding between the anions and the polymer. To verify the statistical validity of these results, transfer curves were obtained in five different devices for each film (see Fig. S7, SI). The five devices exhibited similar characteristics, demonstrating that the EGODTT-based devices exhibit enhanced performance.

Fig. 2 Electrical properties of artificial synapses based on polymer films. Transfer curves of OESTs based on (a) ODTT and (b) EGODTT copolymers. Performed at gate voltage from −3 V to 3 V and drain voltage of −0.2 V. (c) Hysteresis window ΔV_hys, threshold voltage V_th, maximum current I_D,max, and transconductance g_m calculated from the transfer curves of ODTT and EGODTT. (d) Short-term plasticity properties of ODTT and EGODTT represented by changes in EPSC for a single pulse (t_width = 60 ms). Long-term plasticity properties observed when applying 10 pulses (t_width = 60 ms and t_interval = 60 ms) with different amplitudes (−2.0, −2.25, −2.5, and −2.75 V) to (e) ODTT and (f) EGODTT. (g) Schematic diagram showing the de-doping process in the electric double layer domains, amorphous domains, and crystalline domains. (h) Comparison of relaxation time constant values calculated using the tri-exponential function for ODTT and EGODTT. (i) Normalization of retention behavior investigated after applying 10 pulses with a gate voltage of −2.75 V.

To investigate the effect of glycol copolymerization on synaptic plasticity, we applied pulse-formed gate voltages (V_G) and observed changes in EPSCs. To realize short-term plasticity (STP) in ODTT and EGODTT devices, we applied single pulses of different amplitudes (t_pulse = 60 ms, V_G = −2.75, −2.5, −2.25, and −2 V) under V_D = −0.2 V (Fig. 2d). All devices exhibited the highest EPSC peak when a −2.75 V pulse was applied. ODTT-based devices returned to their initial state after all pulse amplitudes were applied. In contrast, EGODTT-based devices showed a larger change in EPSC amplitude and maintained a higher EPSC than the initial state after the application of pulses of −2.5 V or higher, demonstrating the potential for nonvolatile memory characteristics. To confirm these characteristics, we investigated long-term plasticity (LTP) properties by observing EPSC changes when 10 pulses under the same conditions were applied at 60 ms intervals (Fig. 2e and f). The ODTT-based OESTs exhibited a retention of 10.9% relative to the peak EPSC at a pulse of −2.75 V, while the EGODTT-based OESTs showed a distinct trend of EPSC changes at pulses of −2.25 V or higher, exhibiting a maximum retention of 38.3% and significantly improved non-volatile characteristics. These non-volatile characteristics arise due to ion movement at the channel/electrolyte interface. To understand the effect of ions mobility on LTP characteristics, the EPSC reduction in LTP characteristics at a −2.75 V pulse was fitted to a tri-exponential function (see Fig. S8a and b, SI):

I ₀ and I_n are the initial current before applying the spike voltage and the pre-coefficient of EPSC, respectively. τ_n is the relaxation time constant, which is divided into three regions depending on 1, 2, and 3: τ₁ is the EDL depolarization time constant, τ₂ is the dedoping time constant in the amorphous region, and τ₃ is the dedoping time constant in the crystalline region (Fig. 2g).^37,54 In particular, the time constant in the crystalline region has the longest value and is closely associated with long-term nonvolatile characteristics.⁵⁵ The calculation results showed that ODTT has a τ₃ value of 4.66 s, while EGODTT exhibited an increased value of 7.97 s (Fig. 2h). This indicates that the ions migrating to EGODTT exhibit delayed back diffusion behavior due to the glycol groups. Additionally, we examined the EPSC decay behavior over 200 seconds to confirm the LTP characteristics over a longer time period. A comparison of memory levels, which indicates the retention behavior over time relative to the peak values of the two devices, ODTT exhibited a low retention characteristic of 5.57%, while EGODTT exhibited a strong non-volatile characteristic of 24.36% (Fig. 2i). The decay behavior of EPSCs after pulse application was observed by differentiating the EPSC over time to assess ion de-doping levels. The differentiated curve for EGODTT showed a gradual decay toward zero compared to ODTT, indicating relatively slower ion de-doping over time. (Fig. S8c). These results suggest that non-volatile characteristics can be enhanced by controlling the behavior of mobile ions through the copolymerization of the glycol group.

To understand the improvement in non-volatile characteristics due to glycol group copolymerization, changes due to doping were investigated using ultraviolet-visible (UV-vis) spectroscopy. Fig. 3a and b showed the normalized UV-vis spectra of ODTT and EGODTT films, respectively, in situ as a function of applied voltage. The doping state was examined by applying a voltage of 0 to −1 V to each device. In both polymer films, the absorption band at the neutral wavelength (700–900 nm) decreased (reduction in the π–π* peak), and the absorption band at the polaron/bipolaron wavelength (900–1200 nm) increased. Applying a negative bias injects holes into the p-type polymer channel, thereby reducing the population of neutral states and inducing new polaronic energy levels. These effects become more pronounced at higher applied voltages, resulting in the formation of additional polarons and bipolarons.⁵⁶ To compare the degree of doping with respect to voltage amplitude, the reduction in the absorption peak at the neutral wavelength was compared, and a larger reduction in EGODTT with the application of voltage (Fig. 3c). These results indicate that even at the same voltage, stronger doping can be induced, achieving elevated doping efficiency.

Fig. 3 Changes in normalized UV-vis absorption spectra during doping of (a) ODTT film and (b) EGODTT film applying various voltages (0 to −1 V). (c) Comparison of normalized peak absorption. (d) ODTT and (e) EGODTT organic films analyzed by 2D-grazing incidence wide-angle X-ray scattering (GIWAXS). (f) In-plane, q_xy lineouts. (g) Out-of-plane, q_z lineouts. (h) Comparison of d-spacing of (100) and (010) peaks in-plane. (i) Comparison of d-spacing of lamellar structures (100), (200), (300), and (400) peaks.

To observe changes in the crystalline orientation of the film due to copolymerization of the glycol group, we performed grazing incidence wide-angle X-ray scattering (GIWAXS) analysis (Fig. 3d and e). Both polymers exhibited a strong (010) peak corresponding to in-plane (IP) pi–pi stacking and a distinct (h00) peak corresponding to out-of-plane (OoP) lamellar stacking. These results indicate that all films exhibit an edge-on dominant crystalline orientation. Furthermore, to confirm these crystalline differences, in-plane and out-of-plane data were extracted (Fig. 3f and g). EGODTT films exhibited larger q values for all peaks compared to ODTT films. This allowed us to calculate the d-spacing using the following equation: d = 2π/q. EGODTT films showed reduced lamellar and pi–pi stacking distances compared to ODTT films (Fig. 3h). Additionally, it exhibited reduced d-spacing values for (h00), indicating a more compact crystal structure (Fig. 3i). These results are expected to be attributed to the low steric hindrance and high flexibility of the glycol group.^57–59 Furthermore, the improved crystallinity aligns with previous research findings on performance enhancements associated with improved crystallinity.^60–62

Artificial synapse devices mimic the functions of biological synapses that enable humans to learn and remember. To verify whether the devices can perform such biological synapse behaviors, we measured pair-pulse facilitation (PPF), spike-timing dependent plasticity (STDP), and high-pass filtering. PPF is a phenomenon in which applying two consecutive pulses results in a stronger response to the second stimulus (A₂) than to the first spike (A₁). To implement this biological behavior, we applied a pair of consecutive pulses to our device and observed enhanced EPSC responses through the coupling between ions and the film. In Fig. 4a, the EPSC peak when two pulses with a pulse width of 60 ms were applied to the EGODTT-based OESTs at an interval of 90 ms can be seen, with A₂ showing a 165% enhancement compared to A₁. Based on this phenomenon, we examined changes in the PPF index (A₂/A₁ × 100%) as a function of pulse time interval and confirmed that the PPF index decreased as the time interval increased. To analyze the decrease in the PPF index with increasing time intervals, we fitted the PPF index using the following double exponential function:

where Δt is the pulse application interval time, C₀ is the PPF constant when the pulse interval is infinite, C₁ and C₂ are the initial promotion sizes of the rapid and slow phases, respectively, and τ₁ and τ₂ are the relaxation time constants of each phase.⁶³ The fitting results showed that C₀ = 1.08, τ₁ = 16.3 ms, and τ₂ = 297 ms. When comparing these values to those of biological synapses, which have C₀ = 1, τ₁ = 40 ms, and τ₂ = 300 ms, we can confirm that our device behaves similarly to human synapses.⁶⁴

Fig. 4 Synaptic characteristics of EGODTT-based devices. (a) PPF index (A₂/A₁) observed over time (when V_pre = −2.75 V, width = 60 ms, V_ds = −0.2 V). Inset: When time interval (Δt) is 90 ms. (b) Symmetrical Hebbian learning using synaptic weights that change according to two potentiation pulses applied at intervals between pre- and post-synaptic spikes, and (c) anti-Hebbian learning represented using depression pulses. (d) High pass filtering properties are calculated based on the ratio of A₁₀/A₁ at frequencies between 3 and 15 Hz. The inset shows the change in EPSC when 10 pulses are applied at frequencies of 3 Hz and 15 Hz. (e) Changes in the maximum EPSC peak when applying 10 consecutive pulses with increasing width, and (f) when applying pulses with a constant width of 60 ms while increasing the number of pulses.

Spike-timing dependent plasticity (STDP) is an important synaptic characteristic that learns and remembers according to the Hebbian learning rule, whereby synaptic weights are adjusted according to the time interval and sequence of stimuli.⁶⁵ STDP is divided into symmetric Hebbian learning and asymmetric Hebbian learning depending on the order in which pulses are applied. If the postsynaptic pulse (first pulse) is applied before the presynaptic pulse (second pulse), Δt takes on a positive value, and if it is applied later, Δt takes on a negative value. Δt represents the difference between the time when the post-synaptic pulse is applied and the time when the pre-synaptic pulse is applied. When a negative voltage was applied to the EGODTT device, it exhibited symmetric Hebbian learning, and when a positive voltage was applied, it exhibited asymmetric Hebbian learning (Fig. 4b and c). This confirms that the device exhibits LTP behavior and LTD behavior, respectively, depending on whether a negative or positive voltage is applied.⁶⁶

Human synapses can perform dynamic filtering during the process of processing input information. Through this filtering process, high-frequency signals are amplified, low-frequency signals are weakened, and electrical signals at specific frequencies are allowed to pass through. To verify the high-pass filtering behavior of the device, we applied 10 pulse voltages continuously in the frequency range of 3–15 Hz and compared the ratio of EPSC after the 10th pulse application to that after the first pulse application (A₁₀/A₁) (Fig. 4d). As the frequency increases, the EPSC amplification ratio (A₁₀/A₁) also increases, demonstrating that the EGODTT-based devices can mimic high-pass filtering behavior. Additionally, Fig. 4e and f shows the behavior of the EPSC peak as a function of the number and width of pulses and confirmed that the peak increased linearly as the number of pulses increased from 4 to 20 and the width increased from 30 to 240 ms (see Fig. S11c and d, SI). This suggests that synaptic plasticity can be modulated by varying the number and width of pulses applied to the device. The ODTT-based devices also exhibited PPF, STDP, and high-pass filtering behavior under the same conditions, and changes in the EPSC peak after adjusting the number and width of pulses were also observed (see Fig. S10, S11a and b, SI).

Long-term potentiation/depression (LTP/D) characteristics are considered important properties in neuromorphic computing because they lead to the learning and erasing capabilities of synaptic devices. To investigate the effect of glycol copolymerization on the LTP/D characteristics of the devices, we applied 50 potentiation pulse of −2.25 V and 50 depression pulse of +2.25 V to ODTT and EGODTT-based devices and examined their weight update behavior (Fig. 5a and b). When comparing the behavior, the EGODTT-based devices exhibited a more symmetric and linear graph. Particularly in LTD, EGODTT exhibited a more gradual EPSC decay compared to ODTT, which was confirmed by the differentiated curve also gradually decreasing to zero with increasing pulse number (see Fig. S12c, SI). To confirm the performance differences due to copolymerization, we quantitatively compared the dynamic range (G_max/G_min), the number of effective states (NS_eff), and nonlinearity (NL) data extracted from the LTP/D curves in Fig. 5c. The EGODTT-based devices exhibited a larger dynamic range, a higher number of effective states, and lower nonlinearity values in the LTP/D curves. NS_eff refers to the number of states where the conductivity difference between pulses exceeds the 0.5% noise level of G_max − G_min. The nonlinearity (NL) data, which indicates the linearity of conductivity changes in LTP/D, was calculated using the following equation (see Fig. S12d and e, SI):

where G_LTP, G_LTD, G_min, and G_max represent the potentiation, depression, minimum, and maximum conductance of OESTs, respectively. P indicates the number of pulses, P_max indicates the maximum number of pulses, and A_p and A_D are parameters used to calculate the linearity of potentiation and depression conductance.^67,68 Calculations showed that the EGODTT-based artificial synapse exhibited higher G_max/G_min and NS_eff values and lower NL values compared to the ODTT. Furthermore, to verify the stability of the EGODTT-based OESTs during repeated operation, the LTP/D behavior was observed over 10 cycles (1000 pulses) and found that the performance remained similar without any noticeable changes (Fig. 5d). In Fig. 5e, we compared G_max/G_min, NS_eff, NL_LTP, and NL_LTD over 10 cycles. Each parameter values showed no significant degradation, thereby validating the stability of the EGODTT-based devices.

Fig. 5 LTP/D properties of ODTT- and EGODTT-based organic synapse transistors and their application to neuromorphic computing. LTP/D characteristics observed in (a) ODTT and (b) EGODTT devices when 50 potentiation pulses (−2.25 V, 60 ms) and 50 depression pulses (+2.25 V, 60 ms) were applied at V_DS = −0.2 V. (c) Dynamic range (G_max/G_min), number of effective states (NS_eff), and nonlinearity (NL) from the LTP/D curves of ODTT and EGODTT. (d) LTP/D curve of EGODTT during 10 cycles. (e) G_max/G_min, NS_eff, and NL values according to 10 cycles. (f) A multilayer neural network structure consisting of 400 input neurons that receive 28 × 28 pixel images as input, 100 hidden neurons, and 10 output neurons with values ranging from 0 to 9. (g) Confusion matrix for recognition accuracy of the MNIST dataset at epoch 120. (h) Comparison of recognition accuracy over 125 epochs.

To demonstrate that the final OESTs can perform parallel computations in neuromorphic computing, we conducted recognition simulations on MNIST handwritten digits using a multi-layer perceptron (MLP)-based artificial neural network (ANN) (Fig. 5f).⁶⁹ To reflect the characteristics of the device in the computation, we considered values such as dynamic range and NL extracted from the LTP/D curves.⁷⁰ The MLP consisted of 400 input neurons receiving 28 × 28 pixel images, 100 hidden neurons, and 10 output neurons representing numbers from 0 to 9. Fig. 5g shows the confusion matrix for the 10 digits in a dataset of 10 000 MNIST images. This confusion matrix visualizes the output digit corresponding to the target digit, indicating whether the output value accurately represents the desired value without confusion. In the case of 95 epochs, the distribution of accurate output digits according to the target digit clearly forms a diagonal pattern. Based on this principle, when comparing the handwriting recognition accuracy of the two devices over 125 epochs, the ODTT-based OESTs showed up to 91.7% accuracy, while the EGODTT-based OESTs achieved up to 94.1% accuracy, indicating that the EGODTT films can recognize handwriting more accurately than the ODTT films (Fig. 5h). These results suggest that higher recognition accuracy can be achieved by improving the performance of neuromorphic computing through the copolymerization of alkyl groups.

Conclusion

In this study, we achieved improved synaptic properties by copolymerizing glycol groups to enhance the electrical interaction between the polymer channel layer and electrolyte ions. The EGODTT-based devices maintained 24.36% current retention for ∼10² seconds and realized long-term memory characteristics by suppressing ion back diffusion. Through UV-visible spectroscopy and GIWAXS analysis, we found that the copolymerization of glycol groups induces robust interactions with ions and the higher crystallinity of the polymer, thereby improving ion doping efficiency. The fabricated device successfully implemented biological synaptic characteristics, including PPF, STDP, and high-pass filtering. Furthermore, it exhibited high linearity in LTP/D and conductivity, revealing enhanced performance. Finally, we simulated an ANN using the MNIST dataset based on these characteristics and achieved a classification accuracy of 94.1%. These results demonstrate that systematic glycol–alkyl copolymer engineering is a strategy to improve the non-volatile properties of OESTs and provide new possibilities for high-performance artificial synapse devices.

Experimental section

Please refer to the SI for details on the experimental procedure, device fabrication and characterization.

Author contributions

Yoohyeon Jang: conceptualization, data curation, investigation, writing – original draft, formal analysis. Junho Sung: data curation, conceptualization, writing – review & editing, investigation, formal analysis. Suhui Sim: investigation, methodology, conceptualization, formal analysis. Sein Chung: investigation, data curation. Young Un Jeon: investigation. Myeongjin An: data curation. Minju Kim: investigation. Sung Yun Son: investigation. Jaewon Lee: supervision, validation. Eunho Lee: supervision, validation, formal analysis, writing – review & editing.

Conflicts of interest

There are no conflicts to declare.

Data availability

The data supporting this article have been included as part of the supplementary information (SI). Supplementary information is available See DOI: https://doi.org/10.1039/d5nh00623f.

Acknowledgements

This work was supported by the Korea Basic Science Institute (National research Facilities and Equipment Center) grant funded by the Korea government (MSIT) (RS-2024-00404963). This work was also supported by the Korea Planning & Evolution Institute of Industrial Technology (RS-2024-00420537) grant funded by the Ministry of Trade, Industry & Energy (MOTIE).

References

1. P. China Venkanna Varma, K. Venkata Kalyan Chakravarthy, V. Valli Kumari and S. Viswanadha Raju, Big Data Res., 2016, 3, 24–28.
2. Y. LeCun, Y. Bengio and G. Hinton, Nature, 2015, 521, 436–444.
3. Y. Luo, M. R. Abidian, J.-H. Ahn, D. Akinwande, A. M. Andrews, M. Antonietti, Z. Bao, M. Berggren, C. A. Berkey, C. J. Bettinger, J. Chen, P. Chen, W. Cheng, X. Cheng, S.-J. Choi, A. Chortos, C. Dagdeviren, R. H. Dauskardt, C. Di, M. D. Dickey, X. Duan, A. Facchetti, Z. Fan, Y. Fang, J. Feng, X. Feng, H. Gao, W. Gao, X. Gong, C. F. Guo, X. Guo, M. C. Hartel, Z. He, J. S. Ho, Y. Hu, Q. Huang, Y. Huang, F. Huo, M. M. Hussain, A. Javey, U. Jeong, C. Jiang, X. Jiang, J. Kang, D. Karnaushenko, A. Khademhosseini, D.-H. Kim, I.-D. Kim, D. Kireev, L. Kong, C. Lee, N.-E. Lee, P. S. Lee, T.-W. Lee, F. Li, J. Li, C. Liang, C. T. Lim, Y. Lin, D. J. Lipomi, J. Liu, K. Liu, N. Liu, R. Liu, Y. Liu, Y. Liu, Z. Liu, Z. Liu, X. J. Loh, N. Lu, Z. Lv, S. Magdassi, G. G. Malliaras, N. Matsuhisa, A. Nathan, S. Niu, J. Pan, C. Pang, Q. Pei, H. Peng, D. Qi, H. Ren, J. A. Rogers, A. Rowe, O. G. Schmidt, T. Sekitani, D.-G. Seo, G. Shen, X. Sheng, Q. Shi, T. Someya, Y. Song, E. Stavrinidou, M. Su, X. Sun, K. Takei, X.-M. Tao, B. C. K. Tee, A. V.-Y. Thean, T. Q. Trung, C. Wan, H. Wang, J. Wang, M. Wang, S. Wang, T. Wang, Z. L. Wang, P. S. Weiss, H. Wen, S. Xu, T. Xu, H. Yan, X. Yan, H. Yang, L. Yang, S. Yang, L. Yin, C. Yu, G. Yu, J. Yu, S.-H. Yu, X. Yu, E. Zamburg, H. Zhang, X. Zhang, X. Zhang, X. Zhang, Y. Zhang, Y. Zhang, S. Zhao, X. Zhao, Y. Zheng, Y.-Q. Zheng, Z. Zheng, T. Zhou, B. Zhu, M. Zhu, R. Zhu, Y. Zhu, Y. Zhu, G. Zou and X. Chen, ACS Nano, 2023, 17, 5211–5295.
4. S.-A. Oh, H. Yun, K. Park and Y. Kim, Korean J. Chem. Eng., 2024, 41, 337–345.
5. D. Lee, J. Lee and D. Shin, Korean J. Chem. Eng., 2024, 41, 3263–3286.
6. D. Zhang, T. Yuan, H. Zhang, S. Shi, X. Wang, R. Ding and Y. He, Carbon Lett., 2024, 34, 2055–2079.
7. M. Panachikkool and T. Pandiyarajan, Carbon Lett., 2024, 34, 445–475.
8. F. Alibart, S. Pleutin, D. Guérin, C. Novembre, S. Lenfant, K. Lmimouni, C. Gamrat and D. Vuillaume, Adv. Funct. Mater., 2010, 20, 330–337.
9. Y.-C. Lin, W.-C. Yang, Y.-C. Chiang and W.-C. Chen, Small Sci., 2022, 2, 2100109.
10. P. Yao, H. Wu, B. Gao, J. Tang, Q. Zhang, W. Zhang, J. J. Yang and H. Qian, Nature, 2020, 577, 641–646.
11. N. K. Upadhyay, H. Jiang, Z. Wang, S. Asapu, Q. Xia and J. Joshua Yang, Adv. Mater. Technol., 2019, 4, 1800589.
12. B. Yan, B. Li, X. Qiao, C.-X. Xue, M.-F. Chang, Y. Chen and H. (Helen) Li, Adv. Intell. Syst., 2019, 1, 1900068.
13. W. Haensch, A. Raghunathan, K. Roy, B. Chakrabarti, C. M. Phatak, C. Wang and S. Guha, Adv. Mater., 2023, 35, 2204944.
14. L. Shao, Y. Zhao and Y. Liu, Adv. Funct. Mater., 2021, 31, 2101951.
15. S. Chen, E. Li, R. Yu, H. Yang, Y. Yan, Y. Hu, H. Chen and T. Guo, J. Mater. Chem. C, 2021, 9, 9972–9981.
16. J. Sun, Y. Fu and Q. Wan, J. Phys. Appl. Phys., 2018, 51, 314004.
17. F. Akopyan, J. Sawada, A. Cassidy, R. Alvarez-Icaza, J. Arthur, P. Merolla, N. Imam, Y. Nakamura, P. Datta, G.-J. Nam, B. Taba, M. Beakes, B. Brezzo, J. B. Kuang, R. Manohar, W. P. Risk, B. Jackson and D. S. Modha, IEEE Trans. Comput. Aided Des. Integr. Circuits Syst., 2015, 34, 1537–1557.
18. J. Woo, K. Moon, J. Song, S. Lee, M. Kwak, J. Park and H. Hwang, IEEE Electron Device Lett., 2016, 37, 994–997.
19. H. Wang, B. Sun, S. S. Ge, J. Su and M. L. Jin, npj Flexible Electron., 2024, 8, 28.
20. D. S. Jeong, I. Kim, M. Ziegler and H. Kohlstedt, RSC Adv., 2013, 3, 3169–3183.
21. H.-S. P. Wong and S. Salahuddin, Nat. Nanotechnol., 2015, 10, 191–194.
22. C. Zhang, Y.-T. Tai, J. Shang, G. Liu, K.-L. Wang, C. Hsu, X. Yi, X. Yang, W. Xue, H. Tan, S. Guo, L. Pan and R.-W. Li, J. Mater. Chem. C, 2016, 4, 3217–3223.
23. W. Lu, K.-H. Kim, T. Chang and S. Gaba, in 16th Asia and South Pacific Design Automation Conference (ASP-DAC 2011), 2011, pp. 217–223.
24. S. Jiang, S. Nie, Y. He, R. Liu, C. Chen and Q. Wan, Mater. Today Nano, 2019, 8, 100059.
25. S. Kim, C. Du, P. Sheridan, W. Ma, S. Choi and W. D. Lu, Nano Lett., 2015, 15, 2203–2211.
26. A. Paul, S. Yadav, K. A. Rokade, U. V. Shembade, L. K. Hindoliya, M. Dubey, T. D. Dongale, Y.-L. Chueh and S. Mukherjee, J. Phys. Appl. Phys., 2025, 58, 235103.
27. S. Yadav, C. Patel, M. K. Rajbhar, M. Dubey, D. D. Kumbhar, T. D. Dongale, V. Khandelwal, S. Yuvaraja, X. Li and S. Mukherjee, ACS Appl. Mater. Interfaces, 2025, 17, 26871–26880.
28. Y.-B. Leng, Y.-Q. Zhang, Z. Lv, J. Wang, T. Xie, S. Zhu, J. Qin, R. Xu, Y. Zhou and S.-T. Han, Adv. Electron. Mater., 2023, 9, 2300108.
29. J. Bae, J. Won and W. Shim, Nano Energy, 2024, 126, 109646.
30. S. Choi, S. H. Tan, Z. Li, Y. Kim, C. Choi, P.-Y. Chen, H. Yeon, S. Yu and J. Kim, Nat. Mater., 2018, 17, 335–340.
31. S. E. Ng, S. K. Vishwanath, J. Yang, S. S. Periyal, A. Nirmal, N. F. Jamaludin, R. A. John and N. Mathews, Adv. Electron. Mater., 2024, 10, 2300540.
32. F. Zhang, C. Li, Z. Li, L. Dong and J. Zhao, Microsyst. Nanoeng., 2023, 9, 16.
33. E. J. Fuller, F. E. Gabaly, F. Léonard, S. Agarwal, S. J. Plimpton, R. B. Jacobs-Gedrim, C. D. James, M. J. Marinella and A. A. Talin, Adv. Mater., 2017, 29, 1604310.
34. H. Cui, Y. Xiao, Y. Yang, M. Pei, S. Ke, X. Fang, L. Qiao, K. Shi, H. Long, W. Xu, P. Cai, P. Lin, Y. Shi, Q. Wan and C. Wan, Nat. Commun., 2025, 16, 2263.
35. Y. Yang, H. Cui, S. Ke, M. Pei, K. Shi, C. Wan and Q. Wan, Appl. Phys. Lett., 2023, 122, 043508.
36. Z. Tian, Z. Zhao and F. Yan, Wearable Electron., 2024, 1, 1–25.
37. G.-T. Go, Y. Lee, D.-G. Seo, M. Pei, W. Lee, H. Yang and T.-W. Lee, Adv. Intell. Syst., 2020, 2, 2000012.
38. S. Wang, X. Chen, C. Zhao, Y. Kong, B. Lin, Y. Wu, Z. Bi, Z. Xuan, T. Li, Y. Li, W. Zhang, E. Ma, Z. Wang and W. Ma, Nat. Electron., 2023, 6, 281–291.
39. P. Gkoupidenis, N. Schaefer, X. Strakosas, J. A. Fairfield and G. G. Malliaras, Appl. Phys. Lett., 2015, 107, 263302.
40. Z. Liu, C. Liu, Y. Wang, R. Mi and X. Gao, Carbon Lett., 2024, 34, 1287–1299.
41. Y. Yang, X. Zhao, C. Zhang, Y. Tong, J. Hu, H. Zhang, M. Yang, X. Ye, S. Wang, Z. Sun, Q. Tang and Y. Liu, Adv. Funct. Mater., 2020, 30, 2006271.
42. B. Zhou, W. Liu, Y. Xu, C. Jin, J. Yang and J. Sun, J. Phys. Appl. Phys., 2022, 55, 304006.
43. P. Gkoupidenis, N. Schaefer, X. Strakosas, J. A. Fairfield and G. G. Malliaras, Appl. Phys. Lett., 2015, 107, 263302.
44. C. Qian, J. Sun, L. Kong, G. Gou, J. Yang, J. He, Y. Gao and Q. Wan, ACS Appl. Mater. Interfaces, 2016, 8, 26169–26175.
45. Y. Zhao, S. Haseena, M. K. Ravva, S. Zhang, X. Li, J. Jiang, Y. Fu, S. Inal, Q. Wang, Y. Wang, W. Yue, I. McCullocn and D. He, Nano Energy, 2022, 104, 107985.
46. P. C. Harikesh, C.-Y. Yang, D. Tu, J. Y. Gerasimov, A. M. Dar, A. Armada-Moreira, M. Massetti, R. Kroon, D. Bliman, R. Olsson, E. Stavrinidou, M. Berggren and S. Fabiano, Nat. Commun., 2022, 13, 901.
47. A. Giovannitti, D.-T. Sbircea, S. Inal, C. B. Nielsen, E. Bandiello, D. A. Hanifi, M. Sessolo, G. G. Malliaras, I. McCulloch and J. Rivnay, Proc. Natl. Acad. Sci. U. S. A., 2016, 113, 12017–12022.
48. O. Kwon, J. Kang, S. Jang, H. Eom, S. Choi, J. Shin, J. Park, H. Seo, J. H. Kim, S. Park and I. Nam, Korean J. Chem. Eng., 2024, 41, 539–544.
49. M. Moser, L. R. Savagian, A. Savva, M. Matta, J. F. Ponder, T. C. Hidalgo, D. Ohayon, R. Hallani, M. Reisjalali, A. Troisi, A. Wadsworth, J. R. Reynolds, S. Inal and I. McCulloch, Chem. Mater., 2020, 32, 6618–6628.
50. C. B. Nielsen, A. Giovannitti, D.-T. Sbircea, E. Bandiello, M. R. Niazi, D. A. Hanifi, M. Sessolo, A. Amassian, G. G. Malliaras, J. Rivnay and I. McCulloch, J. Am. Chem. Soc., 2016, 138, 10252–10259.
51. H. Lee, H. Jeong, W. Jeong, Y. J. Hwang, B. An, Y. Lee, G. Kim and D.-H. Ha, Korean J. Chem. Eng., 2024, 41, 3371–3393.
52. L. Bynens, K. Zhang, P. Cavassin, A. Goossens, J. Vanderspikken, T. C. H. Castillo, D. Tsokkou, A. Marks, A. Magni, K. Weaver, L. Lutsen, S. Inal, K. Vandewal, N. Banerji and W. Maes, Adv. Funct. Mater., 2025, 2423913.
53. H.-S. Tseng, T. Puangniyom, C.-Y. Chang, J. A. Janardhanan, H. Yu, W.-C. Chen, C.-C. Chueh and Y.-S. Hsiao, Chem. Eng. J., 2024, 486, 150371.
54. J. Sung, S. Chung, Y. Jang, H. Jang, J. Kim, C. Lee, D. Lee, D. Jeong, K. Cho, Y. S. Kim, J. Kang, W. Lee and E. Lee, Adv. Sci., 2024, 11, 2400304.
55. J. Sung, M. Kim, S. Chung, Y. Jang, S. Kim, M.-S. Kang, H.-Y. Lee, J. Kang, D. Lee, W. Lee and E. Lee, Small Struct., 2025, 6, 2400319.
56. I. Zozoulenko, A. Singh, S. K. Singh, V. Gueskine, X. Crispin and M. Berggren, ACS Appl. Polym. Mater., 2019, 1, 83–94.
57. R. J. Sengwa, Polym. Int., 1998, 45, 43–46.
58. M.-J. Sung, D.-G. Seo, J. Kim, H. E. Baek, G.-T. Go, S.-J. Woo, K.-N. Kim, H. Yang, Y.-H. Kim and T.-W. Lee, Adv. Funct. Mater., 2024, 34, 2312546.
59. E. Tan, J. Kim, K. Stewart, C. Pitsalidis, S. Kwon, N. Siemons, J. Kim, Y. Jiang, J. M. Frost, D. Pearce, J. E. Tyrrell, J. Nelson, R. M. Owens, Y.-H. Kim and J.-S. Kim, Adv. Mater., 2022, 34, 2202574.
60. Y. Zhang, G. Ye, T. P. A. van der Pol, J. Dong, E. R. W. van Doremaele, I. Krauhausen, Y. Liu, P. Gkoupidenis, G. Portale, J. Song, R. C. Chiechi and Y. van de Burgt, Adv. Funct. Mater., 2022, 32, 2201593.
61. L. Yang, H. Zhou and W. You, J. Phys. Chem. C, 2010, 114, 16793–16800.
62. R. Halaksa, J. H. Kim, K. J. Thorley, P. A. Gilhooly-Finn, H. Ahn, A. Savva, M.-H. Yoon and C. B. Nielsen, Angew. Chem. Int. Ed., 2023, 62, e202304390.
63. X. Ji, B. D. Paulsen, G. K. K. Chik, R. Wu, Y. Yin, P. K. L. Chan and J. Rivnay, Nat. Commun., 2021, 12, 2480.
64. J. Sun, S. Oh, Y. Choi, S. Seo, M. J. Oh, M. Lee, W. B. Lee, P. J. Yoo, J. H. Cho and J.-H. Park, Adv. Funct. Mater., 2018, 28, 1804397.
65. J. Sun, Y. Fu and Q. Wan, J. Phys. Appl. Phys., 2018, 51, 314004.
66. H. Jang, G. Y. Bae, S. H. Kim, J. Sung and E. Lee, Mater. Horiz., 2024, 11, 4638–4650.
67. H. Lee, J. Cho, M. Jin, J. H. Lee, C. Lee, J. Kim, J. Lee, J. C. Shin, J. Yoo, E. Lee and Y. S. Kim, ACS Nano, 2024, 18, 5383–5395.
68. P. Y. Chen, X. Peng and S. Yu, 2017 IEEE Int. Electron Devices Meet. IEDM 2017, 2018, 6.1.1–6.1.4.
69. N. O. Ogunsola, S. S. Oh, P. R. Jeon, J. L. J. Ling, H. J. Park, H. S. Park, H. E. Lee, J. M. Sohn and S. H. Lee, Korean J. Chem. Eng., 2024, 41, 1923–1953.
70. J. Sung, H. J. Cheon, D. Lee, S. Chung, L. Ayuningtias, H. Yang, B. Jeon, B. Seo, Y.-H. Kim and E. Lee, Mater. Horiz., 2025, 12, 5225–5235.

---

Footnote

† These authors contributed equally to this work.

---