Introduction

Brain-inspired neuromorphic devices provide an ideal buildingblock for next-generation artificial intelligence with promising applications in bionic perception1,2, neuromorphic computing3,4, brain-machine interfaces5,6, and biological prostheses7,8. Bionic sensory devices can convert external stimuli (such as light, pressure, sound, gas, or taste) into electrical signals which, when integrated into the brain, can enable corresponding visual, tactile, auditory, olfactory, and gustatory perceptions9,10,11,12,13. In the human sensory system, approximately 80% of information is acquired through the visual system. There is hence significant interest in mimicking visual neural functions, such as information perception, storage, and preprocessing, using photoelectric neuromorphic devices. Much work in recent years has led to significant advancements in areas such as emotional recognition, image recognition, and associative learning14,15,16. Additionally, neuromorphic computing, when implemented through hardware systems, can emulate the computation similar to human brain, which is hopeful to overcome the limitations of the traditional von Neumann computing architecture17,18,19. Among the various types of artificial neuromorphic devices, transistor-based neuromorphic components stand out for their multi-terminal modulation, rich connectivity between neighboring stimuli, and extremely nature scalability developed by the chip industry since 1950s20,21. The most common neuromorphic transistors incorporate ionic electrolytes to effectively enhance the gate modulation and thereby reduce the operating voltage and power consumption, making them highly promising for future energy-efficient intelligent systems22,23.

Printing is as an ancient manufacturing process and its potential application in fabrication of electronic devices has attracted significant interest in recent years24,25. Its high-throughput, particularly by large-area roll-to-roll production, may enable extremely low cost as compared to traditional microelectronics manufacturing26,27. Furthermore, there are many well-developed printing techniques, including screen printing, inkjet printing, gravure printing, electrohydrodynamic printing, and 3D printing, which offer broad prospects in flexible/stretchable sensors, thin-film transistors (TFTs), solar cells, and other wearable products28,29,30,31,32. Among these, screen printing stands out as one of the most mature and low-cost techniques in industry. However, screen printing typically requires high-viscosity inks (500–10000 cP), which necessitates substantial solid additives such as binders, emulsifiers, plasticizers, and defoamers33,34. This poses a significant challenge in achieving high-performance semiconducting channels with screen-printing inks, as these additives can disrupt the alignment and continuity of the semiconductor components, severely reducing their conductivity35,36.

Oxide semiconductors have seen explosive development since the breakthrough of invention of flexible, transparent oxide transistors in 200437. Oxide semiconductors offer a potential advantage solution-based manufacturing because their conductivity is primarily governed by isotropic metal s-orbital electron orbitals. As a result, even in the presence of solid additives and within a disordered amorphous state, they may maintain high electron mobility and conductivity, as long as the oxide semiconductor particles are not entirely isolated by the additives (e.g., when they form a continuous network-like structure). Moreover, Oxide semiconductors have gained increasing attention as channel materials for synaptic transistors due to their high stability, high carrier mobility, cost-effective solution processability, and inherent persistent photoconductivity (PPC)16,38. Zinc oxide (ZnO) possesses a wide direct bandgap of 3.4 eV and a relatively high exciton binding energy of 60 meV, which enable low leakage current, low static power consumption, highly efficient ultraviolet (UV) photoresponse, and excellent thermal stability39,40. Nevertheless, most solution-processed ZnO inks reported to date are generally precursors that require high temperatures (200~600 °C) for conversion41,42. To date, the very limited number of screen-printed ZnO films demand annealing at 500~600 °C, significantly limiting their suitability for flexible electronics applications42,43. Therefore, low-temperature fabrication of ZnO semiconductors is crucial for flexible electronics, particularly on paper substrates which offer a biodegradable, disposable and eco-friendly alternative44,45.

In this study, we successfully fabricated large-area photoelectronic synaptic TFT arrays on paper substrate using a fully screen-printing technique. Ethylcellulose-dispersed ZnO nanoparticles were used as the active layer and polyethylene oxide/lithium perchlorate (PEO/LiClO4) as the gate insulator. The fully screen-printed, paper-based ZnO TFTs exhibited excellent electrical performance, large-area uniformity, air stability, biodegradability at a maximum processing temperature of 90 °C. Furthermore, the fabricated ZnO synaptic devices demonstrated outstanding mechanical flexibility and typical electrical and optical synaptic behaviors, such as paired-pulse facilitation and depression, high-pass and low-pass filtering, learning, forgetting, relearning, Morse code transmission and image recognition, all achieved at a low energy consumption (3.7 pJ per synaptic event). Moreover, the printed ZnO synapse arrays, consisting of 105 TFTs, showed their potential for applications in visual memory systems and neuromorphic computing, achieving an accuracy of 91.4% based on their promotion and inhibition behavior.

Results

Fabrication and electrical performance of fully screen-printed ZnO TFTs

Figure 1a outlines the fabrication steps for the fully screen-printed ZnO TFTs, featuring coplanar source, drain, and gate electrodes, along with PEO/LiClO4 electrolyte. Further details are provided in the experimental section. The schematic of the fully screen-printed ZnO TFTs with a side gate structure, as shown in Fig. 1b, demonstrates its capability to simulate synaptic neuron functions. Commercial carbon-based conductive ink was used as the electrode material, providing strong electrical and chemical stability and high conductivity, with a sheet resistance of 76.8 ± 3.35 Ω/□. This significantly reduces manufacturing costs compared to gold or silver inks46,47. ZnO nanoparticles, approximately 10 nm in diameter, combined with ethyl cellulose, were selected for the inorganic semiconductor layer to ensure proper dispersion and optimal ink viscosity for screen printing. The viscosities of ZnO inks with varying ethyl cellulose content (0, 5, 9.5, and 13.6%) are detailed in Supplementary Tab. 1, corresponding to 530, 1440, 12,900, and 37,250 cP, respectively. The ink with 5% ethyl cellulose content has been proved to be the most suitable for screen printing. The screen-printed carbon electrodes and ZnO layer were clearly identified and verified via energy dispersive spectroscopy (EDS) elemental analysis, showing thicknesses of approximately 5.0 ± 0.5 µm and 4.0 ± 0.5 µm, respectively (Supplementary Fig. 1). The polymer PEO/LiClO4 was suitable for screen printing, with a viscosity of about 1800 cP. Capacitance measurements were conducted using a typical capacitor structure (Supplementary Fig. 2a). The unit area capacitance of the electrolyte decreases gradually with increasing frequency due to the low mobility of ions, and a large unit area capacitance of 5.7 μF/cm2 at 20 Hz was obtained, demonstrating typical electric double layer (EDL) capacitance characteristics. Supplementary Fig. 2b, c display the phase angle and impedance as functions of frequency. In the low-frequency region (<10 Hz), capacitive behavior (θ < −45°) dominates because the ions have sufficient time to move to the interface and form the EDL with high capacitance. As the frequency increases (>10 Hz), it gradually transitions to resistive behavior (θ > −45°) because insufficient ions reach the interface within each cycle to form the EDL due to low ion mobility. The Nyquist diagram exhibits a semicircular shrinkage effect at low impedance levels, with the ionic conductivity of the PEO/LiClO4 electrolyte calculated to be 10−5 S/cm, using the formula (σ = l/RA (1)). Electrochemical tests were performed using a button cell, where l is the thickness of the electrolyte membrane, R is the bulk resistance, and A is the surface area.

Fig. 1: Fabrication and characteristics of fully screen-printed ZnO TFTs.
Fig. 1: Fabrication and characteristics of fully screen-printed ZnO TFTs.
Full size image

a The fabrication process of fully screen-printed ZnO TFTs. b Schematic diagram of synaptic neurons and the printed side gate ZnO synaptic TFT. c SEM morphology of ZnO film under different processing conditions. d Transfer characteristics of printed ZnO TFTs under different processing conditions. e Zn 2p and f O 1s XPS spectra of printed ZnO films as a function of annealing temperatures.

Fully screen-printed ZnO TFTs with varying ZnO sintering temperatures were prepared on silicon wafers, and the corresponding surface morphologies of ZnO films are shown in Fig. 1c. Agglomerated ZnO particles are densely packed to form channels conducive to carrier transport, benefiting from the strong hydrogen bonding between hydroxyl groups in cellulose and ZnO nanoparticles (Supplementary Fig. 3). Sintering at 400 °C leads to the decomposition of the additives, resulting in relatively large ZnO agglomeration blocks, thereby reducing grain boundaries and the surface roughness compared to those sintered at the lower temperatures (Supplementary Fig. 4). Figure 1d illustrates the transfer characteristics of silicon-based fully screen-printed ZnO TFTs under various processing conditions, all exhibiting large on-state currents (>10−4 A), high switching ratios (~105), and pronounced counterclockwise hysteresis as information storage windows. Detailed performance analysis is presented in Supplementary Fig. 5. X-ray photoelectron spectroscopy (XPS) analyses were performed based on different annealing conditions for the ZnO films (Fig. 1e, f). In the Zn 2p spectrum, two peaks with binding energies of 1021.3 eV and 1044.4 eV correspond to Zn 2p1/2 and Zn 2p3/2, respectively. The Zn-O and Zn-OH peaks are fitted separately, yielding consistent results across different analyses. The O 1s are distinguished by lattice oxygen (~530.0 eV), oxygen vacancies (~531.2 eV), and hydroxyl groups (~532.4 eV), among which the hydroxyl peak corresponding to 400 °C is very weak due to the volatilization of terpineol at high temperature. Notably, ZnO processed at 90 °C for 10 min exhibits a similar chemical composition and device performance to those obtained under high-temperature conditions, making it compatible with paper substrates for environmentally friendly electronics. One reason is that strong and dense hydrogen bond network can form between ZnO and the hydroxyl-rich ethyl cellulose, which facilitates the formation of continuous ZnO channel along the cellulose framework (Supplementary Fig. 3). Another reason is that the additive content is relatively low, only 5% in the solid-state films, and therefore it does not significantly affect the continuity of the ZnO particles. Here, we use a high-viscosity solvent, terpineol, along with an efficient thickening agent, ethyl cellulose, ensuring that even with a low 5% additive content, the ink viscosity (1440 cP) still meets the requirements for screen printing.

Large-area, fully screen-printed ZnO TFT arrays were successfully fabricated on a paper substrate using low temperature and efficient ZnO processing conditions (90 °C for 10 min) (Fig. 2a). An optical image of a single fully printed paper-based ZnO TFT with a channel length of 0.3 mm and a width of 4 mm is presented in Fig. 2b. The typical transfer characteristics of 35 printed ZnO TFTs in the array exhibit uniform performance, with the switching ratio close to 105, large on state current (>10−4 A), low operating voltage (−1–5 V), and a counterclockwise hysteresis window of approximately 1V (Fig. 2c). The histograms in Fig. 2d–j provide a statistical analysis of the key parameters of the 35 devices. The average ION/IOFF ratio, on-state current, threshold voltage (VTH), mobility (μ), subthreshold swing (SS) and hysteresis of these devices are (2.78 ± 2.04) ×104, (1.2 ± 0.29) ×10−4 A, 3.3 ± 0.09 V, 2.94 ± 0.68 cm2/Vs, 0.43 ± 0.06 V/dec, and 1 ± 0.1 V, respectively. The large porosity and roughness of paper substrate do not hinder the electrical properties of the printed TFTs, because of that the thickness of the screen-printed ZnO films is relatively large (~4 μm), making the substrate defects less sensitive. In addition, these fully printed paper-based ZnO devices show high air stability, with their transfer curves remaining consistent even after 60 days of air exposure (Supplementary Fig. 6). The switching ratio and mobility show a slight decrease at 60 days, due to the influences of moisture and oxygen in the air since the absence of surface passivation.

Fig. 2: Electrical characteristics of fully screen-printed paper-based ZnO TFTs.
Fig. 2: Electrical characteristics of fully screen-printed paper-based ZnO TFTs.
Full size image

a Optical photograph of the fully printed paper-based ZnO TFT arrays. b Optical image of a single ZnO TFT. c Transfer curves for 35 ZnO TFTs. Statistical histograms based on the parameters of the 35 TFTs, including d ION/IOFF ratio, e on current, f threshold voltage, g mobility, h subthreshold swing (SS) and i hysteresis.

Optoelectronic synaptic performance of fully screen-printed paper-based ZnO TFTs

Synaptic neurons serve as the fundamental units for information processing in the brain, exhibiting synaptic plasticity, which significantly influences human learning, memory, and the processing of complex tasks48,49. Remarkably, our screen-printed ZnO TFTs can replicate neuronal synaptic plasticity through ion relaxation in the electrolytes during the formation of the EDL. Paired-pulse facilitation/depression (PPF/D) is a common type of short-term plasticity (STP) characterized by the consecutive potentiation or inhibition of postsynaptic currents following closely spaced pulsed stimuli, playing a crucial role in information processing. A pair of positive voltage pulses applied to the gate of the printed ZnO synaptic TFT generates an excitatory postsynaptic current (EPSC) in the ZnO channel, with the amplitude of the second EPSC greater than that of the first. Figure 3a illustrates the PPF response, characterized by a spike width of 40 ms and an interval of 40 ms, where the PPF index is defined as the ratio of the second EPSC amplitude to that of the first (PPF index = A2/A1 × 100% (2)), resulting in a PPF index of 177%. Conversely, when a pair of negative voltage pulses is applied to the gate, an inhibitory postsynaptic current (IPSC) is generated, with the second current peak being more suppressed than the first. The PPD index (PPD index = A2/A1 × 100% (3)) is calculated as 164% at the pulse interval of 40 ms (Fig. 3c). When a positive gate voltage pulse is applied, cations in the electrolyte aggregate towards the semiconductor-electrolyte channel interface, increasing the EDL capacitance and inducing the accumulation of more electrons in the channel and thus EPSC. In contrast, when a negative gate voltage pulse is applied, cations withdraw from the channel interface, reducing the EDL capacitance and leading to fewer electrons in the channel and thus IPSC. Due to the low ion mobility and short pulse interval, the cations accumulated or withdrawn at the channel interface after the first gate pulse cannot fully relax back to their initial positions before the second pulse. As a result, the second EPSC or IPSC peak becomes stronger than the first. The PPF and PPD characteristics at various pulse intervals were measured (Supplementary Figs. 7 and 8), and the corresponding dependencies of the PPF and PPD indexes on pulse intervals are analyzed in Fig. 3b, d. The fittings were performed using the following function: PPF/PPD = C0 + C1 exp(−Δt∕τ1) + C2 exp(−Δt∕τ2) (4), where C0, C1, and C2 represent constants, τ1 and τ2 refer to the relaxation time of the fast and slow phases, respectively. The extracted τ1 and τ2 values are 0.1/0.4 s and 0.1/0.2 s for PPF and PPD, respectively. The results reveal a more pronounced promotion/inhibition effect in artificial synapses at shorter pulse intervals, attributable to reduced ion relaxation in the electrolyte at shorter pulse intervals. As the interval increased, the facilitation or depression effect declines, ultimately diminishing to 100%, which aligns with the STP features of artificial synapses50.

Fig. 3: Synaptic facilitation/depression plasticity induced by electrical stimulation.
Fig. 3: Synaptic facilitation/depression plasticity induced by electrical stimulation.
Full size image

a EPSC triggered by paired spikes with a time interval of 40 ms. b Variation of the PPF index with pulse interval. c IPSC triggered by paired spikes with a time interval of 40 ms. d Variation of the PPD index with pulse interval. e Schematic diagram of high-pass filtering. f EPSC triggered by 20 consecutive positive pulses at a frequency of 12.5 Hz. g Variation of EPSC peak value with pulse number at various frequencies. h High-pass filter gain at various frequencies. i Schematic diagram of low-pass filtering. j IPSC triggered by 20 consecutive negative pulses at a frequency of 12.5 Hz. k Variation of IPSC peak value with pulse number at various frequencies. l Low-pass filter gain at various frequencies. Distribution of electrolyte charges and band diagrams in ZnO TFTs when applying m positive pulses and n negative pulses to the gate.

Due to these promoting and inhibiting behaviors, these artificial synaptic TFTs can function as high-pass and low-pass filters. Figure 3e, i demonstrate high-pass and low-pass filtering schematics, respectively. For the former, high-frequency signals are amplified as they pass through the device, while low-frequency signals are attenuated or eliminated. Whereas for the latter, low-frequency signals are amplified. The EPSC and IPSC responses to 20 consecutive positive and negative gate pulses at various frequencies were investigated using 1–5 V and 3–1 V pulses as the presynaptic spikes (Fig. 3f, j, Supplementary Figs. 9 and 10). The EPSC peak increases with each stimulus and further rises with an increase in pulse frequency (Fig. 3g). Either with increasing positive pulse numbers or frequencies, more cations and electrons accumulate at the channel interface (Fig. 3m), leading to the progressively increased EPSC peak values. The EPSC peak value gain is calculated as: Gain = (A20 − A1)/A1 (5), where A20 and A1 represent the final and initial EPSC values, respectively. It is observed that this gain increases with the increasing frequency (Fig. 3h), because of that with the increased frequency, the pulse interval decreases remarkably, reducing the degree of cation relaxation significantly. In case of negative gate pulses, in principle, fewer cations and electrons accumulate at the channel interface, leading to decreased IPSC peak values, as shown in Fig. 3n. However, only at high frequencies (10–12.5 Hz) or only for the first several pulses at low frequencies (5–8.3 Hz), the IPSC peak value conform to this rule and decrease monotonically with increased pulse numbers. At low frequencies (5–8.3 Hz), after the first several pulses, the IPSC peak values either increase monotonically or remain constant with increased pulse numbers (Fig. 3k and Supplementary Fig. 10). This is suggested because, as the frequency decreases, the time interval of the baseline voltage (3 V) increases, causing the IPSC being unable to counteract the increased current. The increase in the IPSC peak value gain (formula (5)) with frequency can be attributed to a significant reduction in anion relaxation as the pulse interval shortens (Fig. 3l).

Figure 4a presents the optical modulation characteristics of the fully screen-printed ZnO artificial synaptic TFT, with UV pulses serving as the presynaptic stimuli and photogenerated carriers in the channel acting as the postsynaptic response. Due to the wide optical bandgap of the printed ZnO film (3.25 eV, Fig. 4b), the device exhibits a strong response at a wavelength of 365 nm and weak responses at wavelengths of 254 and 405 nm when three light pulses with fixed increasing intensity are applied (Fig. 4c). As the intensity and duration of 365 nm illumination increase, the EPSC peak rises from 0.42 to 1.42 μA and from 0.49 to 1.17 μA, respectively, as shown in Fig. 4d–f. This is mainly attributed to the increased generation of photogenerated charges under higher intensity and prolonged illumination. The relationship between EPSC responses and the PPF index for two optical pulses with varying time intervals is shown in Fig. 4g and Supplementary Fig. 11. Notably, the PPF index reaches 174% at a pulse time interval (Δt) of 0.2 s and remains high at 154% even at Δt = 30 s, the extracted τ1 and τ2 values are 0.43 s and 33.14 s, respectively. Figure 4h illustrates the learning, forgetting, and relearning processes of optical synapses. Initially, 20 light pulses result in a current of 5.7 μA. After a natural forgetting period of 38 s, only 6 identical light stimuli are required to restore the current to its initial level, highlighting the exceptional learning and memory capabilities of this optical synapse. Based on the photosynaptic plasticity of the printed ZnO TFTs, Morse code recognition can be realized through specific presynaptic pulses. As shown in Supplementary Fig. 12, a short pulse (0.3 mW/cm2, 0.5 s) triggered a small EPSC, corresponding to a “dot”, while a long pulse (0.3 mW/cm2, 1.5 s) triggered a large EPSC, corresponding to a “dash”. According to the law of Morse code, the letters H, E, L, P were encoded as H (dot, dot, dot, dot), E (dot), L (dot, dash, dot, dot), and P (dot, dash, dash, dot). A 0.5 s interval separated each adjacent character in the sequence of dots and dashes. Figure 4i displays the EPSC triggered by the letters H, E, L, and P, with corresponding peak values of 8.79 μA (H), 1.53 μA (E), 11.07 μA (L), and 15.01 μA (P), enabling successful decoding and indicating the potential applications of this system in optical information processing and encryption.

Fig. 4: Synaptic plasticity induced by optical stimulation.
Fig. 4: Synaptic plasticity induced by optical stimulation.
Full size image

a Schematic diagram of an optically modulated ZnO synaptic TFT. b Optical bandgap of printed ZnO film. c EPSC responds to 254, 365, 405 nm light at fixed intensities of 0.15, 0.2, and 0.3 mW/cm2, respectively. d EPSC triggered by 365 nm light with intensity increasing from 0.12 to 0.57 mW/cm2 over a duration of 0.2 s. e EPSC triggered by 365 nm light with durations ranging from 0.2 to 1.0 s at a constant intensity of 0.18 mW/cm2. f Relationship between EPSC, 365 nm light intensity, and duration. g The PPF index as a function of optical pulse interval (Δt) with the 365 nm light intensity of 0.23 mW/cm2 and the pulse width of 0.5 s (Inset is the EPSC response with an interval of 0.2 s between two optical pulses). h The learning, forgetting, and relearning process with 365 nm light intensity of 0.23 mW/cm2 and pulse width and interval of 0.2 s, respectively. i Optical responses to the Morse codes of the letter H, E, L, P.

The human brain’s memory process can be divided into three main stages: sensory memory (SM), short-term memory (STM), and long-term memory (LTM) (Fig. 5a). The conversion from STM to LTM can be achieved in our optical synaptic TFTs by adjusting input signals such as gate voltage, frequency, and the number of optical pulses (Supplementary Fig. 13a–c). By increasing the gate voltage from −0.5 V to 1.5 V, or the light frequency from 0.2 Hz to 2.5 Hz, or the number of optical pulses from 6 to 50, a higher EPSC level can be realized. The transition from STM to LTM and the attenuation of normalized conductance changes in the corresponding cases are illustrated in Supplementary Fig. 13d–f, which is mainly attributed to the PPC effect of the printed ZnO films. When the light is removed after illumination, the high density traps (oxygen vacancies and zinc interstitials) in the ZnO film can trap the photogenerated holes and thus hinders or delays the recombination of photogenerated electrons and holes51,52, resulting in sustained photocurrents, which reflects LTM characteristics. Additionally, the peak gain of EPSC ((A10 − A1)/A1) increases with rising UV light pulse frequencies, demonstrating distinct high-pass filtering characteristics (Supplementary Fig. 14).

Fig. 5: Memory characteristics of the ZnO synaptic TFTs.
Fig. 5: Memory characteristics of the ZnO synaptic TFTs.
Full size image

a Human memory characteristics include SM, STM, and LTM. Color mapping of the rose image based on b 30 UV pulse stimulations with forgetting of 0, 5, 20 and 50 s, c 40 UV pulse stimulations with forgetting of 0, 5, 20 and 50 s, and d 50 UV pulse stimulations with forgetting of 0, 5, 20 and 50 s.

To further visualize the memory conversion behavior of the device, a rose image was color-mapped based on the device’s learning and forgetting behavior under UV pulses (365 nm, 0.3 mW/cm2, 2.5 Hz) of 30, 40, and 50 repetitions. The EPSC peak values, corresponding to attenuation at 0, 5, 20, and 50 s are recorded and simulated to the mapped rose color, as shown in Fig. 5b–d and Supplementary Table 2. As the number of pulses increases, the clarity of the rose image improves, and remains visible even after 50 s of forgetting. However, after a forgetting period of 50 s, the rose image corresponding to 30 learning instances becomes relatively faint. These behaviors are consistent with established principles of human learning, memory, and forgetting. Additionally, mechanical flexibility tests were conducted on paper-based printed ZnO synaptic TFTs. Under both flat condition and bending radii of 10, 5, and 2 mm, the transfer characteristic curves of the ZnO TFTs remained stable, with no significant variations in switching ratios, threshold voltages, or mobilities (Supplementary Fig. 15a–c). Notably, even at a bending radius of 2 mm, the device exhibited stable EPSC responses to 365 nm UV light (0.2 mW/cm2, 0.2 s), demonstrating excellent mechanical flexibility (Supplementary Fig. 15d–f). Moreover, the device maintains relatively stable electrical performance after 500 consecutive bending cycles with a bending curvature of 2 mm, demonstrating excellent mechanical flexibility (Supplementary Fig. 15g–i). Energy consumption is another crucial parameter for neuromorphic devices in information processing. The power consumption of the synapse was calculated as Ec = VDS × EPSCpeak × t (6), achieving a remarkably low value of 3.7 × 10−12 J under a single UV light pulse (0.08 mW/cm2, 0.1 s) (Supplementary Fig. 16a). Similarly, a minimum energy consumption of 1.8 ×10−12 J was recorded under a single electrical pulse (0.5 V, 0.1 s), demonstrating its highly competitive performance (Supplementary Fig. 16b).

Artificial visual perception and neuromorphic computing

The human visual system exhibits outstanding learning and memory capabilities. To simulate artificial vision functions, we fabricated three 5×7 screen-printed paper-based ZnO synaptic TFT arrays with excellent uniformity and potential for large-scale application. Figure 6a illustrates the optical imaging process in dark, where an “SDU” mask is positioned above the ZnO TFT arrays, with each letter corresponding to a specific region. UV light at 365 nm and an optical density of 3.91 mW cm2 irradiated through the mask onto the exposed TFTs, enabling image recognition through the variations in the TFT channel currents. The EPSCs of 105 synaptic TFTs were measured at 1, 2, and 5 s during UV stimulation, and 5, 10, and 50 s after light exposure. As expected, the channel currents of TFTs exposed to UV light increased rapidly, while the EPSCs of the TFTs shielded by the mask remained stable, due to the photogenerated carriers in ZnO layers under illumination. As depicted in Fig. 6b–d, with the prolonged illumination, these synaptic TFTs exhibited higher EPSC values, and the contrast of the mapped letters intensified, resulting in a clearer “SDU” image. A higher EPSC value indicates enhanced memory and learning abilities, demonstrating the remarkable learning capacity of the printed ZnO synapses in response to repetitive light stimulation. The conductance of each synaptic unit was further measured after 5, 10, and 50 s of light stimulation (Fig. 6e–g). The contrast of the “SDU” mapping slightly decreased after 5 s, but remained clearly discernible. Over time, the EPSC values gradually declined, leading to a fading of the “SDU” image. Notably the “SDU” image remained recognizable even after 50 s. These results demonstrate that optical information can be effectively perceived and retained in the printed large-scale ZnO synaptic TFT arrays, highlighting its promising potential for emulating human visual memory functions.

Fig. 6: Image learning and memorization applications of the ZnO TFT arrays.
Fig. 6: Image learning and memorization applications of the ZnO TFT arrays.
Full size image

a Schematic diagram of image perception and processing mapped on three 5 × 7 synaptic arrays using a “SDU” shaped mask. “SDU” conductance mapped patterns of optical radiation at b 1 s, c 2 s and d 5 s and after radiation of e 5 s, f 10 s and g 50 s.

Human beings acquire external image information through the visual system, which comprises optical systems, retinas, and visual pathways (Fig. 7a). When light enters the eye and stimulates the retina, the retina captures image information and converts light signals into electrical signals. These signals are then transmitted via the visual nerves to the brain for visual processing53. Within the visual imaging system, thousands of synaptic neurons play a vital role in information transmission. Visual recognition, rooted in the human visual system, enables computers to understand and identify data through comprehensive analysis and processing, which involves crucial steps such as image acquisition, preprocessing, feature extraction, and classification recognition54,55. To evaluate the potential of this hardware device in visual recognition and neuromorphic computing, we constructed an artificial vision system based on printed ZnO synaptic TFT arrays for recognizing handwritten digits obtained from the modified National Institute of Standards and Technology (MNIST) database. This artificial neural network (ANN) consists of 784 input neurons (corresponding to a 28 × 28 pixel input image), 50 hidden neurons, and 10 output neurons (representing digits 0–9) (Fig. 7b). The parameters of the ANN are updated using synaptic enhancement and inhibitory plasticity, where the synaptic weight is defined as the electrical conductance difference between two equivalent synapses. The long-term potentiation (LTP) and long-term depression (LTD) characteristics based on electrical writing electrical erasing, and light writing electrical erasing under 100 pulses are demonstrate in Fig. 7c, d, and Supplementary Fig. 17a, b with detailed training and recognition processes outlined in the experimental section. Following 800 epochs, the recognition accuracy of the handwritten digit 6 reached 90% and 91.4%, respectively (Fig. 7e), showcasing the excellent image recognition potential of the ZnO artificial visual system. In addition, our fully printed devices exhibit good continuous enhancing and inhibiting behavior under complete electrical modulation (Supplementary Fig. 17c). We examined the degradability of paper-based synaptic electronics in NaOH solution. The samples undergo corrosion, dissolution, detachment, and eventual degradation. After just 3 min of stirring, the devices began to detach, and was nearly completely degraded after 60 min under a stirring speed of 300 r/s (Supplementary Fig. 18). Supplementary Tab. 3 summarizes the recent research on synaptic TFTs utilizing metal oxides, organic semiconductors, carbon nanotubes, and two-dimensional materials, most of which are prepared on rigid or plastic substrate and involves vacuum or lithography processes with relatively high temperature30,56,57,58,59,60,61,62,63,64. In this case, our paper-based, fully screen-printed synaptic devices with a maximum processing temperature of 90 °C exhibit overall excellent electrical properties, diverse synaptic functions and degradability, indicating the great potential of our devices for low-cost and efficient artificial neural networks.

Fig. 7: Artificial neuromorphic vision system for image recognition using the ZnO synaptic TFT arrays.
Fig. 7: Artificial neuromorphic vision system for image recognition using the ZnO synaptic TFT arrays.
Full size image

a Schematic diagram of the human visual system. b An artificial neuromorphic vision system based on the printed ZnO synaptic arrays for handwritten digits recognition. Potentiation and depression characteristics based on c electrical writing and electrical erasing (Pulse width and interval of 0.04 s), d light writing and electrical erasing (Pulse width and interval of 0.3 s). e Image recognition accuracy of the ZnO artificial vision system for training images via electrical and optical stimulation.

Discussion

This research developed large-area, paper-based, fully screen-printed ZnO photoelectric synaptic TFT arrays using PEO/LiClO4 polymers as the gate capacitor. The ZnO nanoparticles, dispersed with a small amount of hydroxyl-rich ethyl cellulose as channel materials, exhibited exceptional electrical performance, air stability, and mechanical flexibility at a low temperature of 90 °C. Several key artificial synaptic functions, including PPF/PPD, EPSC/IPSC, STP/LTP, dynamic learning and forgetting, Morse code, and image recognition, were simulated under both electrical and optical modulation with a low power consumption of 3.7 pJ per pulse event. The fabricated ZnO synaptic arrays demonstrated artificial visual learning, forgetting, and information storage capabilities. Furthermore, neuromorphic computation was simulated through the photoelectric co-modulation characteristics of the device, achieving an impressive 91.4% accuracy in handwritten numeral recognition. We have shown that the paper-based transistor arrays are also biodegradable, making them particularly promising for use as disposable devices. These fully screen-printed, paper-based synaptic devices may provide a cost-effective, eco-friendly, and scalable solution for more complex and applicable neuromorphic networks.

Methods

Preparation of ZnO semiconductor inks

The ZnO semiconductor ink comprises ZnO nanoparticles (Jiangsu Xianfeng Nano Material Technology Co., Ltd, 10 nm particle size), ethyl cellulose dispersant (Sigma-Aldrich), and terpineol solvent (Sigma-Aldrich). Initially, 31.6 wt% of ZnO nanopowder and 1.7 wt% of ethyl cellulose powder were thoroughly mixed and then dispersed in terpineol. The resulting mixture was stirred for 12 hours to achieve a uniformly dispersed and homogeneous white viscous solution, which can then be used directly for printing without further processing.

Preparation of PEO/LiClO4 dielectric inks

Screen-printable dielectric inks consist of polyethylene oxide (PEO, Sigma-Aldrich, Mn ~300,000), lithium perchlorate (LiClO4, 99.99%, anhydrous), and methanol solvent. To prepare the solution, 0.1 g of LiClO4 and 0.8 g of PEO are dissolved in 10 ml of methanol, then the mixture is heated to 45 °C and stirred for 2 h until a homogeneous, transparent, and viscous solution with a viscosity of about 1800 cP is obtained. The electrolyte’s viscosity can be adjusted based on the actual demand by varying the amount of methanol and the ink can be used without any further purification.

Fabrication of fully-screen printed ZnO TFTs

The fabrication process of fully screen-printed ZnO TFT array is illustrated in Fig. 1a. The paper substrate (Yalan Yunhe Paper Co., Ltd., 210 × 297 × 0.5 mm, with a 2-um-thick silica coating in the surface) was placed in an oven and prethermal treated at 100 °C for 1 h to relieve stress. Then it is cleaned by dry nitrogen gas. The source/drain and gate planar electrodes were screen-printed through a polyester screen mold (300 mesh, aperture 48 mm) onto the substrate using carbon ink (Jujo Chemical) with the printing speed of 150 mm/s and an average thickness of 5.0 ± 0.5 µm, then cured in an oven at 90 °C for 10 min to ensure sufficient conductivity. For the channel layer, ZnO ink was screen-printed through a mesh model with the printing speed of 150 mm/s and an average thickness of 4.0 ± 0.5 µm, and then heated to evaporate the solvents and achieve better electrical properties. The last step involved screen-printed PEO/LiClO4 electrolyte with the printing speed of 100 mm/s, and repeated 8 times for better insulation, with average thickness of 8.0 ± 0.5 µm followed each printing was cured in the oven at 90 °C for 3 min. Finally, the fully screen-printed ZnO devices can be used directly for testing or neuromorphic simulation without other processing. For the construction of the device array (Supplementary Fig. 19), all the TFT’s sources in the array were connected and grounded using silver paste (JC-2100, Shenzhen Jieyongcheng Technology Co., Ltd.). The gate electrodes were connected as scan row lines and the drain electrodes were connected as the data column lines, using silver paste. The insulation between these grounding, scan, and date lines are using screen-printed insulating layers with the T12 ink (Chongqing Zhiwei Technology Co.).

Characterization and measurements

All functional inks are deposited onto the substrate using a screen printing machine (HY-PV10Y, Huanyu Automation Shenzhen Co. Ltd.). The surface morphology of screen-printed ZnO semiconductors annealed at different temperatures was measured by scanning electron microscope (SEM, Nova NanoSEM 450) and atomic force microscopy (AFM, CSPM5500). The capacitance and impedance of the PEO/LiClO4 electrolyte were measured using an impedance analyzer (Agilent E4980A precision LCR meter). X-ray photoelectron spectroscopy (XPS, Thermo Scientific K-Alpha) using Al Ka radiation was employed to determine the elemental composition and chemical state of ZnO films. The film thickness was measured using stylus profile (DektakXT) and the square resistance was measured by a 4-probe square resistance tester (HPS2662). The ink viscosity was determined by a rotary viscometer (NDJ-4). The electrical characteristics were conducted using a Keithley B2902A parameter analyzer. The fiber coupled laser module (GHUV) was used to provide light pulse signals. The mobility of the device was calculated with the standard linear regime:

$${\rm{\mu }}=\left(\frac{{\rm{d}}{I}_{D}}{{\rm{d}}{V}_{{GS}}}\right)\times \frac{L}{W}\times \frac{1}{{V}_{{DS}}{C}_{i}}$$
(1)

Here, the channel length L is 0.3 mm, and the width W is 4 mm. The value of VDS is 0.5 V, Ci is the capacitance per unit area of PEO/LiClO4 dielectric, approximately 5.7 μF/cm2 at 20 Hz. All the device tests were carried out under an ambient condition at room temperature. The artificial neural network (ANN) structure, comprising an input layer, a hidden layer, and an output layer has the number of neurons of 784, 400, and 10, respectively. The back-propagation (BP) algorithm was utilized for training and recognition processes and a total of 50,000 images were utilized for training, with 10,000 images designated for testing the accuracy. Here, the nonlinearities of the potentiation or inhibition are calculated from the following equations65. Gn and Gn+1 represent the synaptic conductance in the present and updated states, while α and β denotes the step size of the conductance change and the nonlinearity, respectively. Gmax and Gmin are the measured maximum and minimum values of G, respectively.

$${G}_{n+1}-{G}_{n}={\alpha }_{{\rm{P}}}{{\rm{e}}}^{-{\beta }_{{\rm{P}}}\frac{{G}_{{\rm{n}}}-{G}_{\min }}{{G}_{\max }-{G}_{\min }}}({\rm{for\; LTP}})$$
(2)
$${G}_{n+1}-{G}_{n}=-{\alpha }_{{\rm{P}}}{{\rm{e}}}^{-{\beta }_{{\rm{P}}}\frac{{G}_{\max }-{G}_{{\rm{n}}}}{{G}_{\max }-{G}_{\min }}}({\rm{for\; LTD}})$$
(3)