Flexible and Energy-ecient Synaptic Transistor with Quasi-linear Weight Update Protocol by Inkjet Printing of Orientated Polar-electret/High-k Oxide Hybrid Dielectric

Articial synapse by inkjet printing is promising in cost-effective and exible applications, but remains challenging in emulating synaptic dynamics with a sucient number of stable and effective conductance states under ultra-low voltage spiking operation. Hence, for the rst time, a synaptic transistor gated by inkjet-printed hybrid dielectric of electret polyvinyl pyrrolidone (PVP) and high-k Zirconia oxide (ZrO x ) is proposed and thus synthesized to solve this issue. Quasi-linear potentiation/depression characteristics with large variation margin of conductance states are obtained through the coupling of these two dielectric components and the facilitating of dipole orientation, which can be attributed to the orderly arranged molecule chains induced by the carefully designed microuidic ows in droplets. Crucial features of biological synapses including long-term potentiation/depression (LTP/D), spike-timing-dependence-plasticity (STDP) learning rule, “Learning-Experience” behavior, and ultralow energy consumption (< 10 fJ/pulse) are successfully implemented on the device. Simulation results exhibit an excellent image recognition accuracy (97.1 %) after 15 training epochs, which is the highest for printed synaptic transistors. Moreover, the device sustained excellent endurance against bending tests with radius down to 8 mm. This work presents a very viable solution for constructing the futuristic exible and low-cost neural systems.


Introduction
Arti cial intelligence (AI) has attracted increasing attention with the coming big data era and intelligent age in recent years, but the conventional von Neumann architectures have limited data transfer rate for neuromorphic computing due to the physical separation of storage and processor units. The adopted silicon complementary metal-oxide-semiconductor (CMOS) chips are not ideal for emulating the intelligent behaviors in brain-like ways and limited to small systems in regard to energy consumption and design complexity, where at least seven silicon transistors are needed to build an electronic synapse 1 . Our brain has a huge advantage of handling complicated and unstructured issues such as comprehension, determination, recognition, and learning concurrently in a robust and fault-tolerant way with extreme energy-e ciency of only ∼20 W. Synapses are widely recognized as the essential nodes for realizing brain functions by the modulation of their connection strength, which is referred to as the synaptic weight. Although extensive researches on solid-state arti cial synapse in single device level have been carried out to simulate brain functionalities, the realizing of low-cost and exible synaptic devices with improved performances and limited energy consumption remains challenging, thus has received increased attention in the scienti c community 2 .
Recently, the idea of building a cost-effective and exible synaptic device using the drop-on-demand, noncontact and atmospheric processing inkjet printing, instead of sophisticated operation process and vacuum-based physical deposition techniques, has been conceptualized, especially in the fabricating of three/multi-terminal synaptic transistors, which have advantages over two terminal memristors in that they have independent terminals for signal transmission (via drain biased channel ow) and learning operation (via gate biased weight update), thus can eliminate the complex synchronizing algorithms and simplify the learning scheme 3 . The attempts of inkjet printing mainly focus on the electrodes (PEO:P3HT 4 , etc.) and the channel layers (sc-SWCNTs 5,6 , P(VP-EDMAEMAES) 7 , In 2 O 3 8 , ITO 9 etc.) in synaptic transistors, while to fabricate the dielectrics is blocked by the basic need of analog multi-state weight update in biological like protocol. Nonvolatile manipulation of the dielectrics and a further step of achieving su cient states are critical on the e ciency of synaptic functions, for example, the recognition speed and accuracy in image recognition. Bao et al. with Stanford University rst utilized organic electrolyte dielectric (PVDF-HFP) by inkjet printing to prepare a stretchable synaptic transistor, but signi cant impediments still maintain in achieving controllable weight update with nonvolatile and su cient states 10 . Signi cantly, the achieving of su cient dielectric states is still a burden even for noninkjet-printed synaptic transistors, although tremendous contributions have already been made to achieve the nonvolatile behavior based on diverse kinds of dielectric materials (ferroelectric 11,12 , electret 13 Polymer electrets offer a great potential as an exciting candidate for inkjet printing and exible fabrication 35 . The electrostatic interaction between the captured charges and dipoles is bene cial to the acquisition of the semi-stable nonvolatile states 36 . Among all electret-based synaptic transistors, nonpolar electrets are generally employed as e cient chargeable dielectrics, but a high stimulation voltage over 10 V is always required based on the charge trapping/de-trapping mechanism 37 . The use of polarelectrets is limited by the rapid dissipation of the transferred charges and the disorderly arrangement of the dipoles, which leads to an ignorable polarization intensity. High stimulating voltages up to tens of volts for polar-electrets with strong polar side groups on the polymer chains are reported to induce considerable polarization [38][39][40] . However, the orientation of dipoles is still lacking of regulation methods compatible with the inkjet-printing or the non-inkjet-printing devices. Strong dipoles on intrinsic ordered two-dimensional uorographene (FGR) has been proposed in a synaptic transistor to have a large margin weight update under a stimulating voltage as low as 3 V, which demonstrates the signi cance of molecule alignment for energy-e cient synaptic devices based on the polarization mechanism 41 . Although the traits of micro uidic regulation have attracted intense interests in the eld of inkjet printing and are desirable for the alignment of electret molecules, which may be conducive to the orientation of dipoles along the chains, they have not been applied yet in a synaptic transistor. The lack of overlap between these two elds thus far is due to the challenges associated with the developing of new printable dielectric inks that simultaneously meet the needs of micro uidic control and high-e cient synaptic functions.
In this work, a exible synaptic transistor gated by inkjet-printed hybrid dielectric of polar-electret polyvinyl pyrrolidone (PVP) and high-k Zirconia oxide (ZrO x ), which are both low in processing temperature (≤ 200°C), solution processable, and bio-compatible, is rst suggested to emulate synaptic dynamics with a su cient number of stable and effective conductance states under ultra-low voltage spiking operation.
The composites are very promising due to the combination of their predominant advantages, such as the exibility of PVP and the prospective of ZrO x toward low voltage transistors 42 . A co-solvent PVP/ZrO x hybrid ink is developed with carefully designed micro uidic ows, in which the induced compositional Marangoni ows and capillary compensation ows are all outward directions, enhancing the shearing strength on PVP molecules and leading to an orderly arranged morphology of the lm. The variation margin and linearity, which are the key to achieving su cient effective states, are improved in the developed synaptic transistor due to the nonvolatile polarization of a satisfactory number of dipoles divided into discrete regions by the coupling of PVP and ZrO x components. The demonstrated nonvolatile behavior may be attributed to the electrostatic interaction between the dipoles and trapped electrons in oxygen-de cient ZrO x . Hundreds of strong dipoles (polar butyralactam side groups) contained in each of the PVP chains and the alignment of the molecules provide a great opportunity for the regulation of polarization under low voltages. As a result, crucial features of biological synapses with a minimum energy consumption lower than 10 fJ/pulse and an excellent image recognition accuracy up to 97.1 % after 15 training epochs, which is the highest for printed synaptic transistors, are successfully implemented on the device, and sustained excellent endurance against bending tests with radius down to 8 mm. This work has signi cant reference for the inkjet printing of high-performance and low-energyconsumption synaptic transistors applied in the futuristic exible and low-cost neural systems.

Inkjet-printed PVP/ZrO x hybrid synaptic transistor
The learning of neuromorphic system is governed by the synaptic plasticity, i.e. the ability to modulate the strength of connection between neurons (synaptic weight, W) to store and process information. External stimulus or incoming action potentials from the pre-synapse induce the releasing and diffusion of neurotransmitter molecules and depolarize the membrane of the post-synapse, engendering a tunable postsynaptic current (PSC). In a synaptic transistor, channel conductance (G) is analogous to W. The current difference before and after the stimulus (ΔG) represents the change of synaptic weight (ΔW).
When ΔW is increased or decreased by applying positive or negative pulses, synaptic device exhibits excitability and depression, triggering excitatory postsynaptic conductance (EPSC)/inhibitory postsynaptic conductance (IPSC), respectively (Fig. 1a). The proposed device is fabricated on exible PI substrate with PVP/ZrO x hybrid dielectric deposited by inkjet printing (Fig. 1b), which has a regular surface morphology with orderly arranged molecules (Fig. 1c, Supplementary Figure S1). The deposited lms are constructed with multi-droplets released by printer nozzles, where micro-uid ows are unignorable on account of the low viscosity (4 ~ 20 mPa.s) nature for both the developed PVP and PVP/ZrO x inks. When a droplet reaches its equilibrium states after impacting on the substrate, the triple line (air-solid-liquid) will be pinned during the solidi cation process. The uneven evaporation rate along the surface will induce outward compensation ows from the center to the periphery area inside of the droplet, thus making the long chains of PVP molecules arranged in the same direction. The alignment of PVP chains is further enhanced by using co-solvents of 2MOE and EG in the developed inks, which leads to the outward compositional Marangoni ows along the surface of a droplet. The directions of the surface and the interval ows are the same as a result of both the higher boiling point (T b ) and the higher surface tension (σ) of EG (T b = 197.3°C, σ = 46.5 dyne/cm) compared with 2MOE (T b = 124.5°C, σ = 27.6 dyne/cm). The faster evaporation rate will increase the component proportion of the higher boiling point solvent (EG), which make the surface tension gradually increased from the center to the edge of the droplet. The same ow direction ensures the orientation consistency of PVP molecules in the solidi ed lms (Fig. 1d).
The smoothness and uniformity of the PVP/ZrO x hybrid layer provide a good interface basis for the growth of a-IGZO, which is demonstrated by the SEM images shown in Fig. 1e. The gate electrode (Al) is regarded as a presynaptic neuron, and biological synaptic functions can be emulated by applying appropriate gate stimulus to modify ΔG dynamically, which is highly related to the hysteresis effect induced mainly by the dielectric. As indicated in Fig Under positive gate bias, the trapped electrons in ZrO x will reduce the channel current, but the polarization of PVP molecules provides internal dynamics that drive the increased analogue-channel conductanceswitching behavior. Stronger polarization of the dielectric can be induced from its initial state to the saturated state under continuous gate spikes due to the facilitated orientation of dipoles in the aligned PVP chains, which can be inversed simply by the rotation of the polar butyralactam side groups at low voltages (Fig. 1g). As a consequence, the hysteresis window is extremely large for the inkjet-printed PVP/ZrO x synaptic transistor, which presents great potential of reproducing synaptic weight by an intrinsic analogue state of channel conductance with V GS controlled dependence. ms to 500 ms. The PSCs show a sudden saltation during spikes, but decay back slowly due to the nonvolatile switching of the states. The strong EPSC and IPSC responses are consistent with the large hysteresis shown in Fig. 1f. As shown in Supplementary Figure S2, the polarization hysteresis window of the PVP/ZrO x device is much larger than the PVP device, which may stem from the nonvolatile characteristic enhanced by the electrostatic interaction between the dipoles and electrons trapped by the oxygen vacancies in ZrO x . Nonvolatile characteristic makes the modulated weight states more stable, thus multi-states can be easily achieved by gate stimuli with different pulse width.

Basic biological behaviors and non-volatile features
Paired pulse facilitation/depression (PPF/D) depicts a physiologic behavior, in which the subsequent response of synapses is temporally enhanced/depressed by a prior impulse 15 . Such a property can be observed in Fig. 2d  The controllability and reversibility of the weight update under alternating positive (1 V) and negative (-1 V) voltages are explored in Fig. 2f. The enhanced synaptic connectivity under positive pulses can be easily restored when negative pulses continue to be applied, which is similar to the highly repetitive stimulation of the action potential found in the biological nervous system. Figure 2g veri es the shortterm and long-term potentiation (LTP/STP) simulated under consecutive pulses with different amplitudes. The conductance is regulated to a suppressed state before the measurement by applying a -1 V voltage of 50 s. A near three magnitudes potentiation is induced after 10 consecutive pulses (1 V/100 ms), and then slowly decays to a stable equilibrium state, which is still about two magnitudes higher than the initial conductance, indicating a LTP behavior. When the pulse amplitude decreases to 0.5 V, the potentiated conductance decays back to its initial value in only 10 s due to the spontaneous relaxation of the unstable regulated states, indicating a STP behavior. Multi-state controllability of the conductance ascribed to the non-volatile characteristic is further evidenced by symmetric long-term potentiation/depression (LTP/D) behaviors (Fig. 2h). When the repetitive number (N) of the stimulation pulses (2 V/-2 V, 100 ms) increases (N1 to N30), persistent transition to the high/low conductance state of the device can be clearly observed, successfully simulating the LTP/D mechanism of biological synapses, which are presumed to play a signi cant role in learning processes in the brain. The randomly orientated polar side groups on the aligned PVP chains of the initial dielectric state might be necessary to allow dipoles to have a nearly symmetrical response to contrary pulses, as a result that a symmetrical weight update of the potentiation and depression can be achieved. To make the nonvolatile capacity of the synaptic device clear, the retention characteristics of the transistor are measured after 10 gate input spikes with amplitudes from − 1 V to 3 V (V read = 0.05 V). As shown in Fig. 2i, the conductance state keeps longer than 10 4 s after stimuli as low as 2 V and − 1 V for the potentiation and depression, respectively. Although the current after 1 V stimuli decays slowly in 1000 s, the stored data retention is stable enough and suggests a promising feature for neural computing.

Synaptic plasticity for learning functionalities
Physiological learning, forgetting, and relearning processes are emulated by applying low voltage pulses (1 V, 100 ms) to the synaptic transistor (Fig. 3a). Although the PSC decays slowly (forgetting) after 20 pulse trains (learning), it can be enhanced back to the same level in 8 pulse trains (relearning), and afterwards only 2 pulse trains are demanded for the re-learning operation. Interestingly, the persistent time (forgetting) increased from 503 s to 631 s after the re-learning process, which is extremely similar to the re-learning process in our human brain, that is, the old knowledge becomes more and more di cult to forget after repetitive learning. We also con rm the spike timing dependent plasticity (STDP) learning protocol of the synaptic devices, as shown in Fig. 3b. STDP refers to the weight update with relative timing between the presynaptic and postsynaptic spikes (Δt pre−post ), thus has the potential to transform the time information in neural networks into memory storage, and is widely considered as a key feature of a biological neural system. As a biological process in neuromorphic learning, STDP represents the famous Hebbian learning rule and can be divided into asymmetric and symmetric types. It is noteworthy that the LTP/D coexistence in our device enables various types of STDP, while the learning function is determined by the shape of the successive action potentials. Asymmetric learning, where the synapse weight is simultaneously determined by the time difference (|Δt pre−post |) and timing order (Δt pre−post > 0 or Δt pre−post < 0), is mimicked here by using a pair of stimuli that consisted of a positive triangular spike (0 V to 2 V, 300 ms) following a negative triangular spike (-2 V to 0 V, 300 ms) on both of the pre-and postsynaptic parts of the synaptic transistor (Supplementary Figure S4). It is indicated that the potentiation and depression take place when Δt pre−post > 0 and Δt pre−post < 0, respectively. Moreover, the greatest change in synaptic plasticity happens in the ± 10 ms scale, whereas the smallest weight modi cation happens when the time interval between the pre-and postsynaptic spikes is relatively long. According to the empirical relationship of spike time-related plasticity, the dependence of synaptic weight changes on the Δt pre−post can be described as follows 33 :  (Fig. 3c, Supplementary Figure S6). It is also indicated that the variation linearity is improved with a considerable variation margin when the voltages are decreased to 1V. The low voltages are suitable to avoid abrupt switching of the conductance at the premise of excellent state persistence in such a small interval time of 100 ms. Moreover, we assume that the coupling of the PVP and ZrO x components induces discrete regions of the polarization, which allows the dynamic switching of polarization states more linear (Fig. 3d). for most cases) and shows potential of dealing with complex and e cient computing tasks using the developed arti cial synapse.

Flexibility and neuromorphic simulation for image recognition
The exibility of the inkjet-printed PVP/ZrO x synaptic transistor is measured with a bending radius of 8 mm (Fig. 4a, Supplementary Figure S7). The polymer chains of PVP act as connections and buffer between high-k ZrO x molecules without deterioration, and simultaneously improve the leakage and mechanical exibility of the dielectric layer. Basic synaptic functionalities, including PPF/D index as a function of pulse interval (Supplementary Figure S8), LTP/D response to different pulse amplitudes and stimulation numbers (Supplementary Figure S9), and physiological learning experiences (Supplementary Figure S10), are successfully mimicked for the exible arti cial synapse, although higher stimulation voltages are demanded for achieving the similar synaptic behavior before bending. We suppose that the orientation of the polar side groups on PVP molecules (dipoles) is affected by the deformation of the polymer chains and the re-distribution of surrounded electron-trap sites, leading to the slightly decreased retention time of the potentiated state under a positive stimulation voltage of 1 V. The overlapping diagrams of adjacent LTP (1 V, 100 ms) and LTD (-0.8 V, 100 ms) processes (20 weight states) for the devices measured under bend (R = 8 mm) and at conditions are presented in Fig. 4b and Fig. 4c, respectively. The reproducible and consistent LTP/D performance of both the bend and at devices veri es their reliability of the synaptic plasticity, and indicates excellent exibility of the developed arti cial synapse, although a superior linearity is observed for the at device due to its better non-volatile behavior. It can be evidenced by the more linear conductance modulation achieved in Supplementary Figure S11, where an increased stimulation voltage (2 V) is applied to improve the non-volatile characteristic. It is notable that further increasing the voltages will make abrupt conductance switching during the rst few pulses (Fig. 3c), thus non-volatile behavior under low voltages is of great concern for the achieving of linear weight update.
The arti cial neural network (ANN) based on the developed exible and inkjet-printed PVP/ZrO x arti cial synapse before and after bending is constructed to investigate the applicability of arti cial intelligence in pattern recognition (Fig. 4d). The weight update data for the simulation comes from Fig. 4b and Fig. 4c for the bend and at devices, respectively. The images (handwritten digits from "0" to "9" randomly taken from the MNIST database) used to verify the recognition effect are processed and rescaled to "8×8" (small) and "28×28" (large) pixels. The altering values of synaptic weights are calculated by comparing the real output and target output in each training epoch, and using the input signal from the training images. Figure 4e shows the recognition evolution of the bend and at devices for both small and large images. An extremely high pattern recognition e cacy is validated in aspect of both the recognition accuracy (up to 97.1 % after 15 training epochs for the at device) and speed (over 90 % recognition accuracy in only 10 training epochs). Although the non-linearity for both the enhancement and suppression behaviors of the bend device (NL = 0.11/-0.54) is higher than that of the at device (NL = 0.014/-0.045), a minimum conductance difference at nS level ensures all the 20 weight states valid for the simulation, which enhances the fault tolerance of the recognition system and improved the recognition accuracy, as a result that the reliability of the developed devices for exible applications is demonstrated. Importantly, we further investigate the minimum energy consumption demanded for reliable synaptic plasticity, and nd that obvious EPSC behaviors could be mimicked successfully after 1000 gate input pulses with different widths/interval time of 100 ns, 1 µs, 10 µs and 20 µs (Fig. 4f). A near 20 % potentiation of the conductance is observed even for the case when the pulse width and interval time are 100 ns and 100 ns, respectively. An ultra-low energy consumption of about 1.43 fJ per event is estimated by I peak ×t width ×V DS , which exhibits great potential to mimic the real energy consumption of human brain like arti cial synapse (about 10 fJ per event). The ultra-short pulse stimulus is attributed to the fast response of dipole inversion on the aligned PVP chains. As indicated in Fig. 4g, the recognition accuracy and the energy consumption are comparable or superior to most of the reported synaptic transistors 15,24,26,28,29,45−59 . Moreover, we are rst to demonstrate the image recognition capability for exible and inkjet-printed synaptic transistors. This superior performance means that the developed device has the potential to be a promising candidate for exible or even wearable arti cial neural network, in which pattern recognition is regarded as a necessary task.

Conclusion
We rst demonstrate a exible synaptic transistor with PVP/ZrO x hybrid dielectric deposited by inkjet printing, and achieve orderly arranged polar-electret molecules by micro uidic control, which ensures the capability of mimicking biological behaviors based on polarization mechanism. The coupling of PVP and ZrO x dielectric components contributes to the exibility, energy-e ciency and nonvolatile controllability of the synaptic device. Basic biological behaviors and learning functionalities like STP to LTP transition, LTP/D, "Learning-Experience" behavior, STDP learning rule are all successfully mimicked. Reproducible synaptic weight update with cycles of quasi-linear and large margin LTP/D is demonstrated to have a large number of effective states over 200. Nevertheless, the near ideal weight update protocol makes a small number of distinct states (N20) adequate for the simulation of image recognition, which exhibits an excellent recognition accuracy (97.1 %) after 15 training epochs, and is the highest for printed synaptic transistors. Although the conductance variation linearity is slightly deteriorated for the devices after bending (R = 8 mm), a recognition accuracy over 90 % is also resulted. Considering the ultra-low energy consumption of about 1.43 fJ per event for inducing synaptic plasticity, the developed synaptic transistor by cost-e cient inkjet printing technique may be promising to construct futuristic arti cial synapse extremely analogous to that in our brain.

Characterization
The surface morphologies of the lms are characterized by Atomic Force Microscopy (AFM, Asylum Research). The cross-sectional morphologies of the device are investigated by eld emission scanning electron microscope (SEM, NP-O10, ZEISS Gemini 500). The electrical properties of the synaptic devices are measured using a high-precision semiconductor device analyzer (Agilent B1500A) equipped with high-voltage semiconductor pulse generator unit (HV-SPGU, B1525A), which can generate pulses of magnitude up to ± 40 V and width from 10 ns to 10 s. The V DS applied for all electrical measurements is 0.1 V. The exible transistors are peeled off from the glass substrate and then pasted on the stainlesssteel half cylinder. All electrical measurements are performed in the dark and vacuum atmosphere (10 − 3 Pa), which can remove the moisture in the dielectrics and simplify the discussion of the ionization effect of the nonionic PVP-K30 molecules.

Declarations
Data availability The data that support the ndings of this study are available from the corresponding author upon reasonable request.