Inspired by human brain, neuromorphic electronics has attracted increasing attention for the features of low-energy consumption and parallel information processing^[1−3]. Synapse is the connecting site between neurons, responsible for information storage and transmission, which is an important part of biological neural network^[4−8]. Therefore, artificial synaptic devices capable of emulating the synaptic plasticity and other neural functions of biological synapses are important for the successful constructing of neuromorphic systems^[9−12].

Currently, most reported synaptic devices are limited to regulating a single type of charge carrier, thus constraining the process of synaptic weight reconstruction^[13−16]. Inspired by the release of multiple neurotransmitters in response to diverse stimuli observed in regions such as the ventral tegmental area (VTA) in midbrain^{[17, 18]}, a few synaptic devices with p/n heterojunction dual-channels have been designed to emulate the release of two neurotransmitters by controlling electron and hole transport, further expanding multiple synaptic weights^[19−22]. However, the inevitable disparity in trap concentrations between two channels still leads to significant differences in the synaptic plasticity dominated by the two charge carriers, hindering efficient processing of complex information^[23−25]. On the other hand, synaptic devices constructed using intrinsic-ambipolar semiconductors can also simulate the release of two neurotransmitters^[26−28], but often suffer from mismatched n/p polarity strengths or very weak charge carrier transmission capabilities, similarly impeding synaptic weight reconstruction. This underscores the urgent need to develop a synaptic transistor that can regulate two matched charge carriers with relatively small differences in plasticity to meet the varied demands of synaptic weight reconstruction.

Here, we employed F₁₆CuPc small molecules to modify the graphene channel and developed a graphene/F₁₆CuPc synaptic transistor (GFST) for the first time. F₁₆CuPc is an n-type organic semiconductor material with excellent electron transport properties and good air stability^[29]. The modified graphene/F₁₆CuPc channel showcased efficient transmission of balanced p-type and n-type charge carriers to reproduce versatile synaptic plasticity. Leveraging the bipolar property of the graphene/F₁₆CuPc channel, we achieved the emulation of multiplexed neurotransmission by adjusting the polarity of the gate voltage (V_GS) to control the transmission of carriers (electrons or holes) within the channel. Furthermore, we emulated a diverse range of plasticity selections mediated by different neurotransmitters, achieving results across four cases, and established an image encryption-decryption process triggered by two types of spike stimuli. Due to the capability to match the relative amplitudes and plasticity properties of the postsynaptic currents dominated by either holes or electrons, we developed a complementary neural network model triggered by both positive and negative spikes, enabling enhanced weight regulation for improved accuracy in pattern recognition tasks. This strategy of constructing p-type and n-type performance-matched bipolar synaptic devices enhances their applications in neuromorphic computing.

Device preparation: The CVD-grown single-layer graphene (Hangzhou Hangdan Optoelectronics Technology Co., Ltd.) was transferred from Cu foil to SiO₂ substrate by PMMA-assisted wet-transfer method. During the wet-transfer process, Poly(methyl methacrylate) (PMMA, Sigma Aldrich, average Mw ~996 000) was spin-coated onto the surface of single-layer graphene attached on copper foil as a support layer. Then the copper foil was etched with ammonium persulfate solution ((NH₄)₂S₂O₈, 0.2 mol/L in deionized water). The graphene with PMMA was transferred to SiO₂ substrate and dried to allow adhesion on the substrate. Finally, the PMMA was removed by acetone to obtain the single-layer graphene adhered on SiO₂.

The F₁₆CuPc was deposited onto the surface of graphene by vacuum deposition under a pressure of 5 × 10⁻⁴ Pa. Subsequently, gold source/drain electrodes (width 1000 µm; length 200 µm; thickness 60 nm) were thermally deposited onto graphene/F₁₆CuPc channel through the rectangular shadow mask. Finally, the ion gel (mass ratio between Poly(vinylidene fluoride-co-hexafluoropropylene) [PVDF-HFP] and (1-ethyl-3-methylimidazolium bis(trifluoromethylsulfonyl)imide [EMIM-TFSI] was 1 : 3) dielectric layer was transferred onto the channel area. The synaptic transistor with graphene/F₁₆CuPc channel was fabricated.

Characterizations: Raman spectrum of graphene and graphene/F₁₆CuPc films were characterized by TEO SR-500I-A system with 532 nm excitation. XPS spectra of graphene and graphene/F₁₆CuPc films was characterized by ESCALAB 250Xi (Thermo Scientific). AFM images of the films were obtained by Bruker dimension icon microscope. SEM images of the morphology were characterized by Apreo S (Thermo Scientific). All electrical measurements were performed by Keithley 4200A semiconductor parameter analyzer at room temperature and N₂ atmosphere.

The GFST with a p-type and n-type performance-matched bipolar graphene/F₁₆CuPc channel was fabricated on the Si/SiO₂ substrate (Fig. 1(a)). Firstly, we transferred CVD-grown single-layer graphene onto Si/SiO₂ substrate by PMMA-assisted method^[30] (Fig. S1), and vaporized F₁₆CuPc on the graphene surface by vacuum evaporation. Finally, graphene/F₁₆CuPc bipolar synaptic transistor was prepared by transferring ion gel onto the channel (Fig. S2). The ion gel consists of 1-ethyl-3-methylimidazolium (EMIM) and trifluoromethylsulfonyl (TFSI). GFST mimics the structure of a biological synapse, and the graphene/F₁₆CuPc channel enables the emulation of multiplexed neurotransmission.

Raman spectrum of the transferred graphene (Fig. 1(b)) shows G peak at ca. 1580 cm⁻¹ and 2D peak at ca. 2700 cm⁻¹. Meanwhile, the D peak (~1350 cm⁻¹) representing the sp² hybridization defect of graphene was negligible^{[30, 31]}, suggesting high quality of the transferred graphene. The intensity ratio of 2D peak to G peak was used to determine the layers of graphene^{[30, 32]}. Here, the ratio of 2D/G intensity of the graphene film was about 2.75, indicating that the transferred graphene was monolayer. As shown in the scanning electron microscope (SEM) images (Fig. S3), F₁₆CuPc formed a continuous film on the graphene surface, forming a graphene/F₁₆CuPc heterojunction. X-ray photoelectron spectroscopy (XPS) spectra of graphene and graphene/F₁₆CuPc (Fig. S4) showed Cu 2p, F 1s and N 1s peaks on the graphene/F₁₆CuPc film, indicating the successful preparation of F₁₆CuPc. The morphologies and thicknesses of graphene and graphene/F₁₆CuPc films were characterized using atomic force microscopy (AFM). The thicknesses of graphene film and graphene/F₁₆CuPc film are ~2.3 and ~16.7 nm respectively, and the thickness of vapored F₁₆CuPc can be estimated to be ~14.4 nm (Fig. S5). In addition, the R_q (root mean square roughness) values of graphene film and graphene/F₁₆CuPc were 3.32 and 1.11 nm (Fig. 1(c)). The surface roughness of the film was significantly decreased with F₁₆CuPc coverage.

Theoretical studies show that graphene has bipolar property and the Dirac point is located at the position where the gate voltage V_GS is zero. However, the Dirac point of the graphene transistor based on transferred graphene was shifted in the positive direction of V_GS (Fig. S6). This phenomenon could be attributed to the adsorption of impurities such as water, oxygen molecules, and PMMA residues on the graphene in PMMA-assisted transfer process, causing p-type doping^{[31, 33]}. The electronic energy band structure of transferred graphene was demonstrated (Fig. 1(d)). After the evaporation of n-type semiconductor F₁₆CuPc, the transfer curves showed that the n-type semiconductor characteristic of the device was enhanced, resulting in a bipolar synaptic transistor with matched n-type and p-type semiconductor characteristics (Figs. 1(d) and 1(e)). As V_GS increased positively from −5 V, the current decreased as the number of hole carriers decreased, and the value of gate voltage with minimum current was defined as V_Dirac. As V_GS continues to increased, electrons become the dominant carriers and the current gradually increased as the number of electron carriers increased^[34]. Similarly, when the V_GS is scanned from +5 to −5 V, the current underwent a process similar to the forward scanning process (Fig. S7). At the same time, the transfer curves exhibited significant hysteresis behavior, which is the basis for constructing synaptic transistor that can emulate synaptic plasticity. In this device, the hysteresis is mainly attributed to the migration of ions within the ion gel and the capacitive gating effect^{[35, 36]}. The output curves (Figs. 1(f) and S8) of the device were further tested by adjusting the applied V_GS and the source-drain voltage (V_DS). The V_DS was scanned from +0.1 to −0.1 V, while V_GS was varied from +5 to −5 V. Under positive V_GS, the source−drain current (I_DS) in the graphene/F₁₆CuPc channel was mainly formed by the transport of electrons, and when the V_GS was negative, the I_DS in the channel was mainly formed by the transport of holes. The effective operating region of the device can be divided into four cases (Case #1−4). As shown in the output curve of the device (Fig. 1(f)), the operating region was defined as Case #1 when the device was working at positive V_GS and positive V_DS, Case #2 when it was working at negative V_GS and positive V_DS, Case #3 when it was working at negative V_GS and negative V_DS, and Case #4 when it was working at positive V_GS and negative V_DS. This suggested that the GFST has a large working range. The synaptic plasticity of GFST was explored in these four cases.

Neurons in human brain are highly interconnected, and synapses are the connections between neurons that are responsible for information storage and transmission. We constructed the GFST that can mimic biological synaptic plasticity and used V_DS to emulate the cell membrane resting potential (RP) (Figs. 2(a) and 2(b)). Under the stimulation of a presynaptic spike (+4 V), EMIM cations migrated toward the interface between ion gel and graphene/F₁₆CuPc channel, inducing electrons in the channel and forming an excitatory postsynaptic current (EPSC). At −4 V presynaptic spike, TFSI anions accumulated at the interface, inducing holes accumulation in the channel and obtaining EPSC (Fig. 2(c)). At the end of the spike stimulation, the TFSI anions and EMIM cations were returned back to a random distribution and EPSC was decreased slowly. The values of EPSC showed that the electric properties were similar in four cases, which indicated that the n-type and p-type properties of the device were well-matched.

To emulate the selective release of two excitatory neurotransmitters, the spike train with amplitudes varying from +4 to −4 V (0.5 V change in amplitude per spike) was applied to the device. We recorded the currents triggered by each spike at V_DS = ±0.01 V. According to the curves of postsynaptic current (PSC) (Fig. 3(a) and Fig. S9), EPSC and inhibitory postsynaptic current (IPSC) as well as the conversion between EPSC and IPSC could be obtained by varying the spike amplitude. At a positive V_GS, EMIM cations moved toward to the interface of the channel and ion gel and induced electrons in graphene/F₁₆CuPc channel, forming an electron-dominated EPSC. This synaptic potentiation was defined as an e⁻-potentiation. As the V_GS gradually decreased from +4 to 0 V, the amount of EMIM cations accumulated at the interface decreased, and the number of electrons in the channel was also decreased, thus the postsynaptic current decreased as the voltage amplitude decreased. The postsynaptic current less than the baseline current was defined as e⁻-depression. When V_GS became negative, the TFSI anion gradually migrate to the interface to induce holes in the channel. The holes were neutralized with the electrons accumulated previously, leading to a decrease in postsynaptic current and forming IPSC, defined as h⁺-depression. Then, as the negative V_GS gradually increased, more TFSI anions accumulated at the interface, inducing more hole carriers for an EPSC, defined as h⁺-potentiation. In this process, the amount and type of neurotransmitter release were emulated by adjusting the V_GS, enabling the emulation of the selective release of two different excitatory neurotransmitters.

Paired-pulse facilitation (PPF) is an important characteristic of short-term plasticity^[37]. After two consecutive spikes were applied to the gate, the peak value of second EPSC (A₂) was larger than the peak value of first EPSC (A₁). This is due to that in a short time interval (ΔT) between two spikes, the ions in ion gel that had migrated under the stimulation of the first spike had not fully recovered to the equilibrium state of random distribution. At this point, under the second spike stimulation, more ions accumulated at the interface of graphene/F₁₆CuPc channel and ion gel, which in turn induced more carriers in the channel, resulting in a larger EPSC. PPF with different time intervals was recorded (Fig. 3(b)). The PPF index was defined as (A₂/A₁) × 100% and the trend of PPF index with ΔT was explored (Figs. 3(c) and S10). The PPF index exhibited a gradual decline as ΔT increased from 50 to 3200 ms. The above exploration of PPF and the relationship between PPF index with ΔT revealed that continuous spike stimulation as well as the change of time interval between spikes could effectively regulate the synaptic plasticity of the device. Similarly, by altering the number, duration and frequency of spikes, synaptic plasticity could also be effectively modulated.

As the number of spikes gradually increased from 5 to 100, the EPSC peaks also gradually increased, reflecting the spike-number-dependent plasticity (SNDP). An enhancement of long-term plasticity was also successfully achieved in GFST as the number of spikes increased (Fig. 4(a)). We further recorded the trend of ΔEPSC decay within 200 s after successive stimulation, where ΔEPSC is defined as the excitatory postsynaptic current gain (EPSC minus baseline current). As the number of spikes increased, we found that the decay trend of ΔEPSC gradually slowed down (Fig. 4(b)). The ratio of the magnitude of the real-time current to the maximum current generated under the spike stimulation as the retention rate, a value that demonstrates the memory capability of the device. The retention rates of EPSC relative to the peak postsynaptic current were recorded for 200 s after 100 successive spike stimuli in the four cases (Fig. 4(c)). The retention rates of postsynaptic currents were relatively high in all four cases. Among them, the retention rates of postsynaptic currents in case #3 and case #4 were greater than 50%, which were 60.03% and 64.23%, respectively. Moreover, the decreased trends of the curves were slower than case #1 and case #2. This indicated that the device had better long-term plasticity in case #3 and case #4. Additionally, we further explored the spike-duration-dependent plasticity (SDDP) and spike-frequency-dependent plasticity (SFDP) of GFST by adjusting the durations and frequencies of the applied spikes (Figs. S11 and S12). The longer duration and higher frequency facilitated the achievement of higher EPSC and retention rate.

Long-term plasticity (LTP) has an important role in learning memory activities at cellular level and facilitates the formation of biological memories^[38]. At the same time, short-term plasticity (STP) is also the basis for realizing the normal function of the nervous system^[39]. Therefore, the emulation of LTP, STP, and switching between different plasticities in a single synaptic device is crucial. By varying the amplitude of V_GS from +0.5 to +4 V, we could find the transition between STP and LTP in GFST (Fig. S13). Furthermore, we applied a voltage train consisting of 30 spikes with voltage amplitudes periodically switching between −4 and +4 V. The postsynaptic currents were recorded at V_DS = −0.01 V. As shown in the postsynaptic current curves (Fig. S14), the PSC exhibited two exhibited two kinds of long-term plasticity with different memory states (State 1 and State 2). The fast switching of two different synaptic plasticities could be achieved by altering the polarity of the applied spikes. The plasticity switching was found to be very reproducible by repeated application of this spike train. Based on the fast switching of two synaptic plasticities, a 5 × 5 synaptic device array was designed to demonstrate the image encryption−decryption process triggered by two types of spike stimuli. The two different types of synaptic plasticities were used to distinguish the letters 'NKU' in the image. Firstly, 30 consecutive −4 V spikes were set as valid information, while 30 consecutive +4 V spikes were set as noise information. We encrypted the information into the postsynaptic current signal by applying 30 consecutive +4 or −4 V spikes to the synaptic array and recording the current in each device (Fig. S15). The current after 1 s of voltage withdrawal was subsequently extracted and plotted as a heat map (Fig. 4(d)). As shown in the figures, due to two different long-term plasticities induced by two kinds of stimuli, the postsynaptic currents reflect different shades of purple in the heatmap. But we could not read out the encrypted information directly since the interference of the noise information. In order to decrypt the images information, the threshold current value of 90 μA was set as a secret key to decrypt the images information from the postsynaptic current value (Fig. 4(e)). The value of secret key must be located between the postsynaptic currents of the two different plasticity states in Fig. S14. Due to the fast switching of synaptic plasticities, we can quickly encrypt new image information without erasing the previous information.

Good memory property of the device allowed for the emulation of associative learning^{[40, 41]}. We applied two spike trains with different amplitudes to the gate as ringing stimulus (−1 V, 50 ms) and food stimulus (−4 V, 50 ms), respectively, and recorded the currents after stimulation with different spike trains at V_DS = +0.01 V (Fig. S16). A postsynaptic current value of 75 μA was set as the threshold current for salivary secretion. When a spike train as ringing stimulus was applied, the peak value of postsynaptic current was 49.3 μA, suggesting that the conditioned stimulus was unable to induce salivary secretion. Following the action of a food stimulus consisting of a −4 V spike train, the peak value exceeded the threshold current, representing an unconditioned response in which the unconditioned stimulus effectively triggered salivary secretion. During training, the trains of −1 and −4 V were applied alternately at the gate, allowing the ringing stimulus and the food stimulus to correlate with each other. After training, peak value of the postsynaptic current was able to exceed the threshold current under the action of the ringing stimulus alone, eliciting a salivary response, indicating the formation of a conditioned reflex to the ringing stimulus. This GFST successfully simulated the Pavlov’s experiment, reflecting the property of associative learning.

Due to the matched p-type and n-type bipolar properties of GFST, we developed a complementary neural network triggered by both positive and negative spikes, enabling enhanced weight regulation for improved accuracy in pattern recognition (Fig. 5(a)). In this complementary neural network, we performed image recognition simulations on the Fashion-MNIST dataset. The neural network consisted of an input layer, an output layer and a connection weight matrix (Fig. 5(b)). A 28 × 28 pixel Fashion-MNIST image was unpacked into a 1 × 784 row vector as input, where the conductance changes of the synaptic device were used as weights to update the backpropagation algorithm^[42]. In the synaptic potentiation and depression of synapses, the range of variation and linearity of synaptic weights are important factors affecting neuromorphic computation^{[42, 43]}. The synaptic potentiation and depression (Fig. 5(c)) were obtained under Case #1 and Case #2. Synaptic potentiation and depression under Case #1 was triggered by 30 consecutive spike stimuli of +2 V and 30 consecutive spike stimuli of −1 V. While synaptic potentiation and depression under Case #2 was triggered by the spike stimuli of −2.5 and +0.5 V. The electron-dominated and hole-dominated EPSC had different synaptic plasticity, including synaptic weights and linearity. Utilizing this bipolar property, the conductance under Case #1 and Case #2 could be superimposed by the complementary neural network (Fig. 5(d)). We performed image recognition training based on the five obtained conductance changes. After 40 training epochs, the highest image recognition accuracy of 83.23% was achieved for Case 1 + 2 (Fig. 5(e)). This is due to the complementary of p-type and n-type bipolar properties, which enhanced the weight regulation and improved the recognition accuracy. This showed that our device was promising for applications in neuromorphic computing.

In conclusion, a p-type and n-type performance-matched bipolar GFST was proposed in this paper. Based on the bipolar characteristic of the device, the selective release of two different excitatory neurotransmitters as well as the emulation of different neurotransmitter-mediated synaptic plasticity were achieved by adjusting the polarity of gate voltage, which simulated multiplexed neurotransmission. Further by switching the V_GS, fast switching of plasticity was realized. Finally, a complementary neural network for image recognition of Fashion-MNIST dataset was constructed by utilizing the matched relative amplitude and plasticity properties of the GFST dominated by holes or electrons to improve the weight regulation and recognition accuracy, and a recognition rate of 83.23% was achieved after 40 epochs. This work extends the plasticity regulation range of single synaptic device, enriches the functional diversity of device, and provides new perspectives for building more realistic neuromorphic systems.