Neuromorphic electronics have received considerable attention recent years for applications in brain-inspired computation systems and artificial sensorimotor nerves [1-3]. In these systems, artificial synapses are the basic structural and functional units; they emulate basic functions of computation, memory in a brain, and information filtering and delivery functions in the peripheral nervous system (PNS) [4]. Recently, great efforts have been made to design artificial synapses using phase-change materials [5-7], resistive switching materials [8-10], ferroelectric materials [11-13] and ionic materials [14-16]. Synaptic characteristics have also been achieved using floating gate [17-19] and ferroelectric mechanisms [20, 21].

To emulate natural synapses, neuromorphic electronics use synaptic transistors, which have a structure similar to a field-effect transistor, but operate by ion migration and injection in an ion-rich gate-insulator layer. In this structure, ion gels which usually consist of block copolymers and ionic liquids, are an important form that serve as a reservoir of ions [4, 22, 23]. In this structure, block copolymers usually form a physically cross-linked network to serve as a matrix to hold ionic liquids. In this case, the content of ionic liquid can be high, sometimes > 90% by weight. Application of an external electrical field drives migration of ions to form an electrical double layer near the ion gel/semiconductor interface, or injects the ions into the semiconductor, where they can be trapped for a long time to yield long-term plasticity [23-25].

Physical cross-links are usually weak, such as hydrogen bonds, and the formation of these 3D structures usually requires special structures of block copolymers [26]. The synthesis process of block copolymers is usually very strict [27-29], so they are too expensive for industrial production, and the modulation capacity of block copolymers is very limited [28], so they are not easily tuned for different designs and applications. Versatile design of synaptic transistor with easily-tuned properties requires solutions to these problems.

In this paper, we demonstrate a thermally cross-linked matrix for the formation of ion gels instead of traditional physical cross-linked ones. A synaptic transistor was then fabricated using the ion gel to emulate the synaptic cleft, and a thin film of SnO₂ nanoparticles as the conductive channel. The device demonstrated important synaptic functions for neuromorphic computation and neural information transmission, such as paired-pulse facilitation, spike-voltage dependent plasticity, spike-rate dependent plasticity, and spike-number dependent plasticity. The device also achieved logic and high-pass filtering functions. This work provides a versatile strategy for fabrication of robust and versatile ion gels, which are applicable to future design and fabrication of neuromorphic electronic devices.

SnO₂ precursor solution (100 µL) was spin-coated on the SiO₂/Si substrate at 4000 rpm for 30 s, then annealed at 150 ℃ for 30 min. Gold source/drain electrodes were thermally deposited through a multi-branched shadow mask with the channel width and length of 48,400 µm and 100 µm, respectively.

A cross-linkable solution was prepared in a co-solvent (N, N-dimethylformamide: acetonitrile 1:1, v: v) with 5 wt% cyanoethylated pullulan (CEP) (Mw ~489,000, Shin Etsu Chemical Co.) and 1.5 wt% poly(methylated melamine-co-formaldehyde) (PMMF). The solutions were mixed with ionic liquid [EMIM]-[TFSI] (4 times of the weight of CEP), then cast onto the channel region of the synaptic transistor, which was then baked in a vacuum oven at 180 ℃ for 30 min to ensure efficient chemical reactions. The thickness of the ion gel layer is about 100 µm.

FT-IR spectra of solid films were recorded using a VERTEX70. Atomic Force Microscopy (AFM) images were obtained using a Bruker dimension icon microscope in a tapping mode. Electrical characteristics of the electronic devices were measured using a Keithley 4200 semiconductor analyzer in an N₂ atmosphere.

The artificial synaptic device (Fig. 1, right) emulates a biological synapse (Fig. 1, left). A basic biological synaptic structure is a connection between two neurons, including several parts of the presynaptic membrane, the postsynaptic membrane, and the synaptic cleft. To prepare the corresponding artificial synapses, we used metal oxide nanoparticle thin film as a conductive channel on an insulative substrate, then evaporated source-drain electrodes through shadow masks. We used a metal probe as an electrode to supply electrical pulses to the ionic gels (Fig. 1, up).

To evaluate the quality of the SnO₂ thin film, atomic force microscopy (AFM) images were captured using tapping mode (Fig. 2a). The top-view AFM image shows a dense film with no pinholes. The surface of the film was very smooth with a root-mean-squared roughness of 0.782 nm. Dense and smooth thin film is suitable to provide a high-quality interface with stacked functional layers on it. A high-quality interface is good for the elimination of defects and trap sites to improve conductivity of the channel regime.

The AFM image near the edge of the thin film (Fig. 2b), revealed that the height difference between the substrate and the thin film was ~75 nm, which is thus the thickness of the film. Scanning electron microscopy (SEM) image of the cross-sectional view of the thin film confirmed the thickness to be about 75 nm (Fig. 2c).

FTIR spectra were of the ion gel were recorded before and after the thermal crosslinking process (Fig. 2d). Different absorption bands were recorded. The band between 3300 and 3600 cm⁻¹ is assigned to polar functional groups, such OH. After curing, the peaks were significantly reduced and shifted to the higher wavenumbers. The reduction of the peaks implies reaction and reduction of OH groups, and the shift to higher wavenumbers is due to partial break of H-bonds. The absorption band between 2700 and 3000 cm⁻¹ is assigned to C—H bonds. After curing, the peaks were reduced obviously; this change implies removal of side products of CH₃OH during the cross-linking reaction. Thermal crosslinking is a facile process that has a wide selectivity of polymers and crosslinking agents, requiring only very easy thermal baking process. It complements with UV irradiation curing in real applications [30, 31].

The capacitance of the ion gel film as a function of frequency was tested using an Electrochemical Workstation (Fig. S1 in Supporting information). The capacitance of the film reached as high as 138.57 µF/cm² at a low frequency (0.01 Hz). This attributed to sufficient polarization time for mobile ions under external electrical field. However, as the frequency increased, the capacitance of the gel film decreased.

Transmission electron microscope (TEM) images was captured to study the nanostructures of the thin film. Nanocrystallites were observed (Fig. 3a), which showed average diameter of ~5 nm. The lattice structure (Fig. 3b) has a 0.293 nm spacing, which confirms the crystal structure of SnO₂. A fast Fourier transform (FFT) image (Fig. 3c) shows clear (110) lattices. Energy Dispersive Spectrometer (EDS) images of the selected area (Fig. 3d) showed uniform Sn (Fig. 3e) and O (Fig. 3f) element distributions. Transfer curve of the electronic device was recorded (Fig. S2 in Supporting information).

In a biological synapse, action potentials are generated from a preneuron, then propagate from an axon to the presynaptic membrane. The action potentials can cause the release of neurotransmitters into the synaptic cleft. The neurotransmitters diffuse through the cleft, and bind to relevant receptors on the postsynaptic membrane, to open the relevant ionic channel to induce ionic flux that cause postsynaptic current (PSC). Correspondingly, we use a metal probe to deliver voltage spikes (5 V, 50 ms), which are applied to ion gel to cause ion migration similar to the release of neurotransmitters (Fig. 4a and Fig. S3 in Supporting information).

We applied positive spikes to repel cations to the surrounding channel region, and triggered an abrupt change in the current in the channel (Fig. 4b). The current then degraded abruptly, but retained a small current for a short time. This is very similar to that of the postsynaptic current in a biological synapse. This phenomenon underlies basic functions of neural network. After removal of the spike, most of the ions migrate back to the equilium position, whereas the few ions that had adsorbed to the semiconductor/ion gel interface would take longer to spontaneously drift back than the majority of ions due to energy barriers and ion traps at the interface; this delay causes existence of an extended channel current.

Paired-pulse facilitation (PPF) is an important form of short-term plasticity of homosynaptic facilitation in which the second of a pair of action potentials produces a larger EPSC as compared to the first [32]. In response to two successive action potentials (5 V, 50 ms), the second excitatory post-synaptic current (EPSC) peak following is stronger than the first peak. This phenomenon is important for many biological functions. Two successive presynaptic spikes were applied to the ion gel, and triggered two successive EPSC peaks (Fig. 4c); the second one was higher than the first one, as occurs in PPF in a biological synapse. In this process, the first spike induces a certain amount of ions to migrate to the surrounding channel region to cause the first EPSC peak. After the removal of the first spike, most ions diffuse back the equilibrated positions, whereas the few trapped at the interface take longer time than the majority to return. The second spike drives the migration of a similar amount of ions to the channel region, in addition to the remnant ions to induce a stronger effect on the EPSC.

The ratio of the current intensities of the current I₂ generated by the second pulse to the current I₁ generated by the first pulse is called the paired-pulse facilitation index, PPF index. PPF index decreased as the time interval between pulses increased (Fig. 4d). PPF index can be particularly relevant to temporal processing because the amplitudes of excitatory postsynaptic potentials provide temporal information about recent spike occurrence: These differences could be used to encode temporal information. Voltage spikes drive ions to the interface of ion gels and semiconductors, but after removal of external pulse, ions did not immediately drift back to the equilibrium distribution. The application of a second voltage spike could then induce an increased I₂ due to the additional effects of remnant ions.

Using the synaptic device, basic logic functions include "OR" and "AND" were emulated. Herein, EPSC of 1 µA was set as the threshold to define output states. We realized the "OR" logic function under the V of 3 V (Fig. 4e). When the spike signals M1 and N1 were inputted into the postsynaptic membrane respectively, Whether M1 or N1 spike singles were utilized, EPSCs were greater than the threshold value, demonstrating the realization of the "OR" logic function. Different from the "OR" logic function, the "AND" logic function was implemented with the V_DS of 2 V (Fig. 4f). Either M2 or N2 was inputted separately, the EPSCs were less than the threshold; only when M2 and N2 were inputted at the same time, the EPSC was greater than the threshold.

Synaptic strength, as measured by current, and can be sensitive to the number of pulses; this response is spike-number dependent plasticity (SNDP) (5 V, 50 ms, n = 1, 2, 3, …, 10) (Fig. 5a). Synaptic strength is also increased by increased voltage-pulse amplitude; this response is spike-voltage dependent plasticity (SVDP) (V = 0.5 V, 1 V, 1.5 V, …5 V, 50 ms) (Fig. 5b). In the central nervous system and peripheral nervous system, information is usually encoded, transmitted and recognized in the form of trains of spikes at different frequencies.

Spike-rate dependent plasticity (SRDP) is an important form of synaptic plasticity, because frequency information is more reliable and robust than voltage amplitudes. Therefore, emulation of frequency-dependent information processing in neural system is very important. We applied 10 identical presynaptic spikes at different frequencies (f = 3.20 Hz, 3.83 Hz, 4.78 Hz, 6.35 Hz, 9.43 Hz) to the synaptic transistor, which caused distinct output behaviors (Fig. 5c). EPSC peaks as triggered at different frequency was plotted (Fig. 5d and Fig. S4 in Supporting information). The dependence of EPSC on peak frequency was well fitted with the sigmoidal function [33, 34]:

(1)

where p is the order of the function, f_C is the cut-off frequency, a₁ and a₂ are the initial and final amplitude. The fitting gave rise to the value of ~7.4 Hz for f_C. Biologically, synapses can act as a dynamic filter for the processing of the input information, depending on the frequency of the excitation. A synapse with a low probability of vesicle release can act as a high-pass filter, which only respond to high-frequency signals to trigger the neurotransmitter-vesicle release. Therefore, the synapse allows the signals with high frequency exceeding a certain cut-off value to pass through, and immensely weakens low frequency signals. This process is analogous to the high-pass filtering behavior in a biological synapse. The high-pass filtering behavior allows signals at higher frequency than a certain cut-off value to pass through, and blocks low-frequency signals.

To evaluate the short-term and long-term synaptic plasticity of the device, we applied a series of sequential spike to see the EPSC change (Fig. 5e). After only 1 spike, the EPSC was about 1 µA. When additional spikes are applied, the current level increases, and reaches 1400% of the original peak, illustrating good short-term plasticity, which is important for neuromorphic computation. However, after the removal of the spikes, the EPSC immediately drops to the initial current level. These special synaptic behaviors imply fast migration in the cross-linked network, in that the ions could migrate to equilibrated states. The short-term only based synaptic plasticity is nerve signal transmission. Without cross-linking agent, the device was not working properly, which is possibly caused by high electrical leakage of the thin film (Fig. S5 in Supporting information).

To further evaluate energy consumption of the devices, presynaptic spikes with much lower amplitude of 10 mV, 20 mV, 30 mV, 40 mV and 50 mV were applied, and caused obvious change in EPSC (Fig. S6 in Supporting information). After stimulation using different spike numbers, different accumulative EPSC were observed, demonstrating efficient tuning of synaptic plasticity. Recently, great progress has been made to reduce synaptic energy consumption [35]. According to the responsive signals, our device achieved a low-energy consumption of 0.35 fJ per synaptic event, which is below that of biological synapses.

In conclusion, we have fabricated a synaptic transistor using a thin film of SnO₂ nanoparticles as the conductive channel and a thermally cross-linked ion gel as an ion-rich layer to emulate the synaptic cleft, where neurotransmitters were released. The device demonstrated essential synaptic functions for neuromorphic computation and neural information transmission, such as paired-pulse facilitation, spike-voltage dependent plasticity, spike-rate dependent plasticity, and spike-number dependent plasticity. Ultralow energy consumption was achieved below femtojoule, which is so far the lowest among the ion-gel gated synaptic transistors. Logic functions are also realized. Moreover, logic operations and high-pass filtering functions have been achieved using the device. This demonstration of ion gels composed of thermally crosslinked matrix represents a new strategy for fabrication of robust and versatile ion gels, and may provide a useful resource for design and modulation of information processing units in neuromorphic electronics.

Declaration of competing interest

There are no conflicts to declare.

Acknowledgments

We gratefully acknowledge the financial support from the National Natural Science Foundation of China (No. 21601076), the Natural Science Foundation of Liaoning Province (No. 2019-ZD-0266).

Supplementary materials

Supplementary material associated with this article can be found, in the online version, at doi:10.1016/j.cclet.2022.107842.