General concept. Imitating the principle of biological synaptic cleft, GAS with a junction structure of electrolyte/GDY was fabricated to emulate the essential plasticity (Fig. 1). In biological neurons, the communication between cells hinge on the propagation of action potentials on axons (Fig. 1a). When the action potential reaches the front of the junction, the flow of calcium ions (Ca2+) are triggered and the neurotransmitters are subsequently released, and the sodium ions (Na+) eventually flow into the postsynaptic membrane (Fig. 1b). Electrolytes with lithium (Li+) and Na+ are exploited to prepare GASs, referred to as Li-GAS and Na-GAS, respectively. When a set of positive pulses are applied to the top of the artificial synapse (Fig. 1c), the alkali metal ions (Li+ and Na+) are forced to migrate to the gap and accumulate on the surface, or even insert into the interlayer, coinciding the charging process of rechargeable batteries (Fig. 1d).40 The investigation of the kinetics of ions at the interface facilitates the modulation of synaptic response. Furthermore, the information integration of parallel processing can be realized using GAS, by connecting with artificial muscles and completing movement responses on artificial efferent neurons.
Fabrication and operation of GAS. Two-dimensional (2D) GDY is produced from the catalytic coupling reaction of its precursor Hexakis[(trimethylsilyl)ethynyl]benzene (HEB-TMS) through the liquid/liquid interfacial protocol (Supplementary Fig. 1a).26,27 Atomic force microscope (AFM) image of the as-prepared GDY reveals its 2D nanoflake morphology with a thickness of about 3.3 nm (Fig. 2a). The layered structures with highly wrinkled nanosheets of GDY have been observed under the scanning electron microscope (SEM, Supplementary Fig. 1b). The transmission electron microscope (TEM) images (Fig. 2b and 2c) of the GDY nanoflakes clearly demonstrate their lattice fringes with the lattice spacing of about 0.45 nm, indicating the high crystallinity of the prepared GDY samples.32 The selected area electron diffraction (SAED) patterns further illustrate the good crystallinity of the fabricated GDY (Fig. 2d). The high-resolution X-ray photoelectron spectroscopy (XPS) spectra show the deconvoluted C 1s with major contributions from C≡C and C=C species, indicating the sp- and sp2-hybridized carbon atoms of GDY (Fig. 2e). The proportion of sp/sp2 carbon close to 1.5 is also consistent with the chemical composition of GDY. The presence of sp carbon is also demonstrated by Raman spectrum (Fig. 2f), in which the characteristic bands at 1932 and 2131 cm−1 deriving from the vibration of the conjugated diyne linkage are clearly present.44 The flat film of GDY is obtained by spin coating its suspended dispersion in N, N-dimethylformamide (DMF) (Supplementary Fig. 1c and 1d), with the thickness of approximately 400 nm (Supplementary Fig. 1e).
To illustrate the dynamics of the migration of anions and cations, the first ten current-voltage (I-V) sweeps were performed in Li-GAS and Na-GAS, respectively (Fig. 2g and 2h). An obvious negative differential resistance (NDR) phenomenon has been observed in Li-GAS (Fig. 2g), which is probably due to the migration of Li+ to the surface and plane of GDY. In this way, the electrochemical doping process occurs. However, two groups of NDRs appear in the I-V curve of Na-GAS, of which the NDR at the low potential is more pronounced (Fig. 2h). The very obvious NDR in Na-GAS might be caused by the large internal field formed by the interface ions.45 In the positive sweep range, the Li+ cations are doped in GDY and the anions gradually accumulate at the interface to form an internal field. The intensity of this internal field will temporarily exceed the applied electric field, manifesting an NDR phenomenon and a de-doping process. As the scanning window becomes narrower, the NDR phenomenon disappears, manifesting the effect of interface pseudocapacitance (Supplementary Fig. 2a and 2d). The first eight I-V sweeps in Li-GAS were performed to show good reversibility in the low voltage range (Supplementary Fig. 2e and 2f). The response of Na-GAS in the low voltage range is repeatable after the first few sweeps (Supplementary Fig. 2g and 2h). From the analysis of the I-V curve, positive and negative pulses with different amplitudes could have an effect on the synaptic weight under the action of interfacial capacitance and electrochemical doping.
Short-term plasticity and ultra-sensitivity of GAS. Synaptic plasticity is modulated by the presynaptic action potentials, which can trigger the Ca2+ influx and release the excitatory or inhibitory neurotransmitters to strengthen or weaken the information transmission. The presynaptic pulses with different amplitude, duration, frequency and number are processed by the synaptic device to output different types of current signals, which is a spike-dependent plasticity. In our synaptic devices, Li- and Na- GASs, varying degrees of short-term enhancement have been demonstrated. Under the same presynaptic pulse (+2 V, 440 ms), it is easier for Li+ ions to be activated and to migrate in the corresponding synaptic device (Fig. 3a). With the gradual increase of the pulse amplitude (from 0.5 to 4 V) in Li-GAS and Na-GAS, the peak value of postsynaptic current increases stepwise from sub-nA to tens of nanoamps, which is a typical spike-voltage dependent plasticity (Fig. 3a).46 This plasticity can also be observed under the action of negative pulses (Supplementary Fig. 3a). Under a single pulse stimulus, the surge current decayed to the level of ± 0.2 nA in a short time (< 6 s), showing an obvious bidirectional short-term plasticity of GAS (Supplementary Fig. 3b and 3c). Successive pulse stimulation in a short interval (Δt < 3 s) will produce paired current peaks of different heights (A1 and A2). As the time interval increases, the PPF effect (A2/A1×100%) becomes weaker (Fig. 3b), which may result from the rapid diffusion of ions provided by the large pores of GDY network. At the same pulse interval, Na-GAS exhibits a higher facilitation index compared to that of Li-GAS. The facilitation index under negative pulses is also relatively lower, which might be attributed to the volatile effect required for short-term plasticity. Negative pulse sequences with different frequencies have an effect on the peak of postsynaptic current. The higher the pulse rate is, the more obvious the gain (A10/A1×100%) will be (Fig. 3c), demonstrating a typical spike-rate dependent plasticity (SRDP).47 This short-term enhanced plasticity enables GAS to act as a dynamic filter for information transmission.21 It is obvious that Li-GAS shows a better dynamic filtering (Fig. 3d), which is consistent with the previous PPF effect (Supplementary Fig. 4a and 4b). As the number and duration of pulses increase, the current gain becomes less pronounced during the charging process and the current decay is still timely during the discharging process (Supplementary Fig. 4c-4f). The reason for this rapid discharge process could be attributed to the ion migration or interface depolarization.19 Such a rapid volatile process is important for real-time imaging applications through the construction of synaptic array (Fig. 3e). If the letters of “G”, “D”, and “Y” are input into the array (9 × 9), the 81-pixel image formed by the array can be refreshed in a short time, thanks to the exemplary short-term plasticity of the synaptic unit. When continuous nonidentical negative pulses are applied, a repeatable short-term plasticity is observed (Fig. 3f). The pulses with different amplitudes correspond well to the different discharge current peaks, which further indicates the good stability and repeatability of GAS, in comparison with solution-processed reduced graphene oxide.48 It takes only 0.22 seconds to encode the postsynaptic current triggered by the presynapse of different amplitudes into a pattern with obvious contrasts (Supplementary Fig. 5a and 5b). Such a rapid deintercalation of ions should be attributed to the triangular macroporous structure of the GDY network. Moreover, when 10 and 15 consecutive negative pulses (-3.5 and -5 V) were applied in Li-GAS, the peak current was increased and then quickly declined within a few seconds after removing the pulse (Supplementary Fig. 5c). Such an ion diffusion dynamics also occurs in Na-GAS, resulting in a remarkable pulse sensitivity (Fig. 3g). As the pulse amplitude gradually weakens to the range of 80 to 8 mV, such a plasticity can be still achieved and the anverage power consumed by a single synaptic event is 16.7 fW, orders of magnitude lower than the biological level and the lowest value so far in the two-terminal devices (Supplementary Fig. 5d). The sensitivity to presynaptic pulses, ultra-low power consumption, and the significant volatility are integral to construct a bioinspired ionic sensory/motor system.
Integration, parallel processing and artificial efferent neuron actuation of GAS. Electrolyte-based devices can convert biologically ionic/chemical signals into electrical signals, and therefore attractive dendritic integration can be possibly realized.11 Multiple input signals applied to the top of GAS will be integrated and output at the bottom (Fig. 4a). Firstly, the rules of spatiotemporal learning of two synapses are simulated. The two synapses exhibit synaptic connections with different strengths under different amplitude pulse stimulation in Li-GAS and Na-GAS (Fig. 4b). When the two synapses act simultaneously, the nonlinear increases of the output prominently. When one of the two synapses are triggered earlier than the other (ΔT < 0) by less than 4 s, the output postsynaptic current will gradually increase to 340% and 250% of the initial output in Li-GAS and Na-GAS (Supplementary Fig. 6a), respectively. Through the input of multiple preneurons, interesting temporal features of dendritic connections have been observed. This also shows that the synergy of multiple synapses (synapse 1 and synapse 2) can significantly enhance the synaptic weight compared to different numbers of repeated stimulation applied to a single synapse (Fig. 4c). In this way, the gain obtained by parallel processing of multiple synapses is much higher than a single synapse under spike-dependent plasticity. Distributing the weight pressure to a single synaptic unit is the essence of parallel processing. The shared information transmission by multiple synapses resemble better the biological information integration than information transmission by single synapses. Similarly, the gain of the postsynaptic current triggered by presynaptic pulses with different durations in a single synapse is still not as significant as the gain obtained by the integration of multiple synapses (Fig. 4d). Therefore, logic operations can be realized according to the different gains of one synapse and two synapses at different durations (Supplementary Fig. 6b). Meanwhile, the modulation of the neural responses can also be simulated under the synergy of two synapses. Then, shunting inhibition, a mechanism for regulating neural response, can be achieved by applying excitatory and inhibitory pulses to the two synapses respectively (Fig. 4e). With the gradual increase of a presynaptic inhibitory pulse, the postsynaptic current gradually weakens until it disappears. This suggests that the output can be both zoomed in and out during information integration.
The parallel processing of neuromorphic biological signals is of crucial importance in the novel computational paradigm.11 It remains as a challenge for synaptic units to perform real-time parallel processing of external perception information (light and temperature) transmitted by multiple sensory neurons and to make artificial efferent nerves respond differently to the integrated information (Fig. 4f). Here, impulse stimuli of different frequencies are applied to the top electrode of GAS as perceptual information from the outside, and artificial muscles are connected to the bottom electrode of GAS through a circuit to construct artificial efferent neurons (Fig. 4g). Initially, under a single frequency (0.48 Hz) input, the currents generated by two independent synaptic units are close, and their overall outcome is equal to the peak current of the two synapses in the same time period (Fig. 4h). If the two synapses can simultaneously receive signals provided by peripheral devices in the same time period, the signal reflected in the post-neuron is the superposition of the signals obtained by the two independent synapses. The effect of integrating multiple inputs and accumulating outputs is significantly higher than that of a single synapse. Next, the results obtained under different pulse frequency inputs (0.48, 0.8 Hz; 0.344, 0.60 Hz) are consistent, which means the output results of multiple presynaptic pulses in the same time period are accumulated in parallel (Supplementary Fig. 7). Meanwhile, the peak shape of the post-neuronal signal can clearly reflect the time period during which the two synapses receive signals of different frequencies from the peripheral unit, and the frequency of the presynaptic pulse can be inferred from the different peak-to-valley spacing (Supplementary Fig. 7e). Hence, GAS can identify the frequency of presynaptic pulses to a certain extent to infer and analyze the sensory information transmitted from afferent nerves. Finally, as a proof-of-concept, artificial efferent nerves have been constructed to implement the integrated output of multiple synaptic inputs (Supplementary Fig. 8). With only one input terminal, the generated output signal reaches the motor neuron via the efferent nerve to trigger the artificial muscle bending (Fig. 4i). When two sets of inputs are applied, the signals are processed in parallel and output by GAS, and finally transmitted in the efferent nerve to drive the artificial muscle to produce greater bending. Multiple external inputs are integrated in the GAS and the cumulative output drives the artificial muscle. Such a GAS element can process in parallel and integrate the received signals in real time, and can be combined with a multi-functional actuator to sense or control objects, and can also be coupled with a variety of receivers to sensitively reflect external environmental information.1,8,49