Design of the CE-activated artificial afferent

Inspired by the afferent nerve system that the biological stimuli-receptors capture the external touch/stretch/stress/temperature/humidity stimulus and trigger the action/postsynaptic potential to be delivered via dorsal column-medial lemniscus pathway (Fig. 1a)^{29,30,31,32,33,34}, we fabricate the contact-electrification-activated artificial afferents to mimic the function of biological afferent nerves. The triboelectric potential induced by contact electrification activates the synaptic transistor and delivers the relevant mechanical information to functional end terminals (Fig. 1b). Here, the triboelectric potential plays two critical roles: (i) it acts as the power source to drive (gate) the synaptic transistor (Fig. 1c, d); and (ii) it directly correlates the spatiotemporal information of the external stimuli with output signals (i.e., PSCs).

The self-activation component relies on the TENG technique. It is arbitrary with multiple geometries and abundant materials (any material capable of contact-electrification) to reflect spatiotemporal information such as displacement, pressure, or touch pattern. The self-activation component in optional contact-separation (CS), sliding or single-electrode mode enables the recognition of displacement or touch patterns (Supplementary Fig. 1a–c). It is also applicable to the design of a three-point bending architecture (Supplementary Fig. 1d) to monitor external pressure^35,36,37,38, by correlating the pressure information with the displacement in a nonlinear relationship. The external stimuli induce triboelectric potentials through the TENG technique to activate the synaptic transistor, which is a self-powered process according to contact-electrification and electrostatic induction. The driving triboelectric potentials accompanied by CE can also be quantified by the output signals from synaptic transistors, i.e., the external mechanical stimuli are quantifiable by the artificial afferent neuron.

The synaptic transistor is an ion-gel-gated MoS₂ field-effect transistor (FET, Supplementary Fig. 2). The wide band gap and atomically thin body of MoS₂ (Supplementary Fig. 3) can efficiently reduce the direct source-drain tunneling current and improve the transport properties in the channel for high-performance neuromorphic signal transmission and low energy dissipation²⁸. The ion gel is intrinsically a solid-state electrolyte (Supplementary Fig. 4), in which the migration and distribution of ions determine the spatiotemporal postsynaptic electric signals (Supplementary Fig. 5)^5,39,40. Based on the ion-gel-gated FET, we can achieve the gradual decay behavior of the postsynaptic current (75 ms, Supplementary Fig. 5b) according to the slow migration of ions during the unconventional form of the electrical double layers (EDLs)^4,40. This process is critical for the ion-gel-gated synaptic transistors to imitate the working process of biological synapses (Supplementary Fig. 6).

In CE-activated artificial afferents, we consider the contact/separation or sliding forward/backward cycle actions as the activation input pulses. The ion gel (ionic conducting but electrical insulating) can act as a buffer layer to rectify/modify the intrinsically opposite paired pulses of the TENG to minimize (or remove) the influence of the second opposite pulse. Thus, the induced triboelectric voltages can be efficiently coupled to the FET channel through the ion gel and behave like an equivalent gate voltage supply (i.e., triboelectric gate potential, Supplementary Fig. 7)^26,28. Through CE-activation, the artificial afferents (synaptic transistors) show representative biological synaptic behavior, such as the decay time, excitatory postsynaptic current (EPSC), and paired-pulse facilitation (PPF). They are also capable of recognizing spatiotemporal information such as displacements and touch patterns. As a proof of concept, we design a voltage amplifier and processing circuit to convert the corresponding PSCs (related to the spatiotemporal touch information) into output trigger voltages to drive LEDs in relevant logics (representing the virtual excitations in the cerebral cortex). This work demonstrates a coupling of triboelectric charges, ions, and electrons triggered by touch/sliding/pressure, each of which can be converted into triboelectric signals and induce corresponding PSCs in a self-powered fashion to reflect the spatiotemporal information of external mechanical stimuli, i.e., contact-electrification-activated afferent neurons. In contrast to the previous literature, by applying a stimulating or driving voltage to imitate synaptic behaviors^9,41,42, the triboelectric potential triggered by external displacement, pressure, or touch can be considered as a self-powered action potential to complete the artificial somatosensory system. It offers a more direct coupling between external stimuli and PSCs, meanwhile greatly decreasing energy consumption.

Typical synaptic characteristics of the CE-activated artificial afferent

We take the CE-activated artificial afferents in contact-separation mode (CS mode) as an example to demonstrate the neuromorphic behaviors. The CE-activation part is composed of a polytetrafluoroethylene (PTFE) film and an Al electrode (selected according to whether it has enough different electronegativities in the triboelectric series). During the CS process, the induction charges on the Al/PTFE contribute to a Maxwell displacement current and result in a variable electric field (i.e., triboelectric potential or action potentials) related to the separation distance (D, considered as the presynaptic signal in spatial fashion). With PTFE/Al electrode in the CE-activation part connected to the common ground with the source electrode, the triboelectric potential can be effectively coupled to the synaptic transistor and trigger the EPSC (Fig. 2a, Supplementary Fig. 8). During the CS process, the transferred charges (Q) are critical in affecting the EPSC behavior and are closely related to the separation distances (Fig. 2b). A separation distance of 20 μm leads to positive charges of ~0.3 nC transferring to the gate of the synaptic transistor, equivalent to applying a gate voltage of 0.125 V to drive the transistor (Supplementary Fig. 9).

The typical EPSC behavior of the CE-activated artificial afferent is characterized in Fig. 2c. A capacitor with 1 nF is connected in parallel as a buffer to ensure steady triboelectric potential coupling (detailed discussions in Supplementary Note 1, Supplementary Fig. 10, and Supplementary Fig. 11). The separation displacement, acting as the presynaptic input, electrifies the friction layers and couples the triboelectric potential to the ion-gel-gated MoS₂ synaptic transistor. It then induces anion/cation migration to form EDLs in the ion gel, efficiently shifts the Fermi level of the MoS₂ channel, and triggers the EPSC (Supplementary Note 2 and Supplementary Fig. 12). An input displacement at 20 μm results in an EPSC of 0.34 nA, which gradually decays back to the initial value during the contact process between two friction layers of the TENG (Fig. 2c). The separation/contact cycle action (or sliding forward/backward in sliding mode) functionalizes as the activation input pulse of the artificial afferent. The decay time (\(\tau\)) for the CE-activated EPSC (0.34 nA) is estimated to be ~54 ms, indicating that the feature time of ion migration is ~54 ms (Fig. 2d). Notably, the decay time under CE-activation is similar to the intrinsic EPSC (0.87 nA) of the ion-gel-gated synaptic transistor (75 ms, Supplementary Fig. 5b). Compared with previous reports, the decay time of our device triggered by a single pulse is in the range of the sensory memory stage (Supplementary Note 4 and Supplementary Fig. 13a). A survey of the decay time of the previous literature has been listed in Supplementary Table 1 and Supplementary Fig. 13b. The achieved decay time of 75 ms may fluctuate according to the amplitude, duration time, and pulse number of the external stimulation. It has the potential to transform from sensory memory to a short-term memory model. Further engineering on the architecture of the synaptic transistor or developing proper training methods is probably to realize the long-term memory model. The relatively longer decay time compared with that of the synapses in biological afferent nerves (1.5 to 5 ms) presumes the imitation of the learning or memory behavior in the future.

The CE-activation behavior is comparable with the EPSC activation process in a biological synapse. First, a larger displacement (reflecting the spatial information) means a higher action potential in the neural system, which attracts more ions in the synaptic transistor to contribute to the EDLs and induce higher EPSC. As the separation distance increases from 20 to 120 μm, the peak value of the EPSC increases from 0.34 to 136 nA (Fig. 2e). The action potential of the displacement spike (20 μm) gives rise to a similar EPSC at a gate voltage spike of ~0.5 V (Supplementary Fig. 5c).

Second, the duration time of the displacement reflects the temporal information of the external stimuli. To check the ability to recognize the temporal information based on the mechanical displacement, we hold the CE component in the separated state (D = 50 µm) for different amounts of time and monitor the corresponding EPSCs of the artificial afferent. The peak current of the EPSC increases from 0.97 to 20 nA as the duration time of the separated state increases from 0.08 to 1.75 s (Fig. 2f). The longer duration of the separation state gives enough time for the cation accumulation at the interface, more effectively enhances the electron transport in the MoS₂ channel, and leads to a higher EPSC. Notably, the peak value of the EPSC shows a linear increment (for a CS spike duration within 0.6 s) with a subsequent saturation tendency (for a duration ranging from 0.6 to 1.75 s) (Supplementary Fig. 14). The gradually saturated tendency is attributed to the ion having enough time to form EDLs over longer durations, which is consistent with biological synaptic behavior with gradually saturated neurotransmitters under continuous biological action potentials.

Third, the low power consumption of ~10 fJ per spike is the key to realizing complex computations assisted with millions of synapses^14,43. To reduce the power consumption of the artificial afferents (or synapses), it is generally critical to reducing the operating voltage and leakage current of the synaptic transistor. The CE-activated artificial afferent is particularly promising for ultralow power consumption: (i) the gate supply is completely replaced by the triboelectric potential; (ii) the atomic thickness of the MoS₂ channel permits a low operating voltage; and (iii) a thin encapsulation layer on the source-drain electrodes ensures the ultralow gate leakage current²⁸. With the drain voltage scaling down to even 1 mV, the MoS₂ channel still maintains excellent transfer properties and typical synaptic pulse output behaviors with efficient gate modulation ability (Supplementary Fig. 15). The power consumption of this CE-activated artificial afferent can be significantly reduced to 11.9 fJ per spike (Fig. 2g, evaluated from Supplementary Fig. 15c, and Supplementary Note 5), which is comparable to the human synapse (10 fJ), highly approaches the organic nanowire transistor (1 fJ)¹⁴, and precedes most of the artificial counterparts (e.g., the organic and oxide semiconductor^11,22,44,45, CNT^46,47, and 2D materials^4,5,19 based synaptic transistor) in previous literatures (Fig. 2h, Supplementary Table 1). The results show that our device has great potential to simulate the low-power-consuming neuromorphic bioelectronic devices with multiple functions. Note that our estimates for the dissipation energy are based on our specific model for the dissipation energy described above. It is possible that within practical applications additional dissipating channels are present, which can increase the dissipation energy. For instance, the back-end processing circuit and flash LED circuits are part of demonstrations instead of the necessary component of the CE-activated artificial afferent. Their energy dissipations are excluded for evaluation of the ultralow-power consumption of CE-activated artificial afferent.

According to the bijective relation between EPSCs and separation distances (or duration time), the EPSC signals are the direct reflection of the external displacements, which is applicable when deducing the applied mechanical displacements. The sensitivity can be divided into two regions: 7 μm⁻¹ and 45 μm⁻¹ for region I and region II, respectively (Supplementary Fig. 16). The functional integration of distance recognition and signal differentiation is the fundamental feature of the CE-activated artificial afferent. The artificial afferent in different working modes (sliding and single-electrode mode) offers optional ways to recognize different types of displacements or touch patterns as needed (Supplementary Fig. 17, 18). The corresponding sensitivity in the sliding and single-electrode mode is 0.52 μm⁻¹ and 17.8 μm⁻¹, respectively. We can also change the geometric design of the self-activation component with a three-point bending structure for pressure recognition (sensitivity is 12.3 kPa⁻¹, Supplementary Fig. 19 and Supplementary Note 6).

Advanced synaptic characteristics of the artificial afferent

To further characterize the temporal recognition capacity of CE-activated artificial afferents, we apply paired and multiple CS actions as the input stimuli pulses and investigate the corresponding EPSCs (Fig. 3a, b). Under the stimuli of consecutive and paired CS mechanical actions, the self-activation artificial afferent represents typical paired-pulse facilitation (PPF) behavior, i.e., the EPSC evoked by the second spike is increased when it closely follows the previous spike (Fig. 3c). In a typical PPF, the peak current of the EPSC activated by the second contact-separation spike (A₂ = 0.14 nA) is larger than that of the first one (A₁ = 0.1 nA). The demonstrated PPF behavior is a form of short-term potentiation (STP) synaptic plasticity, which is important for decoding temporal information and neuromorphic computations in an artificial biological system. The PPF index (defined as A₂/A₁), which evaluates the facilitation degree of the synapse, is closely related to the interval time between the presynaptic stimuli pulses (△T_pre, defined by the interspike interval time between the paired CS actions). A series of EPSC responses activated by paired CS spikes (D = 50 μm) with different △T_pre values is shown in Fig. 3d. The CE-activated artificial afferent shows an increasing trend of the PPF index with shorter △T_pre (inset of Fig. 3d), which is attributed to the fact that the ions contributing to the formation of EDLs do not have enough time to drift back and recontribute to the formation of the EDLs with the ions triggered by the second CS spike. The tunable capacity of the PPF index initialized by the paired CS actions demonstrates that the CE-activated artificial afferent has excellent short-term synaptic plasticity, which is beneficial for neural information computation and processing.

The CE-activated artificial afferents also possess the function of further potentiation plasticity under multiple consecutive CS action pulses (D = 50 μm, Fig. 3e), represented by a series of triggered EPSCs. The exertion of 60 CS-spike cycles increases the peak current of the EPSCs from 0.03 to 9.7 nA. The multiple consecutive CS spikes trigger multiple triboelectric pulses and continuously activate the MoS₂ channel through the ion gel. According to the gradual saturation of ions when constructing EDLs, the EPSC varies from an obvious increment tendency (initial excitation stage) to a subdued state (the last several cycles, inset of Fig. 3e). This behavior is similar to that of the saturated neurotransmitters under multiple presynaptic pluses. The current gain, defined as A_n/A₁ (A₁ and A_n represent the peak current of the first and last EPSC, respectively), is closely related to the CS action numbers (Fig. 3f). This relation exists because the number of CS actions directly determines the number of ions when constructing EDLs to modulate the MoS₂ Fermi level. The CE-activated artificial afferent shows gradually enhanced potentiation plasticity in real-time with increased stimuli action numbers (from 2 to 40 cycles, Fig. 3g). Excellent reliability and durability of the CE-activated artificial afferents with no obvious baseline drift have also been demonstrated in Supplementary Fig. 20 and Supplementary Fig. 21, which promises more availability and feasibility in practical applications.

Spatiotemporally correlated stimulation from multiple presynaptic terminals can be used to activate postsynaptic currents to establish dynamic logic in an artificial neural network^14,22. The CE-activated artificial afferent is also capable of simulating the spatiotemporal dynamic logic by utilizing the planar dual-gate mode synaptic transistor (with two TENGs coupled to a single MoS₂, Fig. 4a, b), which is similar in geometry with the fore-end of the biological myelinated nerve fiber (i.e., heminode)^48,49,50.

Two distinguishable CS actions with an interval time (△T), considered as the key spatiotemporal trigger sources, are applied to the artificial afferent. The different contact areas of the two TENGs (TENG-1 and TENG-2) induce different output voltages (Supplementary Fig. 22). The CS spike of TENG-1 (D = 30 µm) triggers EPSC-1 with a peak current of ~0.74 nA as the first presynaptic pulse, while the CS spike of TENG-2 (D = 20 µm) triggers EPSC-2 with a peak current of ~0.6 nA as the second presynaptic pulse (Fig. 4c). When the two CS spikes are applied sequentially with an interspike interval (△T_pre2-pre1), the influence of the second presynaptic pulse on the first one is evaluated with the recorded first EPSC. When △T_pre2-pre1 = 0, EPSC-1 and EPSC-2 are triggered simultaneously, resulting in a maximum EPSC of ~1.06 nA in the postsynaptic terminal. When △T_pre2-pre1 = −0.6 s, the CS-2 will superimpose with CS-1 and contribute to EPSC-1, leading to an enhanced EPSC-1. When △T_pre2-pre1 = 0.6 s, CS-2 is applied later than CS-1 and delivers no contribution to EPSC-1, leading to an unchanged EPSC-1. The measured EPSC amplitude at the end of the TENG-1 CS spike (set as t = 0, namely, when the summed EPSC is recorded) is plotted vs. △T_pre2-pre1 to understand the influence of the CS spike on a spatiotemporally correlated presynaptic terminal (Fig. 4d). The amplitude of the EPSC at t = 0 is consistent with the amplitude of EPSC-1 if EPSC-2 is triggered afterward (△T_pre2-pre1 > 0). In contrast, when EPSC-2 is triggered before EPSC-1 (△T_pre2-pre1 < 0), the EPSC at t = 0 is influenced by the superimposition of EPSC-1 and the remaining EPSC-2 activated by CS-1 and CS-2, respectively. The EPSC dynamic logic of the self-activated synaptic transistor in dual-gate mode represents a nonlinear spatiotemporal relationship (Fig. 4d). This is similar to the response of biological hippocampal CA1 pyramidal neurons under spatiotemporally correlated stimuli from different presynaptic terminals.

Except for the dynamic logic correlated with the interval time, it is also of great significance to recognize the frequency and magnitudes of external actions with the CE-activated artificial afferent. We monitor the real-time EPSCs simultaneously activated by the two TENGs at different contact-separation frequencies (Fig. 4e). The combination of the EPSC coordinates of TENG-1 (D = 0.3 mm) and TENG-2 (D = 0.2 mm) is shown in Supplementary Fig. 23. Both the frequencies and magnitudes of the displacements applied to the CE-activated artificial afferent can be recognized after Fourier transforms according to the different spatiotemporal patterns of the postsynaptic currents (Fig. 4f). The monitored results after the Fourier transform are consistent with both the initial and combined data, representing the excellent temporal recognition ability of the artificial afferent. The demonstrated MoS₂ synaptic transistor can combine with and recognize the triboelectric potential inputs from two pre-TENGs, exhibited as the EPSCs comprising the temporal information at two corresponding frequencies. The CE-activated artificial afferent in dual-gate (or multigate) mode mimics the way multiple presynaptic neurons input action potentials to the dendrites of a postsynaptic neuron, which is critical for abundant tactile sensation by the biological SA-I afferent nerve in advanced animals⁴⁸. Furthermore, the contact-electrification induces coupling of the triboelectric potential to the MoS₂ postsynaptic transistor in a self-activated way without any gate supply, which is beneficial for the low-power-consuming decoding of mechanosensation signals containing spatiotemporal information.

Dynamic logic recognition of spatiotemporal touch patterns

As a proof of concept, the dynamic logic function of the artificial afferent is demonstrated on a flexible substrate with the assistance of flash LED circuits. The system consists of a CE-activated artificial afferent neuron (Supplementary Fig. 24), a microcontroller unit (MCU), and two groups of LEDs in series (Fig. 5a). The dynamic logic is reflected through the flash sequences of the LEDs by recognizing different afferent signals induced by contact-electrification. When two different triboelectric potentials are coupled to the synaptic transistor (rectifier and filter), different EPSCs are triggered and converted into relevant bias voltage outputs through a bleeder circuit (Fig. 5b and Supplementary Fig. 25). The voltage outputs captured by the processing circuit are recognized as the threshold signals and drive the corresponding flash LED circuit (Supplementary Fig. 26). When a touch action is applied to TENG-1 (2 × 2 cm²), the triggered voltage is >2 V with a corresponding EPSC > 30 nA. The converted voltage trigger signal (>0.5 V) through the bleeder circuit can be acquired and recognized to induce a series of red lights to illuminate (Fig. 5d). When a similar touch pattern is applied to TENG-2 (1.5 × 1.5 cm²), the triggered voltage is <2 V with corresponding EPSC < 30 nA and trigger voltage <0.5 V, which can induce a series of green lights to illuminate (Fig. 5e, Supplementary Movie 1). More complex external touch patterns applied in the sequence are also distinguishable by the CE-activated artificial afferent (Supplementary Movie 2), representing the excellent capacity of the self-activated artificial afferent to recognize the spatiotemporal information of touch patterns.