An organic electrochemical neuron for a neuromorphic perception system

Significance

Bioelectronic-inspired neuromorphic perception systems have the potential for efficient sensing and processing environmental stimuli. However, artificial neurons and synapses constructed with silicon and inorganic materials are constrained by their limited compatibility with biological systems. Organic electrochemical transistors (OECTs) have emerged for constructing organic electrochemical neurons (OECNs) because their working mechanisms resemble those of various biological processes. However, expanding the range of neuron responses that OECNs can mimic remains challenging. Here, we introduce an OECN concept using OECTs with an innovative semiconductor and vertical architecture, achieving a broad, adjustable firing frequency range of 0.13 to 147.1 Hz in a compact design. Additionally, we demonstrate a neuromorphic perception system integrating mechanical sensors with OECNs and an artificial synapse for tactile sensing.

Abstract

Human perception systems are highly refined, relying on an adaptive, plastic, and event-driven network of sensory neurons. Drawing inspiration from Nature, neuromorphic perception systems hold tremendous potential for efficient multisensory signal processing in the physical world; however, the development of an efficient artificial neuron with a widely calibratable spiking range and reduced footprint remains challenging. Here, we report an efficient organic electrochemical neuron (OECN) with reduced footprint (<37 mm²) based on high-performance vertical OECT (vOECT) complementary circuitry enabled by an advanced n-type polymer for balanced p-/n-type vOECT performance. The OECN exhibits outstanding neuronal characteristics, capable of producing spikes with a widely calibratable state-of-the art firing frequency range of 0.130 to 147.1 Hz. Leveraging this capability, we develop a neuromorphic perception system that integrates mechanical sensors with the OECN and integrates them with an artificial synapse for tactile perception. The system successfully encodes tactile stimulations into frequency-dependent spikes, which are further converted into postsynaptic responses. This bioinspired design demonstrates significant potential to advance cyborg and neuromorphic systems, providing them with perceptual capabilities.

Biological systems consistently surpass their electronic counterparts in interacting with the dynamic real world, owing to their unparalleled sensorimotor capabilities (1, 2). Emulating the functions or structures of natural systems presents an opportunity to address the limitations of current digital technologies (3, 4). Unlike conventional systems, biological systems operate through distributed computing paradigms, characterized by adaptive, plastic, and event-driven sensory neuron networks, which inherently provide superior fault tolerance and energy efficiency (5, 6). By mimicking these neuronal processes, it becomes possible to achieve biological-like perceptual abilities, facilitating seamless integration of these technologies into natural systems (7–10). To develop a biologically inspired perception system, integrating sensors with artificial neurons is essential. While artificial neurons based on solid-state silicon (Si) circuits hold significant promise (11–13), they face limitations such as mechanically rigid interfaces, restricted biocompatibility, and high system complexity (14). Similarly, volatile nonlinear devices based on memristors and spin torque oscillators have been explored to fabricate artificial neurons with high integration density (15–17), but have limited compatibility with biological environments persists.

Synthetic organic semiconductors are attractive alternative materials for neuromorphic devices due to their structural kinship with biomolecules, enabling integration in artificial synapses, nerve electronics, and neural interfaces (14, 18–21). Organic semiconductor-based devices offer significant potential for interfacing with biological systems due to their mechanically soft nature, tunable physical properties, potential biocompatibility, and ability to directly interact with ions in aqueous electrolytes (22–27). State-of-the-art organic artificial neurons, often implemented using multiple ring oscillators, demonstrate impressive performance in neurorobotics and neuroprosthetics (28, 29), however, their application is constrained by the high supply voltages required and the complexity of their circuit designs. Recently, artificial neurons based on organic electrochemical transistors (OECTs) have emerged for simulating spiking-based neural behavior with low voltage supplies and simple designs (30–34). Nevertheless, duplicating the range of biological spiking frequencies (<1.9 Hz and 6 to 40 Hz) is severely limited by the intrinsic performance of current OECT elements, presenting a challenge to operate within the wide frequency range (from <1 Hz to >100 Hz) of biological neurons (35–38). Advances in materials and structural design for OECT-based neuronal circuitry are therefore clearly needed.

Here, we demonstrate a “leaky integrate-and-fire (LIF)” organic electrochemical neuron (OECN) concept for neuromorphic perception systems based on vertical OECT (vOECT) complementary circuits. Essential to this realization is the introduction of an advanced n-type ionic-electronic polymeric material enabling closely balanced p-type and n-type vOECT pairs and thereby, complementary circuits with excellent performance. The resulting circuits provide reduced footprints (4.9 × 10³ µm² channel area), balanced inverter performance (>84% for V_DD/2 in both low and high noise margins), and rapid response (up to 500 Hz). Such performance yields OECNs with excellent neuronal characteristics, producing spiking signals within a wide calibratable frequency range from 0.13 to 147.1 Hz, enabling flexible and adjustable neuronal activity duplicating the firing patterns of biological neurons (31). This enables a complete neuromorphic perception loop by integrating these OECNs with mechanical sensors and artificial synapses for tactile signal processing.

Fig. 1 illustrates a scenario in which a human is shaking hands with a robot, highlighting the marked difference between a biological perception system and an artificial perception system responding to tactile stimuli, as reported here. In the biological perception system, mechanoreceptors and neurons play a crucial role in converting external tactile stimuli into electrical spikes. These electrical signals are then transmitted to the cerebral cortex for further processing and interpretation (39). In the OECN-based artificial perception system, we integrate mechanical sensors (pressure and strain sensors), which function as the mechanoreceptors, within artificial skin. These sensors convert external tactile stimuli into electrical signals. The OECN acts as a neuron, encoding various types of mechanical sensory signals into electrical spikes, enabling subsequent communication with other neurons through an artificial synapse. By mimicking the mechanisms of the biological perception system, the artificial system demonstrated here offers the capability to perceive and respond to tactile stimuli in a manner similar to biological organisms.

Fig. 1.

Conceptual schematics of biological and organic artificial perception systems. (A) Schematic of biological perception system. Physical stimulus is converted by subcutaneous mechanoreceptors and neurons into electrical impulse signals, which are then transmitted to the cortex for further processing. (B) Schematic of a synthetic neuromorphic perception system. The OECN combined with mechanical sensors (pressure and strain sensors) encodes external tactile stimulus into electrical spikes, which are further transmitted to other neurons through artificial synapses.

Results and Discussion

Building Blocks for an OECN.

To achieve OECNs with the required functionalities (Fig. 1) it is mandatory to fabricate high-performance organic complementary circuit building blocks. We recently demonstrated vOECTs having ultra-high performance and small footprints (40), providing significant advances in constructing basic logic circuits. Nevertheless, complementary circuitry requires highly balanced p- and n-type enhancement-mode transistors in which the turn-on voltages (V_ONs) and on-currents (I_ONs) are comparable, which was not achieved in that previous work. Here, an advanced electron-transporting mixed ionic-electronic conducting organic polymer, Homo-gDPPTz (see chemical structure in Fig. 2A, synthetic details in SI Appendix, Supplementary Text and Figs. S1–S6http://www.pnas.org/lookup/doi/10.1073/pnas.2414879122#supplementary-materials), is synthesized, purified, and used to fabricate n-type vOECTs. Complementary p-type vOECTs incorporating the hole-transporting mixed ionic-electronic polymer gDPP-g2T, are then integrated with the above n-type vOECTs to construct complementary inverters (Fig. 2A). The schematic structure and top view optical image of the inverter are shown in Fig. 2B. Details of vOECT and complementary inverter fabrication can be found in http://www.pnas.org/lookup/doi/10.1073/pnas.2414879122#supplementary-materialsSI Appendix, Supplementary Text and Figs. S7 and S8http://www.pnas.org/lookup/doi/10.1073/pnas.2414879122#supplementary-materials. Briefly, since all vOECTs are fabricated by sandwiching the channel between Au source/drain electrodes in the vertical direction, vOECT-based complementary inverters can be constructed by placing the n-type vOECT directly on top of the p-type vOECT. Note, to pattern the semiconductor in the channel, the redox-active semiconductor (p-type gDPP-g2T or n-type Homo-gDPPTz) is blended with a photocurable polymer component [Cinnamate-Cellulose (Cin-Cell)]. Finally, a Cin-Cell encapsulation layer is spin-coated and photopatterned on the inverter, leaving the channel area open.

Fig. 2.

Design and performance of vOECTs and complementary inverters. (A) Layout diagram of the vOECT complementary inverter, along with the semiconducting polymer chemical structures. (B) Exploded structure and topview optical image of the vOECT-based complementary inverter. Representative transfer and transconductance curves of vOECTs based on: (C) p-type gDPP-g2T and (D) n-type Homo-gDPPTz (electrode width W = 70 µm). Voltage transfer curves (E) and the corresponding (F) voltage gains of the vOECT-based inverter with V_DD varying from 0.3 V to 0.7 V. (G) Comparison of voltage gain as the function of V_DD in this work to literature reports. (H) Switching behavior of the vOECT-based inverter with V_IN switching between 0 V and 0.7 V at frequencies ranging from 1 Hz to 500 Hz; the top and bottom curves indicate V_IN and corresponding output voltage V_OUT, respectively. (I) Voltage amplitude of input sinusoidal voltage with amplitude of 5 mV and frequency of 10 Hz (red line) and output signals (green line) using the inverter as a signal amplifier. (J) Voltage gain of the amplifier as a function of input frequency, where the sinusoidal signal amplitude is fixed at 5 mV. V_DD = 0.7 V. The insert illustrates the amplifier configuration.

Before fabricating vOECT-based complementary circuits, the p- and n-type vOECT characteristics were first individually evaluated. The representative p-type gDPP-g2T vOECT (electrode width W = 70 μm) transfer plot shown in Fig. 2C demonstrates excellent performance, achieving negligible hysteresis, a large on/off current ratio (I_ON/I_OFF) of ~10⁶, and peak transconductance (g_m,p) of 359.71 ± 12.1 mS, along with V_ON = 0.1 V. Likewise, Fig. 2D presents the representative transfer plot for the corresponding n-type Homo-gDPPTz vOECT which, for an n-type device, achieves an unprecedented I_ON/I_OFF >10⁶ and g_m,p of 259.42 ± 20.7 mS, along with V_ON = −0.1 V. Note that the two devices are well balanced, reflecting the favorable molecular engineering of the n-type polymer semiconductor in which a thiazolyl ring replaces the conventional thienyl group (40, 41), stabilizing the lowest unoccupied molecular orbital (LUMO) by ~0.3 eV (−4.15 vs. −3.88 eV, respectively) as assessed electrochemically (SI Appendix, Fig. S9).

Atomic force microscopy (AFM) and grazing-incidence wide-angle X-ray scattering data indicate that Homo-gDPPTz has a more favorable morphology and microstructure for electrolyte exchange versus Homo-gDPP (SI Appendix, Figs. S10 and S11 and Table S1http://www.pnas.org/lookup/doi/10.1073/pnas.2414879122#supplementary-materials). Specifically, the Homo-gDPPTz crystal-coherence lengths (CCLs) for the (100)/(010) reflections are smaller than those of Homo-gDPP (59.94/45.89 Å vs. 106.49/50.08 Å, respectively). Additionally, AFM imaging indicates that Homo-gDPP forms well-defined fibrils, in agreement with the greater texturing, and is rougher than the Homo-gDPPTz film (RMS roughness = 1.82 vs. 0.56 nm). Note that the greater Homo-gDPP CCLs and increased fibril densities can restrict electrolyte transport since electrolyte uptake occurs primarily in amorphous polymer domains, and polymer crystallites and aggregates can serve as electrolyte ion traps (42–44). Thus, the favorable LUMO energy and morphology/microstructure of Homo-gDPPTz make it an efficient mixed ionic-electronic conductor. The highly symmetric transfer characteristics (similar I_ONs, I_ON/I_OFF, and symmetric operating voltages) in the p- and n-type vOECTs are thus established, affording complementary inverters with the following characteristics: 1) Rail-to-rail output at a supply voltage (V_DD) ranging from 0.7 to 0.3 V (Fig. 2E), delivering an ideal logic level in digital integrated circuits. 2) Switching voltage [V_SW, input voltage (V_IN) that corresponds to peak voltage gain] is located at approximately V_DD/2 (SI Appendix, Table S2). Large high and low noise margins (NM_H and NM_L) as shown in SI Appendix, Fig. S12, and the total NM with respect to V_DD approaches 87% (V_DD = 0.7 V), which is among the highest reported for OECT-based complementary inverters (45). 3) An ultra-high voltage gain of 433 is obtained under V_DD = 0.7 V (Fig. 2F and SI Appendix, Table S2), which exceeds previously reported values for either unipolar, ambipolar, or complementary inverters based on electrolyte-gated transistors and carbon nanotube-based CMOS circuits (Fig. 2G and SI Appendix, Table S3). 4) V_DD can be lowered to less than 0.1 V, where good inverter performance with high voltage gains is still obtained, along with ultra-low static power consumption of 1–15 nW with V_DD from 0.05 to 0.2 V (SI Appendix, Fig. S13 and Table S2http://www.pnas.org/lookup/doi/10.1073/pnas.2414879122#supplementary-materials).

The dynamic switching characteristics of the inverter were further evaluated by applying a square-wave V_IN (switching between 0 V and 0.7 V) with frequencies varied from 1 Hz to 500 Hz (Fig. 2H). Note that for frequencies up to 500 Hz, the inverter still shows good logic inversion, reaching a high logic state of “1” with V_OUT of ~0.7 V, and low logic state of “0” with V_OUT of ~0 V. The rise (τ_rise) and fall (τ_fall) transient times are fitted to ~215 µs and 204 µs (SI Appendix, Fig. S14), respectively. Regarding cycling stability, stable voltage output behavior is recorded over 50,000 switching cycles (SI Appendix, Fig. S15). Moreover, the ultra-high performance of the present inverter is ideal for ion-sensing over a wide concentration range, from 10⁻⁵ M to 1.0 M (SI Appendix, Fig. S16A). From the measured transition voltage as a function of ion concentration, an ion sensitivity of 122 mV/dec is derived by linear (dashed line) fitting of the data (SI Appendix, Fig. S16B)—far higher than sensitivities reported to date for organic complementary circuits (<100 mV/dec) (40, 46).

Although typically implemented as digital devices, inverters can also be used as analog amplifiers for electrophysiological signal sensing (47). A major advantage of complementary inverters in the push–pull configuration as amplifiers is that they benefit from summation of the g_m values of both p-type and n-type transistors (45, 48). The present vOECT-based push–pull amplifier shows a gain of 78.4 (38 dB) when amplifying a 5 mV (amplitude) sinusoidal signal with a frequency of 10 Hz under a DC offset ~0.35 V (Fig. 2I shows the voltage amplitudes for comparison). Gains of the voltage amplifier with respect to different input frequencies (from 0.1 Hz to 1,000 Hz with a 5-mV input signal) are further demonstrated in Fig. 2J and SI Appendix, Fig. S17. The amplification gains are 99.4 (40.0 dB) and 20 (26.0 dB) for 0.1 Hz and 100 Hz input signals, respectively, indicating superior performance versus those in the current OECT literature (gain < 30 V/V) (45, 47, 49). Thus, this vOECT-based complementary inverter design can serve as a realistic high-performance component for diverse organic electronic circuits.

OECN for Reproducing Biological Neuronal Signals.

The design goal of the present OECN is to replicate the operational mechanism of biological nerve cells (Fig. 3A). In biological neurons, dendrites receive input signals from presynaptic neurons and transmit them to the soma, where the signals are integrated. When the membrane potential surpasses a threshold, an action potential is triggered. The action potential variation induced by ion exchange can be divided into several distinct identifiable states (see details in SI Appendix, Fig. S18) (50). Here, the action potential spikes generated by the nerve cell can be clearly mimicked based on the present LIF OECN. As shown in Fig. 3A, the present OECN is constructed according to an Axon Hillock circuit (31, 51), which includes a two-stage vOECT complementary inverter serving as the noninverting amplifier block, an n-type vOECT acting as the reset transistor (T_reset), and two capacitors for the membrane capacitance (C_mem) and the feedback capacitance (C_f). Fabrication details of the OECN are shown in SI Appendix, Fig. S19. In this configuration, the input current (I_IN) is integrated by the capacitor (C_mem), leading to a gradual increase in the voltage (V_mem). Once V_mem reaches a specific threshold, a nerve-like pulse is fired at V_OUT. The amplifier block exhibits a sharp gain increase as V_mem surpasses the transition voltage (V_T), causing V_OUT to rise. When V_OUT becomes sufficiently high, T_reset activates, allowing C_mem to discharge and V_mem to decrease. As V_mem drops back to V_T, the feedback loop engages, producing a rapid drop in V_OUT with a minor change in V_mem. Consequently, V_OUT returns to zero, T_reset switches off, and the pulse ends, then restarting the cycle. Note that, due to the reduced footprint of the present vOECT and vOECT-based complementary inverter, the OECN is very small (W × L < 4.5 mm × 8.2 mm) (Fig. 3A). Note that additional miniaturization could be achieved by reducing the capacitor size and more accurate patterning of the semiconductor.

Fig. 3.

OECN fabrication and characteristics. (A) Schematic of a biological neuron cell and its analogy to the present vOECT-based OECN. The right side is an optical image of the OECN with a scale bar of 1 mm. (B) OECN spiking patterns under a constant input current I_IN of 1 μA with four different capacitances (C_mem = C_f = 1, 10, 100, and 1,000 nF). An aqueous electrolyte (PBS) is employed here to better simulate biological media. (C) OECN spiking patterns with 7 different input currents (I_IN = 0.1, 0.5, 1, 5, 10, 50, and 100 μA) with C_mem = C_f = 100 nF. (D) OECN spiking frequency modulation as a function of I_IN with four different capacitance configurations (C_mem = C_f = 1, 10, 100, and 1,000 nF). (E) Spiking frequency range of the corresponding human biological neurons from 0.1 Hz to 1,000 Hz.

The membrane capacitance significantly influences the conduction speed of action potentials in biological neurons. Lower membrane capacitance allows for faster action potential propagation, as fewer ionic charges are required to induce changes in the membrane potential (39). Accordingly, the spike frequency and width of the OECN can be controlled by adjusting the circuit capacitances. Fig. 3B shows the OECN spiking patterns with different capacitances (C_mem = C_f = 1 nF, 10 nF, 100 nF, and 1 μF) for a constant input current (I_IN) of 1 μA. The frequency and full-width-at-half-maximum (FWHM) variations in spikes with the capacitance changes in the circuit are summarized in SI Appendix, Fig. S20. The firing frequency approaches 1.09 Hz with a 1 μF capacitance and increases to 16.77 Hz with a 1 nF capacitance as lower capacitances reduce the charging time to reach the spiking threshold voltage and lead to higher spike frequencies. Similarly, the spike FWHM increases at higher capacitances since the peak width depends on the discharging time of the capacitors through the resetting transistor (T_reset). Additionally, since the amplifying gain of the complementary circuit determines the sharpness (slope) of the spikes, the ultra-high gain enables the generation of sharp spikes, which are crucial for simulating the spiking signals of biological systems.

The range of spiking frequencies that an OECN can achieve is of great significance as it determines the range of biological neural firing that it can mimic. The output response of the present OECN was further investigated by varying I_IN (from 50 nA to 100 μA) for different capacitances (1 nF to 1 μF). As shown in Fig. 3C, the OECN generates spikes with a specific current threshold of ≥100 nA when the capacitance is 100 nF, as lower I_IN cannot charge the capacitors sufficiently to reach the spiking threshold. This behavior mirrors the leaky characteristics of biological neurons, where the membrane voltage must exceed a certain threshold to trigger a pulse. The spiking frequency can be adjusted from 0.7 Hz at 100 nA to 100 Hz at 100 μA, as higher I_IN charges the capacitors more quickly, bringing V_mem closer to the threshold in a shorter time. The frequency-current (f-I) correspondence of the OECN with different capacitances is shown in SI Appendix, Figs. S21–S23 and summarized in Fig. 3D. The frequency range of the OECN can be modulated from 0.13 Hz at 200 nA with a 1 μF capacitance to 147.1 Hz at 100 μA with a 1 nF capacitance. Note, the OECN spiking frequency does not increase further by enhancing I_IN and/or reducing C_mem and/or C_f limited by the OECT transient response time (31). According to the transient responses of the p-/n-type OECTs used in the present OECN (SI Appendix, Fig. S24), the total delay induced by these transistors is approximately 5.8 ms. Thus, the maximum theoretical frequency should be ~170 Hz, a value close to the experimental upper limit. Since the response time is a function of the device capacitance, which depends on the channel material and geometry, the spiking frequency of our OECN could be further increased by improving the resolution of the encapsulation and scaling down the device size. However, the present measured frequency range (from 0.13 Hz to 147.1 Hz) is more than 50× broader than reported A-H neural circuits based on conventional OECTs (<2.4 Hz) (31, 52, 53). Such a wide and biologically plausible frequency range should be able to mimic a greater range of neural firing rates (Fig. 3E), including below 1 Hz for the human vasoconstrictor neurons in muscle or skin, or higher firing rates of more than 100 Hz for the highly active spiking neurons in the human neocortex (36, 37).

Organic Neuromorphic Perception System.

Humans perceive tactile sensory signals from the external environment through specific receptors, which encode the signals into spikes and transmit them to the cerebral cortex, enabling perception and learning. Similarly, the spiking-based OECN described above can be integrated with various sensors to facilitate tactile sensing. The external tactile receptors are simulated by using pressure and strain sensors, which are then connected in series with the OECN (configured with 100 nF capacitors) and a DC power supply (Fig. 4A). A pressure sensor fabricated with conductive carbon nanotube (CNT) foam is integrated with the OECN to realize pressure signal sensing (fabrication details in Materials and Methods). When external pressure is applied, the conductive foam deforms varying the electrical resistance. An applied pressure ranging from 0 to 5 kPa alters the sensor resistance from 286.7 kΩ to 117.7 kΩ (SI Appendix, Fig. S25). This change in resistance regulates the I_IN of the OECN, effectively mapping the sensory signal to the neuron's spiking frequency through f-I modulation. Specifically, the real-time spiking-based response of the OECN is recorded based on the external pressure applied to the pressure sensor (Fig. 4B and Movie S1). As the pressure applied by the finger on the conductive foam increases, the output spiking frequency of the OECN also increases. Fig. 4C summarizes the relationship between spiking frequency and applied pressure, illustrating a monotonic increase in spiking frequency from 3.2 Hz to 6.7 Hz as the pressure rises from 0 kPa to 5 kPa. Additionally, an artificial perception unit for strain signals is realized by integrating the OECN with a strain sensor printed on a rubber glove (Fig. 4D). The resistance of the strain sensor increases along with the finger curvature (SI Appendix, Fig. S26), which in turn lowers I_IN and the neuron spiking frequency. As shown in Fig. 4E and Movie S2, It is observed that the spiking frequency of the sensory neuron decreases with increasing finger bending. This OECN-based sensing system enables simultaneous sensing and spike conversion in a straightforward manner, benefiting from the compact footprint and efficient circuit design of the OECN, which offer advantages over earlier pioneering methods (31, 54). This enables achieving different perception modalities by integrating various sensors, highlighting the attraction of the present concept.

Fig. 4.

OECN for pressure and strain sensing. (A) Schematic of an OECN applied to the perception of tactile signals. The conductive foam-based pressure sensor or the printed strain sensor is combined with the OECN and provides the input current signal I_IN. (B) Different pressure signals are applied by pressing on the pressure sensor, and the corresponding spiking responses of the OECN. (C) The effect of pressure on spiking frequency of the OECN. (D) Different strain signals are provided by bending the finger with the printed strain sensor, and the resulting spiking responses of the OECN. (E) The effect of strain under different bending angles on spiking frequency of the OECN. (F) Schematic of the spike-based neuromorphic perception computing system. Each pixel value in a handwritten digital image is treated as pressure. The pressure image is translated into spikes by 784 artificial neurons and further processed by a three-layer spikes neural network (SNN). (G) Evolution of the test accuracy during training process. (H) Confusion matrices of training results. The corresponding image was recognized successfully when a large value appeared along the diagonal line.

To illustrate the potential of the present OECN in the fabrication of large-scale SNN, simulations were performed based on the experimental data. Fig. 4F shows a diagram of a spike-based neuromorphic sensory computing system for MNIST-based pressure image classification. A fully connected artificial neural network (784 × 196 × 10) is used for training, with additional details provided in SI Appendix, Supplementary Text. The SNN is trained online using backpropagation, incorporating the experimentally measured electrical characteristics of the devices. This establishes a relationship between the OECN spiking frequency and applied pressure. Fig. 4G illustrates the evolution of test accuracy during training, reaching 96.16% after 10,000 epochs. The average spike counts of the output layer neurons posttraining are shown in SI Appendix, Fig. S27, where the columns correspond to the input image labels, rows represent the 10 output neurons, and color bars indicate the average spike count. After training, most pressure images are correctly identified. The confusion matrix in Fig. 4H shows that the majority of pressure distribution images are classified correctly, with columns representing the actual categories, rows representing the classification results, and color bars indicating the number of instances. These results demonstrate the OECN circuit's ability to convert tactile stimuli into spikes and perform complex tasks effectively.

A complete neuromorphic tactile perception system can encode mechanical stimuli into action potential spikes and convert them into postsynaptic responses for further communication between neurons in a neural network (SI Appendix, Fig. S28). Consequently, the present organic artificial synapse was developed and integrated with the OECN sensing setup to convert the electrical signals from the presynaptic neuron into postsynaptic currents (PSC) (SI Appendix, Fig. S29). Here, the organic artificial synapse was fabricated with a PEDOT:PSS channel layer and an ion gel electrolyte in a vOECT configuration. Due to the ion permeation in the channel layer of OECTs, the high volumetric capacitance allows the synaptic transistor to operate at relatively low voltages, making it well-suited for mimicking biological synaptic functions (21, 55). The PEDOT:PSS-based synaptic vOECT can simulate neurotransmitter release (in the form of ions) from the presynaptic membrane (gate surface), which in turn leads to the excitation or inhibition of the postsynaptic membrane (channel) in response to electrical signals (Fig. 5A). The injection or extraction of ions from the PEDOT:PSS channel induces doping/de-doping in the channel, enabling modulation of the ΔPSC. The short-term plasticity (STP) of the synaptic vOECT was initially investigated. Paired-pulse facilitation (PPF), as a crucial form of STP, plays an important role in decoding temporal information (56). PPF is elicited by two consecutive excitatory pulses (Inset of Fig. 5B). Following the application of presynaptic pulses with varying voltage pulse intervals (Δt), the relationship between the PPF indices with Δt is summarized in Fig. 5B. Further fitting analysis of the PPF is provided in SI Appendix, Supplementary Text. Additionally, we explored spiking-rate-dependent plasticity of artificial synapses related to the filtering behaviors in biological systems. As illustrated in SI Appendix, Fig. S30, the ΔPSC strongly depends on the pulse frequency (0.5 to 49 Hz). The PSC gain, defined as A₁₀/A₁ is used as the performance indicator, where A₁₀ and A₁ represent the ΔPSC of the 10th peak and the 1st peak, respectively. According to Fig. 5C, the PSC gain increases from 1.0 to 2.8 as the frequency of spikes is varied from 0.5 to 49 Hz. This response is akin to biological high-pass filtering behavior (57). To assess the capability of integrating the synaptic transistor with the OECN sensing setup, simulated presynaptic pulses triggered by pressure stimuli with varying extent and duration were applied (Fig. 5D). The memory level, defined as the steady-state change in channel current before and after the application of a voltage pulse, along with the ΔPSC peak value, is summarized in Fig. 5 E and F. The postsynaptic responses are influenced not only by the spiking frequency (corresponding to the pressure stimulus extent) but also by the duration of the pressure stimulus. Importantly, although the duration of the pressure stimulus does not alter the voltage output frequency from the OECN, its effect on the postsynaptic response is detectable. A longer duration time increases the voltage pulses applied to the postsynaptic transistor, thus a higher ΔPSC peak value and memory level, are observed. This phenomenon is analogous to the function of neural synapses, where multiple action potential spikes in the presynaptic neuron trigger the release of a large number of neurotransmitters, thereby increasing the transmission efficiency between neurons. This is known as the spiking number-dependent plasticity of artificial synapses. Finally, the voltage output from the OECN pressure sensing setup is connected to the gate of the synaptic transistor to simulate the tactile perception process in biological system (Fig. 5G). The ΔPSC peak increases from 2.22 mA to 3.17 mA, and the memory level increases from 52.18 ± 2.36 μA to 124.09 ± 13.58 μA as the applied pressure increases from 0 kPa to 5.1 kPa. Constructed in this manner, a complete artificial perception system is able to process tactile stimuli and demonstrates significant potential for direct interfacing with biological systems.

Fig. 5.

Organic neuromorphic perception system. (A) Schematic representation of the synaptic vOECT in analogy to a biological synapse. (B) Pair-pulsed-facilitation (PPF) behavior of the synaptic transistor as a function of voltage pulse interval (Δt). The inset is the PSC as a function of time with a pair of presynaptic pulses (pulse voltage V_P = 0.7 V, pulse duration t_P = 100 ms) applied at the gate electrode. (C) Spiking-rate-dependent gains as a function of increasing spiking frequency. Inset: typical PSC triggered by 10 consecutive spike trains at 10 Hz. V_P = 0.7 V, t_P = 20 ms. (D) PSC response under simulated pressure-triggered spikes at different frequencies (1 Hz, 2 Hz, 3 Hz, 5 Hz, and 10 Hz, fixed duration: 4s) and durations (1 s, 2 s, 4 s, 8 s, and 10 s, fixed frequency: 5 Hz). The peak value and memory level of PSC variation as a function of (E) the frequency and (F) duration of the presynaptic spikes. (G) Schematic of an artificial perception system for pressure signal processing. (H) The postsynaptic response depending on different pressure stimulus applications (0 kPa, 2.8 kPa, and 5.1 kPa) for a fixed duration (5 s).

Discussion

Inspired by the perception process of biological systems, we have developed an OECN for constructing a neuromorphic perception device capable of encoding tactile stimulations into frequency-dependent spikes and subsequently converting them into postsynaptic responses. The OECN fabricated with high-performance vOECT complementary circuits is demonstrated to have a reduced footprint and wider calibratable spiking range than previously reported for artificial neurons based on organic semiconductors, as well as having far fewer elements in comparison to traditional Si-based circuits. The design and implementation of an advanced n-type OECT is essential for achieving the circuit performance. The OECN output spiking frequency range can be modulated by simply changing the input current and capacitance. This feature offers facile integration with different types of sensors, mimicking biological receptors and artificial synapse transmission of signals to neural networks. We demonstrated this possibility experimentally by integrating the OECN with mechanical sensors and synaptic transistors to yield an artificial perception system which can respond to tactile stimulations with specific spiking signals and generate postsynaptic responses, thereby achieving both sensing and communication capabilities, thus mimicking the functionality of biological neurons. We believe that the results on the present OECN-based artificial perception system can open possibilities for a wide range of applications in next-generation intelligent soft robotics, cyborg design, and brain–machine interfaces. The potential for further exploration and utilization of this technology is significant, paving the way for exciting advances in many fields.

Materials and Methods

Organic Artificial Perception System Fabrication.

OECN fabrication.

The fabrication process for the OECN, as illustrated in SI Appendix, Fig. S19, was carried out on a precleaned Si/300 nm SiO₂ wafer. Initially, a 3 nm Cr layer and a 150 nm Au layer were thermally evaporated through a shadow mask to create the bottom electrodes (V_DD, I_IN). A p-type polymer blend was spin-coated at 3,000×g rpm for 20 s. Selective patterning for the p-type channels in amplifier block was achieved by UV crosslinking for 60 s with a shadow mask, followed by immersion in chloroform for 3 s and subsequent blow drying. A 150 nm Au layer was then thermally deposited through a shadow mask to define the middle electrode (V_OUT). Subsequently, an n-type polymer mixture was spin-coated and photopatterned to form the n-type channels in the amplifier block and T_reset regions, following the same protocol as the p-type material. The top electrode (GND) was fabricated by thermally evaporating another 150 nm Au layer through a shadow mask. To expose specific regions, including the active channel areas, gate electrodes, and capacitor connection points, pure Cin-Cell solution was spin-coated at 5,000×g rpm for 20 s, UV crosslinked for 60 s, and developed in chloroform for 3 s. Gate electrodes for the amplifier block and T_reset were coated with Ag/AgCl paste (ALS, 011464) and vacuum-dried for 30 min. Capacitors were attached using conductive silver paint (HUMISEAL, 948-06G). Finally, a drop of 1× PBS electrolyte was applied to the gate electrodes and their adjacent active channel regions to complete the fabrication process.

Organic artificial synapse fabrication.

The vOECT-based artificial synapse was fabricated on a precleaned Si/300 nm SiO₂ wafer. A 3 nm Cr layer and a 150 nm Au layer were thermally evaporated through a shadow mask to form the source and gate electrodes. The channel layer was created by spin-coating a PEDOT:PSS solution at 1,000 rpm for 60 s, followed by annealing at 150 °C for 2 h in an inert environment. A 150 nm Au drain electrode was then deposited by thermal evaporation through a shadow mask. Finally, an ion-gel solution was spin-coated and UV-crosslinked for 60 s to complete the device.

Organic neuromorphic perception system fabrication.

The pressure sensor was fabricated based on a conductive sponge. A polyurethane sponge (10 mm × 10 mm × 20 mm) was washed by ultrasonication with acetone, ethanol, and deionized water for 10 min each and dried. The sponge was dipped in a CNT aqueous ink (5 wt%) for 60 min and then vacuum dried at 60 °C for 60 min to obtain a conductive sponge. Next, two conductive copper wires were connected on either side of the conductive sponge, one conductive wire was connected to a DC power source, and the other one was connected to the I_IN of the organic artificial neural circuit mentioned above. The strain sensor was fabricated on a nitrile rubber glove. Specifically, a CNT aqueous ink was brushed onto the surface of the glove’s finger part and then dried under vacuum. This procedure was repeated three times. After that, two conductive copper wires were connected to the top and bottom sides of the finger. Similarly, one conductive wire was connected to a DC source, while the other one was connected to the I_IN of the organic artificial neural circuit. The organic artificial perception system for processing pressure stimuli was achieved by integrating the above-mentioned setup for pressure sensing with the organic artificial synapse. The V_OUT in the organic artificial neural circuit was connected with the side gate electrode of the organic artificial synapse.

Device Characterization.

Transistor measurements.

Electrical characterization of the vOECTs and complementary circuits was carried out with a Keithley 4200A-SCS Parameter Analyzer in ambient. The voltage sweeping speed was 0.1 V/s for the vOECT measurements. For the inverter cycling tests, the voltage pulse was generated by a Keysight waveform generator (AFG3252C), while the voltage variation was monitored with a Keithley 4200A-SCS Parameter Analyzer. For the characterization of the inverter frequency, a constant V_DD of +0.7 V was applied with a Keithley 4200A-SCS Parameter Analyzer, and the V_OUT was monitored by an oscilloscope (Tektronix, MDO3014). For characterization of the push–pull voltage amplifier, sinusoidal signals with a frequency from 0.1 Hz to 1,000 Hz at amplitude = 5 mV were applied as V_IN, by a Keysight waveform generator (AFG3252C) and a homemade voltage follower circuit, and the V_OUT was monitored by an oscilloscope (Tektronix, MDO3014). Note, the decibel (dB) of voltage amplifier is defined as

$$dB=20log\left|\frac{\partial V_{\mathrm{OUT}}}{\partial V_{\mathrm{IN}}}\right|.$$ [1]

For the organic artificial neural circuit characterization, constant input current (I_IN) signals from 100 nA to 100 μA were utilized while a constant V_DD of + 0.7 V was applied with a Keithley 4200A-SCS Parameter Analyzer and the V_OUT was monitored by an oscilloscope (Tektronix, MDO3014). All measurements were carried out in ambient air. For the organic artificial synapse characterization, the gate voltage pulse was generated by a Keysight waveform generator (AFG3252C), while the PSC was monitored with a Keithley 4200A-SCS Parameter Analyzer.

Sensing measurements.

For characterization of the organic artificial neural circuit for pressure sensing, the DC power source connected with the pressure sensor was set at 0.1 V, then an increasing pressure was applied on the conductive sponge part of the pressure sensor by a human finger and the V_OUT was monitored by an oscilloscope (Tektronix, MDO3014). Meanwhile, the resistance variation of the conductive sponge during the pressing process was recorded by a Keithley benchtop digital multimeter (DMM7510). For the characterization of the organic artificial neural circuit for strain sensing, the DC power source set at 0.1 V was connected with the strain sensor attached to the finger, and the V_OUT was monitored by an oscilloscope (Tektronix, MDO3014) during finger bending. Meanwhile, the resistance variation of the sensor under different curvatures was recorded by a Keithley benchtop digital multimeter (DMM7510). For the characterization of the organic artificial perception system for processing pressure stimuli, the V_OUT in the above pressure sensing setup was directly connected with the side gate of the organic artificial synapse made of PEDOT:PSS-based vOECT, the PSC was monitored with the Keithley 4200A-SCS Parameter Analyzer.

Supplementary Material

Appendix 01 (PDF)

Movie S1.

A OECN for pressure signal sensing.

Movie S2.

A OECN for strain signal sensing.

Data, Materials, and Software Availability

All study data are included in the article and/or supporting information.