Confinement of ions within graphene oxide membranes enables neuromorphic artificial gustation

Significance

The development of fluidic neuromorphic components and perceiving systems for emulating the function of biological nervous systems is imperative and challenging. This study introduces a graphene oxide ionic memristive device capable of performing both sensory and computing functions. Experimental and theoretical simulations indicate that the prolonged retention of ions within graphene oxide channels accounts for the memristive characteristics. Importantly, this device can be used to construct a reservoir-computing gustatory system, which can effectively perceive sweet, salty, bitter, and sour flavors. Our system-level implementation of artificial gustation—although it is a proof-of-concept demonstration—should represent a substantial advancement toward future intelligent perceptions and/or taste reconstruction.

Abstract

Introducing neuromorphic computing paradigms into taste-sensing technology will bring unprecedented opportunities for developing new hardware architectures with perceptual intelligence. Constructing the biomimetic gustatory system, however, remains a challenge due to the scarcity of suitable components operating under wet conditions. Here, we report that ion confinement within the layered graphene oxide membranes can be used to develop a memristive device capable of implementing both synaptic function and chemical sensing. The continuum model and ion dynamics characterizations demonstrate that interfacial adsorption–desorption slows down ion transport and leads to memristive behavior. Based on this nanofluidic device, we built an artificial gustatory system in the physiological environment, which can efficiently classify different flavors according to the reservoir computing algorithm. Our results suggest a paradigm for in-sensor computing in liquid.

Separating sensory elements and computing units usually leads to time latency, low processing efficiency, and high power consumption; thereby, considerable efforts have recently been devoted to integrating both functions to achieve neuromorphic perceptual intelligence (1, 2). Due to the rapid development of artificial neural algorithms and neuromorphic devices, a large number of integrated systems have been created to mimic biological visual (3, 4), auditory (5), tactual (6), olfactory (7), and gustatory (8–11) systems, displaying wide-range applications from machine vision (4) to motion recognition (12) and health monitoring (13). Most of these hardware architectures are based on solid-state building blocks, which deal with physical signals via electrons and cannot respond directly to chemical stimuli in aqueous environments (14, 15). The lack of suitable fluidic neuromorphic components would impact the design of fluid-based intelligent systems, for example, the bioinspired gustatory system, which needs to operate under wet conditions. To address this problem, current artificial gustatory architectures usually employ the ionic component to sense while the electronic counterpart to process. Although they succeeded in emulating taste perception for salt concentration and different flavors (8, 9), significant disparities in materials and structures of separated modules prevent their further integration. In addition, the biological gustatory system relies purely on the dynamics of ionic charges under physiological conditions, and therefore, to fully “replicate” the functions of the human taste, it is indispensable to develop neuromorphic components and systems by manipulating the ions in the fluidic circumstances (16–18).

With delicate control of interfacial interactions and/or ion concentration distributions within confined nanofluidic space (19), diverse ionic devices such as sensors (20, 21), rectifiers/diodes (22, 23), transistors (24), and their assembled circuits have been developed (23, 25). Nonetheless, these ionic components face great challenges in performing neuromorphic computing. Until very recently, several studies have observed the resistive-switching memristive phenomenon in confined nanofluidic channels (26–28). For example, the analytical theory predicts that the ionic charges will assemble into elongated clusters in the quasi–two–dimensional slits, which leads to nonlinear ion transport and hysteretic conduction (26). In the active carbon (sub)nanofluidic channels, the interfacial processes, such as ionic self-assembly or surface adsorption, slow down the transport of ions, which allows the author to develop the nanofluidic synapse for implementing Hebbian learning (27). Grafting polyimidazolium brushes on the glass micro- or nanopipette produces highly charged confined spaces, in which the interaction between polyelectrolyte and ion also results in the current hysteresis and memristive response (28). However, these ionic memristive devices can only emulate basic synaptic behaviors and implement simple neuronal functions, there is still a lack of system-level consideration to mimic the full perceptual functionalities of biological ones. Although artificial gustatory systems have been demonstrated by assembling both electronic and ionic components (8, 9), integrating both sensory and computing functions in the pure aqueous environment, especially on one nanofluidic device, remains a significant challenge.

Results

Design and Electrical Characteristics of the GO-ISMD.

Here, we report a bioinspired neuromorphic artificial gustatory system by using reconstructed layered graphene oxide (GO) membranes in the physiological environment (Fig. 1). The GO nanofluidic device can not only function as a flavor sensor but also act as a memristor to classify the sour, bitter, salty, and sweet flavors according to the reservoir computing algorithm. To fabricate this GO ionic sensory memristive device (GO-ISMD), the GO membrane is first prepared by vacuum filtration. This process facilitates a layer-by-layer assembly of GO sheets and the formation of laminated structures (see the characterization of both GO sheets and GO membrane in SI Appendix, Figs. S1–S7) (29). The GO membrane is then placed onto the silicon substrate with a 100 nm-thick floating SiN_x window, distributing 36 perforated holes (ca. 3 μm, see Fig. 2A and SI Appendix, Fig. S8, and Materials and Methods for the detailed procedures). For consistency, GO-ISMDs in this report are fabricated using GO membranes of negatively charged GO sheets unless otherwise noted. The electrical characteristics of the fully hydrated GO-ISMD are subsequently monitored by using the DAQ card (NI PCIe-6361) and Low Noise Current Preamplifier (SR570) in the saline solution (SI Appendix, Fig. S9 and see Materials and Methods for more details).

Fig. 1.

Schemes of the biological and GO-ISMD-based gustatory systems. In the biological gustatory system, chemical stimuli first trigger the electrical responses through various taste receptor cells distributed on the tongue. These electrical potentials are subsequently encoded and transmitted via nerves to the cerebral cortical neurons for processing and taste recognition. In the GO-ISMD-based system, the sensing signals generated by feeding different saline flavors are first encoded and then imported to a dynamic GO-ISMD-based reservoir computing system for processing and perceiving.

Fig. 2.

Electrical characteristics of the GO-ISMD. (A) Schematic illustration of the device configuration. The size of the GO membrane is 4 mm × 4 mm. The scale bar in the SEM image is 500 nm. (B) Current transient (red trace, Right y-axis) of the device under square voltage (black trace, Left y-axis). Inset: magnified current spikes. (C) Typical current-voltage characteristics of the device measured in 10 mM KCl (f = 4 mHz). Inset: the equivalent circuit and the current-voltage characteristics of the device with a positively charged GO membrane. Arrows in the graphs indicate the sweeping directions. (D) Dependence of the normalized area of the positive half (see Inset) on the triangular voltage pulse frequency. The black curve is the fitting line according to Eq. 7. The thickness of the GO membrane is 850 nm. (E) The normalized conductance transients after potentiated by 40 square pulses (amplitude/width: 1 V/1 s) for different GO membrane thicknesses. The black lines are based on single-exponential fitting. Inset: memory time of the device with different membrane thicknesses. The black line is a quadratic fitting. The error bars are based on three independent devices. (F) The conductance (black curve, Left y-axis, measured under 0.15 V) and the normalized current-voltage loop area (red curve, Right y-axis, f = 10 mHz) as a function of saline concentration. (G) Normalized area versus voltage pulse frequency at 0.1, 1, and 10 mM KCl solution. The solid lines are the fittings according to Eq. 7. The thickness of the GO membrane is 550 nm. The error bars are based on five independent devices. (H) Histograms of the memory time for the device with different cations. The thickness of the GO membrane is 550 nm, and the saline concentration is 10 mM. This memory time is calculated based on the fittings of area-frequency curves. The error bars are based on three independent devices.

The time-dependent current is first measured by applying a square voltage (Fig. 2B). Apparent current rectification is observed, which is due to the asymmetric voltage-driven ionic charge flux through GO membranes (SI Appendix, Fig. S10). No apparent water flow through GO-ISMD is found under pressure, which indicates that the flow rate of ions is probably not promoted through the GO-ISMD at least under 0.8 bar (SI Appendix, Fig. S11) (22, 30). The emergence of current spikes ($\mathrm{\Delta }t<\; <0.1\mathrm{s}$) at the edges of voltage transitions originates from the charging processes occurring at the solid–liquid interfaces, showing non-negligible parasitic capacitance within the device (31). Importantly, the current has increased (decreased) monotonously under high (low) constant potential. This feature, similar to the potentiation and depression behaviors of the synapse, indicates the memristive characteristics of the device (32). This memristive phenomenon is also confirmed by recording the current-voltage characteristics in which the hysteretic window appears (Fig. 2C), and a cross-point emerges at a positive nonzero bias (32). Replacing the KCl electrolyte with CaCl₂ saline solution, similar memristive behavior is observed (SI Appendix, Fig. S12). In addition, in the device with positively charged GO (covalently connecting quaternary ammonium cation on the GO) membrane (33), a reversed current rectification but the same memristive effect is demonstrated (Fig. 2 C, Inset). The flipped current rectification indicates the surface charge-controlled ion transport with asymmetric gating (the surface-charge effect is evidenced by elemental mappings as shown in SI Appendix, Figs. S13 and S14 and Table S1) (34). The cross-points appearing at the nonzero voltages (ca. 0.08 V and -0.08 V for the negatively and positively charged GOs) are probably induced by the capacitive effect (see the Inset of Fig. 2C for the equivalent circuit) (35, 36).

To understand the underlying physics of this memristive phenomenon, comprehensive experiments are subsequently performed. Since the memristive device dominates the current at the positive bias (for the negatively charged GOs), the current-voltage loop area at this half versus triangular voltage pulse frequency is first studied. The normalized area exhibits a maximum value at a frequency of ~7 mHz (Fig. 2D), corresponding to a memory time of ~140 s for the 850 nm-thick GO membrane device. This timescale is approximately 10⁵ times larger than that expected for ion transport through a similar length with a diffusion coefficient in an aqueous solution (ca. 0.001 s), which indicates that solvated ions in our device transport with an extremely low migration coefficient. In addition, the nonmonotonic dependence on the frequency can be well-fitted by Eq. 7 (see Materials and Methods), which is derived from the adsorption–desorption model (27, 37). This suggests that similar surface-charge-governed ion transport also exists in our device, and interfacial adsorption–desorption might be the key factor.

Next, the membrane thickness-dependent current transient is monitored after the removal of a square pulse string (Fig. 2E and SI Appendix, Fig. S15). The normalized conductances decay exponentially according to the equation $\mathrm{G}=\mathrm{a}\ast \mathrm{e}\mathrm{x}\mathrm{p}(-t/\tau )$. The time constant τ of different membrane thicknesses extends from tens to hundreds of seconds which is consistent with the value calculated from Fig. 2D. Moreover, the memory time follows a quadratic dependence on the membrane thickness (see the Inset of Fig. 2E), which also conforms to the adsorption–desorption model (see Materials and Methods, Eq. 6).

In another experiment, we find that the conductance of our memristive device has increased with the saline concentration, whereas the normalized current-voltage loop area shows a reversed dependence (Fig. 2F and SI Appendix, Fig. S16). Increasing electrolyte concentration can provide more ions participating in conduction, but attenuates and/or conceals the contribution of memristive ions. Consequently, the cross-point on the current-voltage loop has disappeared at high concentrations (e.g., at 1 M, SI Appendix, Fig. S16G). Importantly, we also find that the peak frequencies (~20 mHz, cf. memory time, ~50 s for the 550 nm GO membrane) of the normalized current-voltage loop area have not changed at the concentration range from 0.1 to 10 mM (Fig. 2G). This indicates that our memristive device is robust under a wide range of physiological conditions.

To gain further insight into the interfacial adsorption–desorption process, the memory time of the device with different cations is studied (Fig. 2H and SI Appendix, Fig. S17). Cations with the same valence have recorded similar memory timescales. Divalent cations record significantly longer memory time compared to monovalent ones. This difference should be attributed to the interaction between the cations and the surface charges within GO membranes. Normally, the divalent cations display stronger Coulomb interactions (20, 38), leading to an extension of the memory time. This is consistent with the adsorption–desorption model and can be further verified by monitoring the current responses to the constant voltage (SI Appendix, Fig. S18). Interestingly, this finding also suggests the potential use of our memristive device to detect different chemicals.

Synaptic Functions of the GO-ISMDs.

The memristive characteristics of our device indicate that the emulation of synaptic functions is possible. First, by applying a string of square pulses with increasing amplitude, the synaptic weight, defined by ${(I}_n-I_1)/I_1$, increases and tends to saturate (Fig. 3A), demonstrating an effective potentiation behavior. Remarkably, this potentiation process can be observed at a small pulse amplitude of 0.2 V, which indicates a low threshold to switch on the device and also a possibility of low energy consumption for neuromorphic computing systems (39). We also find that our memristor is volatile. Potentiation of the device with different numbers of positive pulses (from 1 to 40, amplitude/width: 1 V/1 s), all of them decay exponentially and return to the initial state (Fig. 3B). As the increase of pulse number, the retention time has also increased (see the Inset of Fig. 3B). Then, we emulate the paired-pulse facilitation (PPF) to evaluate the capability of our device in processing spatiotemporally correlated signals. The change of synaptic weight [cf. PPF index, defined by ${(I}_2-I_1)/I_1$], where the I₁ and I₂ are the currents under two consecutive voltage pulses) has exponentially decayed over the time interval of the paired pulses (Fig. 3C).

Fig. 3.

Synaptic functions of the GO-ISMD. (A) The synaptic weight (${(I}_n-I_1)/I_1$) as a function of square pulse numbers with the pulse amplitude increased from 0.1 V to 1 V. Inset: dependence of synaptic weight on the pulse amplitude. (B) The current decays after potentiated by 1, 5, 10, 20, and 40 square pulses (amplitude/width: 1 V/1 s). The currents are measured by applying square voltage pulses (amplitude/width: 0.15 V/1 s). The solid lines are the single exponential fittings, where the parameter τ denotes the decay rate. Inset: dependence of τ on the pulse number. (C) Evolution of the PPF index with the time interval between paired voltage pulses (amplitude/width: 1 V/1 s). The black line is a single exponential fitting. The error bars are the SD of ten independent measurements. Inset: scheme of paired pulses. (D) Potentiation and depression under 50 positive (amplitude/width: 1 V/1 s) and 50 negative (amplitude/width: -1 V/1 s) pulses. The conductance is read by a pulse of amplitude 0.15 V (width: 1 s) with a time interval of 3 s after the write or erase pulse. Inset: schemes of pulses applied. The typical current transients at the intervals between the operating and reading pulses are shown in SI Appendix, Fig. S19. The impact of the reading voltage on the conductance state of a GO-ISMD is shown in SI Appendix, Fig. S20. (E) Endurance test of a GO-ISMD (GO membrane thickness 1,700 nm) by placing continuous potentiation (red markers, Left y-axis) and depression (blue markers, Right y-axis) pulse sequences for 89 cycles. 30 square pulses (amplitude/width: ±1 V/1 s) are applied for each potentiation/depression cycle. All data are acquired by using 850 nm-thick GO membrane devices at 10 mM KCl solutions unless otherwise noted. I_n is the current measured under the N_th voltage pulse (in graph A) or the N_th reading pulse (in graph D).

The primary function of a synapse is to reversibly modulate the connecting strength between neighboring neurons. In our device, such modulation is achieved by applying a string of positive pulse followed by a string of negative one (Fig. 3D). The weight of our synaptic device, defined by ${(I}_n-I_1)/I_1$, has increased to ~20% after the potentiation by 50 positive pulses (Note: The GO membrane thickness is 850 nm). Under a successive string of 50 negative pulses, it quickly returns to the initial state. In fact, due to the volatile feature, our device can go back to the original conduction state at a slower pace without any external electrical stimuli. Apparently, this characteristic is useful for reducing both time and power consumption. In addition, the potentiation and depression process can also be demonstrated in the device with a much thicker membrane (1,700 nm). Importantly, the alternate switching can be reversibly modulated for up to 89 cycles (Note: We only record 89 cycles and the device is still working), where no apparent performance degradation is observed (Fig. 3E).

Theoretical Simulations.

Based on our experimental observations, a notable difference between our GO-ISMDs and other ionic diodes is their significantly low ion migration rate within the confined GO spaces (Fig. 2, ~10⁻¹² m²/s vs 10⁻⁹ m²/s) (27, 40). The dynamic adsorption–desorption interfacial interactions slow down the rearrangement of ionic charges upon physical stimuli, leading to an extremely prolonged retention time of the conductance states (Fig. 4A), which are established through ion depletion or accumulation within GO membranes under voltage biases (16, 41). This feature is essential for the manifestation of the memristive phenomenon in an ionic device with substantial parasitic capacitance (42). Otherwise, the synaptic behaviors would be dominated by the capacitance effects of the device, where no dynamic resistive switching can be observed (35, 43, 44).

Fig. 4.

Theoretically predicted electrical characteristics of the GO-ISMD. (A) Scheme of the ion dynamics within GO layers. (B) Simulated currents as a function of time under the constant voltages of 0.6 V (Left y-axis) and −0.6 V (Right y-axis) with the ion diffusion coefficient decreased from $3.92\times {10}^{-12}{\mathrm{m}}^2/\mathrm{s}$ (black curves) to $0.98\times {10}^{-12}{\mathrm{m}}^2/\mathrm{s}$ (red curves). (C) Simulated current-voltage characteristics (f = 50 Hz). The Inset shows the scheme of the accumulation (positive I−V half) and depletion (negative I−V half) process. (D) Frequency dependence of the normalized area of the positive half of the simulated current-voltage loops. The black curve is the fitted result based on Eq. 7. (E) Normalized current lapses under 0.15 V for different numbers of GO layers. The solid lines are based on single exponential fitting. Inset: memory time as a function of the number of GO layers. The black line is a quadratic fitting. Note: The memory time obtained from the simulation is about 10 ms, which is consistent with the experimental value of a GO-ISMD with comparable GO layers. (F) The dependence of synaptic weight (${(I}_n-I_1)/I_1$) on the pulse amplitude (red markers, Bottom x-axis and Left y-axis) and the dependence of τ on the pulse number (blue markers, Top x-axis and Right y-axis). (G) Simulated PPF index vs time intervals between two consecutive pulses (amplitude/width: 0.5 V/0.5 ms). The black line is a single exponential fitting. (H) Simulated endurance performance of the device by placing continuous potentiation (red markers, Left y-axis) and depression (blue markers, Right y-axis) pulse sequences for 100 cycles. The thickness of the device: 5 GO layers. The amplitude/width of the squared voltages: ±0.5 V/0.5 ms.

Accordingly, the “slow” ion dynamics is subsequently simulated by using a continuum charge-transport model combining Nernst–Planck diffusion and Poisson’s equations (42, 45). This model (SI Appendix, Fig. S21 and see Materials and Methods for details) transforms the adsorption–desorption interfacial interaction into the compromised ion diffusion coefficient (1.96 × 10⁻¹² m²/s, ~1/1,000 of that in liquid) (27). Upon placing a constant positive/negative voltage, both the simulated current (Fig. 4B) and ion concentration (SI Appendix, Fig. S22, recording at point A in the model, see SI Appendix, Fig. S21) have increased/decreased. At the steady state, distinct ion distributions are finally established within the confined spaces (SI Appendix, Fig. S23). The time scale extends to ms, which is thousands of times longer than that in liquid. In addition, decreasing the diffusion coefficient simulates a longer retention time (memory time), which is consistent with a stronger Coulomb interaction of divalent cations (Figs. 4B vs. 2H and SI Appendix, Figs. S17 and S18). Importantly, the model qualitatively reproduces the typical current-voltage characteristics of our GO-ISMD (Figs. 4C vs. 2C). The asymmetric gating effect is also demonstrated (rectifying phenomenon), and the rectification ratio exhibits a positive correlation with surface charge density (SI Appendix, Fig. S24). With the positively charged GOs, a flipped current rectification is simulated (Fig. 2 C, Inset vs. SI Appendix, Fig. S25). The normalized current-voltage loop area (accumulation half) can be well-fitted by Eq. 7 (Figs. 4D vs. 2D). The quadratic dependence of memory time on the number of GO layers is also consistent with Eq. 6 (Figs. 4E vs. 2E). Moreover, as the increase of electrolyte concentration, a significant decrease in the current-voltage loop area is simulated, which is attributed to the reduction of the relative contribution of memristive ions (SI Appendix, Fig. S26). Finally, the model also qualitatively reproduces all of the experimentally recorded synaptic plasticity: i) the pulse amplitude-dependent synaptic weight (Fig. 4F and SI Appendix, Fig. S27 vs. Fig. 3A), ii) the dependence of decay time on the pulse number (Figs. 4F and SI Appendix, Fig. S28 vs. Fig. 3B), iii) the paired-pulse facilitation (Figs. 4G vs. 3C), and iv) the continuous potentiation and depression (Figs. 4H vs. 3E).

Classification of Four Different Flavors.

Inspired by the distinct responses to different cations (Fig. 2H and SI Appendix, Figs. S17, S29, and S30), next, we try to combine both the sensory and computing functions in our memristive device to develop a neuromorphic artificial gustatory system. This system is composed of three distinct components: the sensing input, the reservoir layer, and a single-layer fully connected neural network (Fig. 1). This combination is learned from the biological gustatory system in which the chemical signals are intercepted, encoded, transmitted, and finally processed to perceive different flavors (Fig. 1). As a concept demonstration, four different chemicals, CH₃COOH (sour), MgSO₄ (bitter), NaCl (salty), and Pb(CH₃COO)₂ (sweet), representing four different flavors are selected [Note: Pb(CH3COO)₂ is a typical ionic compound of a sweet taste. Attention should be paid to the lead exposure, which could result in internal organ damage, neurotoxicity, or hematologic toxicity] (46, 47). The current transients upon different tastants are recorded and demonstrate unique characteristics (SI Appendix, Figs. S31 and S32, and Materials and Methods for more details). These current transients can be classified with an accuracy of about 98.5% by using a modified artificial neural network (SI Appendix, Fig. S33) (48). A homemade dataset is constructed from these sensing curves, serving as the sensing part of the artificial gustatory system (see Materials and Methods for more details). The reservoir layer, consisting of 4 GO-ISMDs, is used to abstract the features from our dataset. The volatile characteristics of our device indicate that the historical write operation will affect the current write performance (see PPF). Thereby, the fading characteristics of GO-ISMDs can be used to mimic the neuronal nodes with recursive connections. In the framework of reservoir computing (49, 50), such a volatile device can be utilized to form a computing node. Subsequently, each sample of the dataset is flattened into four vectors, each with 192 elements, after being augmented through a mask matrix of size 8 × 24, which is randomly generated, containing only -1 or 1 as elements (50, 51). It should be noted that once generated, the mask matrix remains consistent across all samples within the dataset. Following this step, four voltage pulse sequences are created based on these vectors. These sequences maintain a pulse width and interval of 1 s, with the pulse amplitude determined by the corresponding element values within the vectors. Thereafter, the voltage pulse sequences will be applied to GO-ISMDs in the reservoir layer individually. Currents will be measured at these voltage pulses, yielding 4 × 192 reservoir states (Fig. 5A and SI Appendix, Fig. S34). These reservoir states are then classified by the single-layer fully connected neural network to perceive the final class of the unknown taste.

Fig. 5.

Gustation classification based on GO-ISMDs. (A) Part of the state-transition of computing nodes in the reservoir. The current can be thought of as a descriptor for the state of the computing node. Each line represents an independent node. (B) The binary classification of the CH₃COOH (Top) and MgSO₄ (Bottom) on the training (Left) and test (Right) datasets. The gradient-colored lines and the dashed black lines are the target and predict results, respectively. (C) The accuracy of binary classification on the training and test datasets for four tastants. (D) Quaternary classification results on the training (Left) and test (Right) datasets. The hollow-red circles represent the target class of the samples, while the blue squares are the final predicted results.

To make the categorization more intuitive, we break down a four-class classification into four binary problems. Each binary classification can be used to determine whether the unknown taste can be classified into its class. As the dataset is divided into the training part and the test part, the four single-layer neural networks are trained by 160 samples and tested by 40 samples. The performances on the whole dataset are shown in Fig. 5B and SI Appendix, Fig. S35. The accuracy for the test parts reaches 90% (CH₃COOH), 80% (MgSO₄), 75% (NaCl), and 90% (Pb(CH₃COO)₂), respectively. Similar accuracies on the training and the test datasets are demonstrated (Fig. 5C and SI Appendix, Fig. S36), which indicates that our system can perceive samples that have not been “seen” before. Moreover, the four binary results can be converted into a quaternary classification (Fig. 5D). The four original outputs can be sorted as the probability of each salt. In this way, 87.5% and 72.5% accuracies on the training and the test datasets are achieved.

Discussion

The prediction results indicate that the gustation classification based on our GO-ISMDs can be performed on the in-liquid computing system. The capability of this system in perceiving complex flavors—drinks of coffee, coke, and their mixtures is also demonstrated with remarkable accuracy (SI Appendix, Figs. S37–S40), which is comparable to Pannone’s artificial neural network (48). These performances should be attributed to the slow dynamics of ionic charges within confined circumstances. The significantly extended memory time is far beyond the reach of common charging and discharging processes induced by parasitic capacitances, which provides a guideline for the development and/or improvement of future memristive devices in liquid. In addition, our demonstration of integrating both sensory and computing functions on one device suggests that neuromorphic in-sensor computing in liquid is possible. On the other hand, there are also notable limitations in terms of energy consumption and detection sensitivity, largely due to the system’s bulky physical architecture (SI Appendix, Table S2). Device miniaturization/improvement and circuit integration could be the next steps to develop a more efficient perceptual system (9, 52). Looking forward, we envision our preliminary work could inspire the research on biomimetic gustation, which is not only important for intelligent perceptions but also capable of reconstructing taste for taste-loss people.

Materials and Methods

Materials.

The water dispersion of the negatively charged graphene oxide [GO (−)] was procured from Graphenea Inc. The positively charged graphene oxide [GO (+)] was synthesized according to the procedure reported in our previous research (33). All salts with a purity exceeding 95% were purchased from Aladdin Chemistry.

Preparation of GO Membranes.

The commercial GO (−) solution (4 mg mL⁻¹) was first diluted to a concentration of 0.4 mg L⁻¹. The diluted solution was then washed three times with DI water with the assistance of centrifugation (6,000 rpm, 15 min). 20 mL (for 850 nm-thick membranes, 0.4 mg L⁻¹) washed GO was vacuum filtered by using a mixed cellulose membrane (MCE, diameter: 50 mm and pore size: 0.22 μm). The GO/MCE membrane was then dried in a vacuum oven at 60 °C for 6 h. The dried GO/MCE membrane was cut into square pieces (4 mm × 4 mm) and immersed in dimethylformamide (DMF) for 4 h to remove the MCE. The GO laminates were thoroughly rinsed with DI water and fully hydrated by soaking in DI water for 48 h. The thickness of the GO membranes was controlled by adjusting the volume of the GO solution. A similar procedure was used to fabricate the positively charged GO membranes by using 0.1 mg L⁻¹ GO (+).

Materials Characterization.

The zeta potential of the GO sheets was measured by using a Zetasizer Nano ZS (SI Appendix, Fig. S1). The Raman spectrum was recorded on a microscopic Raman spectrometer (Renishaw inVia plus) (SI Appendix, Fig. S2). The Fourier-transform infrared spectrum (FTIR) was conducted on an FT-IR Spectrometer (Excalibur 3100) (SI Appendix, Fig. S3). The X-ray photoelectron spectrum was recorded on an X-ray photoelectron spectroscopy (ESCALAB 250XI), with the photon energy calibrated relative to the adventitious carbon C1s peak (284.8 eV) (SI Appendix, Fig. S4). The morphologies of graphene oxide membrane cross-sections (SI Appendix, Fig. S5) and device geometries (SI Appendix, Fig. S8), as well as elemental mappings (SI Appendix, Figs. S13 and S14 and Table S1), were characterized by using a field emission scanning electron microscope (Hitachi S4800) equipped with a Horiba EMAX X-ray detector. The interlayer spacing of GO sheets within GO membranes was characterized by X-ray diffraction (Panalytical X’Pert Pro Multipurpose Diffractometer) (SI Appendix, Fig. S6). The thickness of GO membranes was measured by using a Dektak XT Stylus Profiling System (Bruker) (SI Appendix, Fig. S7).

Device Fabrication.

To fabricate GO-ISMD, a floating SiN_X window (75 μm × 75 μm) was first fabricated at the center of a SiN_X (100 nm)/Si (500 μm)/SiN_X (100 nm) substrate (10 mm × 10 mm). This was achieved by photolithography, dry etching of the SiN_X layer with a mixture of O₂ and CHF₃ gases, and wet etching of the Si layer at 80 °C by saturated KOH solution. Subsequently, 36 holes (diameter: 3 μm) were drilled through the floating SiN_X window by photolithography and dry etching (23). Before transferring the GO membrane, the hydrophilicity of the as-prepared SiN_X substrate was improved by immersing it in a mixture of H₂SO₄:H₂O₂ (7:3 v/v) under 80 °C for 30 mins. This step could enhance the adhesion of the GO membrane on the SiN_X substrate. To further enhance the adhesion of the GO membrane on the SiN_X substrate, an additional thin GO film (thickness: 50 nm and size: 4 mm × 4 mm) was transferred onto the substrate and dried with N₂. Finally, a fully hydrated GO membrane (size: 4 mm × 4 mm) was wet-transferred onto the thin GO film and gently dried with N₂. The SEM image of a GO-ISMD is shown in SI Appendix, Fig. S8.

Experimental Setup.

Electrical measurements were carried out on custom-built units (SI Appendix, Fig. S9). The unit for emulating synaptic functions (SI Appendix, Fig. S9A) consisted of two Teflon cells. Both cells were drilled to create columnar cavities with a diameter of 5 mm and a length of 8 mm, which served as solution reservoirs. A hole (diameter: 3 mm) was drilled on top of each cell for the filling of solutions. GO-ISMDs were clamped between two PDMS O-rings and fixed between the Teflon cells. For chemical sensing, a hole (diameter: 1.5 mm) was drilled at the bottom side of the left Teflon cell, serving as a diffusion exit for the target chemical solutions. This unit was immersed in a 500 mL bath filled with the base saline electrolytes (SI Appendix, Fig. S9B). Before electrical measurement, the internal chambers of Teflon cells were first rinsed with isopropanol to improve hydrophilicity. Saline solutions were immediately added into the reservoirs and replaced by fresh solutions three times to avoid dilution by residual isopropanol. After this step, fresh saline solutions were added to the Teflon cells again for measurements. To perform chemical sensing, 20 μL of target chemical solutions were added to the left Teflon cell. Two Ag/AgCl wires with a diameter of 1 mm served as the electrodes for electrical tests. The right Teflon cell was always grounded while the left Teflon cell, directly contacting the topside of GO-ISMDs, was connected to voltage sources throughout all experiments in this study. A combination of a DAQ card (NI PCIe-6361) and a Low Noise Current Preamplifier (SR570) was used to characterize the basic synaptic functions and the chemical sensing behaviors of GO-ISMDs. During electrical measurements, currents were recorded at a rate of 100 samples per second. For square pulse tests, the amplitude of a current response to a voltage pulse was determined by averaging the last quarter of the current response. Furthermore, signal processing of the gustation classification systems was executed on the custom electronic circuits and software.

Construction of the Dataset for the Bioinspired Gustatory System.

For the training and testing of the bioinspired gustatory system, a homemade dataset was built. The dataset consisted of dynamical current responses to different target chemical solution drops. These sensing signals were collected by GO-ISMDs with a GO membrane thickness of 550 nm. The electrical measurements were performed in the unit shown in SI Appendix, Fig. S9B. In this system, the solution bath and Teflon cells were filled with 10 mM KCl electrolytes. Ag/AgCl electrodes were placed in the right Teflon cell and the solution bath. A constant voltage of 0.6 V was applied across the device. To collect a dynamical current response to a target chemical, a drop of 20 μL chemical solution (100 mM) was dropped into the left cell, and currents were instantaneously sampled at 100 samples/s for 300 s. For a subsequent test, it was necessary to flush the left Teflon cell with 10 mM KCl electrolytes to promptly switch the device to its original conductance state. The sensing of target chemicals of CH₃COOH, NaCl, Pb(CH₃COO)₂, and MgSO₄ was randomly conducted with different GO-ISMDs. The current-time curves contained 30000 points. Subsequently, each current-time curve would be converted into a temporal waveform of unit amplitude. This was accomplished by first subtracting the steady current level, defined as the mean value of the segment prior to the peak of the sensing curve, and then scaling the amplitude to 1, which aligns with the voltage amplitude employed in the experiment. Next, sequences of 4 points were directly extracted from these temporal waveforms at an interval of 80 s and used as data samples within the dataset. Finally, a dataset containing 200 samples was created, 50 samples for each class. The dataset was randomly divided into the training part and testing part with a ratio of 4:1.

Derivation of the Frequency Dependence of the Normalized Area of the Positive Half of the IV Loop.

We derived the fittings of normalized areas of the positive half of the IV loops to the voltage scan rate based on the work reported by P. Robin et al. (27). Briefly, for a pinched I − V curve obtained from a GO-ISMD under sinusoidal voltage pulses ($\mathrm{V}=V_{A}sin\omega t$, where $V_A$ represents the amplitude and $\omega$ denotes the angular frequency of the pulse), the area of the IV loop could be calculated according to:

$$A\left(\omega \right)={\int }_{-V_A}^{V_A}\left|I_B-I_F\right|dV,$$ [1]

where $I_B$ and $I_F$ are the backward and forward currents, respectively. From the equivalent circuit model of GO-ISMDs in the Inset of Fig. 2C (also see SI Appendix, Fig. S9C), the current of the device can be expressed:

$$I=I_C+I_M,$$ [2]

where $I_C$ and $I_M$ are ionic currents through the parasitic capacitance and the equivalent memristor of the GO-ISMD, respectively. The IV loops display pronounced asymmetrical characteristics, with a rectification ratio exceeding 10. Furthermore, the positive half exhibits noncapacitive hysteretic characteristics. This suggests that the current passing through the equivalent memristor is pivotal at the positive voltage bias. To investigate the effect of synaptic ionic dynamics, we focused on the positive half of the IV loop. In this case, the area of the positive half of the IV loop is

$$A\left(\omega \right)={\int }_0^{V_A}\left(I_B-I_F\right)dV={\int }_0^{V_A}\left(I_{\mathit{BM}}-I_{\mathit{FM}}\right)dV-{\int }_0^{V_A}\left(I_{\mathit{BC}}-I_{\mathit{FC}}\right)dV.$$ [3]

The subscripts $\mathrm{C}$ and $\mathrm{M}$ are labels for the ionic currents $\mathrm{I}$ through the parasitic capacitance and the equivalent memristor, respectively. Ref. 27 has conducted a profound analysis of their memristive devices with active carbon nanofluidic channels based on the adsorption–desorption model. Considering the cases in this study, we adapt their theoretical results and get

$${\int }_0^{V_A}\left(I_{\mathit{BC}}-I_{\mathit{FC}}\right)dV=\frac{1}{2}\pi C_p{V}_A^2\omega \left(ideal capacitance\right),$$ [4]

$${\int }_0^{V_A}\left(I_{\mathit{BM}}-I_{\mathit{FM}}\right)dV=V_A{I}_{\mathit{AM}}\frac{\omega {\tau }_m}{1+{\omega }^2{\tau }_m^2}\left(\text{adsorption}-\text{desorption model}\right),$$ [5]

$${\tau }_m\sim \frac{k}{\lambda }\frac{L^2}{D}\left(\text{adsorption}-\text{desorption model}\right),$$ [6]

where $C_p$ is the value of parasitic capacitance, $I_{\mathit{AM}}$ is the amplitude of currents through the equivalent memristor, $\mathrm{k}$ and $\lambda$ are the ion adsorption and desorption rates in GO membranes, $\mathrm{L}$ is the thickness of GO membranes, and $\mathrm{D}$ is the diffusion coefficient of ions in the liquid. Particularly, ${\tau }_m$ could be regarded as the memory timescale of the equivalent memristor. In the present research, it was observed that parasitic capacitance exerted a negligible influence at the positive bias when the voltage scan rate was maintained at a minimal level (below 1 Hz) (Fig. 2 B–D). We could obtain $I_{\mathit{AM}}\sim I_A$, and $I_{\mathit{CM}}=C_p{V}_A\omega \ll I_A$, with $I_A$ the amplitude of current through the GO device and $I_{\mathit{CM}}$ the amplitude of current through parasitic capacitance. Then, upon normalization using $I_A\times V_A$ as the factor and the application of Eqs. 3–5, the normalized area of the positive half of the IV loop of the device will be

$$A^{\ast }\left(\omega \right)\sim \frac{\omega {\tau }_m}{1+{\omega }^2{\tau }_m^2}+C\left(\ll \frac{\pi }{2}\right),$$ [7]

with $c\left(\ll \frac{\pi }{2}\right)$ denoting a constant value much smaller than $\frac{\pi }{2}$. This Eq. 7 could be applied to GO-ISMDs under triangular voltage waves as well (27).

Ion Dynamics Simulation.

The ion transport within the GO-ISMDs is modeled by using Poisson–Nernst–Planck (PNP) equations. The simulations are conducted via COMSOL Multiphysics. The KCl electrolyte concentration is set as 1 mM, and the size of the rectangular system is 200 nm in length and 500 nm in width. In this rectangular system, a central pore of 5 nm × 20 nm facilitating ion transportation is created (SI Appendix, Fig. S21). To simulate the graphene oxide membrane, several laminated porous, thin layers (5 layers of GO sheets) are placed onto the central pore (Note: Due to the limited processing power of our computing system, only 5 layers of GO sheets are used for theoretical simulation which is approximately one-hundredth of an 850 nm-thick GO membrane). Pores of 10 nm in width are uniformly distributed across these layers for ion migration within the simulated membrane. The distance between adjacent layers is maintained at 1 nm. A homogeneous surface charge density of −1 × 10⁻⁴ C/m² is assumed on the porous, thin layers. To clarify the memristive behaviors, we have conducted simulations of the evolution of concentration at point A (SI Appendix, Fig. S21) and the corresponding currents upon application of voltages.

The dynamic voltages are applied on the top edge of the system while the bottom is always grounded. Within this two-dimensional framework, the migration of charged species (K⁺ and Cl⁻) through the membrane is described by PNP equations: (42, 45)

$$\frac{\partial c_i}{\partial t}=-\nabla ·J_i=-\nabla ·\left(-D_i\left(\nabla c_i+\beta c_i{q}_i\nabla \phi \right)\right),i=+,-,$$ [8]

$$-\nabla ·\epsilon \nabla \phi +c_{+}e-c_{-}e=0,$$ [9]

where $c_{+}$, $c_{-}$ represent the concentration for K⁺ and Cl⁻ ions, $J_i$, $D_i$ are the total ion flux and diffusion coefficient for the two ion species, $\beta =1/k_{B}T$ is a thermodynamic parameter with Boltzmann constant $k_B$, T is absolute temperature, $\phi$ is the electric potential, and $\epsilon$, $\mathrm{e}$ are the dielectric permittivity and elementary charge, respectively. This numerical model incorporates the impact of adsorption–desorption interfacial interaction by assigning the diffusion coefficient with compromised values, which are only a few thousandths of their counterparts in liquid (1.96 × 10⁻⁹ m²/s and 2.03 × 10⁻⁹ m²/s for K⁺ and Cl⁻, respectively).

The ion distributions, potentials, and currents can be numerically simulated once the initial boundary conditions are performed. This work employs three forms of applied voltages on the system’s top boundary: 1) a fixed voltage, 2) a sinusoidal waveform, and 3) square pulses. Fixed voltages of 0.6 V and -0.6 are used to study the time evolution of current and concentration polarization in the model (Fig. 4B and SI Appendix, Figs. S22–S24). Sinusoidal voltage waves $\mathrm{V}=V_A\text{sin}\left(2\pi ·t/T\right)$, with $V_A$ the amplitude, $\mathrm{T}$ the period, and $\mathrm{t}$ the instantaneous time, are used for the current-voltage curve simulation (Fig. 4 C and D and SI Appendix, Figs. S25 and S26). Square voltage pulses of width 0.5 ms are used to simulate the synaptic behaviors of the GO-ISMDs (Fig. 4 E–H and SI Appendix, Fig. S27).

Supplementary Material

Appendix 01 (PDF)

Code S01 (TXT)

Data, Materials, and Software Availability

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