Neuromorphic ionic computing in droplet interface synapses

Abstract

Ionic devices with memory capabilities can emulate neural functionality, enabling neuromorphic computing and biomedical applications. In this study, we report an ionic spiking synapse based on aqueous droplet interface bilayer assembly. Under stepwise triangular voltages, the device displays coupled memcapacitive-memristive behavior, showing noncrossing pinched hysteretic I-V loops. This hysteretic ion dynamics can be regulated by modifying bilayer components, reconstituting protein channels, or adjusting droplet assembly configuration. Droplet interface synapses (DIS) exhibit fundamental neuromorphic behaviors such as paired-pulse facilitation/depression, spike rate–dependent plasticity, Hebbian learning, and short-term associative learning under classical conditioning. We also used reservoir computing with DIS to implement two learning algorithms: a classification algorithm that recognizes handwritten digits and a reinforcement learning algorithm that learns to play a board game of tic-tac-toe.

Droplet interface synapses emulate neuromorphic ionic learning and computing and enable pattern recognition and board game play.

INTRODUCTION

Conventional computing platforms have been reaching their physical and operational limits despite remarkable advances in their speed and miniaturization, prompting researchers to explore alternative materials and architectures (1). The human brain, which consists of billions of neurons and synapses, is much slower in comparison (2), yet it can perform complex parallel tasks with remarkable efficiency using highly interconnected network of synaptic connections (3). Neuromorphic computing research (4, 5) dates back 35 years (6), but notable advancements began 25 years ago with the demonstration of synaptomimetic functionality in a semiconductor memristor and other memory elements (7–9). Since then, substantial research efforts have focused on developing devices, architectures, chips, and algorithms that use these memory elements for neuromorphic computing (10, 11).

Most existing memory devices are based on solid-state components and rely on electrons or holes as information carriers (12, 13), which results in relatively high energy consumption (14). In contrast, brain signaling uses highly energy-efficient ion transport in an aqueous environment (3). The emerging field of ionic computing draws inspiration not only from the brain’s architecture but also from its reliance on ionic, rather than electronic, signaling. The introduction of the analytical framework (15) and subsequent experimental demonstration (16) by Robin et al. brought rapid progress to development of various memory devices using nanofluidic channels and exhibiting synaptic plasticity (17–22). However, most of these systems still incorporate at least some solid-state components, which limit their scalability, flexibility, and biocompatibility.

Researchers have used droplet assemblies to implement various biomimetic functionality (23–26). Synaptomimetic behavior in these assemblies was discovered by Sarles, Collier, and colleagues (27–30), who also showed memristive behavior and sensory adaptation in droplet assemblies doped with ion channels (27, 31, 32). However, the memristive functionality in those devices remained unstable due to the uncontrolled and inconsistent insertion of ion channels in the droplet interface bilayers (27, 33). Najem and colleagues (28, 34) also demonstrated synaptic functionality in purely memcapacitive droplet assemblies, where capacitive memory arose from reversible geometrical changes in lipid bilayers. They subsequently used these effects to implement reservoir computing applications such as digit recognition and high-dimensional mapping (35, 36). These advances in droplet-based ionic devices have recently coalesced into an emerging field known as “dropletronics” (37) with potential applications in biointerfaces (38), biobatteries (39), and iontronic devices and circuits (37).

In this study, we explore neuromorphic ionic computing functionality in droplet assembly–based droplet interface synapses (DIS). They exhibit combined memristive and memcapacitive behavior with the ionic current signals, notably enhancing system stability and simplicity compared to previous designs. We systematically investigate the neuromorphic potential of DIS, demonstrating robust synaptic plasticity and associative learning beyond prior demonstrations. Moreover, we successfully implement reservoir computing algorithms using DIS, enabling tasks such as handwritten digit recognition and strategic gameplay against an ideal opponent and demonstrating a substantial expansion of the computational capabilities of droplet-based synapses.

RESULTS

Droplet interface synapses

The DIS platform relies on a small patch of lipid bilayer (40) formed at the interface between two aqueous droplets (~600 nl each) coated with synthetic 1,2-diphytanoyl-sn-glycero-3-phosphocholine (DPhPC) lipid and suspended in hexadecane oil (Fig. 1A; see also Materials and Methods and fig. S1 for details). When we applied a triangular voltage waveform to this droplet assembly, we observed a highly nonlinear I-V curve which displayed a pinched loop (Fig. 1B and figs. S2 and S3) and a strong conductance hysteresis (Fig. 1B, inset) at low sweep frequencies. While the maximum power consumed by the device (P_max) showed a monotonic increase with the sweep frequency, the loop area of the ion current hysteresis (P_hys) went through a maximum at ~0.1 Hz (Fig. 1C, inset). Consequently, the normalized loop area (P_hys/P_max) also went through a maximum at ~0.04 Hz and then monotonically decreased at higher sweep frequencies (Fig. 1C).

Fig. 1. DIS I-V hysteresis.

(A) Droplet interface bilayer schematics, showing a bilayer region formed between two lipid-coated aqueous droplets in oil. Two Ag/AgCl electrodes hold the droplets, apply the external voltage (V_s), and record the ionic current (I_s). Inset: Optical micrograph of the droplet assembly showing the bilayer at the droplets’ interface. (B) I-V characteristics of the DIS under a periodic triangular voltage sweep. Top inset: Triangular voltage sweep with 0.1-Hz frequency. Bottom inset: The corresponding conductance (G)–voltage characteristics. The droplets were filled with 100 mM KCl at pH 7.5, and the sampling frequency was 1 kHz. (C) Normalized I-V loop area (P_hys/P_max) as a function of voltage sweep frequency. Inset: The I-V loop area (P_hys) and the maximum power $[P_{\text{max}}=2V_{\mathrm{A}}(I_{\text{max}}-I_{\text{min}})]$ as a function of sweep frequency. (D) I-V loops recorded for different triangular wave amplitudes (V_A). The sweep frequency was 0.1 Hz. Inset: Normalized I-V loop area as a function of V_A. (E) Charge (Q_M)–voltage characteristics of the DIS under different sampling frequencies (sweep frequency was kept at 0.1 Hz). Inset: The corresponding I-V characteristics under different sampling frequencies. (F) Relative DIS bilayer capacitance (C_M) as a function of applied dc voltage. The red curve represents a quadratic fit to the data. Inset: Optical micrographs of the DIS under different dc voltages.

The observed pinched I-V hysteresis loop and the reduction in loop area with increasing sweep frequency align with two key fingerprints of a memristor, as defined by Chua’s theory (7, 41). However, rather than collapsing into a single line as predicted by Ohm’s law at high voltage sweep frequencies (the third fingerprint of a memristor), the pinched hysteresis loop transitions into a nonpinched, hyperbolic-shaped loop (fig. S3), characteristic of memcapacitive behavior. This deviation from ideal memristive response reinforces our conclusion that DIS functions as a coupled memristor and memcapacitor (42, 43), similar to behaviors previously observed in plant tissues (44, 45).

Further analysis identified the DIS as a unipolar memory element, characterized by a nonself-crossing I-V curve at the origin (Fig. 1B and fig. S3). This observation rules out asymmetry-induced ion adsorption and desorption as the underlying ion memory mechanism, which would have resulted in a diode-like self-crossing I-V loops (16, 18). We also did not observe a clear threshold voltage for pronounced I-V hysteresis, and the normalized loop area remained unaffected by changes in voltage amplitude V_A (Fig. 1D). This behavior suggests that I-V hysteresis is not triggered by exceeding a certain voltage threshold, ruling out ionic bandgap effects, such as the second Wien effect (15, 18) or spatial resistance (46), as the underlying mechanisms of ion memory. In addition, no saturation in ion current was observed within the 250-mV amplitude range, which was insufficient to induce strong ion concentration polarization across the bilayer (47), ruling out this memory mechanism as well.

Following the Hodgkin-Huxley framework for lipid membranes (43, 48, 49), the DIS can be modeled as a parallel arrangement of a capacitor (C_M) and a resistor (R_M) (see text S1 for details). At the low sampling frequency used in our experiments, the applied triangular voltage effectively becomes a stepwise voltage (fig. S2; see also text S1), and the ion current reflects the combined effects of capacitive and resistive behaviors. Thus, low sampling frequencies accentuate the memristive behavior in the DIS, which shows a memristor-like nonpinched charge-voltage loop, whereas high sampling frequencies accentuate memcapacitive behavior, characterized by a pinched hysteresis loop in both the charge-voltage and capacitance-voltage plots (Fig. 1E and fig. S4) (28, 50, 51).

Previous studies (27, 28, 49, 52) have shown that applying a nonzero voltage leads to bilayer electrostriction, including an increase in the bilayer area due to electrowetting and a decrease in bilayer thickness due to electrocompression. We also observed an increase in total bilayer capacitance in the steady state as a result of these geometric changes (Fig. 1F and fig. S5). The bilayer capacitance follows the empirical relationship $C_{\mathrm{M}}/C_0=1+\mathrm{\beta }{V}_{\mathrm{s}}^2$ , where C_M represents the total capacitance of the bilayer, and C₀ is the capacitance at zero voltage. Our measurements yield a nonlinearity coefficient of β ~ 14.8 V⁻² for DPhPC bilayers in hexadecane, which is in precise agreement with previously reported values (27, 28, 49, 52). Instant mesoscale geometric changes in bilayer area and thickness (Fig. 2A and movie S1) directly influence the I-V characteristics through the memcapacitive ion current.

Fig. 2. DIS ion memory mechanisms.

(A) Schematic of the main DIS ion memory mechanisms. Triangular voltage actuation leads to electrostriction and electroporation, changing in the resistance (R_M) and capacitance (C_M) of the DIS. (B) I-V characteristics of DIS with different numbers (N) of α-HL channels (inset) in the bilayer, compared to the control without the channels. (C) Fluo-8H dye fluorescence intensity changes (top) near the bilayer (region a, inset) and deep within the droplet containing the dye (region b, inset) in response to triangular voltage (bottom), when the second droplet contained CaCl₂ solution. Small periodic fluctuations in fluorescence intensity are caused by bilayer electrostriction altering optical path. Inset: Fluorescence image of the droplets used for the experiment. (D) I-V characteristics of the DIS containing different KCl concentrations. (E) I-V characteristics of the DIS with different bilayer compositions: DPhPC lipids (control), 75% DPhPC + 25% cholesterol (Chol.), and 50% DPhPC + 25% cholesterol +25% sphingomyelin (SM). Inset: Schematics of bilayer structure and lipid domain formation in the presence of Chol. and SM. DPhPC and Chol. are distributed in both the leaflets, and SM is enriched in the outer leaflet. (F) I-V characteristics for three droplets connected in series under triangular voltage with amplitudes (V_A) of 200 and 400 mV and a sweep frequency of 0.1 Hz. Inset: Optical microscope images of the DIS with three droplets in series. Control: DIS containing 100 mM KCl and pure DPhPC lipid, V_A = 200 mV.

Reconstitution of α-hemolysin (α-HL) channels into the bilayer (Fig. 2B, inset; see also fig. S6) provides more clues to the ionic memory mechanisms. α-HL channels add a parallel conductance element to the bilayer, which should increase the overall ion current and reduce the hysteresis loop due to strong suppression of transmembrane potential and electrostriction. Addition of α-HL channels to the DIS produced substantial increase in the device conductance and strongly suppressed the hysteresis (Fig. 2B). The normalized I-V loop area also decreased substantially as the number of α-HL channels in the bilayers increased (fig. S6). Additional experiments that monitored the fluorescence intensity of a membrane potential–sensitive dye incorporated into the bilayer at the droplet junction (fig. S7) reinforced these conclusions. As we cycled the voltage, the dye response registered corresponding swings of the membrane potential, indicating that the membrane is getting capacitively charged, with the dye response reducing significantly after addition of α-HL channels.

However, this capacitive response cannot account for the major features observed in the I-V curves, especially those at lower sampling frequencies that demonstrate a pronounced memristive component to the device response. Several additional mechanisms, such as (de)solvation, densification, reorientation of lipid head groups (43), and temporary electroporation of the bilayer (Fig. 2A) (53) at relatively high voltage values used in our study could contribute to this memristive behavior. To further probe the ion transport dynamics in DIS, we conducted a fluorescence assay using a Ca²⁺-sensitive dye commonly used to monitor ion channel activity (54). Under a triangular voltage waveform applied to the bilayer, we observed an increase in fluorescence near the bilayer region (Fig. 2C), consistent with transient pore formation.

Further insights into the ion memory mechanisms come from experiments where we altered the key parameters that can influence ion dynamics. While changes in ion concentration did not markedly affect the shape of the I-V curves (Fig. 2D), higher ion concentrations led to an increase in peak current and a slight enlargement of the hysteresis loop area (fig. S8). Introduction of cholesterol, which enhances the membrane rigidity and surface tension (55), produced larger I-V hysteresis loops (Fig. 2E) and an increased normalized loop area (fig. S9). Cholesterol also restricts lipid movement and facilitates a stable lateral structure during electrocompression (55), which should reduce the memcapacitive behavior. We observed that cholesterol addition amplified the I-V hysteresis in our experiments. Instead, we speculate that cholesterol may enhance the probability of transient pore formation under high voltages, given its greater negative curvature compared to DPhPC (49).

In a different set of experiments, we added both cholesterol and sphingomyelin (SM) (56). The addition of SM pulls the cholesterol molecules into ordered “lipid raft” domains (Fig. 2E, inset) and decreases the bilayer’s in-plane elasticity (57), which should negatively affect both memristive and memcapacitive contributions to hysteresis. We did observe the gradual registration of the bilayer under a triangular voltage sweep, which likely diminished the electrostriction-induced memcapacitive behavior. We saw a substantial reduction in both the ion current amplitudes and, crucially, the values of the I-V hysteresis (Fig. 2E).

As replicating the full functionality of biological neural networks requires the cooperative interaction of multiple synapses, we explored the properties of two configurations of three-droplet assemblies (Fig. 2F and figs. S10 and S11). As expected, connecting two bilayers in series reduced the overall ion current under a triangular voltage sweep with an amplitude of 200 mV (Fig. 2F and fig. S10). The hysteretic response was restored when we increased the sweep amplitude to 400 mV. Expectedly, DIS junctions connected in parallel exhibited I-V hysteresis similar to that of a single DIS (fig. S11).

Synaptic functionality and plasticity in DIS

Synaptic plasticity plays a crucial role in the formation of learning and memory in biological synapses, where it plays a key role in decoding, processing, and transmitting temporal information (Fig. 3A). As a unipolar memory device, a DIS should allow both SET and RESET operations at the same voltage polarity (58). To enable these operations, we held the DIS at a preset voltage value of +130 mV to allow it to equilibrate before applying different voltage spike patterns (Fig. 3).

Fig. 3. DIS synaptic plasticity.

(A) Schematic of the DIS device mimicking biological synapses. (B) PPF and PPD behaviors under paired-pulse voltages (0.1-s duration) at difefrent time intervals (Δt). The solid lines are biexponential fits to the PPF or PPD index (I₂/I₁): $I_2/I_1=A_0+A_1{e}^{-\mathrm{\Delta }t/{\mathrm{\tau }}_1}+A_2{e}^{-\mathrm{\Delta }t/{\mathrm{\tau }}_2}$ , where I₁ and I₂ are currents induced by first and second voltage spikes, and τ₁ and τ₂ are two time constants; A₀ is a constant, and A₁ and A₂ are weights. Inset: Example ionic currents for PPF (200 mV) and PPD (0 mV). (C) Relative ionic current (I_N/I₁) under successive 200 mV spikes for facilitation (Fac.) and 0 mV for depression (Dep.), each 0.1 s, at varying intervals. I₁ is the first-spike current; I_N is the Nth. (D) Relative ionic current (I_N/I₀) recorded over five consecutive cycles, where each cycle included 200 write spikes (200 mV, 0.1 s) and 200 erase spikes (0 mV, 0.1 s), with a read pulse between each spike. I₀ is the initial current; I_N is the current after the Nth spike. (E) Time traces of relative ionic current decay after a facilitation (200 mV, 10 s) and depression (0 mV, 10 s) voltage spikes. The current was measured before and after the spike with successive read pulses. I₀ is the ionic current reading before the spike, and I_t is the current reading at subsequent time points. (F) STDP signals, implemented with Hebbian rules, showing ionic current changes (ΔI/I₀) as a function of relative activation timing (Δt). After the activation of two neurons, A and B, a series of erase (0 mV, 0.1 s) and write (200 mV, 0.1 s) spikes were applied, respectively. ΔI = I₁ − I₀ represents the change in ionic current read before (I₀) and after (I₁) activation. Inset: Example spike sequence with neuron A activating first (Δt = 0.5 s). All experiments used 100 mM KCl droplets and a preset voltage of +130 mV. All read pulses were applied at ΔV_s = +10 mV for 0.04 s.

To mimic short-term plasticity (STP) electric pulse patterns, we first applied paired voltage spikes to the device and recorded the resulting ion current spikes. Paired-pulse facilitation (PPF) and paired-pulse depression (PPD) refer to the phenomena where the response amplitude of the second stimulus spike (I₂) is either higher (PPF) or lower (PPD) than that of the first spike (I₁) (59). Two consecutive spikes of +200 or 0 mV applied to the DIS induced either an increase (PPF; Fig. 3B, top inset) or a decrease (PPD; Fig. 3B, bottom inset) in ion current (see also fig. S12 for the DIS response to voltage spikes). This behavior reflects the long ionic memory scale seen in the DIS I-V curves (Fig. 1B; see also text S2 and figs. S13 and S14 for additional analysis of the ion dynamics in the DIS).

The values of the PPF and PPD index (I₂/I_I) are approaching unity as the interval between the voltage spike interval (Δt) increases and the memory effects diminish (Fig. 3B). The ion current change is strongly correlated with the spike interval, following a biexponential relationship, characterized by two distinct time constants. For PPF, those fitted time constants τ₁ and τ₂ were 28 and 857 ms, respectively (τ₁ and τ₂ were 167 and 840 ms, respectively, for PPD). The different time constants for PPD and PPF suggest asymmetric changes in the DIS states around the preset voltage. These time constant values align with those observed in biological systems, where facilitation and depression of synaptic weights occur on a similar timescale, with τ₁ and τ₂ for PPF being ~40 and ~300 ms, respectively (59).

STP plays a crucial role in temporal filtering in the brain, where synaptic weights change in response to multiple spikes with varying rates—a phenomenon known as spike rate–dependent plasticity (SRDP). We have replicated this behavior in the DIS by applying 50 presynaptic voltage spikes with an intensity of +200 mV/0 mV, a duration of 0.1 s at varying intervals corresponding to frequencies ranging from 0.9 to 5 Hz (Fig. 3C and fig. S12). In these experiments, the postsynaptic current decreased or increased as the pulses were close to each other while showing minimal changes when the pulses were far apart in the time domain. These differences in the dynamic response of the DIS also mirror the changes in synaptic weights that are characteristic for biological STP phenomena (59).

Synaptic behavior in the DIS is highly repeatable. Whereas a “write” voltage spike would increment the DIS ion current, which could then be reset to its original value with an “erase” spike, the stored value could also be accessed with a smaller “read” pulse. When we applied periodic writing spikes, we observed that the synaptic weight rapidly increased and then saturated with the number of writing pulses (Fig. 3D). Conversely, the synaptic weight decreased quickly when erase spikes were applied. This modification process was fully reversible, with the device maintaining its performance even after 2000 write-erase cycles (Fig. 3D). We also conducted dynamic monitoring of ion current changes following full facilitation/depression (write/erase) voltage spikes (10-s duration) (Fig. 3E). After a write (erase) spike, ion current jumped (dropped) and then gradually returned to the original value in the absence of further spikes, indicating a DIS memory retention timescale of ~20 s.

Reseaerchers demonstrated spike timing–dependent plasticity (STDP) as a Hebbian synaptic learning rule in various neural circuits in the brain, where the connection strength between pre- and postsynaptic neurons changed on the basis of the timing and order of incoming spikes (Δt) (60, 61). To simulate this behavior, we applied a voltage time series (16) designed to emulate the behavior of two neurons to the DIS (Fig. 3F and fig. S15). The activation of presynaptic neuron A was represented by a weak positive voltage pulse. Upon activation, we flipped the voltage, applying negative voltage spikes until the postsynaptic neuron B was activated, at which point positive voltage spikes were applied again. If neuron B activated first, then the sequence was reversed: Positive spikes were applied first, followed by negative spikes after neuron A was activated. We then compared the baseline ion current of the DIS to the current measured after five successive activation of neurons A and B, at different relative spike timing. When the presynaptic neuron spiked before the postsynaptic neuron, the ion current of the DIS increased, simulating the strengthening of synaptic connections between those neurons (Fig. 3F). Conversely, when the postsynaptic neuron spiked first, the ion current decreased, simulating the weakening of synaptic connections. In addition, longer interval between the spikes of the pre- and postsynaptic neurons produced smaller changes in ion current. This behavior is similar to that observed in nanofluidic memristors with graphite nanochannels (16) and closely resembles spatial perception tuning behavior in biological vision systems (62).

Measurements of the SRDP behavior across several droplet assemblies (fig. S16) indicate that normalized ionic current changes are preserved across a range of droplet volumes (~0.3 to 1.0 μl) and remain highly robust. This result suggests that the DIS could potentially be scaled down to enable further miniaturization (37). However, droplets that are too small may compromise system longevity due to droplet evaporation (63). In addition, we showed that synaptic weights can be efficiently modulated during facilitation and depression by adjusting the preset voltage value (fig. S17) where a lower preset voltage increased the facilitation behavior, and a higher preset voltage enhanced the depression behavior. Preset voltage values also regulated the asymmetry of Hebbian STDP rules, with a net bias toward facilitation at low preset voltages. We could also tune the synaptic plasticity of DIS by altering the lipid composition or droplet assembly configuration (fig. S18). This behavior is consistent with the ion dynamics modulation described in Fig. 2 and also mirrors biological synapses, where the lipid composition of neuronal cell membranes influences their function and, ultimately, the brain’s ability to reconfigure neuronal connectivity (64). Last, we note that, in contrast to previously reported solid-state or fluid-based synaptic devices that require high operating voltages (8, 9, 16, 17), DIS devices operate at voltage much closer to that of the biological synapses (200 mV versus ~110 mV), demonstrating that our devices could potentially function as bioelectronic device components.

Associative learning with DIS

Pavlovian conditioning, also known as classical conditioning or associative learning, is a form of implicit memory that operates unconsciously or procedurally (65, 66), enabling biological species to learn and remember relationships between unrelated stimuli. Ivan Pavlov discovered this form of conditioning through his experiments with dogs (65). In Pavlov’s experiment (Fig. 4A), feeding food serves as the unconditioned stimulus (US), while the ringing bell and salivation are the neutral stimulus (NS) and unconditioned response (UR), respectively. Initially, the bell ringing does not induce salivation, whereas the presence of food (US) does. During training, the dog is repeatedly exposed to the sound of the bell (NS) followed by the presentation of food (US), creating an association between the two. Eventually, the bell alone [now termed the conditioned stimulus (CS)] triggers salivation, known as the conditioned response (CR). This process is called acquisition. However, if the CS is repeatedly presented without the US, then the association weakens, leading to extinction. Pavlovian conditioning is characterized by three key features (i) acquisition, where the CS and US are paired closely in time; (ii) extinction, where the CS is presented alone, leading to a decrease in the CR; and (iii) recovery, where retraining after extinction occurs more rapidly than the initial acquisition.

Fig. 4. Associative learning and reservoir computing in DIS.

(A) Schematic of a Pavlov’s dog associative learning experiment. The bell and food correspond to low voltage spikes (100 mV, 0.1 s) and high voltage spikes (200 mV, 0.1 s), respectively. The preset voltage was 0 mV, and read pulses (ΔV_s = +10 mV, 0.04 s) were used to measure the currents between spike intervals (0.3 s). See the main text for symbol explanation. (B) Pavlov’s dog experiment with DIS. The salivation threshold was set at I_t/I₀ = 1.2. (C) The relative ionic currents (I_N/I₀) of DIS under 16 different 4-spike voltage inputs denoting different 4-bit strings. Here, a 0 and 1 correspond to erase (0 mV, 0.5 s) and write spikes (200 mV, 0.5 s), respectively. The preset voltage was +130 mV, and read pulses (ΔV_s = +10 mV, 0.04 s) were used to measure the currents before and after each spike. Inset: An example of input voltage for “1001.” (D) Handwritten digit identification with DIS. An example digit (6) from the MNIST database downsampled to 20 × 22 pixels was sectioned into rows of four pixels, yielding 110 “4-bit” strings. These were sequentially applied to the single DIS, where the current responses formed the reservoir states. (E) Confusion matrix for the digit identification algorithm in (D). (F) Tic-tac-toe gameplay with DIS. Board state information X and O positions were encoded into binary strings with optional redundancy and applied to the single DIS. The corresponding current responses form the reservoir state vector for the gameplay. (G) Game performance of different agent architectures, represented as a percentage of the draws from the multiple games played against an ideal opponent.

To mimic Pavlovian conditioning behavior, we applied a voltage spike (100 mV) to the DIS to represent the “bell” (NS/CS), while a larger voltage spike (200 mV) represented the “food” (US) (Fig. 4A). Ion currents were recorded using a read pulse (10 mV, 0.04 s) where we defined a 20% ion current change as the threshold for the “salivation” response. We observed that the DIS exhibited associative learning and extinction behavior (Fig. 4B). Before training, applying the NS alone produced synaptic weight changes below the threshold, indicating that the DIS did not respond to the NS. In contrast, applying the US alone generated a synaptic weight above the threshold, triggering the UR. After pairing the bell (NS) and food (US) signals, the synaptic weight increased above the threshold, establishing an association between the NS and US. Following this training, the NS alone (now CS) would initially produce a synaptic weight above the threshold. After ~11 s, the synaptic weight gradually decayed below the threshold, leading to extinction of the associative behavior. As the association faded, the CS reverted to NS, and the synaptic weight returned to its original state. We also note that applying the NS (bell) immediately after an unpaired US (food) resulted in an ionic current above the threshold for ~6 s (Fig. 4B and fig. S19). This time was shorter than the ~11-s memory time after paired training, supporting the formation of true associative memory between the CS and US.

Reservoir computing with DIS

The short-term synaptic plasticity of DIS makes it an ideal candidate for reservoir computing: a brain-inspired algorithmic framework that leverages a fixed dynamic network, or “reservoir,” to transform complex temporal data into easily classifiable outputs (67). For a device to be suitable for reservoir computing, it must have two critical properties: short-term memory and nonlinear dynamics. As we demonstrated in previous sections, DIS exhibits both. In this work, we implemented a hybrid reservoir computing system, where a single DIS provided the reservoir functionality and the readout layer was implemented in software (20, 68–72). The response of the DIS under different voltage pulses formed unique device states, which we then fed into the software-implemented readout layer to perform computing tasks. Unlike traditional neural networks, which require training the entire network, reservoir computing trains only a simple readout function to interpret the DIS’s output.

We began by processing 4-bit strings using a single DIS. We converted all 16 possible 4-bit combinations into voltage spikes, each lasting 0.5 s, with “0” and “1” represented by voltages of 0 and +200 mV, respectively (Fig. 4C). The spikes were spaced with 0.5-s intervals, during which a preset voltage of +130 mV was maintained. Because of the nonlinear PPF/PPD characteristics of synaptic plasticity, the DIS generated 16 distinct reservoir states reflected in ion current changes, with minimal variability across devices, as indicated by the SD from five separate runs (Fig. 4C). The final state of the reservoir was influenced not only by the last input but also by the memory of its previous state, suggesting the presence of lateral connections within the DIS reservoir. Different spike durations (0.1- or 1.0-s on/off times) were also used to process 4-bit strings, producing distinct ionic current states depending on the interplay between learning and forgetting behaviors (fig. S20). We estimate the energy consumption of a single DIS per spike to be 4 to 8 pJ, accounting for both capacitive energy storage and resistive dissipation (text S3). Although this value is higher than that of biological neural spikes (~0.1 to 10 fJ) (73), it is comparable to, or even lower than, most previously reported electron- and ion-based artificial synapses used in reservoir computing (table S1). Further reductions in energy consumption may be achieved by decreasing droplet size (which reduces bilayer capacitance) and minimizing spike duration.

With this DIS reservoir, we first implemented a benchmarking protocol (20, 68–72) for classifying handwritten digits (0 to 9) from the Modified National Institute of Standards and Technology (MNIST) database (74). We encoded each 22 × 20 pixel MNIST image into 110 strings of four bits, which generated analog reservoir states from the response of a single DIS (see text S4 and also Fig. 4D and fig. S21, digit “6” as an example). We then implemented an in silico, single-layer fully connected neural network with a 110 × 10 architecture to map these reservoir states to digit labels (0 to 9). The network was trained on 40,000 handwritten digits from the MNIST training dataset. During each epoch, we adjusted the readout weights to improve classification accuracy on the training set, while we monitored the performance separately on the 4000-digit validation set. We then used a separate 4000-digit test set to evaluate the final accuracy. The system requires far less synaptic weight parameters due to a 4× decrease in the number of inputs after DIS-based reduction in data dimensions. During training, the DIS-based recognition system converged rapidly, achieving 86.2% accuracy on the test set after only 20 epochs and minimal computational time (Fig. 4E and fig. S21). This accuracy is comparable to, and in many cases exceeds, previously reported results using similar protocols with single-layer neural networks in semiconducting memristor-based systems (83 to 91%) and solid-state nanochannel memristor-based systems (81%) (table S1). Most misclassifications occurred for digits “5,” “8,” and “3,” a challenge that human observers also faced at the 22 × 20 resolution. The system’s performance can be further modified by adjusting stimulation parameters such as spike amplitude, duration, and frequency. We explored the effects of different spike durations (0.1 and 1.0 s), achieving classification accuracies of 82.4 and 84.3%, respectively (fig. S21). We found that shorter training resulted in slower convergence rates but did not notably alter the accuracy, suggesting that DIS can facilitate fast and energy-efficient computing.

Handwritten digit recognition is relatively stable to pixel errors—if you flip a bit here or there, the digit’s overall shape does not change. We were interested to see whether we can use DIS reservoir–based encoding in a task where each individual bit value is critically important. Specifically, we looked at the game of tic-tac-toe (Fig. 4G and fig. S22). We used a simulated network of DIS units to encode each board position and trained an agent to play the game using the outputs from the network to select the next move (see also text S5 for details). We modeled each unit in the network based on the measurements of single physical DIS unit. Each square of the board can have three states: “X,” “O,” and empty. We capture these states in a series of 18 bits: writing 0 for each X and 1 otherwise and once for Os.

Our first intuition was to encode each board square information once. However, DIS noise can produce different output current values for the same DIS configurations. To mitigate this issue, we added redundancy in the representation by repeating each row’s bits R times. We varied R from 1 and 4, which produced between 5 and 18 (N) 4-bit vectors and the corresponding DIS states to represent a single board configuration. For each 4-bit vector, we obtained our DIS response and built a single-layer fully connected neural network of size N × 10 to map the states to individual squares on the board. The agent then selected an empty square with the highest value from the neural network to make the next move. The 10th network output (omitted from the figure) is needed for training and produces the expected outcome of the game (Fig. 4F and see also text S5).

We defined and trained the network in silico using the Proximal Policy Optimization (PPO) algorithm with invalid action masking (see text S5 for details) (75, 76). This algorithm is a reinforcement-learning method, where the network learns to maximize the reward—in this case, winning the game. We trained the agent to play the game by playing it against an optimal tic-tac-toe agent (OA), which always selects the best move by doing a brute-force search to the end of the game. The PPO algorithm adjusts the network weights after each move. Because of the random noise in the DIS readouts, the same board position can result in different output values, which can affect the game. To mitigate the overall noise effect, we evaluated the trained agent performance in a 10,000 games match between the agent and the OA and measured the success by the percentage of games that ended in a draw (since forcing a draw against an OA is the best possible outcome). Each match takes ~6.5 s on a MacBook Pro. Agents with redundancy R set to 3 and 4 learned to play the game perfectly, matching the OA agent’s performance (Fig. 4G and fig. S22B). In contrast, systems with R set to 1 and 2 did not fare as well, drawing only 89 and 93% of the games, respectively. To further examine the effect of the DIS noise on the network, we compared the performance of our network with a separate version of the agent where we did not add noise and did not repeat any bits. Unexpectedly, the system without the input noise performed the worst, drawing only 71% of the games, indicating that the noise of the DIS network was highly beneficial for the learning process.

DISCUSSION

In this work, we have demonstrated that a simple ionic synapse based on a droplet interface bilayer, which shows coupled memristive and memcapacitive characteristics, can be configured to exhibit sophisticated neuromorphic behaviors. Nonequilibrium ion dynamics in these droplet interfaces, governed by the voltage-dependent capacitance and resistance of the bilayer, creates the ionic memory effects that enable neuromorphic functionality. Further experimental and modeling efforts should help to fully elucidate the ion transport pathways and complex dynamics at the droplet interface and better understand the origin of the observed I-V hysteresis and corresponding memory mechanism. These devices also demonstrate stable and reliable modulation of ion dynamics, which can achieve continuously adjustable multiple states and replicate various functions of biological synapses, such as PPF/PPD and SRDP. We also demonstrated that this system implements basic learning algorithms such as Hebb’s rule as well as more sophisticated learning behaviors such as “Pavlov’s dog” associative learning with the salient features of learning and forgetting. Furthermore, we showed that DIS can be used for reservoir computing where temporal signals encoded as voltage spike trains were processed by DIS and subsequently classified in silico using a simple readout function. Unexpectedly, we found that DIS-based reservoir performance was comparable to that of conventional solid-state platforms, and the noise inherent to this device actually produced enhanced performance. We performed these demonstrations using a single DIS (i.e., one droplet pair) as the analog computing element, generating reservoir states for computing tasks. Notably, the platform is inherently scalable to networks of droplets. Expanding to multiple DIS units (e.g., interconnected droplets or parallel arrays of DIS elements) could considerably increase computational complexity and parallelism, analogous to increasing the number of neurons and synapses in a biological neural network.

Further advancements are necessary to improve the performance, stability, scalability, and energy efficiency of DIS for neuromorphic ionic computing. One key challenge lies in synaptic plasticity retention and sensitivity: The current memory duration remains much too short for long-term plasticity, and the potentiation and depression responses are neither rapid nor pronounced enough to support high-speed computing. A promising strategy to address these limitations is the integration of engineered memristive ion channels into the bilayers to optimize ion conductance (27, 46). However, stable and controlled incorporation of these channels is essential to preserve system robustness and reproducibility. Another limitation arises from the inherently liquid nature of the droplet assemblies in oil, which compromises long-term device durability. Potential solutions include the use of hydrogel scaffolds, polymer-stabilized bilayers, or encapsulated oil phases to enhance structural integrity and extend operational lifetimes (23, 37). Scalability also presents a critical challenge. Current droplet-based systems require precise manual assembly and time-consuming deposition. Emulsion-based self-assembly (77) offers a promising route for fabricating dense, interconnected droplet networks. However, managing network complexity and ensuring reproducibility will be critical for reliable operation in such high-density architectures. Last, energy consumption remains a critical benchmark for artificial synapses and must be a central consideration during system optimization. Strategies such as miniaturizing droplet size to reduce membrane capacitance (37), incorporating efficient ion channels to enhance ionic conductance, and constructing parallel DIS arrays to distribute computational load can collectively enable operation at lower voltages with shorter pulse durations, thereby minimizing both charging energy and dissipative losses per synaptic event. Addressing these critical challenges will be essential for advancing DIS networks toward scalable, energy-efficient, and high-performance neuromorphic ionic computing platforms.

MATERIALS AND METHODS

Materials

The DPhPC lipids and SM were obtained from Avanti Polar Lipids Inc., USA. Potassium chloride (KCl), calcium chloride (CaCl₂), tris base (tris), EDTA, cholesterol, α-HL from Staphylococcus aureus, hexadecane, low-melt agarose gel, and silver wires were sourced from Sigma-Aldrich, USA. The Fluo-8 sodium salt dye was obtained from AAT Bioquest, USA. The Di-8-ANEPPS dye was obtained from Thermo Fisher Scientific, USA.

Fabrication of DIS

Droplet interface bilayer was formed between two aqueous droplets immersed in hexadecane oil and lined with a lipid monolayer using the “lipid in” technique. First, to prepare the ~100-nm-diameter DPhPC large unilamellar vesicles (LUVs), 2 mg of DPhPC lipids dissolved in chloroform was placed into a glass vial. For different bilayer compositions, the total amount of DPhPC, cholesterol, and SM was kept at 2 mg while varying their mass ratios. The solvent was evaporated under an air stream and further dried overnight in a vacuum desiccator. Then, 1 ml of buffer was added to the desiccated lipid film to achieve a final lipid concentration of 2 mg/ml after a 30-s bath sonication. Unless otherwise noted, the buffer solutions in the droplets were 100 mM KCl, 10 mM tris, and 1 mM EDTA at pH 7.5. The mixtures were incubated at ambient temperature for 30 min. To form unilamellar vesicles, the samples underwent 7 freeze-thaw cycles, involving rapid freezing in liquid nitrogen and subsequent thawing at 50°C. The samples were then extruded through 100-nm pore-sized polycarbonate membranes 21 times using a mini-extruder (Avanti Polar Lipids).

Next, two 100-μm-diameter Ag/AgCl electrodes with ball-ended tips were made hydrophilic by coating with low-melt agarose in KCl buffer (3%, w/v). The electrodes were affixed to micromanipulators (NMN-21, Narishige) mounted onto an inverted optical microscope (Leica DMi1) and connected to a patch-clamp amplifier headstage input and ground. Approximately 600-nl droplets of LUV solution were carefully placed on the electrodes in the hexadecane oil bath using a micropipette. For reconstitution of α-HL in the bilayer, a diluted α-HL stock solution (0.5 mg/ml reduced to 1 μg/ml) was added to the LUV solution before droplet formation.

The droplets were incubated for at least 5 min to allow the formation of a self-assembled lipid monolayer. During this process, the droplets sagged slightly away from the electrode, becoming relatively free from strong electrode adhesion. Subsequently, the droplets were gently brought together to form a bilayer at the interface, which was confirmed by optical microscopy imaging and membrane capacitance measurements under an applied triangular voltage wave. The relative freedom of the droplets from the electrode ensures that the electrode-droplet interfaces do not interfere with the bilayer geometry or its structural response under the applied voltages (movie S1).

Electrical measurements

All currents were recorded using an Axopatch 200B patch-clamp amplifier coupled with a 1550B data acquisition system (Molecular Devices), operating at a low band-pass filter frequency of 1 kHz. For all measurements, droplets and measurement probes were placed under a laboratory-made Faraday cage to minimize environmental noise.

To measure the steady-state bilayer capacitance, a low-amplitude triangular voltage wave (fig. S1) with a sampling frequency of 100 kHz was supplied. For each preset voltage, the triangular wave voltage was centered at the preset voltage, and measurements were performed after the bilayer had sufficiently stabilized.

To record the I-V characteristics, custom Axopatch protocols were used to apply periodic triangular voltage waves (e.g., fig. S2). The ion current was then obtained by averaging the ion current traces using a custom Igor Pro script. The sampling frequency during measurements was 1 kHz unless otherwise noted.

For the voltage spike studies, custom Axopatch protocols were used to apply different voltage spike patterns [e.g., potentiation and depression (fig. S12) and Hebbian rule (fig. S15)]. The sampling frequency during these measurements was 1 kHz unless otherwise noted, and the peak value of the generated ion current for each voltage spike was used for the synaptic plasticity analysis.

Calcium dye assay

Two types of LUV solutions were prepared: one containing 1 M KCl, 10 mM tris, and 1 mM EDTA (pH 7.5), and the other containing 0.67 M CaCl₂, 10 mM tris, and 1 mM EDTA (pH 7.5). Fluo-8 (40 μM) was added to the KCl-containing LUV solution before droplet deposition. After the DIS was formed, the left droplet contained 1 M KCl and 40 μM Fluo-8, while the right droplet contained 0.67 M CaCl₂ (to maintain osmotic balance between the two droplets). Fluorescence excitation was performed at 488 nm using a Nikon D-LEDI Fluorescence light-emitting diode system. Images were acquired using an inverted microscope (Nikon Eclipse Ti2-A, CFI60 Plan Fluor 20× objective) and a DS-10 camera. Exposure times ranged from 5 to 100 ms.

4-Bit string measurements

All 2⁴ = 16 unique 4-bit voltage spike sequences were applied to the DIS, with a 30-s interval between each sequence. Each voltage train consisted of four spikes (0.5 s each), with 0.5 s between spikes. Logical 0 and 1 were represented by erase (0 mV, 0.5 s) and write (200 mV, 0.5 s) spikes, respectively. A preset voltage of +130 mV was used before spike application. To measure the ionic current response, five short read pulses (ΔV_s = +10 mV, 0.04 s each) were applied before and after each spike. The average peak current from the five read pulses was taken as the ionic current response of the DIS. This procedure was repeated for five independent DIS devices. The total estimated time for the full measurement protocol was ~2800 s. The normalized ionic currents and corresponding SDs were used to construct the reservoir states that were input into the readout layer for downstream tasks. We also process the 4-bit strings at two additional timescales (fig. S20). In the short-time scenario, each voltage spike lasted 0.1 s with 0.1-s intervals; a single read pulse was used before and after each spike. In the long-time scenario, each spike lasted 1.0 s with 1.0-s intervals; 10 read pulses were used before and after each spike.

Reservoir computing implementation

We implemented a hybrid reservoir computing framework using a single DIS device as the physical reservoir and a software-based readout layer. Terminology related to reservoir computing is summarized in table S2. We emulated multinode behavior by sequentially applying input-encoded voltage-spike patterns and then recording the resulting ionic current responses as reservoir states. The only training was the digital readout, enabling efficient nonlinear processing with flexible output design.

To evaluate classification performance, we used the MNIST dataset and trained a digit classifier. Specifically, we downsampled MNIST digit images (28 × 28 pixels) to 22 × 20 binary pixels and then flattened each image into a 440-bit sequence. We partitioned this sequence into 110 4-bit strings. Each 4-bit input was encoded as a voltage-spike pattern and applied to the DIS. We recorded the ionic current response after each application to generate 110 analog reservoir states per image. We passed these states to a fully connected, single-layer neural network (110 × 10) trained via supervised learning. The readout function used softmax regression and was optimized using cross-entropy loss. We trained for 20 epochs on 40,000 MNIST samples and then evaluated 4000 samples. The training required ~7.3 s on an Intel Core i7-13700F CPU.

For the reinforcement learning task, we encoded the board states from a 3 × 3 tic-tac-toe game into binary strings representing X and O positions. These strings were converted into 4-bit segments and sequentially applied to the single DIS. The corresponding current responses formed the reservoir state vector. This state vector was input into a policy network trained using the PPO algorithm with invalid action masking. The network produced move probabilities and was optimized using advantage-based policy gradients. Reinforcement learning over 1 million steps (about 140,000 games) took ~20 min on an M1 MacBook Pro.

Supplementary Materials

The PDF file includes:

Supplementary Text

Figs. S1 to S22

Tables S1 and S2

Legend for movie S1

References

Other Supplementary Material for this manuscript includes the following:

Movie S1