Neuromorphic tissues: Soft biomolecular networks for brain-inspired temporal computing Authors: Nicholas X. Armendarez, Ahmed S. Mohamed, Md Sakib Hasan, Joseph S. Najem Published: 2026 ABSTRACT Brains achieve extraordinary efficiency in processing temporal information through dense interconnectivity and recurrent feedback among neurons. Inspired by this principle, we introduce neuromorphic tissues—soft biomolecular networks comprising cell-sized aqueous compartments interconnected by lipid membranes containing voltage-gated ion channels. When a compartment is electrically stimulated by current injection, the membranes separating it from neighboring compartments polarize until channel activation occurs, transiently transforming the interface into a conductive synapse that couples adjacent nodes. These dynamics generate intrinsic physical recurrence, enabling the network to encode, propagate, and reconstruct time-dependent signals without external feedback circuitry. Experiments and modeling demonstrate nonlinear, fading-memory, and recurrent dynamics characteristic of reservoir computing, enabling accurate prediction of nonlinear and chaotic sequences such as NARMA-10 and the Lorenz attractor. This work suggests that spatial interconnectivity can enhance the computational capabilities of physical reservoirs and highlights soft, self-assembled materials as a promising platform for implementing such interconnected systems. INTRODUCTION Brains achieve extraordinary efficiency in temporal processing through neuronal interconnectivity and recurrence (1, 2). Recurrence, defined as the continual feedback of signals through coupled pathways (Fig. 1A), endows neural networks with memory and nonlinear dynamics essential for prediction, learning, and control (3, 4). Inspired by this principle, reservoir computing, a subset of recurrent neural networks (5, 6), harnesses nonlinear dynamical systems as untrained “reservoirs” that project inputs into high-dimensional states that are linearly combined at the readout (7, 8). This approach offers an energy-efficient route to time-series inference and classification (9, 10). Comparison of parallel and interconnected reservoir architectures, highlighting biological neural network inspiration and an increase in state-space dimensionality. (A) Biological neural network in the prefrontal cortex, where densely interconnected neurons form thousands of synapses, enabling recurrent dynamics and high-dimensional state spaces (58). The illustration is intended as a high-level functional and structural inspiration and does not imply biological equivalence to our system. (B) Parallel reservoir computing architecture in which uncoupled nodes with fading memory respond independently to inputs (voltage inputs in most cases); corresponding reservoir states are read out through a trainable linear layer. (C) Interconnected reservoir architecture inspired by brain-like recurrence, where nodes are physically coupled to enrich spatiotemporal dynamics and increase the reservoir’s dynamical order. (D) Network of biomolecular memristive devices consisting of cell-sized aqueous compartments interconnected by ion channel–doped lipid bilayers, enabling analog, recurrent dynamics. Each compartment acts as a soma-like node that integrates inputs, while the lipid membranes between droplets serve as adaptive, memory-bearing junctions that mediate nonlinear coupling across the network. Inset: Schematic of the ion channel–doped lipid bilayer, showing phospholipids and inserted peptide channels that form the dynamic interfacial element responsible for coupling between neighboring compartments. Created in BioRender. Najem, J. (2026) https://biorender.com/r28rczr. Physical reservoir computing (PRC) aims to realize this directly in hardware by exploiting the natural dynamics of physical systems (11, 12). Photonic, spintronic, and memristive devices have all been explored (13–15), with memristors in particular attracting interest for their compactness, low power, and tunable nonlinear dynamics (10, 15, 16). However, existing memristor-based reservoirs overwhelmingly use parallel, uncoupled arrays, in which the devices respond independently to inputs (Fig. 1B) (10, 15, 17–22). This parallel architecture is in stark contrast to the highly coupled echo state networks and liquid state machines that are the basis of reservoir computing (7, 23). While parallel systems capture fading memory and nonlinearity—key features for temporal processing—they lack coupled recurrence. Many intrinsic nonlinear dynamics arise only through the coupled interaction of ordinary differential equations (22). For example, in autonomous continuous-time systems, stable limit cycles cannot arise when the number of coupled states is less than two, and chaotic dynamics cannot arise with less than three coupled states (24). In contrast, each uncoupled device operates as an independent first-order dynamical system. As a result, even large arrays do not generate internal state interactions or collective dynamics and therefore lack the higher-order state mixing that characterizes reservoir computation (see Supplementary Texts S1 and S2 as well as fig. S1 for further discussion). A natural path to overcoming this limitation is physical coupling between nodes (Fig. 1C). Interactions among connected nodes introduce internal feedback, richer spatiotemporal dynamics, and enhanced temporal memory capacity (25), inspired by the brain’s massively recurrent wiring (Fig. 1A). Conventional solid-state substrates, being crossbar-based, are not natively configured to couple dynamics between nodes. However, through external circuitry and feedback connections, coupling may be imposed at the system level (26, 27). Furthermore, attempts have been made to introduce intrinsic coupling into memristive materials via nanowire networks (28, 29) and percolation networks (30, 31). These systems are typically modeled as resistive networks, in which the reservoir states are described by voltage measurements at selected probe points within the memristive network. Notably, these works have argued that spatial coupling of physical reservoir states improves upon purely temporal coupling (31) and enables collective dynamics (29). By contrast, soft materials—especially biomolecular, membrane-based memristive devices—offer a similar functionality yet a fundamentally different structural framework (10, 18, 21, 32). Their capacity for compartmentalization, self-assembly, and flexible interfacing enables architectures in which connectivity emerges from material organization and can be adaptively tuned and where both memristive circuitry and capacitive circuitry arise directly from the material composition and architecture. Here, we demonstrate a physically interconnected biomolecular memristor network for reservoir computing. These assemblies consist of tissue-like, cell-sized aqueous compartments connected by ion channel–doped phospholipid membranes. Information injected as current into a compartment polarizes the shared membranes, analogous to signal integration at neuronal somas. When a membrane’s voltage reaches its threshold, ion channels open, transforming the membrane into a dynamic, conductive synapse with memory. Through these transient synapses, neighboring compartments exchange charge and information, allowing signals to propagate, interfere, and recur (Fig. 1D). Through experiments and modeling, we demonstrate that this coupling enables accurate prediction of nonlinear and chaotic signals (e.g., NARMA-10 and Lorenz attractor) that were previously unattainable with parallel memristor-based reservoirs (see Supplementary Text S3 and fig. S2) (9, 10, 15, 17, 19, 21, 33–36). By leveraging self-assembly, modularity, and biocompatibility, these networks overcome long-standing challenges of scalability in membrane-based systems (fig. S1) and achieve functional diversity, pointing toward deployable, energy-efficient computing for adaptive sensors, soft robotics, and biointegrated systems. Our results support physical interconnectivity as a useful design principle for soft neuromorphic systems and introduce “neuromorphic tissues” as a material platform in which recurrence emerges from composition and architecture. This demonstration of a recurrent physical memristive reservoir suggests a pathway toward realizing recurrent neural network–like dynamics directly in material form, with potential relevance to brain-inspired computation at the interface of living and artificial systems. RESULTS Design and modeling of current-driven biomolecular networks To establish the foundation of our neuromorphic tissues, we first characterized the current-driven dynamics of single biomolecular memristive devices formed as droplet interface bilayers (DIBs) doped with voltage-gated Alamethicin (Alm) peptides (fig. S3) (32, 37). These devices remain electrically insulating until the applied voltage exceeds the activation threshold of pore-forming Alm peptides, which usually depends on the peptide concentration and membrane composition (Fig. 2A) (32, 37). Once activated (Fig. 2B), the channels insert into the membrane from the cis droplet (where the positive voltage electrode is inserted) to the trans droplet (where the ground electrode is connected) and exhibit strongly nonlinear current-voltage characteristics with volatile memory hallmarks of memristive behavior (Fig. 2C) (32, 37). Such fading memory and nonlinearity are fundamental requirements for reservoir computing (38–40). Our prior studies demonstrated that Alm-based devices, when organized into parallel-node architectures, can function as physical reservoirs capable of solving standard temporal benchmark problems (9, 10, 18, 20, 34, 41). Moreover, the modular nature of DIBs allows structural and compositional asymmetry between cis and trans droplets (fig. S4) (41, 42), as well as independent tuning of channel concentration and lipid formulation, allowing heterogeneous and versatile reservoirs. Current-driven networks. (A) Single device at subthreshold bias (vmvth), Alm peptides insert into the bilayer, forming voltage-gated ion channels that increase conductance; the membrane can be modeled as a memristive element M(Na,Vm) in parallel with Cm. Electrical measurements are made through embedded Ag/AgCl electrodes (see Materials and Methods for more details). (C) Voltage-driven current-voltage (I-V) curves of single devices at 1 and 2 M Alm concentrations, showing threshold-like channel activation and nonlinear current increases. (D) In a parallel reservoir of uncoupled devices, voltage input and transmembrane current readout are practical because each node’s accessible state, Na,i, can be inferred from current. (E) Current-driven network and equivalent circuit representation. The network is operated with a current input and voltage readout using the membrane capacitance of each bilayer (transparent droplets shown behind the circuit). The accessible reservoir states are the nodal voltages Vi measured with respect to a common ground. Each membrane is modeled as M(Na,1−2,ΔV1−2) with parallel capacitance and resistance, and the nodal voltages are recorded via ADC channels under current stimulation. (F) Current-driven I-V responses under the same concentrations as (C), highlighting memristive behavior and overshoot resulting from capacitance charging. (G) A DPhPC-Alm membrane is stimulated with a constant 2-nA current input. The response is characterized by fast, stochastic oscillations around the steady state. (H) A BTLE-Alm device is likewise stimulated with a 2-nA current. The response is noticeably more stable because of the longer lifetime of pores in the BTLE membrane. Created in BioRender. Najem, J. (2026) https://biorender.com/r28rczr. In earlier implementations (10, 18, 32), Alm-based devices, either isolated or arranged in parallel (Fig. 2D), were operated in voltage-driven mode, with output current used to calculate the areal pore density Na, the primary state variable. This approach is practical for uncoupled devices, where each current directly reflects the reservoir state; however, it becomes impractical for the interconnected networks we introduce here. In parallel architectures, each Alm-device is addressed in series with its own current path, allowing Na,i to be inferred from its local current (Fig. 2D). In contrast, in our coupled networks, the compartments and their membranes share conductive pathways, preventing electrical isolation of individual junction currents and rendering Na,i experimentally inaccessible (fig. S5). To address this limitation, we use a current-driven operation inspired by neuronal models, where we encode input information as current injections (Fig. 2E). Although the Alm channels themselves are not directly activated by current, the lipid membrane is intrinsically capacitive (32, 42). Figure 2F compares current-voltage responses in current-driven mode with conventional voltage-driven operation. Injected current polarizes the membrane until the transmembrane voltage surpasses the insertion threshold, triggering channel formation and memristive switching. For single-device characterization, we examined setups with Alm-doped membranes of two lipid compositions: 1,2-diphytanoyl-sn-glycero-3-phosphocholine (DPhPC) and brain total lipid extract (BTLE) (32). Voltage measurements showed that DPhPC membranes, which lack cholesterol, exhibited pronounced stochastic fluctuations under a 2-nA step current (Fig. 2G). In contrast, BTLE membranes, which naturally contain cholesterol, showed a sharper overshoot but more stable steady-state behavior (Fig. 2H). This is consistent with the well-established role of cholesterol in increasing bilayer thickness and stiffness (43), thereby extending the mean lifetimes of both subconductance states and single-channel bursts (44). These longer channel lifetimes suppress stochastic fluctuations and explain the improved agreement of BTLE membranes with model predictions during step-sequence stimulation (fig. S6). In this current-driven framework, we describe each membrane by two coupled state variables: pore density, Na, which governs the conductance, and transmembrane voltage, vm, which we can directly measure at each droplet node. Therefore, while Na remains a hidden internal state, vm provides an experimentally accessible proxy for network dynamics. Beyond improved observability, the current-driven mode offers two additional benefits. First, it minimizes the risk of electroporation or membrane rupture, which can occur when voltages exceed ∼200 to 300 mV (45). Because current injection allows voltages to gradually rise to the insertion threshold on a timescale comparable to channel formation, abrupt spikes are unlikely to occur that could cause the membrane to fail. Second, this configuration simplifies network readout: Reservoir states can be captured directly as node voltages, eliminating the need for transimpedance amplifiers that were previously required to convert currents into voltages in parallel architectures (10). To capture the current-driven operation of single Alm devices and an arbitrary interconnected network of these devices, we developed a physics-based state-space framework that unifies their electrochemical and biophysical dynamics (full derivations and parameter definitions are provided in Supplementary Text S4). For a single device, modeled as a capacitor in parallel with an Alm memristor, the dynamics reduce tov˙m=hεA(Iinj−vmGuNaA)(1)N˙a=α(vm)Na1−Na/Nss(vm)Z(2)where vm is the transmembrane voltage of the bilayer, Iinj is the injected current, A is the membrane area, h is the membrane thickness, ε is the permittivity, Gu is the mean single-pore conductance of Alm (37), and α(vm),Nss(vm), and Z are voltage-dependent channel-insertion parameters. Here, α(vm) describes the voltage-dependent pore insertion rate, typically increasing exponentially with voltage (10, 20, 32, 37); Nss(vm) is the corresponding steady-state pore density following a sigmoidal voltage dependence (18, 20); and the exponent Z controls the asymmetry of the insertion kinetics (20). For a network of L droplets with node voltages Vi referenced to ground, we define the state vectorx=[VN],x˙=[V˙N˙](3)where V=[V1,V2,…,VL]T contains the voltage of each droplet i, and N contains the nonzero areal pore densities Nij, where the first index i denotes the droplet for which the equation is describing and the second index j denotes its neighboring droplet across a shared membrane. The voltage dynamics followV˙=M−1B(4)with the capacitance-weighted coupling matrixMij=∑k=1LεAikh,i=j,−εAijh,i≠j,0,if nodeiis fixed to ground or unconnectedwhere the diagonal of M is given by the summation of the capacitance of existing membranes connected to a node. We use the summation index k to define the summation over the rows of A. The driving vector is given byBi=Ii−∑j=1LGuAij(Nij+Nji)(Vi−Vj) Each membrane side obeysN˙ij=αij(Vi,Vj)Nij1−Nij/Nss,ij(Vi,Vj)Zij(5)where αij(Vi,Vj) and Nss,ij(Vi,Vj) are the local voltage-dependent insertion rate and steady-state pore density, respectively. The matrix Aij serves as an area-weighted adjacency matrix (see fig. S7), zero where no membrane exists and equal to the physical interfacial area where a connection is present, thereby governing both capacitive and memristive couplings. This formulation naturally collapses to the single-device case (Eqs. 1 and 2) when only two nodes are present, highlighting the nonlinear coupling of voltage and pore-density states in larger networks. Unlike parallel configurations, where each element evolves independently, these nonlinear couplings enable richer, higher-order dynamics essential for reservoir computation. To validate the biomolecular network model, we assembled a six-node chain of Alm devices (Fig. 3A) with the membranes consisting of synthetic lipid DPhPC. We then stimulated the network with a discrete current input signal, Iu. The same signal was also applied in numerical simulations of the model. Figure 3A shows the conceptual design of the droplet chain, while Fig. 3B presents a microscope image of the experimental network (see Materials and Methods for fabrication details). The input (Fig. 3C, top) was designed to mimic a discrete-time series commonly used in reservoir computing benchmarks (10, 15) and was defined asIu(kΔt:kΔt+Δt)=γu(k)+δ(6)where k is the discrete index, u(k) is a uniform random variable between 0 and 0.5, Δt is the pulse width, γ is the scaling factor, and δ is the bias. Here, Δt was set to 3 ms, γ to 170 nA, and δ to 0. Figure 3C compares the applied input (top), the simulated network response (middle), and the experimental response (bottom). Comparison between simulated and experimental biomolecular networks voltage outputs. (A) The conceptual biomolecular network has six measurable voltage states and current injection. (B) The experimental biomolecular network achieves individual voltage probing via Ag/AgCl electrodes. (C) The input signal encoded between 0 and 170 nA is shown in the top panel. The middle panel shows the simulated response of the six-droplet network to the input stimulation. The bottom panel shows the experimental results of the same network construction. Note that voltages are measured relative to the common signal ground, so they stack additively for membranes spatially far from the ground node. Created in BioRender. Najem, J. (2026) https://biorender.com/r28rczr. The simulated and experimental voltage outputs across the six nodes showed close agreement, particularly at higher voltages, with an average normalized mean squared error (see Supplementary Text S6) of ∼0.41 across 100 step changes in Iu. Although single DPhPC membranes typically exhibit substantial stochastic fluctuations [corresponding to a normalized mean squared error of 1.15 between simulation and experiment for a similar input signal (fig. S8)], the chain response appears markedly less noisy, suggesting that internode coupling averages out random activity. Notably, this stochasticity originates from the inherent stochastic opening and closing of Alm pores (32, 37), which dominate the steady-state response at the single-membrane level. As a control, we simulated the response of a chain of parallel resistor-capacitor (RC) compartments representing undoped DIBs with no Alm pores (fig. S8). For appropriately chosen Rm and Cm values that reflect the electrical properties of DIBs in hexadecane oil (41, 42), the RC-only model exhibited continuous, noisy charging behavior in response to the random input current. Because the membrane charging timescale is much slower than that of the input variations, the RC response effectively tracked only the root mean square of the signal rather than its rapid fluctuations. By contrast, the Alm-doped device chain produced voltage responses that closely followed changes in the input, as the input timescale (1 to 10 ms) matched that of ion-channel insertion and removal dynamics. This comparison confirms that the experimentally observed dynamics arise from the memristive pore kinetics rather than from passive RC charging effects. Together, these results demonstrate that the current-driven device model faithfully reproduces the dynamics of coupled biomolecular networks and that stochastic pore activity, although evident in single devices, is substantially reduced in coupled chains. Time-series prediction using neuromorphic tissues NARMA-2-10 time-series prediction task With an experimentally validated model of the current-driven biomolecular network in hand, we evaluated its utility as physical reservoir layers by simulating nonlinear autoregressive moving average (NARMA) benchmark tasks (46, 47). NARMA tasks are standard time-series prediction problems used to assess reservoir computing performance at increasing levels of difficulty. The order-R NARMA task is defined asy(k+1)=0.3y(k)+0.05y(k)∑i=0Ry(k−i)+1.5u(k−R+1)u(k)+0.1(7)where k denotes the discrete time-step index, u(k) is a random input sequence uniformly distributed in [0, 0.5], and R is the task order. Successful prediction requires the reservoir to retain the memory of the previous R states while also performing nonlinear transformations of the input, capabilities that generally require recurrence within the reservoir (7, 47). This makes higher-order NARMA tasks, such as NARMA-10, particularly challenging for first-order, parallel-node PRC architectures (48). To our knowledge, memristor-based reservoirs, implemented as parallel arrays, have been limited to the intrinsic order of single devices, typically solving only NARMA-2 or a similarly difficult SONDS (second-order nonlinear dynamical system) task (12, 49). We encoded u(k) as a discrete current waveform injected according to Eq. 6, with the input bounded between IL and IU. Many encoding parameters (pulse width, duty cycle, and amplitude range) mirror those used in parallel-node PRCs; however, interconnectivity introduces additional hyperparameters, including topology, interfacial membrane area between nodes, node-specific Alm concentration, and injection sites. Although our framework supports arbitrary topologies, here, we used a one-dimensional chain with current injection at each node and a grounded terminal at the end, similar to the setup in Fig. 3 (A and B). Chains comprised N nodes with approximately constant interfacial areas; Alm concentrations followed a repeating 1, 2, and 3 μM pattern. The primary variables were the injected current range, the number of spatial nodes, and the number of virtual nodes. Because the reservoir exhibits long transients, we drove it with 3000 training and 3000 testing inputs, utrain(k) and utest(k), but retained only the final K=800 steps of each for training and evaluation. We defined the reservoir states as the node voltages we sampled at the end of each pulse Δt, forming a state matrix X(i,k), where i is the node index and k is the time step. To enhance separability, we augmented features with squared states, which is an established strategy (48) with NARMA tasks, and quantified performance by the testing normalized root mean square error (NRMSE; see Supplementary Text S6). Following Appeltant et al. (38), we also introduced virtual nodes via temporal masking (36) and subdivided each input pulse into M segments of length θ, with a predefined mask modulating each segment (Fig. 4A). We fixed θ to be 1 ms such that the full masked pulse width is M×θ. Doing so increases the number of effective states per input from N to N×M by leveraging delayed self-dynamics. Increasing M boosts dimensionality but introduces a memory/decorrelation trade-off: Overly long masks degrade the recall of past inputs, which is detrimental for high-order NARMA tasks (36, 38). Simulated benchmarking NARMA tasks with interconnected versus parallel reservoirs. (A) Encoding scheme for NARMA tasks. The discrete signal u(k) is a random sequence generated uniformly between [0, 0.5] and mapped to a continuous-time signal Iinj in the range [Imin, Imax]. This signal is temporally multiplexed with a binary mask: A value of 1 leaves the signal unchanged, while a value of 0 flips the encoding range to [Imax, Imin]. The multiplexed current signal is applied to the simulated reservoir, and reservoir voltage states are sampled at discrete times corresponding to changes in the input. A trained weighted sum of these states produces the prediction yˆ(k) of the true discrete output y(k) from the NARMA task equation. (B) Training and testing results for NARMA-10 using the connected network reservoir. (C) Training and testing results for NARMA-10 using a simulated parallel reservoir. (D) Comparison of connected (blue) and parallel (red) reservoir architectures. For higher-order tasks, the connected reservoir dynamics substantially outperform simple parallel reservoirs. The training and testing NRMSE values are obtained from the data in (B) and (C), respectively. (E) Contour map from a coarse hyperparameter search, comparing NARMA-10 testing NRMSE across mask lengths and network sizes for Iinj∈(0,0.1) nA. The lowest NRMSE (red box) corresponds to the data in (B). Contour maps for other current ranges are shown in fig. S9. Created in BioRender. Najem, J. (2026) https://biorender.com/r28rczr. The optimal configuration for the NARMA tasks consisted of N = 50 spatial nodes and M = 3 virtual nodes (150 total states). On NARMA-2, the connected reservoir achieved training and testing NRMSEs of 0.0103 and 0.0124, respectively. In contrast, the parallel reservoir architecture (see Supplementary Text S5) produced substantially higher errors, with training and testing NRMSEs of 0.433 and 0.458, respectively. On NARMA-10, the connected reservoir reached training and testing NRMSEs of 0.236 and 0.327, respectively (Fig. 4B), while the parallel reservoir produced much higher errors of 0.546 and 0.627, respectively (Fig. 4C). These results highlight the superior memory capacity of the connected network, specifically its ability not only to retain information about past inputs but also to reconstruct them from its current state. Such reconstructive memory is essential for reservoir performance, as NARMA prediction requires recalling and transforming multiple past inputs to generate the correct output. By comparison, parallel memristor-based reservoirs are fundamentally capped by the intrinsic order of a single device, which severely limits their effective memory capacity and prevents them from tackling higher-order tasks. Figure 4D presents a comparative summary of the training and testing NRMSEs for the NARMA-2-10 tasks. It is worth noting that we identified the optimal configuration through a coarse parameter sweep of possible network hyperparameters, including mask length, the number of spatial nodes, and the input signal encoding range. Figure 4E shows the NRMSE contour plot for the NARMA-10 task as a function of mask length and the number of nodes at the optimal current range [Iinj∈(0,0.1)nA], while we provide corresponding contours for other current ranges in fig. S9. Next, we compared the heterogeneous reservoir described earlier, composed of droplets with a repeating pattern of 1, 2, and 3 μM Alm concentrations, to homogeneous reservoirs, where all nodes had a fixed concentration of either 1, 2, or 3 μM (three separate cases). This comparison enables a direct assessment of how heterogeneity and device modularity influence the computing performance of our devices. As shown in fig. S10, heterogeneous reservoirs consistently outperformed their homogeneous counterparts on the NARMA tasks. This finding aligns with a prior memristor study (10) and can be understood in terms of information processing capacity: Heterogeneity reduces redundancy among node dynamics, yielding a richer, less overlapping set of nonlinear transformations (i.e., broadens the nonlinear state space), thereby expanding the system’s effective computational capacity (25, 39). To test whether this improvement requires a structured heterogeneity pattern, we also constructed fully random heterogeneous reservoirs, where each droplet concentration was independently sampled between 1 and 3 μM. As shown in fig. S11, these random networks achieved comparable performance to the structured 1-2-3 μM pattern, suggesting that beyond a certain threshold, additional heterogeneity provides diminishing returns. This observation is consistent with the view that the primary benefit of heterogeneity is to avoid redundancy and maximize the span of independent functions available for projection (50). Lorenz63 chaotic time-series prediction task While the NARMA benchmark tasks are widely used to assess memory and nonlinear processing capabilities of reservoir computing systems, they are fundamentally scalar and synthetic, with well-defined input-output dependencies. Although solving higher-order NARMA tasks (e.g., NARMA-10) demonstrates that a reservoir can retain and nonlinearly process information over extended timescales, these tasks do not fully capture the complexity of real-world dynamical systems, especially those with higher-dimensional, coupled states (46). To more rigorously evaluate the computational richness of our neuromorphic tissues, we turn to the Lorenz63 task, a canonical chaotic time-series prediction benchmark that involves the simultaneous prediction of three tightly coupled, nonlinear state variables. Unlike NARMA tasks, the Lorenz system requires the reservoir to process and integrate multidimensional inputs in a coordinated way, making it a more stringent test of the network’s ability to model intertwined temporal dependencies and reproduce complex attractor dynamics. Success in the Lorenz63 task indicates that the reservoir not only has long memory and nonlinearity but also the spatial interconnectivity required to model multi-input, multivariate dynamical systems, a capability that purely parallel device architectures inherently lack. To this end, we specifically examined the Lorenz63 chaotic time-series prediction task. The goal of this benchmark is to perform a one-step-ahead prediction of a system governed by three coupled, nonlinear ordinary differential equations that are highly sensitive to initial conditions. Notably, this task relies on the three coupled states of the system as input to predict the next step in the system’s evolution. This is in contrast to NARMA tasks, where the input can be encoded as a single waveform. We expect uncoupled state reservoirs to perform poorly in this task, as they lack the coupled dynamics inherent in the ground-truth equations. However, we hypothesize that a sufficiently complex coupled reservoir should be able to linearly map the dynamics of the present state to the one-step-ahead predicted state. The equations give the ground truth for the Lorenz63 systemdxdt=σ(y−x)(8)dydt=x(ρ−z)−y(9)dzdt=xy−βz(10)where x, y, and z represent the three states of the system, and constants σ, ρ, and β are set to 10, 28, and 83, respectively. With these parameters and the nonlinear coupling of the equations, the Lorenz63 system displays autonomous, chaotic oscillations. Using a fourth-order Runge-Kutta numerical integration scheme with a time step size of 10−4, we generated two datasets: one for training with initial conditions x = 1, y = 1, and z = 1 and one for testing with initial conditions x = 1.01, y = 0.99, and z = 1.01. Because of the sensitivity to initial conditions of Lorenz63, these 1% changes in initial conditions lead to substantial changes in the oscillation of the chaotic system despite the same basic attractor dynamics. We trained a hexagonal lattice reservoir network with the first dataset by injecting the Lorenz states as nanoampere-level currents and subsequently evaluated the trained readout following the workflow illustrated in Fig. 5A. Lorenz system forecasting across x, y, and z states. (A) Input encoding scheme. Ground-truth Lorenz trajectories are generated by numerical integration, finely discretized, and linearly scaled into three current signals Iinj,x, Iinj,y, and Iinj,z. These signals are injected into three locations of the simulated connected reservoir [detailed in (B)]. Reservoir voltage states are recorded, and a trained weighted sum of these states predicts the next time step xˆ, yˆ, and zˆ. (B) Training dataset showing both true and predicted responses for all three states. (C) Testing dataset demonstrating that readout weights learned during training generalize to new inputs. (D and E) Three-state attractors from training (D) and testing (E) show consistent agreement with ground-truth dynamics. (F) A comparison across network sizes demonstrates that larger networks with more states generally improve training and testing performance. Created in BioRender. Najem, J. (2026) https://biorender.com/r28rczr. Unlike the NARMA tasks, there are three distinct inputs to the Lorenz63 task. These three inputs need to be injected into the physical reservoir simultaneously. While the Lorenz63 system is in arbitrary units, we encode the input signal such that 1 time unit of Lorenz63 maps to 1 s of injected current. Furthermore, the injected signals were linearly scaled to a current range suitable for the biomembrane devices: −50 to +50 nA. We used a branched network topology, with three current injection nodes (one for each signal) and one grounded central node. Figure 5A (under “Reservoir”) shows a schematic representation of the network used for this task. To map the network states to the next Lorenz states x, y, and z, we used Ridge regression at the readout. After training, the network is reset and stimulated with the state of the testing dataset. During this phase, the reservoir output at each time step represents the predicted next-time-step value of the Lorenz system. Figure 5 (B and C) shows the predicted and actual states of the Lorenz 63 system for training and testing, respectively, demonstrating that the predicted outputs closely match the ground truth over a 40-s test period. The NRMSEs for x, y, and z were 0.071, 0.183, and 0.174, respectively, indicating that the reservoir could successfully learn and reproduce the complex, coupled dynamics of the Lorenz system. Figure 5 (B and C) contains the training and testing errors for the optimal 43-node network. The consistently low NRMSE values across all three Lorenz states demonstrate that the connected reservoir can reliably learn and generalize coupled, nonlinear dynamics when provided with a sufficiently large network size. In addition, to visualize the quality of the predictions in the reconstructed phase space, Fig. 5 (D and E) compares the predicted and ground truth trajectories of the Lorenz attractor. During both training and testing, the predicted trajectories closely tracked the true attractor structure, faithfully capturing the system’s nonlinear oscillations and chaotic switching between lobes. Notably, the reconstructed attractors retained the geometric features of the ground truth, confirming that the reservoir captures not only pointwise predictions but also the underlying structure of the Lorenz dynamics. To evaluate the effect of network size on prediction accuracy, we compared reservoirs with 13, 25, and 43 nodes (Fig. 5F). Each smaller network was obtained by removing one or more outer hexagonal cycles from the 43-node lattice, thereby preserving the same local connectivity while reducing the overall system size. As expected, increasing the number of nodes improved prediction performance by providing a larger, higher-dimensional state space. This improvement was most pronounced for the y and z states, where testing NRMSEs decreased from 0.244 and 0.222 in the 13-node network to 0.183 and 0.174 in the 43-node network, respectively. The x state also showed a marked improvement, reaching a testing NRMSE of 0.071 in the 43-node case. These results highlight how scaling the physical reservoir directly enhances representational capacity and prediction accuracy for complex coupled systems. While scaling clearly improved performance, even the smallest connected networks produced reasonable predictions, underscoring the fact that internode coupling is the critical enabler of computation in these systems. To highlight this point, we also evaluated a simulated parallel (uncoupled) reservoir on the Lorenz task on the basis of the simulation framework established in our previous work (see Supplementary Text S5) (10, 20). As shown in fig. S12, the parallel architecture performed very poorly, with NRMSE values exceeding 0.6 to 0.8 for the training set and degrading further on the testing set. The reconstructed attractors quickly diverged from the true trajectories, failing to capture the underlying chaotic dynamics. These results confirm that without physical recurrence, parallel architectures are fundamentally inadequate for solving tasks with multiple coupled inputs such as the Lorenz system. Similar to the NARMA tasks, we also examined the role of heterogeneity in the Lorenz prediction task. In the main results (Fig. 5, B to F), we used a repeating pattern of 1, 2, and 3 μM Alm concentrations across the droplets. To probe this further, we tested three homogeneous networks composed entirely of 1, 2, or 3 μM Alm droplets (figs. S13 to S15). These homogeneous networks performed notably worse, with testing NRMSE values much higher than those of the heterogeneous reservoir. We also examined networks with randomly assigned Alm concentrations between 1 and 3 μM (fig. S16), which yielded performance comparable to the structured 1-2-3 pattern. These results reinforce heterogeneity as a key design principle for soft-matter reservoirs. By diversifying local node dynamics, heterogeneity reduces redundancy in the reservoir state space, enabling a richer, higher-dimensional projection of inputs. This enhanced dynamical diversity enhances the reservoir’s ability to reconstruct complex, coupled signals, such as the Lorenz system. More broadly, our findings suggest that incorporating heterogeneity will be a central guiding principle in the future design of interconnected reservoirs and will be essential for realizing scalable, high-performance soft neuromorphic systems. DISCUSSION Our work establishes physical interconnectivity as a design principle for reservoir computing in soft materials, demonstrating that biomolecular networks can serve as an effective substrate for brain-inspired, recurrent computation. By transitioning from parallel, isolated devices to lattices of aqueous compartments coupled by phospholipid membranes (Fig. 2), we embed recurrence and higher-order dynamics directly within the network’s physics while preserving the soft-matter advantages of self-assembly, modularity, and biocompatibility. Unlike conventional crossbar arrays, these architectures feature circuitry intrinsic to the material itself, emerging spontaneously from droplet interfaces rather than being externally fabricated. A current-driven operating scheme, combined with a state-space model coupling voltage and pore density, enables direct voltage readout at each node and eliminates the need for per-node amplifiers typically used in parallel architectures. Experimental measurements on custom biomolecular networks closely matched simulations (Fig. 3), validating this framework across scales. These advances enable benchmarks that were previously unattainable for memristor-based reservoirs. Our connected networks solved NARMA-10 (Fig. 4) with a lower error than parallel baselines and accurately reproduced the three coupled states of the Lorenz63 chaotic system (Fig. 5). Together, these results demonstrate that coupling confers reconstructive memory and richer state evolution, both of which are essential for higher-order temporal inference and multi-input dynamical prediction. From these findings, several design principles emerge. Heterogeneity in node dynamics reduces redundancy and broadens the nonlinear transformations available to the readout, systematically improving performance (figs. S10 to S16). Scaling the number of nodes expands representational capacity and reduces error, while accurate prediction in small lattices underscores that coupling, not size alone, is the critical enabler. Inherent compartmentalization further permits asymmetric compositions and node-level customization, offering a degree of flexibility unavailable to rigid solid-state devices (15, 19, 26, 27). Looking forward, opportunities exist at multiple levels. At the device level, systematic exploration of lipid compositions and ion-channel libraries can extend the tunability of individual nodes, with cholesterol or alternative peptides providing diverse kinetics and nonmonotonic behaviors (41, 51). Leveraging Nernst potential gradients could introduce excitatory-inhibitory balance and spiking-like behaviors (37, 52). At the network level, understanding how topology shapes performance will be critical, especially as three-dimensional (3D) architectures and current droplet printing methods (53) enable richer recurrence and higher volumetric density, more closely resembling brain-like tissue organization. At the system level, compact, low-power drivers and in-material readouts could reduce reliance on rigid electronics. At the same time, online learning and closed-loop adaptation may enable real-world applications such as chaotic forecasting, temporal classification, and bio-interfacing. Together, these advances position our neuromorphic tissues as a promising material platform for embedding computation directly in matter, offering a potential path toward scalable and adaptive brain-inspired computing systems at the interface of living and artificial intelligence. MATERIALS AND METHODS Biomolecular network experimental apparatus fabrication The following methodology is adapted from Armendarez (54). To fabricate the main container for the connected reservoir network, we used a two-sided molding technique with 3D-printed molds (55). The mold design was created using Autodesk Inventor (formerly AutoCAD). Slicing for 3D printing was done using Lychee Slicer (Mango3D). All molds in this work were printed using high-resolution stereolithography digital light projection 3D printers (Mono S3, Anycubic). The printer used has a 34.4-μm x/y resolution and as little as 10-μm z resolution. All molds were printed using Grey Craftsman Resin (Anycubic). All molds were printed flat relative to the print bed, with features in the x/y plane. This orientation restriction was required to prevent aliasing caused by the discrete layers of the 3D print when printed at nonorthogonal orientations. The mold designs are shown in fig. S17 (A and B). The main features of the top mold are the rows of “egg crates” features on the bottom of a relatively high-walled container. These egg crates held one aqueous droplet each, and the surrounding container was filled with oil. The bottom mold features narrow pillars that align with the bottom of each egg crate in the top mold. These pillars will be the vias that allow Ag/AgCl connections to be made to each droplet individually. The process of 3D-printed mold fabrication contains four steps: printing, washing, postcure, and postbake. The first three steps are standard stereolithography printing procedures: The mold is printed, removed from the build plate, washed in 99% isopropyl alcohol, allowed to dry in ambient conditions, and then postcured at 60°C in exposure to 405-nm light for 1 hour. This postcure step was done in an ultraviolet cure station (Form Cure, Formlabs). The final major step of mold preparation is to bake at an elevated temperature for a more extended period, specifically at 70°C, for at least 12 hours. This final step is critical for polydimethylsiloxane (PDMS) casting, as, without proper postbaking, leachates from the 3D-printed resin can prevent the curing of PDMS (56). The 10:1 PDMS mixture is poured into the top and bottom molds, which are then degassed in a vacuum oven for 1 min to prevent air bubbles from creating voids in the final structure. The molds are then removed from the vacuum and assembled. This assembly is placed in a bench vise and tightened until no PDMS leaks from the molds. The bench vise is then placed in an oven at 60°C for 2 hours to cure PDMS. The assembly is then removed from the vise and separated to obtain the main structure of the connected reservoir apparatus. These major steps are visualized in fig. S17. Separately, 11 silver wires (150-μm diameter; Sigma-Aldrich) were cut to ∼5-cm lengths. These wires were then held to an open flame on one end to cause the silver to bead up to ∼300 μm at one end. The wires were then dipped in bleach (Clorox) for 30 min to produce a layer of Ag/AgCl at the beaded tips. These wires were manually fed through the vias in the egg crates left by the bottom mold pillars such that the balled-up and chlorinated ends were within the egg crates. Figure 3B shows the wires inserted into the egg crate structure. These wires were then adhered into place with ultraviolet-cured epoxy, and the entire assembly was set onto a glass substrate. The silver wires were connected to the voltage-controlled current source (CS580, Stanford Research Systems), the multichannel 24-bit analog-to-digital converter (ADC; NI-9202, National Instruments), or the common ground via alligator cables. Experimental biomolecular network preparation Single-device experiments using the hanging-droplet method were performed following the procedures described in detail in (18, 32, 57). The apparatus described in the previous section was prepared for droplet insertion by filling the main container with hexadecane oil. Three concentrations of Alm-containing solution were prepared using a base stock of DPhPC (2 mg/ml), 1 M KCl, and 10 mM Mops buffer using procedures described in detail in (18, 32, 57). The three solutions contained 1, 2, and 3 μM Alm concentrations, respectively. Next, ∼500 nl of aqueous solution was pipetted into empty extra egg crates. After waiting at least 10 min past droplet deposition, the droplets were moved to an electrode-containing egg crate. This wait time was to allow lipid monolayer formation in the oil. When the droplets were transferred onto the electrode-containing egg crates, they came into contact with each other and formed DIBs. Droplets were dispensed such that the following sequence of droplets was repeated: 1, 2, and 3 μM. The full sequence of 10 droplets used for experiments was [1, 2, 3, 1, 2, 3, 1]. As shown in Fig. 3A, the current source is connected to the first droplet in the sequence, and the ground connection is at the last droplet in the sequence. Each electrode in between is connected to an ADC in reference to the signal ground. A custom Python script was used to generate a square-wave input signal with a sampling rate of 10 kHz, a range of 0 to 1.7 V, and a pulse width of 3 ms. The script then outputs this signal to a voltage DAC (NI-9264, National Instruments) connected to the voltage-controlled current source. The current source accepts a voltage input between 0 and 2 V and converts it to a current signal with a selected scaling coefficient. For these experiments, we selected a range of 100 nA per 1 V such that the output current range was 0 to 170 nA based on the 0- to 1.7-V input to the current source. Simultaneously, the script controls the input voltage measurements from the multichannel ADC input from the six droplet electrodes. These voltages are read directly in the 0- to 10-V range without amplification or filtering.