A Brain-Inspired Tissue for Forecasting Chaos

A Cluster of Droplets Under Glass

On a bench at Penn State sits a shallow bath of hexadecane oil. Inside it floats a honeycomb of water droplets. Each one, though, is about the size of a large cell. Silver wires rise through the floor of the container. Their tips are beaded and chlorinated, so each wire can carry signals into and out of one droplet.

Where two droplets touch, they do not merge. Instead, the fatty molecules coating each one zip together into a single membrane. That membrane is two molecules thick, the same architecture that wraps every living cell.

Then a tiny current arrives, measured in billionths of an ampere. Charge accumulates. Voltage climbs. At a threshold, peptides embedded in the fat spring open into conductive pores. So the wall between two droplets stops being a wall. Signals cross, ripple outward, bounce off neighbors, and return. Moments later the pores close and the memory of the event fades.

The Claim in One Paragraph

That returning signal is the whole point. Researchers at Penn State and the University of Mississippi wired their droplets to one another through shared membranes. Nothing, therefore, sits in isolation. As a result, the material can forecast a chaotic system called the Lorenz attractor. Arrays of uncoupled devices have not matched that accuracy. The authors call this material a neuromorphic tissue. By that they mean a soft substance that computes the way a network of neurons does, through local interactions and fading memory. Their results appeared in Science Advances on 23 July 2026, after a review that ran from October 2025 to June 2026.

What Makes This Different From a Chip

Most neuromorphic hardware reproduces brain behavior using transistors, memristors or spiking circuits etched into silicon. This work, however, takes another route. It uses a soft droplet network that assembles itself. The electrical interactions inside that network supply the memory, the nonlinearity and the coupling. A digital processor moves information through fixed logic gates made in advance. This system lets the physics of the material carry out part of the computation instead.

That property suits it to reservoir computing. The plain-language version goes like this. First, the material does the hard part, scrambling incoming signals into a rich set of internal states. Then a small output layer gets trained on top. No backpropagation through millions of weights. No gradient descent inside the substrate. The physics handles the transformation and a simple linear readout handles the answer.

Why droplets, though? The appeal is not simply their small size, though. They assemble themselves into a network whose connections live in the material rather than in a fabrication step. Better still, the behavior of any single junction responds to the chemistry of one droplet.

Why Recurrence Matters More Than Node Count

Brains do not compute by lining up neurons side by side. Signals travel in loops and feed back through coupled pathways. A cortical circuit therefore carries traces of inputs that arrived many steps earlier. That property is recurrence. It therefore gives neural systems memory and the nonlinear mixing that prediction and control require.

The Weakness in Parallel Arrays

However, most memristor based reservoirs share a structural weakness. The devices sit in parallel arrays. Each one responds to the input on its own, with no electrical conversation between neighbors. Every device behaves as an independent first order dynamical system.

The authors lean on a standard result from nonlinear dynamics here. Coupled differential equations need a minimum number of interacting states before certain behaviors appear at all. Self sustaining oscillation requires at least two. Chaotic motion requires at least three. Adding more uncoupled elements raises the number of available measurements. By that argument, though, it does not raise the dynamical order of the collective. Growing the array buys breadth rather than depth.

The proposed fix is physical coupling. Nodes must therefore share conductive pathways directly. External feedback wiring bolted on at the circuit board level does not count. Nanowire meshes and percolating nanoparticle films already move in this direction. Soft biomolecular materials offer something extra, namely compartments that assemble themselves. Their interfaces also adjust droplet by droplet.

The Building Block

The elementary unit is a droplet interface bilayer doped with alamethicin. Alamethicin is a peptide that stays quiet below a gating threshold. Above that voltage, the molecules insert themselves into the fat and form voltage gated channels. Conductance then rises sharply. Take the voltage away and the channels retract.

What remains is a device with two useful traits. First, it shows strongly nonlinear current voltage behavior. It also holds fading memory. Reservoir computing depends on exactly that pair. The membrane is more than a wire that opens and closes. It carries a record of recent activity in the number of open pores, and that record decays on its own.

Why the Engineers Flipped the Drive Scheme

Earlier work from the same groups drove these devices with voltage and read out current. Current, after all, maps cleanly onto pore density, the internal state that matters. That approach works when each device sits on its own isolated path. A coupled lattice breaks it. The droplets and their membranes share conductive routes, so individual junction currents become experimentally inaccessible.

So the engineers flipped the scheme. Input information now arrives as injected current, in the manner of a neuron receiving synaptic drive at its soma. The reservoir state gets read as the voltage at each droplet node. Three benefits follow.

First, node voltages are directly measurable. A hidden internal variable becomes an observable one.

Second, gradual charging keeps membranes away from the 200 to 300 millivolt range where electroporation and rupture occur. This benefit rests on a nice piece of engineering judgment. Indeed, the membrane behaves as a capacitor. A fixed injected current therefore fills it at a steady pace rather than slamming it with a voltage step. That pace happens to match the timescale of pore formation. Voltage climbs smoothly until channels open and start carrying charge across. Opening pores then bleed off the buildup before the membrane ever reaches the range where it would tear. The failure mode gets designed out of the drive scheme rather than managed after the fact.

Third, reading voltages directly removes the per node transimpedance amplifiers that parallel architectures required. That matters a great deal for anyone hoping to scale the approach.

Chemistry as a Design Knob

Membrane composition also tunes behavior. Synthetic DPhPC lipid produced noisy voltage responses under a steady two nanoampere current, with pores flickering open and shut. Brain total lipid extract behaved differently. It naturally contains cholesterol, which thickens and stiffens the bilayer. Each open channel therefore lives longer, and the response settled into a much steadier state.

Modeling the Whole Lattice

To capture a network of arbitrary shape, the authors built a physics based state space framework. Each membrane carries two coupled variables, namely the areal pore density and the transmembrane voltage.

Node voltages then evolve according to a capacitance weighted coupling matrix. An area weighted adjacency matrix encodes which droplets touch and how large each shared interface is. That single matrix governs both the capacitive and the memristive coupling. Such economy follows from a simple fact. The circuit is the material. Reduce the network to two nodes and the equations collapse back to the single device case.

Testing the Model on Six Droplets

Validation came from a six droplet chain. Engineers built it with the alternating one, two and three micromolar alamethicin recipe. They fed it a random current sequence encoded as three millisecond pulses spanning zero to 170 nanoampere. The same input then ran through the simulation. Predicted and measured node voltages then tracked each other closely. The average normalized mean squared error landed near 0.41 across a hundred step changes.

Notably, the chain behaved far more calmly than a single membrane under a comparable signal. There the same error metric reached 1.15. Coupling between droplets appears to average out the stochastic chatter of individual pores. That is a welcome surprise for a technology built from thermally jostled molecules.

A control simulation ruled out passive charging as the explanation. The authors modeled a chain of undoped droplet pairs as plain resistors and capacitors. That chain, however, could only track the root mean square of the input. Passive charging is far too sluggish to follow millisecond fluctuations. The doped chain followed the input closely instead, because channel insertion and removal happen on that same one to ten millisecond timescale. The dynamics come from the pores, not from the plumbing.

Benchmark Results, With One Important Qualifier

Two standard tests followed. Both ran in simulation, using the model validated against the six droplet chain. Neither ran on a physical lattice of the size the results describe. Readers should hold the numbers below in that light. The step from six measured droplets to fifty or more simulated ones remains the part of this work awaiting experimental confirmation.

The NARMA Tasks

The NARMA family asks a system to reproduce a nonlinear autoregressive sequence. Task order sets how far back the system must remember. Higher order versions have stayed out of reach for the parallel memristor reservoirs surveyed in the authors’ comparison. Those systems tend to top out near second order behavior, in the authors’ framing.

The winning configuration used fifty spatial nodes plus three virtual nodes created by temporal masking, for 150 states in total. On NARMA-2, the connected reservoir hit a testing error of 0.0124. In contrast, their parallel baseline managed 0.458. On the harder NARMA-10 task, the coupled network reached 0.327 against 0.627. The gap reflects reconstructive memory. Such memory does more than hold past inputs. It rebuilds them from the current state.

The Lorenz Attractor

The Lorenz63 task is the more telling result. Three coupled differential equations produce the famous butterfly shaped attractor. Yet forecasting it one step ahead means processing three interdependent input streams at once.

Engineers scaled the Lorenz variables into currents spanning negative fifty to positive fifty nanoampere. They injected them at three separate sites in a 43 node hexagonal lattice, with one grounded node at the center. Ridge regression at the readout mapped node voltages to the next state. Testing errors came in at 0.071 for x, 0.183 for y and 0.174 for z. The reconstructed trajectories also held the geometric shape of the true attractor, including its chaotic switching between lobes.

Worth dwelling on is what the network had to do to get there. Three separate signals entered the lattice at three separate points at the same moment. The correct forecast for any one of them depends on the other two. Coupled droplets, though, blend those streams on their own. Charge injected at one site propagates through shared membranes and interferes with charge arriving elsewhere.

A parallel array has no such route. Each node sees only its own input and evolves in isolation. Mixing independent streams therefore requires feedback circuitry wired in externally at the board level. Such wiring puts the coupling outside the reservoir rather than inside it. Precisely this is where the parallel comparison came apart. Its errors exceeded 0.6 and its trajectories diverged from the real ones almost immediately.

Size and Variety

Two design principles fell out of the sweeps. First, larger lattices help. A 13 node version gave errors of 0.244 and 0.222 on the y and z states, and scaling to 43 nodes brought those down. Even so, the small lattices still worked reasonably well. Coupling rather than raw size appears to do the heavy lifting.

Second, variety helps too. Networks with mixed alamethicin concentrations consistently beat uniform ones. Varied local dynamics reduce redundancy and broaden the set of independent functions the readout can call upon. Randomly assigned concentrations performed about as well as the structured one, two, three pattern. Diversity therefore yields diminishing returns past a certain point.

What Fabrication Actually Involves

Practically speaking, the apparatus is refreshingly low tech. The authors printed two sided molds on a consumer resin printer. They cured and baked them so residual chemicals would not spoil the silicone casting. They then cast PDMS to produce a container with rows of egg crate wells. Eleven silver wires got flame beaded and dipped in household bleach to grow a silver chloride layer. The wires then threaded up through vias in the well floors, held in place with glue.

Next, half microliter droplets went into the wells by pipette. They sat for ten minutes to grow their lipid coats, then got nudged into contact. A Python script generated the input waveform. From there, a commercial current source converted it into current. Six node voltages went straight into a 24 bit converter, with no amplification or filtering.

Why the Simplicity Matters

No cleanroom. No lithography. The complexity lives in the chemistry and the topology rather than in the fabrication line. That carries a practical consequence worth stating plainly. Cheap assembly plus coupling that arises on its own means a graduate student can test a new topology in an afternoon. The same experiment on a silicon neuromorphic platform would require a mask set and a fabrication run. Iteration speed, in short, is itself a research advantage.

Where This Could Go

Several application paths look credible over the next decade.

Forecasting Messy Systems

Chaotic forecasting is the most direct. Turbulent airflow over a wing resists conventional modeling. So do cardiac rhythm irregularities, local weather microclimates and power grid oscillations. All of them, however, share the structure the Lorenz task stands in for. A physical reservoir running on nanoampere currents needs no amplifier per node. Such a device could sit inside a sensor rather than shipping data to a server.

Robots That Think With Their Bodies

Soft robotics is another path. Robots built from elastomers and gels currently think with rigid silicon bolted to their frames. A computing substrate of water, oil and lipid could be cast into the same body it controls. It would then process strain and pressure signals at the point where they arise.

Talking to Living Tissue

Biointegration is the longest shot and the most interesting one. These membranes use the same phospholipids as cell membranes, along with real ion channels. They therefore speak the ionic language of tissue directly. Prosthetic interfaces, closed loop neural implants and diagnostic patches that classify signals locally all become conceivable. The same material would interface with living cells on one side and a readout circuit on the other.

The Research Agenda

The authors sketch next steps at three levels. Swapping in other lipids and channel proteins would broaden the available kinetics at the device level. Exploiting ion concentration gradients could add excitatory and inhibitory balance along with spiking behavior. Moving up to the network level, three dimensional droplet printing methods already exist. Those methods would allow denser, more brain like connectivity than a flat lattice permits. Reaching the system level, compact drivers and in-material readouts would cut the reliance on rigid electronics. Finally, online learning schemes would let the reservoir adapt in place.

The Honest Caveats

A few things deserve flagging. Engineers built and measured the six node chain. The NARMA and Lorenz results, though, came from simulation using the validated model. Indeed, no fifty node or 43 node physical lattice appears in the paper. Scaling from six droplets to dozens raises assembly and stability questions the authors do not settle. Droplet interface bilayers also live in oil baths and have finite lifetimes. Long term operation outside a laboratory therefore remains an open matter. The parallel baselines are simulations too, though the authors take them from their own prior published frameworks.

Still, the core claim holds up well. Spatial interconnectivity changes what a physical reservoir can compute, not merely how well it computes. Adding devices to a parallel array does not produce chaos prediction on the evidence gathered here. Wiring devices to each other through shared material does. That distinction matters for anyone designing neuromorphic hardware, whatever the substrate.

The platform is compelling. Even so, it remains early stage. Read it as a proof of principle for soft, physically coupled computing rather than as a technology ready to leave the bench.


Source

Nicholas Armendarez, Ahmed Mohamed, Md Sakib Hasan and Joseph Najem. Neuromorphic tissues: Soft biomolecular networks for brain-inspired temporal computing. Science Advances, volume 12, issue 30, article eaed0971 (2026). DOI 10.1126/sciadv.aed0971. Published 23 July 2026 under a Creative Commons Attribution NonCommercial License 4.0. Source code and datasets sit in a Zenodo repository, DOI 10.5281/zenodo.17296788.

Financial support came from the Army Research Office under grant W911NF-23-1-0314. National Science Foundation grants CCF-2403560, CMMI-2314559 and CCF-2403559 added to that total. Seed funding came from the Convergence Center for Living Multifunctional Material Systems. On 7 July 2026, The Pennsylvania State University filed a provisional utility patent application with the USPTO. It covers apparatus and method for providing soft biomolecular networks for brain inspired temporal computing.

Related Articles

PHP Code Snippets Powered By : XYZScripts.com
★

Track This Company

Create a free account to add companies to your watchlist and get notified when patents, funding, or key developments change.