Akshay DineshA computer chip usually treats memory and calculation as separate jobs. Data moves from memory into a processor, gets changed, and travels back again. That arrangement is dependable, but the constant shuttling takes time and energy. Researchers at the Massachusetts Institute of Technology have now demonstrated a very different possibility: a tiny device in which a soft polymer physically remembers recent electrical inputs and helps process the next one.
The device does not think, learn, or reproduce a whole biological neuron. It copies two narrower behaviors that make neurons useful for information processing: accumulating a series of signals and producing a strong response only after those signals cross a threshold. The work, reported in Science Advances by Peter Satterthwaite, Sarah Spector, Farnaz Niroui, and their colleagues, shows how the mechanics of a material can perform part of the computation that would normally require extra electronic components.
Why computers spend energy moving information
Most modern computers follow an architecture in which the processor and memory are physically distinct. That division lets engineers improve each part, but it also creates what is often called the memory bottleneck. A processor may be able to calculate quickly while waiting for instructions and data to arrive from memory. Moving those bits back and forth can consume more energy than the calculation itself, especially in systems that repeatedly analyze streams of sensor data.
Neuromorphic computing takes inspiration from nervous systems, where storage and processing are not divided so neatly. A biological neuron changes its response according to signals that arrived moments earlier. It integrates those inputs over time, leaks away some of their effect, and fires when the combined stimulus becomes strong enough. Engineers do not need to duplicate every detail of a living cell to use this strategy; they need hardware whose present behavior depends on its recent past.
Many artificial-neuron circuits reproduce that behavior with transistors, capacitors, and supporting circuitry. The MIT team asked whether a material could supply the temporary memory by itself. Their answer uses a familiar silicone called polydimethylsiloxane, or PDMS, in an unfamiliar setting: a film only about two nanometers thick.
A soft layer becomes a nanoscale spring
The experimental device places the ultrathin PDMS film between two gold electrodes. One electrode can move. When voltage is applied, electrostatic attraction draws the metal surfaces closer and compresses the polymer between them. The gap becomes smaller, so electrons can tunnel across it more readily and the measured current rises.
At this scale, simply bringing two metal surfaces close together creates a problem. Adhesive forces can make them stick permanently, ending the device’s ability to change state. The PDMS layer prevents that collapse. It acts like a nanoscale spring, providing a restoring force that helps the upper electrode return after the voltage is removed.
PDMS is also viscoelastic. An elastic material snaps back immediately after a force disappears; a viscous material changes more slowly. A viscoelastic polymer does both. Its delayed recovery means the device retains a physical trace of recent compression. If another voltage pulse arrives before the polymer has relaxed, that new pulse begins from a different mechanical state than the first.
That short-lived state is the device’s memory. It is not a stored file or a permanent bit. It is closer to the fading dent in a cushion just after someone stands up: the material’s current shape reveals something about the force it recently experienced. By changing the polymer chemistry and cross-linking, engineers can tune how quickly that trace fades.
How compression produces neuron-like firing
A neuron is often described with a leaky integrate-and-fire model. Incoming signals build toward a threshold, while their effect gradually leaks away. Signals that arrive close together are more likely to trigger firing than the same signals spread far apart. The polymer device recreated this timing-dependent response mechanically.
In one demonstration, the researchers sent a sequence of 2.5-volt pulses into the device. Each pulse compressed the PDMS a little more before it had fully recovered. The team then used smaller read pulses to measure the changing current. After 16 input pulses, the junction reached a threshold and its current rose sharply, producing the artificial equivalent of a firing event.
The important feature is not the number 16. Engineers can adjust thresholds and timing for a particular task. What matters is that integration, temporary memory, and threshold response arose inside one nanoscale junction rather than from a separate capacitor and a larger surrounding circuit. The paper reports a 200-by-200-nanometer active region, with the polymer spacer itself only about two nanometers thick.
The researchers also connected two devices so that firing in one suppressed the response of the other. This behavior, called lateral inhibition, is common in nervous systems. It helps biological networks emphasize contrast and lets competing signals influence one another. Demonstrating inhibition does not create a brain-like network, but it shows that the devices can interact in a way that may support temporal classification and other neuromorphic tasks.
What the energy numbers really show
Brain-inspired hardware is often promoted as a route to low-energy computing, but prototype measurements need careful reading. The complete experimental device consumed about 54 nanojoules during a firing event. Most of that total came from electrical leakage and the practical limits of the laboratory setup, not from the minimum energy needed to compress the polymer and change the junction.
The researchers calculated that an optimized version could store roughly 2 femtojoules of electrical energy and 0.1 femtojoules of mechanical energy per firing event. A femtojoule is one millionth of one billionth of a joule. On that basis, a smaller, better-insulated device could potentially operate more efficiently than a biological neuron while occupying less than 0.04 square micrometers.
Those values are projections, not the measured performance of a finished chip. Turning the idea into a useful computing system would require reducing leakage, fabricating many devices consistently, connecting them into larger arrays, designing ways to train the network, and proving that it can solve real tasks reliably. The seconds-long memory demonstrated here is also much slower than the microsecond or millisecond timescales used by many artificial neurons, though it could suit sensors that need to recognize patterns unfolding over longer intervals.
There are other open questions. Soft materials can age, deform, and vary with temperature. A device that relies on tiny mechanical changes must keep those changes predictable across millions or billions of cycles. Its unusually long relaxation time can be useful for temporal context, but it may also limit speed. These are engineering challenges to investigate, not details already solved by the proof of concept.
Where material-based computing could fit
The most promising uses may be far from a large data center. Edge devices such as wearable sensors, environmental monitors, robots, and prosthetics often need to interpret signals where they are collected. Sending every measurement to a remote processor adds delay and uses power. A material that senses, remembers, and filters a recent sequence locally could reduce the supporting hardware required for a simple decision.
Imagine a prosthetic surface receiving a changing pattern of pressure. A conventional design might digitize every reading and pass it through a processor. A future material-based system could begin sorting the timing and strength of those signals in the sensor itself. The same principle could help a monitoring patch distinguish a sustained change from a brief fluctuation, or allow an environmental sensor to respond to a sequence rather than a single measurement.
These examples remain possibilities, and the MIT device is not ready to be placed in a wearable or prosthetic system. Its larger contribution is a design idea: computation can come from a material’s physical dynamics, not only from logic gates arranged around passive matter. The polymer’s slow rebound, normally just a mechanical property, becomes a way to encode time.
That shift changes the question engineers can ask. Instead of choosing a material only because it holds components in place, they can ask what useful transformation the material itself can perform. A soft layer that bends, relaxes, and remembers may never replace the general-purpose processor. It could, however, become a compact specialist for the kinds of signals that arrive as patterns through time.
