Press a soft material, release it, and watch it recover. If recovery takes time, its present shape contains a trace of what just happened. Researchers at MIT are turning that ordinary physical idea into an unusual component for computing.
In a September 16 account of new research, MIT describes tiny devices whose deformable polymer layer provides short-lived memory and neuron-like firing. This is a laboratory device, not a working artificial brain or a commercial processor. The interesting possibility is that a material’s motion could perform part of the information processing itself.
A memory made from delayed recovery
According to MIT, the device places polydimethylsiloxane, or PDMS, between metal electrodes. Voltage draws the electrodes together, compressing the polymer and changing the current. PDMS is viscoelastic: its recovery is delayed, so the electrical response depends on recent stimulation. Accumulated inputs can trigger a threshold response resembling a neuron firing.
MIT reports polymer films as thin as two nanometers. That describes the film, not the dimensions of an entire computing system. The university points to possible sensors, wearable devices and robotic applications; those remain prospective uses, not demonstrated products.
The underlying study is Viscoelastic nanomechanical devices for neuromorphic information processing. The Johnson group’s publication list identifies it as accepted by Science Advances, while the Niroui group’s list also records related conference work on a viscoelastic artificial neuron. Those listings corroborate the research project, not its performance independently.
Why forgetting can be a feature
A conventional file should remain unchanged when nobody is using it. A device processing a stream of sensations may need a different kind of memory: enough persistence to relate the present input to the recent past, followed by decay so old inputs stop dominating.
Imagine, as a possible design goal rather than a demonstrated application, distinguishing one isolated tap from a rapid sequence. A system must retain something about the first tap long enough to interpret the next. Software can track that history explicitly. A physical component with a suitable relaxation time could instead supply part of that behavior through its own dynamics.
This is the attraction of neuromorphic computing: borrowing useful principles from nervous systems without claiming to recreate their full complexity. A component that responds at a threshold is one building block. It does not establish perception, learning or intelligence on its own. Those require organized systems and evidence at the level of the task.
The missing evidence is system-level
The paper’s full text returned an access error during this review. This explanation therefore relies on MIT’s attributed account and the authors’ publication records; it does not claim an independent inspection of the complete measurements or supplementary experiments. We are not assigning an energy-saving percentage, operating lifetime or commercial readiness level.
For an eventual application, our next questions would be practical. Do nominally identical devices respond consistently? Does repeated deformation alter their behavior? Can the useful timescale be controlled? And does any advantage remain after the electronics needed to supply inputs and read outputs are included? These are evaluation questions, not findings that this study has answered.
That boundary connects this story to Vastkind’s reporting on measuring useful computation, whole-system power constraints and the difference between neural signals and usable control.
The decisive demonstration would put these soft components inside a complete sensing task and show what their mechanical memory contributes—beyond the behavior of a single device.
