South Korean researchers build a stacked chip that remembers motion as it happens

A team of South Korean researchers has built a stacked semiconductor chip that can tell not just what a wearable sensor is detecting right now, but what it detected moments or seconds earlier — a capability that has typically required extra memory circuits and constant power draw. The device, described by engineers at the Korea Advanced Institute of Science and Technology (KAIST) working with colleagues at UNIST and POSTECH, layers transistors tuned to different response speeds into a single three-dimensional structure, letting the hardware itself hold a short-term trace of past signals rather than outsourcing that job to software.

The research, led by professor Jimin Kwon of KAIST's Department of AI Systems and first author Haksoon Jung, was announced by KAIST this week and appears in the journal Advanced Materials, in a paper titled "Monolithic 3D-Integrated All-Solid Ion-Gated Carbon Nanotube Transistors With Tunable Ionic Conductance for Multi-Timescale Reservoir Computing." Independent science-press coverage of the announcement, including a summary published by Tech Xplore, confirmed the same benchmark results the institute reported. The work is also indexed on the National Library of Medicine's PubMed database. Co-authors include Yong-Young Noh of POSTECH's Department of Chemical Engineering and Hyeongjun Kim of UNIST, along with Hanbin Cho, Yongwoo Lee, Seunghun Baek, Yun Goo Ro, Hyunhyub Ko and Joonki Suh.

Building a chip with a built-in sense of time

The core problem the team set out to address is a familiar one in wearable electronics: telling apart a stumble from a jog, or a tremor from a steady hand, requires comparing a signal to what came before it. Conventional digital chips do this by writing data to memory and running repeated calculations, a process that consumes power and space that small, battery-limited devices like smartwatches or health patches do not have to spare.

The KAIST-led group instead turned to ion-gated transistors, devices in which an applied voltage pushes charged ions through a material to control the flow of current, rather than relying solely on an electric field the way a standard transistor does. Because ions move and settle more slowly than electrons, the transistor retains a faint trace of a previous signal for a stretch of time after the voltage that produced it is gone — a kind of built-in, fading memory.

To make that property useful, the researchers first had to solve a fabrication problem. Ion-conducting materials are usually liquid, which is difficult to integrate into standard chip manufacturing. The team converted the material into a solid thin film, called an ionogel, in which ions are embedded in a polymer matrix. By adjusting the concentration of ions and the thickness of the film, they could tune individual transistors to respond quickly and forget fast, or respond more slowly and hold onto a signal's imprint longer.

Stacking layers with different response speeds on top of one another, on carbon nanotube transistors, produced what the team calls a multi-timescale reservoir — fast layers picking up the most recent change in a signal, slower layers carrying information about what happened earlier, all processed simultaneously within one piece of hardware.

Testing whether the chip actually remembers

To check whether the stacked design worked as intended, the researchers ran two kinds of tests. In the first, they fed the device four consecutive on-or-off electrical pulses in varying sequences and asked whether the chip's output could distinguish among all 16 possible orderings — a direct test of whether the hardware was sensitive to the order of events, not just their presence. The device told all 16 patterns apart.

In the second test, the team used Moving MNIST, a benchmark data set of handwritten digits that drift across a frame over time, commonly used to evaluate how well a system handles sequences rather than single, static images. Feeding the moving digits into the chip at different playback speeds — with individual frames changing roughly every one to ten milliseconds — the device classified the digits with validation accuracy above 90 percent at multiple timescales, according to the researchers.

The team also fabricated the transistors on both rigid four-inch silicon wafers and on flexible substrates suitable for skin-worn sensors, and reported that devices kept stable electrical characteristics for 55 months after fabrication — evidence, the researchers said, that the underlying ionogel material does not degrade quickly under normal conditions, a longstanding concern for ion-based electronics.

"These devices can be fabricated on large-area substrates using existing thin-film semiconductor processes, and they can also be stacked in multiple layers," Kwon said in the KAIST announcement.

An old computing idea, built into new hardware

The chip is a physical implementation of an idea called reservoir computing, a concept that has circulated in machine learning research since the early 2000s. In its original form, a reservoir computer feeds input signals into a large, fixed network with rich, complex internal dynamics — the "reservoir" — and trains only a simple output layer to read patterns out of it, sidestepping the heavier computational cost of training an entire deep neural network from scratch. Researchers have spent the past decade trying to build that reservoir directly into physical materials — memristors, spintronic devices, photonic circuits and other ion-gated transistors among them — on the theory that letting a material's own physics do part of the computing could cut the energy cost of processing sequential data, such as motion, speech or physiological signals, on small devices rather than sending it to a data center.

KAIST's device fits into that broader push toward what is often called physical or in-materials reservoir computing, and toward edge computing more generally — running artificial intelligence tasks locally, on the device collecting the data, instead of over a network connection to a remote server. For a fitness tracker or a continuous health monitor, that distinction affects both battery life and how quickly a device can flag something like an irregular gait or a fall.

The research was supported by South Korea's National Research Foundation and by the Korea Planning & Evaluation Institute of Industrial Technology, according to the KAIST release, reflecting continued government investment in semiconductor and AI-hardware research as the country competes with the United States, China and Taiwan in advanced chipmaking.

What remains unresolved

The demonstrations so far involve laboratory benchmarks — a data set of drifting digits and sequences of on-off electrical pulses — rather than a finished wearable product tested on human movement in real-world conditions. Moving from a working transistor stack to a chip embedded in a commercial smartwatch or medical patch typically involves additional engineering: integrating the reservoir with conventional digital circuitry for the simple "readout" layer reservoir computing still requires, packaging the device for everyday wear, and testing it against the noisy, inconsistent signals real bodies produce, as opposed to a clean benchmark data set.

The 55-month stability figure, while notable, describes the material's electrical behavior under the conditions the researchers tested, and does not by itself establish how the full stacked device performs after years of the bending, sweat and temperature swings a wrist-worn sensor would encounter. The researchers have not published a timeline for moving the technology beyond laboratory prototypes, and outside groups working on competing approaches to physical reservoir computing — including memristor- and graphene-based designs described in recent research — have yet to weigh in publicly on how the ion-gated carbon nanotube approach compares on power consumption or manufacturing cost at scale.

Still, the approach adds to a growing body of evidence that transistors built around slow-moving ions, rather than only fast-moving electrons, can be engineered to do more than switch on and off — they can hold a trace of recent history, a property that ordinary digital logic has to simulate at a cost in power and complexity that small, always-on wearable devices are generally built to avoid.