Researchers at KAIST, working with UNIST and POSTECH, have built a stacked neuromorphic transistor that classifies time-varying signals with verification accuracy above 90%. The device, called CNT sIGT, appeared in the journal Advanced Materials on June 30. KAIST announced the result on September 29.

What is a CNT sIGT?

The device pairs carbon nanotube channels with a solid thin film holding mobile ions. Apply voltage and the current shifts as ions drift through the film. Remove voltage and the ions take time to return to their rest state.

That lingering drift carries a short memory of past signals, much as ripples persist on a pond after a stone lands. The team, led by Professor Kwon Ji-min of the Department of AI Systems, built a CNT sIGT. It is a carbon nanotube version of a solid-state ion-gated transistor.

Previous ion-gated transistors relied on liquid or gel electrolytes. Those resist the fine patterning and multilayer stacking that chip production demands. The team converted the electrolyte to a solid film and tuned both ion content and thickness.

How does the stack read time?

Stacking two sIGTs of different speeds creates a layered reader. The faster layer tracks newly arriving signals. The slower layer carries residue from earlier pulses. Together they encode the temporal sequence, not just the present state.

The team drove the stack with four successive on-off pulses and recorded sixteen distinct timing patterns. They then fed the device measurements from moving digit images and reported verification accuracy above 90% across varied playback speeds. The exercise amounts to a textbook demonstration of reservoir-style temporal encoding on a thin-film device.

Why does this matter for neuromorphic computing?

Neurons in the brain read time-staggered electrical spikes and ignore static noise. Silicon, in general, does not. Neuromorphic chips aim to copy that trick while sipping power, and temporal encoding is a known weakness in conventional digital neural accelerators.

Jeong Hak-soon, research professor at KAIST and first author of the paper, said: "CNT sIGT can be utilized in neuromorphic semiconductors, which are considered next-generation semiconductor technology." That pitch matters because wearable AI is the most credible near-term market for low-power neuromorphic silicon.

What could it become?

Professor Kwon said the device can be fabricated on large-area substrates using existing thin-film semiconductor processes and stacked in multiple layers. He added that the work could evolve into an AI chip that reads motion and bodily signals in power-constrained gear such as smartwatches.

The carbon nanotube channel keeps the device compatible with back-end-of-line integration rather than demanding a leading-edge fab node. That puts the technology within reach of mature production lines that already handle large-area flexible substrates. The path matters for wearable form factors.

Where does this fit in Korea's neuromorphic push?

Two Korean groups have chased related work this year. A Sungkyunkwan University team in June unveiled a device that selectively forgets old data. GIST and Seoul National University in July reported a single neuromorphic device aimed at cutting leakage current.

The KAIST result adds a third Korean entry, this one focused on reading timing rather than storing state. The paper does not quote a partner foundry or switching endurance numbers. Production timing remains a roadmap rather than a confirmed capacity figure.

What the work does confirm is that temporal neuromorphic behavior can be built from standard thin-film process lines. The next measure of progress will be integration density, switching endurance, and a real foundry partner willing to commit wafers. Without those, the 90% figure remains a lab result rather than a shipped device.