'It remembers how you moved': KAIST develops skin-attachable AI semiconductor

- Joint research by KAIST, UNIST and POSTECH

- Device identifies motion sequences with more than 90% accuracy

A concept illustration — generated by AI — showing the principle of stacking semiconductor devices with different response speeds to read changes over time, and the potential application in skin-attachable devices. [KAIST]

South Korean researchers have developed an AI semiconductor that does more than detect human movement — it remembers how the body moved moments earlier and uses that context to make decisions. The technology is expected to serve as a core component in ultra-low-power wearable AI devices such as smartwatches and skin-attached sensors.

KAIST announced Tuesday that a research team led by Professor Kwon Ji-min of its Department of AI Systems, working with researchers from UNIST and POSTECH, had developed a neuromorphic AI semiconductor capable of processing information that changes over time. The chip achieves this by stacking multiple layers of semiconductor devices with different response speeds.

Accurately distinguishing between two motions — say, walking and briefly swinging an arm — requires more than a snapshot of a single moment. The chip must read the flow of time, tracking how one movement leads into the next.

The research team built that capability directly into the semiconductor itself. The key was turning what had long been considered a drawback — the slow movement of ions inside a device — into an advantage.

When voltage is applied to the new semiconductor, ions inside it shift and alter the flow of current. Even after the voltage is cut, the ions take time to return to their original state, leaving a trace of the previous signal in the device — much like the way a bell continues to ring after it has been struck.

The joint research team behind the study. From left: Baek Seung-hun, UNIST researcher; Lee Yong-woo, KAIST doctoral researcher; Kwon Ji-min, KAIST professor; Kim Hyeong-jun, UNIST researcher; and Jeong Hak-sun, KAIST doctoral researcher. [KAIST]

The team fabricated the ion-containing material into solid thin films, then tuned the quantity of ions and the film thickness to vary how long each layer retains a signal. By stacking these layers, the device can process information from different time windows simultaneously — drawing on both short-lived and long-lasting memory at once.

The researchers confirmed that the device can distinguish among 16 distinct patterns generated by the on-and-off sequences of four consecutive input signals.

The team also demonstrated the technology's potential for large-area and flexible applications. They successfully fabricated the semiconductor on a 4-inch wafer and implemented it on a bendable substrate. The electrical properties of the resulting devices remained stable across 55 months of measurement.

The researchers expect the technology to enable real-time analysis of movement and biosignals in wearable devices with limited power supplies, such as smartwatches and skin-attached sensors. However, AI performance and power savings when the chip is actually integrated into such devices still require further verification.

"An advantage of this approach is that it can be fabricated on large substrates using existing thin-film semiconductor processes and can be stacked in multiple layers," Kwon said. "It has the potential to develop into an AI semiconductor that reads movement and bodily signals in devices that need to consume very little power, like a smartwatch."

The team's next goal is to advance the packaging and integration technology that connects the device, sensors and peripheral circuits, ultimately developing an ultra-low-power, high-density neuromorphic AI hardware platform in which the semiconductor processes temporal information directly.

The findings were published in the international journal Advanced Materials.

nbgkoo@heraldcorp.com