Researchers crack key barrier to vertically stacked AI chips

South Korean researchers have developed a technology that controls where semiconductor crystals begin to form, steering the starting point to a precise, predetermined location. Because the technique allows high-quality semiconductor channels to be grown directly where a circuit requires them — rather than fabricated elsewhere and transferred — it is seen as a potential breakthrough for next-generation stacked AI chips.

KAIST announced Thursday that a research team led by Professor Kang Ki-bum of its Department of Materials Science and Engineering, working jointly with UNIST, Hanyang University and TDS Innovation, had developed a new growth technique for precisely controlling the "nucleation" site — the point at which a 2D semiconductor crystal first forms.

The findings were published Thursday in the international journal Nature.

Two-dimensional semiconductors are only a few atomic layers thick yet retain excellent electrical properties. Because devices can be stacked in multiple layers on top of existing semiconductor circuits, the material is regarded as a leading candidate for next-generation high-density AI chips.

When AI processes vast amounts of data, moving that data within a chip consumes considerable time and power alongside the computation itself. Stacking logic and memory devices vertically so they sit closer together shortens the distance data must travel, boosting processing speed and cutting power consumption.

The challenge is that the sites where 2D semiconductors begin to grow are determined probabilistically. When crystals nucleate simultaneously at multiple locations and eventually meet, grain boundaries form between them. Those boundaries impede electron movement and degrade semiconductor performance.

The research team overcame this limitation by achieving what they call "deterministic nucleation" — making crystals nucleate one at a time at designated positions. Rather than simply defining the region where a semiconductor will grow, the technique specifies the exact starting point at which the crystal first forms. The method has been named etching-flux-mediated single-center nucleation, or EF-SCN, and can be applied to selective nucleation and single-crystal growth of 2D semiconductors.

The key lies in oxygen components released from oxide barriers — such as hafnium oxide (HfO₂) — surrounding the growth region. These components suppress the formation of crystal nuclei in molybdenum disulfide (MoS₂), one of the most representative 2D semiconductors.

The nucleation-suppression effect is strongest near the edges closest to the oxide barrier and weakest at the center, farthest from it. The team exploited this gradient to induce crystal growth at a single central point, while simultaneously preventing new crystals from forming in the surrounding area — allowing one large single crystal to develop unimpeded.

In experiments, the team successfully grew a single MoS₂ crystal at the designated central location in more than 99 percent of 400 test patterns. The researchers also succeeded in selectively coating only targeted portions of the 2D semiconductor with ruthenium metal.

As the technology matures, it is expected to enable "bottom-up" manufacturing — growing 2D semiconductor channels directly at the locations where circuits are designed, rather than relying on a transfer process that moves pre-formed single-crystal semiconductors to the required positions. The approach is also expected to suit large-area, high-density semiconductor manufacturing by simplifying the overall process.

Beyond high-density logic and memory devices, the technique could be extended to a range of next-generation semiconductor platforms, including self-aligned contact devices using lateral heterojunctions, single-channel multi-device architectures and position-selective doping devices.

"In thin-film growth, where a crystal begins to form has traditionally been probabilistically determined — the core significance of this research is that we can now control that starting point to wherever we want," Professor Kang said. "This will serve as the foundation for a stacked-channel process that grows high-quality 2D semiconductors directly where they are needed and stacks semiconductor devices vertically."