Instead of random grains, one MoS₂ crystal grows where we want it

On a semiconductor wafer lies a pattern of fields. In each of them a single atom-thick crystal of molybdenum disulfide is supposed to grow. The old problem is that the crystal nucleus appears wherever it likes. Then several crystals meet, and at the grain boundaries the electronics get worse. A KAIST team showed in Nature that the nucleus can be forced to start exactly at the center of the field. The tool is not a laser or a needle, but an etching flux from an oxide barrier.

The paper “Spatially deterministic nucleation of 2D semiconductors by etching flux” was published on 7 October 2026 in Nature (vol. 658, pp. 398–406). Among the first authors are Jeongwon Park, Jiyun Kim, Sumin Kang, Jieun Oh and Youngmin Sunwoo. The corresponding author is Kibum Kang of the Quantum Materials and Devices Laboratory at KAIST. Kang is also co-CEO of the Korean spin-off TDS Innovation, which builds equipment and materials for two-dimensional semiconductor devices. Nature added a separate Research Briefing; The Register covered the result on 11 October.

Engineering the place where a crystal begins

Two-dimensional materials are not a metaphor. They are a layer of atoms. Graphene is the most famous example, but it lacks a natural band gap, so it is hard to make a classical switching transistor from it. Transition-metal dichalcogenides, including MoS₂, do have such a gap. That is why they have been tested for years as channels for future transistors and as building blocks for stacks that bring logic and memory closer together.

Growth of such layers on a wafer usually looks like this: inside a patterned window the crystal starts at a random point. Or at several points at once. When grains meet, boundaries form. Boundaries hurt carrier mobility and device-to-device uniformity. Classical methods can limit the number of nuclei or confine growth to a designated region. They do not, however, dictate where inside that region a single nucleus appears. The authors call this spatial stochasticity of nucleation.

Kang and colleagues designed a process they call EF-SCN: single-centred nucleation steered by an etching flux. An oxide barrier — most often HfO₂, and ZrO₂ in a low-temperature variant — releases oxygen. That oxygen acts as a local etching flux near the pattern edge. Nuclei at the rim of the field are suppressed. At the geometric center, one survives. Systematic experiments established design rules for three regimes: zero nuclei, one nucleus, and many nuclei. That is the language of a process engineer, not magic.

What was measured on a chip, not only on a slide

The process produced MoS₂ single crystals on the scale of about 10 micrometres. Supplementary descriptions also mention triangular patterns around 7 µm and lines 10 µm long. Field-effect transistors can be built on the same crystal. The highest reported carrier mobility in the FETs reached 117 cm²/V·s. That is higher than in many earlier reports on selectively grown MoS₂, though we are still in the lab, not on a factory line.

The strongest scaling argument is a test on a two-centimetre substrate. Single crystals grew in 397 of 400 patterned fields. That is 99.3 percent. The authors also write of single-crystal yield exceeding 99 percent with uniform selectivity over a larger area. A separate process variant runs at 430 °C with a ZrO₂ barrier. Lower temperature matters in practice: it is easier to fit with other semiconductor process steps.

The process is compatible with MOCVD, a family of tools chip factories already understand. The authors also showed more than “one pretty crystal in the middle.” They decoupled nucleation from later growth. That let them grow line-shaped single crystals, several transistors inside one large common crystal, and aligned 2D lateral heterostructures for self-aligned contacts. This is no longer only growth chemistry. It is a set of integration tricks.

Who needs this outside the lab

Kang is clear about the destination. If you can impose where a crystal should begin, you can place high-quality 2D single crystals where they are needed and build a next-generation process around that. In an interview with The Register, TDS Innovation co-author Jony Jung gives a horizon: commercial use of 2D semiconductors around 2030. Timing will depend on further validation. Demand from AI, including physical AI and robotics, could — according to the company — pull adoption forward, but that is a business speculation, not a measurement result.

Why do factories look at atomic layers at all? Because they can be stacked. IBM and Intel have already shown CFET concepts in which electron and hole transistors sit on top of each other rather than side by side. Two-dimensional materials have been in that conversation for years. What was missing was a repeatable way to avoid a mosaic of grains instead of one channel. EF-SCN attacks exactly that point: not “is the material interesting,” but “can you put it where the designer drew the window.”

Nobody in Poland took part in this paper. There are, however, groups that have worked for years on two-dimensional dichalcogenides and growth processes — among them ITME in the Łukasiewicz network and teams at Warsaw University of Technology. I am not attributing any role in EF-SCN to them. I only note that the language of this paper — oxide barrier, MOCVD, FET mobility, pattern selectivity — is a language those labs already know. The shortest path to checking the result is repeating the process, not a press release.

How a day on the growth line looks

Picture one wafer. First lithography defines the windows where material should grow. Then the oxide barrier appears. In the MOCVD chamber atmosphere the barrier releases oxygen. Near the window edge that oxygen acts as a nucleation brake. Metal and sulfur adatoms do not build a stable nucleus there. Closer to the center the brake weakens. At the geometric center the window “wins” one nucleus. From then on the crystal grows sideways as a single domain.

That is the heart of the trick: you do not need to aim a laser at the start point. You design a chemical gradient that itself chooses the center. The authors showed that the regimes can be switched. Too weak a flux or wrong conditions — random nuclei return. Too strong — the window stays empty. In a narrow parameter window you get exactly one nucleus. A process engineer likes such maps: they know which knob to turn.

Why does 10 micrometres matter? Because that is already a scale where a transistor channel, or several devices side by side in one grain, can fit. A random mosaic at that scale means some devices hit grain boundaries and some do not. The electrical result then dances. One nucleus in the middle means the designer knows what to expect.

What circuit designers gain

The paper does not end with a pretty microscope image. The authors separated the moment of nucleation from later growth. That allows single crystals to grow as lines, not only triangles. On a long channel they built a FET with mobility of 117 cm²/V·s. They also showed several transistors in one common large crystal and lateral 2D heterostructures aligned for contacts. In practice that means a metal contact can “hit” a planned place in the material, not a random boundary.

For people outside the chip industry: imagine printing traces that must meet one continuous atomic layer. If the layer is a patchwork, each trace has a different quality. If the layer is one crystal with a known center, the traces can be drawn with a map in hand. EF-SCN gives exactly that map at the growth stage.

Kang talks about bringing logic and memory closer. In today’s AI systems energy cost often goes up and down between chips, not only in the arithmetic itself. Stacking functions closer together makes energetic sense. Two-dimensional materials are one candidate for such stacking in the third dimension. Without nucleus control, though, you are left with beautiful chemistry and hard manufacturing.

What this paper does not promise

This is not a phone chip for 2027. A two-centimetre substrate is still not a production wafer. Mobility of 117 cm²/V·s refers to the best devices in the report, not a factory-wide average. The 2030 horizon comes from the spin-off, not from the Nature abstract. The authors solved spatial randomness of the nucleus inside a patterned window. They did not solve every problem of integrating 2D materials with silicon, contacts, reliability or cost.

It is also worth remembering that the “etching flux” is not a nozzle in a school lab. It is a chemical oxygen gradient released from the barrier. The mechanism was checked by comparison: when the oxygen source disappears, random multiple nucleation returns. When the flux is present, adatoms gather in the center and one nucleus remains. Elegant, but it still needs control of temperature, composition, pattern geometry and barrier quality.

We do not yet know how the process will behave on a full 200 or 300 mm wafer, with hundreds of thousands of windows and the tolerances a fab demands. We also do not know whether the same scheme will transfer smoothly to other dichalcogenides beyond MoS₂. The paper opens a door. It does not claim we are already on the other side.

Maturity: TRL 3–4

On the technology readiness scale this is TRL 3–4. There is a working lab process, transistors, nucleation design rules, high yield on a small substrate and a spin-off with an explicit commercial target. What is missing is validation on a full production wafer, line statistics and proof that the same scheme holds for other dichalcogenides and in a full CFET stack. That is still a different level from another pretty micrograph of random growth.

The main lesson is engineering. For a long time we could tell a material: “grow in this window.” Now a second sentence can be added: “start exactly here.” In semiconductors that second sentence often decides whether a lab result becomes a process — or only a paper.

Bibliography

1. Park J., Kim J., Kang S., Oh J., Sunwoo Y., Oh S. et al. Spatially deterministic nucleation of 2D semiconductors by etching flux. Nature 658, 398–406 (2026). DOI: 10.1038/s41586-026-11095-1
2. How to control where semiconductor crystals start growing (Research Briefing). Nature (2026). DOI: 10.1038/d41586-026-02934-2
3. Robinson D. Researchers work out how to control 2D semiconductor growth for future chips. The Register, 11 October 2026. theregister.com