| Jun 19, 2026 |
AI system monitors 2D semiconductor manufacturing at the atomic layer
Researchers combined low-temperature plasma processing with machine learning to synthesize and etch 6-inch MoS2 and WS2 wafers and predict film thickness in real time.
(Nanowerk News) A research team at the Korea Institute of Machinery and Materials (KIMM) has built a system that reads the light and gas signals given off during plasma processing and uses machine learning to estimate how thick a 2D semiconductor film is as it forms (Advanced Materials, "Electron Release via Internal Polarization Fields for Optimal S‐H Bonding States").
|
|
The method lets 2D semiconductor manufacturing run at low temperature on 6-inch wafers, with a model tracking film growth in real time instead of halting the process to measure it.
|
Key Findings
- KIMM combined low-temperature plasma processing with machine learning to synthesize and etch 6-inch MoS₂ and WS₂ wafers at the atomic-layer level.
- The system reads real-time light-emission and gas-mass signals during processing and uses machine learning to predict film thickness at the atomic-layer level without interrupting the run.
- Because it taps the optical viewports already built into commercial equipment, the method can be added to existing production lines without hardware changes.
|
|
A team led by Senior Researcher Hyeong-U Kim at the Semiconductor Manufacturing Research Center of KIMM, headed by President Seok-Hyeon Ryu, developed the synthesis and etching processes for 6-inch MoS₂ and WS₂ semiconductors and built them into an AI-based intelligent system.
|
|
Differences in process conditions have often caused quality variations that made uniform production difficult, and the team's aim was to keep film quality consistent across a full wafer by tying plasma processing directly to data analysis.
|
|
The fabrication relies on two standard plasma tools. The team used plasma enhanced chemical vapor deposition (PECVD) to grow the films at low temperature and a reactive ion etcher (RIE) to remove material precisely. Both run cooler than the high-temperature methods 2D semiconductor work has typically required, which is what makes them compatible with equipment already installed on factory floors.
|
|
To follow what happens inside the chamber, the researchers measured the light emitted and the variation in gas mass during processing in real time, then analyzed those signals with machine learning to infer the state of the process. They collected time-series data from several diagnostic instruments at once and fed it into machine learning models to predict semiconductor thickness at the atomic-layer level.
|
|
Three monitoring tools supplied that data. Optical emission spectroscopy (OES) reads the light given off by the plasma, while time-of-flight mass spectrometry (ToF-MS) and quadrupole mass spectrometry (QMS) track the gases present. Combining these different data types gave the models enough information to estimate how many atomic layers had formed without stopping to measure directly.
|
|
Senior Researcher Hyeong-U Kim pointed to the role of combining different data streams. "By applying multimodal data-based machine learning technologies, we simultaneously achieved process prediction and optimization, significantly improving both process reproducibility and productivity," he said.
|
|
Conventional 2D semiconductor processes have mostly used high-temperature methods, which sit poorly alongside existing production lines and struggle to coat large areas evenly. Atomic-layer etching has also tended to be slow and low in output. KIMM's low-temperature plasma process works with mass-production equipment and performs atomic-layer etching in a single step, raising both efficiency and throughput. Kim described that combination as the core result. "This research is meaningful in that it successfully implemented 2D semiconductor processes on 6-inch wafers at the atomic-layer level under low-temperature conditions," he said.
|
|
The technology could be used across several parts of the semiconductor industry, including AI chips, next-generation electronic devices, and displays. Its industrial appeal rests partly on convenience: because it uses the OES viewports already present on commercial equipment, it can be installed without structural changes to the machines.
|
|
What the system demonstrates today is real-time diagnosis and thickness prediction rather than hands-off operation. The team frames autonomous, intelligent manufacturing as the goal it is building toward as more process data accumulate, and Kim set out that ambition directly. "We plan to expand this technology into a core platform for the AI intelligentization of next-generation semiconductor manufacturing framework," he said.
|
|
The work was carried out by a KIMM team of graduate and postdoctoral researchers, who have produced more than 30 SCI journal papers across five research projects while advancing the technology. KIMM is a non-profit government-funded research institute under the Ministry of Science and ICT, established in 1976 to conduct R&D in machinery and materials and to commercialize the technologies it develops. The findings were published in Advanced Materials in September 2024.
|
|
The step that matters here is moving thickness measurement inside the running process. By inferring atomic-layer growth from signals the equipment already produces, the KIMM system turns a fixed plasma recipe into one that can be checked against live data on every wafer, the groundwork the researchers say a self-adjusting process would need.
|
