KAIST Professor Kwon Ji-min (from left), and researchers Jeong Hak-sun and Lee Yong-woo. [Provided by KAIST]
KAIST Professor Kwon Ji-min (from left), and researchers Jeong Hak-sun and Lee Yong-woo. [Provided by KAIST]

A painstaking step in 2D semiconductor research — manually scanning microscope images to locate usable samples one by one — is set to be automated.

KAIST announced Thursday that a research team led by Professor Kwon Ji-min of the Department of Electrical Engineering and the Department of AI Systems has developed a technology that automatically identifies suitable 2D semiconductor flakes from optical microscope images and links the results directly to transistor fabrication. The work was carried out in collaboration with UNIST, Hanbat National University, Hanyang University, and Washington University in St. Louis.

Two-dimensional semiconductors are ultra-thin materials just a few atomic layers thick. Because they can achieve smaller, lower-power operation than conventional silicon chips, they are often called "dream semiconductors." Silicon-based chips are approaching a physical ceiling: as circuits shrink further, power loss and heat generation grow worse. Two-dimensional semiconductors are widely regarded as a next-generation material to overcome that barrier, with potential applications in AI chips, smartphones, data centers, wearable devices, foldable and stretchable electronics, and miniaturized medical sensors.

The challenge has been that 2D semiconductor flakes produced through solution processing vary widely in position, size and thickness, forcing researchers to hunt for suitable samples under a microscope one at a time. Electrode patterns then had to be designed manually around each located flake — a time-consuming process that made it effectively impossible to analyze thousands of devices at once.

A schematic diagram of the research findings (AI-generated image). [Provided by KAIST]
A schematic diagram of the research findings (AI-generated image). [Provided by KAIST]

The research team used molybdenum disulfide (MoS₂), one of the most widely studied 2D semiconductor materials. By exploiting the fact that RGB (red, green, blue) brightness values observed under an optical microscope shift with layer thickness, the team trained a computer system to automatically identify suitable flakes and generate electrode designs without human intervention. Validation using atomic force microscopy confirmed the system could accurately distinguish thickness differences as fine as three to eight atomic layers.

Using this approach, the team screened more than 120,000 semiconductor flakes and successfully fabricated and analyzed 1,615 transistors.

The researchers say the study's broader significance lies not just in automating fabrication, but in shifting 2D semiconductor research from experience-driven intuition to a data-driven methodology. The team plans to fabricate and analyze larger numbers of devices more rapidly, identify higher-performing materials, and ultimately develop AI-assisted semiconductor design.

"This approach — assessing sample thickness and quality through optical inspection alone and feeding that assessment directly into device design — can serve as a rapid sorting and quality-control tool at the research and prototyping stage for 2D material devices," Kwon said. "Our long-term goal is to refine a model that predicts device characteristics from optical signals, enabling pre-fabrication sample selection and contributing to higher yields."

The findings were published in Advanced Functional Materials, an international journal in the field of materials science.


nbgkoo@heraldcorp.com