Patterns only become visible at scale. With thousands of devices, you see the actual relationship.
For generations, the search for better semiconductors has demanded the patience of a craftsman — one researcher, one microscope, one flake at a time. A team at KAIST has now automated that entire process, using optical imaging and machine learning to screen over 120,000 material samples and fabricate 1,615 transistors, revealing statistical truths about two-dimensional semiconductors that were simply invisible at the scale human hands could manage. The work marks a quiet but profound shift: from science guided by intuition to science guided by data, opening a path toward AI-designed chips that could carry computing beyond the physical limits of silicon.
- Silicon is running out of room — conventional chips are hitting thermal and physical walls, and the semiconductor industry is racing to find what comes next.
- Two-dimensional materials like MoS₂ have long promised a breakthrough, but their development has been strangled by the sheer slowness of hand-screening thousands of microscopic flakes one by one.
- KAIST researchers cracked the bottleneck by teaching a computer to read the subtle color shifts in microscope images that betray a flake's thickness, then automating the entire path from detection to working transistor.
- The resulting flood of data exposed a hidden trade-off: thicker 2D semiconductors carry current better but lose switching precision — a relationship statistically invisible until 1,615 devices could be compared at once.
- The system is already pointing toward its own successor — feeding this data to AI to design entirely new semiconductors, compressing the timeline from laboratory discovery to commercial chip.
For decades, finding the right semiconductor material meant standing at a microscope, hunting by eye through thousands of candidates, sketching electrode designs by hand, and fabricating devices one at a time. The slowness was not merely inconvenient — it was a ceiling on knowledge, because you can only learn what your sample size allows you to see.
A team led by Professor Jimin Kwon at KAIST, working with partners across South Korea and Washington University in St. Louis, has dismantled that ceiling. Their system uses optical microscope images alone to automatically identify two-dimensional semiconductor flakes, design electrode patterns, and fabricate working transistors — no human hand required at each step. Where researchers once tested dozens of devices in a study, this platform has already screened more than 120,000 flakes and produced 1,615 transistors for analysis.
The material at the center of the work is molybdenum disulfide, or MoS₂, one of the most promising so-called dream semiconductors. These two-dimensional materials — just a few atomic layers thick — could allow chips to run cooler and more efficiently than silicon, enabling next-generation AI hardware, wearables, and medical implants. The problem is that solution-processed flakes are all different, and finding usable ones has always demanded human judgment. The KAIST team discovered that a microscope's color readings shift subtly with material thickness, and trained an algorithm to read those shifts — distinguishing layers as fine as three to eight atoms with confirmed accuracy.
The scale unlocked by automation revealed something previously out of reach: a clear statistical relationship showing that thicker MoS₂ conducts current more easily but switches less precisely — a trade-off long suspected but never demonstrable with the small samples hand-screening allowed. That finding alone reframes how researchers will think about designing these devices.
Published in April in Advanced Functional Materials and selected as a cover article, the work points toward a near future where this accumulated data trains AI systems to design new semiconductors from scratch — shortening the distance between scientific discovery and the chips that power the world.
For decades, the hunt for better semiconductors has been a painstaking affair. Researchers would peer through microscopes, hunting for the right flakes of material among thousands of candidates, marking their positions by hand, sketching electrode designs on paper or screen, then fabricating devices one at a time. It was slow work—the kind that limited what you could learn because you could only afford to test a handful of samples. Now that era is ending.
A team at KAIST, South Korea's premier science and technology university, has automated the entire process. Working with collaborators at UNIST, Hanbat National University, Hanyang University, and Washington University in St. Louis, researchers led by Professor Jimin Kwon developed a system that uses optical microscope images alone to identify two-dimensional semiconductors, then automatically designs and fabricates transistors from them. The breakthrough is not just faster—it is fundamentally different in kind. Where researchers once analyzed dozens of devices, this system has already processed over 120,000 semiconductor flakes and fabricated and tested 1,615 transistors.
Two-dimensional semiconductors are materials only a few atomic layers thick. They are called dream semiconductors because they promise to overcome the hard physical limits that silicon has begun to hit. As conventional chips shrink, they generate more heat and waste more power. Two-dimensional materials could enable smaller, cooler, more efficient devices—the kind that could power next-generation AI chips, foldable phones, wearable sensors, and medical implants. The catch is that when you make these materials through solution processing, every flake is different. Each one has a different position, size, and thickness. Finding the ones you want requires human judgment and patience.
The KAIST team focused on molybdenum disulfide, or MoS₂, a representative two-dimensional material. They noticed that the color values a microscope captures—the red, green, and blue brightness—shift subtly depending on how thick the material is. A computer can see those shifts. By training an algorithm to recognize them, the researchers enabled automated identification of flakes with specific thickness ranges. The system then automatically designs the electrode patterns needed to turn each flake into a working transistor. Atomic force microscopy confirmed that the system can distinguish even tiny differences—between three and eight atomic layers—with accuracy.
The scale of analysis made possible by this automation revealed something that had been hidden in plain sight. The team discovered a clear statistical relationship between thickness and electrical performance: as the material gets thicker, current flows more easily, but the ability to switch electricity on and off—a critical function for any transistor—actually declines. This trade-off had been suspected but never clearly demonstrated, because testing enough samples to see the pattern statistically was simply not feasible when everything was done by hand.
What matters most about this work is not the speed, though speed matters. It is the shift in how research happens. Two-dimensional semiconductor science has relied on human intuition and experience—the researcher's eye and judgment. This automation transforms it into data-driven research. Thousands of devices yield thousands of data points. Patterns emerge that no human could spot by testing a few dozen samples. The next step, already visible on the horizon, is to feed this data to AI systems and let them design new semiconductors from scratch.
The research was published in April in Advanced Functional Materials, a leading journal in materials science, and was selected as a cover article in the 2D Materials & Electronics section. The work was funded by South Korea's National Research Foundation and the Korea Planning & Evaluation Institute of Industrial Technology. For the semiconductor industry, which has been searching for materials to replace silicon as the physical limits of conventional chips draw near, this represents a significant acceleration. The path from discovery to commercialization just got shorter.
Citas Notables
The greatest significance of this study is that it transformed two-dimensional semiconductor research from human experience-based to data-driven research.— KAIST research team
La Conversación del Hearth Otra perspectiva de la historia
Why does it matter that researchers can now test thousands of devices instead of dozens?
Because patterns only become visible at scale. When you test thirty samples by hand, you see noise and variation. You can't tell if something is a real trend or just chance. With 1,615 transistors, you can see the actual relationship between thickness and performance with statistical confidence. That's the difference between guessing and knowing.
The system uses color values from a microscope image to identify thickness. How does that work?
Light behaves differently depending on what it passes through. When you look at a two-dimensional material under a microscope, the thickness changes how much red, green, and blue light gets reflected back. The computer learns to read those color shifts the way a human eye learns to read a face. It's pattern recognition, but at a scale humans can't match.
You said this shifts the field toward data-driven research. What was it before?
Craft. Intuition. A researcher would develop a feel for which samples looked promising, which positions on the substrate were likely to work well. That knowledge was real and valuable, but it couldn't scale. It couldn't be shared precisely. Data-driven means the knowledge is in the numbers, reproducible, shareable, and available to be fed into AI systems.
What's the practical consequence? When does this matter to someone buying a phone or using AI?
Not immediately. This is foundational research. But two-dimensional semiconductors are the leading candidate for what comes after silicon. If this automation accelerates the discovery of better materials and faster paths to manufacturing, you're looking at chips that are smaller, use less power, and generate less heat. For AI systems running in data centers, that's enormous. For wearables and medical devices, it's the difference between feasible and impossible.
The team found that thicker semiconductors conduct current better but switch worse. Why is that a problem?
A transistor is a switch. You need current to flow when it's on, but you also need to turn it off quickly and completely. If thickness helps one but hurts the other, you have to choose. Now that researchers know this trade-off exists and can measure it precisely, they can start designing around it—finding materials or structures that don't have the same limitation.
What happens next?
The team has built the infrastructure. The next phase is using this data to train AI systems to predict which materials will perform well, and eventually to design new materials from scratch. You're moving from discovery to design. That's when the real acceleration begins.