SK hynix (000660.KS), the world's second-largest memory chipmaker, has secured a new artificial intelligence technology capable of effectively detecting defects in semiconductor wafers to improve yield. The company plans to accelerate the construction of an AI fab — a semiconductor factory that maximizes production efficiency through AI — by developing new technologies and partnering with big tech companies.
A joint research team from SK hynix and Korea University published a "super-resolution model for wafer transmission electron microscopy (TEM) images" in Scientific Reports, a Nature sister journal, on May 30 local time, according to industry sources on Monday. The research was directly funded and supported by SK hynix.
A transmission electron microscope is a device that transmits electrons instead of light through a specimen, producing far more precise images than conventional optical microscopes. The AI model is expected to analyze TEM images of wafers and identify circuit defects and material properties at nanometer-level precision during the manufacturing process, replacing human inspectors.
However, TEM images differ from optical images in analysis methods, suffer from a shortage of training data, and involve complex semiconductor characteristics. As a result, even image models developed by big tech firms such as Google have shown limited accuracy.
The research team developed an AI model specialized in wafer image analysis through a proprietary training method and achieved the highest accuracy among existing models in comparison tests. The model scored 68.2% on the mean Intersection over Union (mIoU) metric, a key indicator of image analysis accuracy, averaging 10.3 percentage points higher than nine major competing models.
In semiconductors, a 1% difference in yield can translate into tens of billions of won in revenue differences, making the improvement of AI-driven defect analysis a top priority for AI fabs.
The research is part of SK hynix's broader plan to build an AI fab. The company aims to establish an autonomous fab by 2030 — a facility that learns and makes decisions on its own, dramatically shortening the transition from design to mass production. The goal is to cut the time required for key tasks such as defect analysis by more than 50%, thereby expanding production capacity for flagship products like High Bandwidth Memory (HBM) while significantly reducing costs.