KAIST "Applying Semiconductor Inspection Method to Batteries, Achieves Drastic Reduction in Defect Rate"

An image recognition technology that applies inspection methods used in semiconductor processes to the battery field has been developed, significantly reducing defect rates. This technology focuses on accurately predicting information such as the elemental composition and charge-discharge cycles of batteries solely from the surface morphology of the battery through artificial intelligence learning.

KAIST announced on the 2nd that Professor Hong Seungbeom of the Department of Materials Science and Engineering at KAIST, the Electronics and Telecommunications Research Institute (ETRI) of Korea, and Drexel University in the United States have jointly developed a methodology that inspects the main elemental composition and charge-discharge status of batteries with 99.6% accuracy, the first of its kind in the world.

(From left) Professor Hong Seungbeom, Dr. Oh Jimin, Dr. Yeom Jiwon, Department of Materials Science and Engineering. Provided by KAIST

The methodology developed by the joint research team involves training artificial intelligence based on a convolutional neural network (CNN) with scanning electron microscope (SEM) images of NCM (Nickel-Cobalt-Manganese) cathode materials of various compositions and different charge-discharge cycles, aiming to reduce battery defect rates. A convolutional neural network is a type of multilayer feed-forward artificial neural network used for analyzing visual images.

In semiconductor processes, SEM is mainly used for defect inspection of wafers. However, in battery processes, the use of SEM is rare, and for degraded battery materials, reliability must be predicted from fractured and broken particle morphologies.

The joint research team proceeded with the study inspired by these field conditions. First, they judged that if battery processes, like semiconductor processes, could reduce defect rates by inspecting the cathode material surface with automated SEM to verify whether the desired composition was synthesized and whether a reliable lifespan could be expected, it would be revolutionary.

Based on this judgment, the joint research team trained a convolutional neural network-based artificial intelligence, applicable to autonomous vehicles, with surface images of battery materials, enabling prediction of the main elemental composition and charge-discharge cycle status of cathode materials.

They also confirmed whether this methodology could be applied to cathode materials containing additives, finding that while the composition was predicted quite accurately, the accuracy for charge-discharge status was relatively lower.

Accordingly, the joint research team plans to train the AI with morphologies of battery materials produced through various processes in the future, aiming to utilize it for composition uniformity inspection and lifespan prediction of next-generation batteries.

Professor Hong Seungbeom of KAIST stated, “The significance of this joint research lies in developing an AI-based methodology that quickly and accurately predicts the main elemental composition and charge-discharge status from scanning electron microscope images of material structures.” He added, “The microscopy image-based method for distinguishing battery material composition and status developed in this study will play an important role in improving the performance and quality of battery materials in the future.”

Meanwhile, the joint research was conducted with participation from co-first authors Dr. Oh Jimin and Dr. Yeom Jiwon, graduates of the Department of Materials Science and Engineering at KAIST, co-author Dr. Kim Kwangman of ETRI, and Professor Agar of Drexel University in the United States. The research results were published in the international journal npj Computational Materials (May 4 issue).