Pukyong National University

PKNU Research Team Develops Physics-Based AI Analysis Technology

- Prof. Lee Seunghun’s Team from the Department of Physics Presents a Physics-Informed Strategy for Maximizing Learning Efficiency

Professor Lee Seunghun’s research team from the Department of Physics at Pukyong National University (President Bae Sang-hoon) has developed a machine learning-based technology capable of analyzing the properties of superconductors rapidly and accurately within just tens of milliseconds.

Professor Lee Seunghun, along with lead author Lee Dong-ik (master’s program), published their study titled “Rapid analysis of point-contact Andreev reflection spectra via machine learning with physics-guided data augmentation” in <Materials Today Physics> (Impact Factor: 9.7), a prestigious international journal in the field of applied physics.

This study has also been highly regarded academically, as it proposes a strategy to maximize model learning efficiency based on a solid understanding of physics.

Superconductors are materials that exhibit zero electrical resistance, making them essential for various applications such as lossless power transmission, high-field medical equipment (e.g., MRI), and as core materials for quantum computers. With the recent spotlight on high-temperature superconductors following the LK-99 controversy and the growing interest in next-generation quantum computers based on topological superconductors, the need for technology that can quickly and accurately distinguish between various types of superconductors has become increasingly important.

Professor Lee Seunghun’s research team adopted machine learning technology to significantly improve the accuracy and reduce the analysis time of point-contact spectroscopy (PCS), a technique used for analyzing superconductors. While traditional spectrum analysis could take anywhere from several hours to days, the newly developed model enables highly accurate analysis in under 0.1 seconds.

Professor Lee Seunghun explained, “Training an AI model is similar to teaching a baby what a pig is. By repeatedly showing images that emphasize defining features―like a pig’s snout―and saying, ‘This is a pig,’ the baby naturally learns that the snout is a key clue in identifying a pig” (see Figure 1).

Professor Lee Seunghun’s research team designed a model that generates large volumes of theoretical spectra for training and incorporated distorted data emphasizing key spectral features―based on physics knowledge―to enhance the model’s learning. This strategy maximized learning efficiency and significantly improved real-world performance, including analysis accuracy (see Figure 2).

Professor Lee Seunghun stated, “This research is significant not only because it drastically reduced analysis time, but because it presents a physics-guided strategy to maximize machine learning efficiency.” He added, “This technology is expected to accelerate new superconductor research and be broadly applicable to data analysis in fields such as materials science, biomedical engineering, and sensor technology.”

(https://doi.org/10.1016/j.mtphys.2025.101792)