Choi Jun-hyuk, a master's student, Park Se-jin, a master's student, and Won Yeo-ji, an undergraduate research student (all from the Department of New Materials Engineering, Artificial Intelligence Materials Design Laboratory, advised by Professor Ki Sub Cho) recently won the Outstanding Paper Award (Jeong In-sang Award) at the 2025 Korean Society of Heat Treatment Spring Academic Conference, demonstrating their excellent research capabilities.

Choi Jun-hyuk presented a paper titled “Potential Detection and Segmentation of Atomic-Scale STEM Images Using U-Net-Based Focused Region Training,” proposing an AI-based image analysis model that overcomes the inefficiencies and errors associated with traditional manual-based potential analysis methods. In particular, by emphasizing areas where dislocations are concentrated using the Focused Region Training (FRT) technique and utilizing an ensemble strategy, he achieved model performance capable of accurately detecting even minute defects.

Park Se-jin presented on the topic of “Development of a Full-Cycle Design Framework for Fe-based Soft Ferromagnetic Alloys Using LLM-based Data Extraction and Semi-Supervised Learning.” He constructed a refined database of Fe-based magnetic materials' composition, processes, and magnetic properties through a large-scale language model (LLM)-based automatic extraction system from literature and patents. Based on this, he developed a prediction model combining Autoencoder-based semi-supervised learning and Gaussian Process-based Active Learning, deriving the optimal composition that simultaneously satisfies actual manufacturing performance and prediction reliability.

Won Yeoji, an undergraduate research student, presented research on “OCR-based automatic extraction of data from PDF literature on nickel-based superheat-resistant alloys and accuracy evaluation using multimodal LLM.” We developed a system that automatically extracts text, tables, and images from PDF-format literature on nickel-based superheated alloys using OCR, and then quantitatively verifies the consistency of the extracted data using a multimodal LLM. Through evaluations using BLEU metrics and a confusion matrix, we achieved higher accuracy compared to manual processing, demonstrating the potential for expanding automated data processing in material science.

This award reaffirms the National University of Korea's Artificial Intelligence Materials Design Research Laboratory as a leading research group in AI-based materials design and data-driven automation technology, and is expected to contribute to various materials industries in the future.