Park Se-jin, a Ph.D. candidate, and Won Hye-ji, an undergraduate research assistant, both from the Artificial Intelligence Materials Design Laboratory in the Department of New Materials Engineering at Kookmin University (advisor: Professor Cho Ki Sub), demonstrated their outstanding research capabilities by winning the Outstanding Oral Presentation Award and the Outstanding Student Poster Presentation Award, respectively, at the Artificial Intelligence Materials Science session of the 2026 Spring Conference of the Korean Society of Metals and Materials, held recently.

Ph.D. candidate Park Se-jin delivered an oral presentation titled “OPSI: An On-premises LLM-based Schema-Inductive Framework for CPSP Database Construction of Ni-based Superalloys,” while undergraduate research assistant Won Ye-ji presented a poster titled “Research on LLM-based Automation of Table Data Extraction and Prompt Self-Improvement in an On-premises Environment.” Both studies are significant in that they propose methodologies for reliably extracting composition, process, microstructure, and property information from literature data on nickel-based superalloys and utilizing this information for the construction of AI-based materials databases and advanced alloy design.

Recently, there have been active efforts in the field of materials research to systematically collect and analyze existing literature data in order to reduce experimental time and costs and accelerate alloy design. However, in the case of nickel-based superalloys—where table structures in papers are complex and individual data points hold high value—existing automatic extraction methods alone had limitations in securing accurate data. Furthermore, cloud-based large language models (LLMs) pose security constraints due to the risk of sensitive research data being transmitted to external servers, leading to a growing need for on-premises LLM environments that can ensure both security and reproducibility.

In response, Ph.D. candidate Park Se-jin designed the on-premises OPSI framework, which performs the entire data processing workflow within an internal environment, and analyzed the numerical hallucination issue of on-premises LLMs from the perspective of the model’s internal information processing and inference processes. Based on this, he improved the existing single-prompt-based extraction method and proposed a three-stage pipeline consisting of item identification, source text verification, and JSON structuring. As a result, the OPSI framework extracted approximately twice as many items as the existing GPT-4o-based extraction method and also demonstrated improved extraction accuracy.

Undergraduate research student Won Hye-ji proposed a methodology for exploring optimal prompts by applying GP-UCB-based active learning in an on-premises LLM environment. In the study, prompts were structured with three slots and five candidates per slot, creating a total of 125 combinations, and GP-UCB was utilized to strategically evaluate promising combinations. Additionally, the team designed a loss function that imposes a high penalty for hallucination, thereby increasing extraction reliability. As a result, they reached the top-performing prompt—equivalent to that achieved through a full search—with just 10 evaluations, demonstrating the efficiency of active learning-based prompt optimization.

This award recognizes the excellence of the research being conducted by Kookmin University’s Artificial Intelligence Materials Design Laboratory in the areas of AI-based materials database construction, prompt engineering, and automated materials design. In particular, it demonstrates that an on-premises LLM environment serves as a practical research infrastructure capable of simultaneously ensuring research security and data extraction efficiency, and is expected to play a crucial role in the future construction of high-precision databases for the design of high-temperature structural materials and advanced alloys.