Ishikawa, Japan --High-entropy alloys are promising advanced materials for demanding applications, but discovering useful compositions is difficult and expensive due to the vast number of possible element combinations. Now, researchers have developed a novel AI-driven framework that integrates experimental data, computational modeling, and cross-disciplinary expert knowledge extracted from scientific literature. By combining these sources in a way that accounts for uncertainty, their approach can make reliable predictions even for poorly studied alloy compositions, outperforming conventional data-driven machine learning methods that rely on training data alone.
Progress in modern technologies relies on advanced materials, such as alloys used in aircraft engines and components capable of resisting corrosion and heat in industrial settings. In this context, high-entropy alloys (HEAs) have emerged as one of the most promising areas of study in materials science. By combining several elements in near-equal amounts, these materials can achieve exceptional strength, stability, and durability. However, discovering useful HEAs is exceptionally difficult and expensive, as each additional element dramatically increases the number of possible combinations. With growing demand for sustainable energy technologies and next-generation electronics, accelerating the discovery of advanced materials has become increasingly urgent.
Researchers worldwide have turned to artificial intelligence (AI) as a powerful aid in materials research—but this comes with its own limits. Most machine learning models are good at interpolation, meaning they can make predictions for materials that closely resemble those already in their training data. When researchers move beyond familiar territory and consider truly new compositions, the models’ accuracy drops. Meanwhile, decades of expert knowledge about how elements interact and substitute for one another in HEAs are buried across the scientific literature, with no clear way to integrate that expertise into data-driven AI tools.
Against this backdrop, a research team led by Professor Hieu-Chi Dam from Japan Advanced Institute of Science and Technology (JAIST), Japan, has developed a new AI-driven framework for HEA discovery. Their study, published in the journal Digital Discovery on December 19, 2025, was co-authored by JSPS’s Researcher Dr. Minh-Quyet Ha and doctoral student Dinh-Khiet Le at JAIST, Dr. Viet-Cuong Nguyen from HPC Systems, Japan, Professor Hiori Kino at the Institute of Statistical Mathematics, Japan, and Professor Stefano Curtarolo from Duke University, USA. The team set out to combine experimental and computational materials data with cross-disciplinary expert knowledge extracted directly from scientific papers, creating a system designed specifically to work in data-scarce and unexplored regions.
At the core of the approach is a well-known idea in alloy design called elemental substitution. Under the optimal conditions, chemically similar elements can substitute one another without significantly affecting a material’s properties and stability. The researchers first identified substitution patterns directly from large materials datasets by comparing alloys that differ by just one element. They then used state-of-the-art large language models (LLMs), including GPT-4o, GPT-.5, Claude Opus 4, and Grok3 to extract expert judgments in the literature pertaining to five key scientific disciplines: metallurgy, solid-state physics, materials mechanics, materials science, and corrosion science.
Each source of information provided a piece of evidence rather than a final answer, and these pieces were combined using a mathematical framework known as Dempster–Shafer theory. Unlike standard probability methods, this framework can explicitly represent uncertainty and even ignorance, as Prof. Dam explains, “ Traditional classifiers force binary ‘yes-or-no’ predictions even when information is insufficient. Our approach explicitly quantifies uncertainty, allowing ‘we cannot tell’ as a legitimate scientific outcome. ” Simply put, the proposed system does not pretend to know more than it does when exploring unknown territory.
When tested on several alloy datasets, the team's framework consistently outperformed conventional machine learning models, especially when little information was available. . Most strikingly, it was able to predict the behavior of alloys containing elements that were completely absent from the training data, achieving accuracy rates ranging from 86% to 92%. The researchers also validated their approach against 55 experimentally confirmed quaternary alloys from the literature and showed that it outperforms far more computationally expensive methods, such as free-energy models. Beyond individual predictions, the proposed method can produce compositional maps that show where predictions are reliable and where uncertainty remains high. This enables researchers to focus experiments on the most promising and informative regions of the compositional space.
According to Prof. Dam, the broader significance of this work lies in showcasing how AI can be used for scientific discovery. “ LLM-based extraction combined with formal evidence fusion can transform decades of dispersed expert knowledge into searchable, comparable, and quantitatively usable resources, which are particularly valuable for interdisciplinary problems where relevant insights span multiple fields ,” he remarks. Notably, the same approach used in this study could accelerate drug discovery, guide battery development, and help optimize catalysts. In each case, the framework’s ability to quantify uncertainty would help research teams prioritize the most informative experiments, potentially reducing discovery timelines and costs.
Overall, this work demonstrates a path forward for AI in scientific discovery—one where machine learning serves not to replace expert judgment, but to systematically extract and combine it with experimental evidence to accelerate innovation across disciplines.
###
Title of original paper:
Beyond interpolation: integration of data and AI-extracted knowledge for high-entropy alloy discovery
Authors:
Minh-Quyet Ha, Dinh-Khiet Le, Viet-Cuong Nguyen, Hiori Kino, Stefano Curtarolo and Hieu-Chi Dam*
Journal:
Digital Discovery
DOI:
About Japan Advanced Institute of Science and Technology, Japan
Founded in 1990 in Ishikawa prefecture, the Japan Advanced Institute of Science and Technology (JAIST) was the first independent national graduate university that has its own campus in Japan. Now, after 30 years of steady progress, JAIST has become one of Japan’s top-ranking universities. JAIST strives to foster capable leaders with a state-of-the-art education system where diversity is key; about 40% of its alumni are international students. The university has a unique style of graduate education based on a carefully designed coursework-oriented curriculum to ensure that its students have a solid foundation on which to carry out cutting-edge research. JAIST also works closely both with local and overseas communities by promoting industry–academia collaborative research.
About Professor Hieu-Chi Dam from Japan Advanced Institute of Science and Technology, Japan
Dr. Hieu-Chi Dam is a Professor at Japan Advanced Institute of Science and Technology (JAIST) and at Tohoku University’s International Center for Synchrotron Radiation Innovation. He received Master’s and Ph.D. degrees from JAIST in 2000 and 2003, respectively. His research focuses on data science and materials informatics, integrating first-principles calculations, machine learning, and diffraction physics to study magnetic materials, superconductivity, and strongly correlated systems. He has published over 90 papers on these topics.
Funding information
This work was supported by the JST-CREST Program (Innovative Measurement and Analysis), under Grant number JPMJCR2235 and the JSPS KAKENHI Grant Numbers 20K05301, JP19H05815, 20K05068, 23KJ1035, 23K03950, and JP23H05403. S.C. acknowledges support from US-DoD (ONR MURI program number N00014-21-1-251). H. K. acknowledges support from the Japan Science and Technology Agency (JST) ASPIRE Program under the project “International Collaborative Research Network for Advanced Atomic Layer Processes.” The authors thank Dr Huan Tran and Dr Xiomara Campilongo for fruitful discussions.
19-Dec-2025