Advancing AI for science: extracting and fusing cross-disciplinary expert knowledge with data to accelerate alloy discovery
News Release - 13-Feb-2026
Ishikawa, Japan
High-entropy alloys promise serious performance in extreme environments, but discovering useful compositions is slow and costly. A team led by researchers at the Japan Advanced Institute of Science and Technology (JAIST) has built an AI-for-Science framework that fuses experimental data with expert knowledge extracted from literature-while quantifying uncertainty at every step.
Instead of betting on models that only interpolate around known data, the framework uses two independent evidence streams and a formal method to combine them. The result: reliable predictions even for poorly studied compositions, and clear signals on where current knowledge falls short.
Why this matters for materials R&D
HEAs mix multiple elements in near-equal ratios to achieve strength, stability, and durability. But each added element explodes the search space, making brute-force exploration impractical.
Conventional machine learning can help, but performance drops as you move away from well-studied regions. Meanwhile, decades of substitution rules and mechanistic insights are scattered across papers-hard to use at scale.
What the team built
The study, published in Digital Discovery (19-Dec-2025), integrates materials data with AI-extracted knowledge from five disciplines: metallurgy, solid-state physics, materials mechanics, materials science, and corrosion science. The team includes researchers from JAIST, HPC Systems (Japan), the Institute of Statistical Mathematics (Japan), and Duke University (USA).
- Two evidence sources for elemental substitution: (1) materials datasets where near-identical alloys imply substitutable elements; (2) judgments extracted from literature using large language models (GPT-4o, GPT-.5, Claude Opus 4, Grok3).
- Evidence fusion via Dempster-Shafer theory, which explicitly represents uncertainty and ignorance.
- Uncertainty-aware predictions and compositional maps that highlight where the model is confident versus where information is thin.
Performance highlights
- Accurate predictions for alloys containing elements absent from the training data, with 86%-92% accuracy.
- Validation against 55 experimentally confirmed quaternary alloys from the literature.
- Outperformed conventional ML baselines in data-scarce settings and beat more computationally expensive free-energy models.
How it works (at a glance)
- Identify substitution patterns by comparing alloy pairs that differ by one element in large datasets.
- Query LLMs across five disciplines to extract cross-checked expert judgments from scientific papers.
- Convert each evidence stream into belief functions and fuse them using Dempster-Shafer theory.
- Score candidate alloys with explicit confidence and uncertainty; generate maps to guide experiments.
As the team puts it: "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."
What this enables for researchers
- Explore under-sampled composition spaces without overfitting to familiar chemistries.
- Plan experiments where they'll be most informative, using uncertainty maps to prioritize.
- Systematically integrate dispersed expert knowledge with data-useful well beyond alloys.
- Extend the same workflow to drug discovery, batteries, and catalysts where substitution, mechanism, and data scarcity are common.
Publication and reference
Peer-Reviewed Publication
Digital Discovery - DOI: 10.1039/D5DD00400D
Title: 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
Credit: Hieu-Chi Dam from Japan Advanced Institute of Science and Technology, Japan.
About JAIST
Founded in 1990 in Ishikawa prefecture, JAIST is a leading graduate university in Japan with a coursework-oriented curriculum designed to support high-impact research. Diversity is central to its mission; about 40% of alumni are international students. The university collaborates closely with industry and global partners.
About Professor Hieu-Chi Dam
Professor Hieu-Chi Dam holds appointments at JAIST and Tohoku University's International Center for Synchrotron Radiation Innovation. His work spans data science and materials informatics, integrating first-principles calculations, machine learning, and diffraction physics across magnetic materials, superconductivity, and strongly correlated systems, with 90+ publications.
Funding
Supported by JST-CREST (JPMJCR2235); JSPS KAKENHI (20K05301, JP19H05815, 20K05068, 23KJ1035, 23K03950, JP23H05403). S.C. acknowledges US-DoD ONR MURI N00014-21-1-251. H.K. acknowledges JST ASPIRE "International Collaborative Research Network for Advanced Atomic Layer Processes." The authors thank Dr Huan Tran and Dr Xiomara Campilongo for discussions.
For teams scaling AI skills
If your lab is building LLM-based extraction or uncertainty-aware workflows, you may find curated AI course paths useful: Latest AI courses.
Your membership also unlocks:
