Researchers in South Korea have used AI to mine hundreds of scientific papers and narrow a search space of roughly 150 million possible material compositions down to two experimentally validated lead-free dielectrics that hold up at high temperatures. The approach, developed at Seoul National University and published in Nature Communications, replaces trial-and-error materials discovery with a data-driven workflow that could extend to other classes of materials.
A literature problem, not just a chemistry problem
Dielectric materials store electrical charge and are key components in multilayer ceramic capacitors (MLCCs), which sit inside smartphones, electric vehicles, and power electronics. The challenge: the best-performing dielectrics often contain lead, and the relevant data for lead-free alternatives is scattered across thousands of papers in different formats-text, tables, and graphs-with inconsistent measurement conditions.
The research team, led by professor Ho Won Jang of the Department of Materials Science and Engineering, built a dataset of 1,202 dielectric-property records from 448 papers. First author Kwanwoo Song, an integrated M.S./Ph.D. student, led the work from data extraction through experimental validation.
The team used large language models to organize composition and processing information from text and tables, then converted graph images into numerical data to capture how dielectric properties change with temperature. Twenty-two physical descriptors were incorporated to unify information across publications, and 30 independently trained machine-learning models were combined to predict three indicators related to dielectric constant and temperature stability.
From millions of candidates to two materials
After applying performance targets and physicochemical constraints to roughly 150 million virtual compositions, the framework flagged 37 candidates. The researchers then adjusted component ratios within the most promising compositional family and selected two compositions for synthesis-one with 1% tin substitution, the other with 2%.
Both materials delivered strong performance. The samples showed room-temperature dielectric constants of 3,422 and 3,307, respectively, and maintained stability across temperature ranges set by the international X5R, X6R, and X7R MLCC standards-which the materials exceeded. The 1% and 2% tin substitutions improved temperature stability without sacrificing dielectric constant, thanks to tin "expanding the crystal framework and increasing electrical heterogeneity at the atomic scale."
The team's analysis used piezoresponse force microscopy, Raman spectroscopy, and atomic-resolution electron microscopy to verify the physical mechanism behind the performance gains.
The methodology matters beyond this particular set of materials. For researchers working on materials R&D, the study shows it's possible to systematically integrate fragmented scientific data and use model agreement as a selection tool.
"The significance of this study lies not simply in predicting performance with machine learning, but in integrating information scattered across multiple papers into a training dataset under a project that considers both physical laws and consistency among model predictions to narrow the search all the way down to candidates that could actually be synthesized," Jang said.
It's worth remembering the recruiting and synthesis pipeline still required human judgment. The machine-learning models narrowed a massive space; the team then synthesized two materials and tested them thoroughly.
The research group's previous work applied similar machine-learning techniques to discover a tungsten single-atom catalyst for water electrolysis, also published in Nature Communications.
Why this matters for materials scientists and AI researchers
This study is worth reading if you're building AI pipelines for scientific discovery or working on high-performance dielectrics. The team's method-pairing multimodal literature mining with physics-informed constraints-could apply to any domain where data lives in scattered, unstructured formats including functional oxides and thin-film materials.
The work also demonstrates that combining multiple machine-learning models, rather than relying on a single model's predictions, can produce results that are physically plausible enough to validate in the lab. That's a useful lesson for anyone designing automated discovery workflows for materials and chemistry. For practical applications, the lead-free dielectrics developed here could find their way into high-temperature capacitors for vehicles, power electronics, and aerospace systems.
Your membership also unlocks:
