AI literature mining accelerates discovery of heat-stable lead-free dielectrics

A team at Seoul National University has demonstrated an artificial-intelligence workflow that extracts scattered data from hundreds of research papers to design new lead-free dielectric materials with high-temperature stability. The study, published in Nature Communications, was led by Professor Ho Won Jang of the Department of Materials Science and Engineering, with integrated M.S./Ph.D. student Kwanwoo Song as first author.

Dielectrics are insulating materials that store electric charge and are essential components of multilayer ceramic capacitors (MLCCs) used in smartphones, electric vehicles, and aerospace electronics. For practical use, these materials must maintain a high dielectric constant — the ability to store energy — across a wide temperature range. Conventional trial-and-error searches are impractical because the number of possible lead-free compositions is virtually limitless, and relevant experimental data are dispersed across text, tables, and graphs in diverse publications.

To overcome this, the researchers built a dataset of 1,202 dielectric-property records from 448 papers. They used large language models to parse composition and processing details from text and tables, and converted graphical data into numerical temperature-dependent properties. The team then incorporated 22 physical descriptors, such as elemental composition and microstructure, and trained 30 independent machine-learning models to predict three key performance indicators simultaneously.

The framework assessed agreement among the models' predictions to prioritize candidates with higher confidence. After applying predefined performance targets and physicochemical constraints to approximately 150 million virtual compositions, the search space was narrowed to 37 candidates. The researchers then fine-tuned component ratios within the most promising compositional family and selected two materials for experimental synthesis.

The two synthesized compositions, designated SNBTS1 and SNBTS2, contained 1 mol% and 2 mol% tin (Sn) substitution, respectively. Experiments confirmed room-temperature dielectric constants of 3,422 and 3,307. Both materials maintained high dielectric constants over a broad temperature range and satisfied the international X5R, X6R, and X7R stability standards for MLCCs, which require the dielectric constant to remain within ±15% of its 25°C value up to 85°C, 105°C, and 125°C, respectively.

Compared with barium titanate (BaTiO₃), a widely used dielectric whose constant changes sharply around 125°C, the new materials showed far more stable performance. The machine-learning predictions closely reproduced the experimentally observed trends. Additional characterization using piezoresponse force microscopy, Raman spectroscopy, and atomic-resolution electron microscopy revealed that the small Sn addition expands the crystal framework and increases electrical heterogeneity at the atomic scale, enhancing temperature stability without substantially reducing the dielectric constant.

The researchers emphasize that the significance of the work lies in the methodology: integrating fragmented literature data into a unified training set, embedding physical laws, and using model consensus to identify synthesizable candidates. They suggest the approach could be extended to other functional oxides and thin-film materials where data are similarly scattered. The lead-free dielectrics validated here are expected to enable high-temperature MLCCs for electric vehicles, power electronics, and aerospace systems.