AI Mines Research Papers to Discover New Material: SNU Team Develops High-Temperature-Stable Lead-Free Dielectric

- Dataset of 1,202 records built from 448 papers; approximately 150 million virtual compositions screened to identify candidate materials

- Kwanwoo Song led the entire research process, from dataset construction and AI model development to candidate screening and experimental validation

- Expected to support development of materials for next-generation MLCCs and electric-vehicle electronics

- Published in the international journal Nature Communications

▲ Conceptual image illustrating how a machine-learning model trained on data constructed through multimodal literature mining and physical knowledge explores the vast compositional space of lead-free dielectrics and identifies candidate materials for experimental validation.

Artificial intelligence has analyzed data scattered across hundreds of research papers to discover new lead-free dielectric materials that maintain stable performance even at high temperatures. The study presents a new approach that could transform materials discovery from a trial-and-error process into a data-driven one.

Seoul National University College of Engineering announced that a research team led by Professor Ho Won Jang of the Department of Materials Science and Engineering has developed a technology for designing lead-free dielectric materials by combining data extracted from scientific literature with physics-informed machine learning. Kwanwoo Song, an integrated M.S./Ph.D. student, served as the first author and led the overall research, while integrated M.S./Ph.D. student Youngmin Kim and postdoctoral researcher Jaehyun Kim participated in the collaborative study.

Dielectrics are insulating materials that prevent electricity from flowing directly while storing electric charge, and they are key materials in multilayer ceramic capacitors (MLCCs) used in smartphones, electric vehicles, and other electronic devices. The higher the dielectric constant, the more electrical energy a component of the same size can store. For practical use in electronic devices, however, dielectric performance must also remain stable at high temperatures.

The research team combined multimodal literature mining, which automatically extracts information distributed across the text, tables, and graphs of scientific papers, with physics-informed machine learning to develop an inverse-design approach that first identifies compositions with a high likelihood of meeting targeted performance requirements. After constructing a dataset of 1,202 dielectric-property records from 448 papers, the researchers explored a virtual compositional space of approximately 150 million possibilities and narrowed it down to 37 candidates. They then synthesized two of these compositions and experimentally confirmed both high dielectric constants and excellent high-temperature stability.

The findings were published in the internationally renowned journal Nature Communications.

As more electronic devices operate at high temperatures—including electric vehicles, power electronics, and aerospace equipment—the importance of dielectric materials that maintain stable performance despite temperature fluctuations is increasing. In particular, relaxor ferroelectrics, whose electrical response changes relatively gradually with temperature, have the potential to combine high dielectric constants with stability across a broad temperature range. Yet even when the search is restricted to lead-free compositions, the number of potential candidates is virtually limitless because of the many possible combinations of elements and mixing ratios, making trial-and-error exploration costly and time-consuming. Moreover, relevant data are scattered across the text, tables, and graphs of different papers, while measurement conditions such as temperature, frequency, and sample characteristics vary from study to study, making the data difficult to use directly for machine-learning training.

To address these challenges, the researchers developed a machine-learning framework that integrates information distributed across multiple papers into a unified format while incorporating physical laws to screen for materials that can realistically exist. The team used large language models to organize composition and processing conditions from the text and tables of research papers, while converting graphs into numerical data to extract temperature-dependent dielectric properties. This process yielded 1,202 records covering composition, processing conditions, and dielectric properties from 448 papers. The researchers then incorporated 22 physical descriptors, including elemental composition and microstructure, to integrate information scattered across different publications into a unified training dataset. Next, they combined 30 independently trained machine-learning models to simultaneously predict three key indicators related to dielectric constant and temperature stability. The framework was also designed to assess agreement among the models’ predictions, allowing candidates with higher predictive confidence to be prioritized.

After sequentially applying predefined performance targets and physicochemical constraints to approximately 150 million virtual compositions, the team narrowed the search space to 37 candidate materials. The researchers then finely adjusted component ratios within the compositional family containing the largest number of remaining candidates and selected two compositions for experimental testing.

Experiments showed that the two samples, in which 1 mol% and 2 mol% of tin (Sn) were substituted, exhibited high room-temperature dielectric constants of 3,422 and 3,307, respectively. The small amount of Sn substitution created a favorable trade-off, improving temperature stability without substantially reducing the dielectric constant. Compared with barium titanate (BaTiO₃), which is widely used in today’s multilayer ceramic capacitors, the newly developed materials were also found to maintain high dielectric constants more consistently over a broader temperature range. Both samples satisfied the high-temperature stability requirements of the international X5R, X6R, and X7R standards for multilayer ceramic capacitors and recorded among the highest dielectric constants when compared with previously reported data for materials in the same compositional family.

*X5R, X6R, and X7R: Temperature-stability classifications for dielectric materials used in multilayer ceramic capacitors, generally indicating whether the dielectric constant remains within ±15% of its value at 25°C from −55°C up to 85°C, 105°C, and 125°C, respectively.

The researchers combined insights into the machine-learning model’s decision-making with results from piezoresponse force microscopy, Raman spectroscopy, and atomic-resolution electron microscopy. Their analysis revealed the underlying mechanism by which a small amount of Sn expands the crystal framework and increases electrical heterogeneity at the atomic scale, thereby enhancing temperature stability.

▲ Temperature-dependent performance of the lead-free dielectrics (left) ‘SNBTS1,’ containing 1 mol% tin (Sn), and (right) ‘SNBTS2,’ containing 2 mol% Sn. The colored solid lines represent experimental measurements, while the black dashed lines indicate machine-learning predictions. Both materials maintained high dielectric constants—their ability to store electrical energy—over a broad temperature range, while the predicted results closely reproduced the trends observed experimentally. The lower curves indicate the degree of electrical energy loss.

▲ (Left) Performance comparison between lead-free dielectrics previously reported in the literature and the materials developed in this study. The stars indicate the reference material without tin (SNBT) and the materials containing 1 and 2 mol% Sn (SNBTS1 and SNBTS2), respectively. Materials positioned farther toward the upper right exhibit both higher dielectric constants and better temperature stability. (Right) Comparison of the temperature-dependent dielectric constants of the materials developed in this study with those of barium titanate (BaTiO₃), a representative dielectric material. While the dielectric constant of barium titanate changes sharply at around 125°C, SNBTS1 and SNBTS2 maintain relatively stable dielectric constants over a broad temperature range.

The study is significant because it demonstrates a research methodology that uses AI to integrate and analyze data scattered throughout the scientific literature and then applies those data to the design of new materials. The researchers believe the approach could also be applied to the development of a wide range of other materials, such as functional oxides and thin films, for which relevant data are distributed across numerous publications. The lead-free dielectrics validated in this study are expected to find applications in the development of high-temperature multilayer ceramic capacitors as well as electronic components for electric vehicles, power electronics, and aerospace systems.

▲ From left: Kwanwoo Song, integrated M.S./Ph.D. student in the Department of Materials Science and Engineering at Seoul National University and currently with Samsung Electro-Mechanics; Youngmin Kim, integrated M.S./Ph.D. student; Jaehyun Kim, postdoctoral researcher; and Professor Ho Won Jang.

Professor Ho Won Jang said, “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 and then considering both physical laws and consistency among model predictions to narrow the search all the way down to candidates that could actually be synthesized.”

He added, “We hope the strategy presented in this study—combining multimodal literature mining with physics-informed machine learning—will extend beyond dielectric materials to the discovery of other functional oxides and thin-film materials, where data are scattered across numerous papers and formats and therefore require systematic integration.”

Kwanwoo Song, the first author of the study and an integrated M.S./Ph.D. student, led the entire research process, from constructing the literature-derived dataset and developing the machine-learning models to screening candidate materials and conducting experimental validation. He is currently conducting research on machine-learning-based discovery of new materials, expanding his work into a range of electronic materials, including new lead-free dielectric and MLCC compositions as well as oxide channel materials for semiconductor transistors. Building on this research experience, he plans to continue R&D on high-performance electronic and dielectric materials.

Meanwhile, Professor Ho Won Jang’s research team previously used AI to discover and experimentally validate a tungsten single-atom-based, non-precious-metal water-electrolysis catalyst for green hydrogen production in a study led by postdoctoral researcher Jaehyun Kim as first author. That research was also published in Nature Communications.

This research was supported by the National Research Council of Science & Technology (NST) and the Ministry of Science and ICT (MSIT) (GTL25021-230); the National Research Foundation of Korea (NRF) and MSIT (RS-2024-00421181); the MSIT InnoCORE Program (1.250021.01); and the Nano & Materials Technology Development Program of the NRF and MSIT (RS-2024-00405016).

[Reference Materials]

Title / Journal: Machine-learning-guided inverse design of lead-free relaxors enabled by multimodal literature mining, Nature Communications

DOI: 10.1038/s41467-026-74376-3

[Contact Information]

Professor Ho Won Jang, Department of Materials Science and Engineering

Seoul National University / hwjang@snu.ac.kr