Artificial intelligence has dramatically accelerated the search for new electronic materials by screening 150 million virtual compositions and identifying two promising lead-free dielectric candidates that outperform current industry standards. A team at Seoul National University combined machine learning with data extracted from hundreds of scientific papers to design materials that remain stable at high temperatures, a critical requirement for electric vehicles and aerospace systems.

Why Finding New Electronic Materials Matters?

The electronics industry faces a growing challenge: as technologies like electric vehicles, power electronics, and aerospace equipment operate at higher temperatures, existing materials struggle to maintain performance. Dielectrics, which are insulating materials that store electrical charge, are essential components in multilayer ceramic capacitors (MLCCs) found in smartphones, electric vehicles, and countless other devices. The problem is that discovering new materials traditionally requires expensive, time-consuming trial-and-error testing across an almost limitless number of possible chemical combinations.

The Seoul National University research team, led by Professor Ho Won Jang of the Department of Materials Science and Engineering, tackled this challenge by combining multimodal literature mining with physics-informed machine learning. Rather than starting from scratch, the researchers extracted data from 448 scientific papers to build a foundation for their AI models.

How Did Researchers Use AI to Accelerate Materials Discovery?

• Literature Mining: Large language models extracted compositions and processing conditions from text and tables across 448 scientific papers, while graphs were converted into numerical data to recover temperature-dependent dielectric properties.
• Data Integration: The team assembled 1,202 records of dielectric properties and added 22 physical descriptors to make data from different publications comparable, addressing inconsistencies in experimental conditions and reporting formats.
• Physics-Informed Screening: Thirty independently trained machine learning models predicted three important measures related to dielectric performance simultaneously, with the system prioritizing compositions backed by greater predictive confidence across multiple models.
• Candidate Reduction: After applying performance requirements and physicochemical constraints to approximately 150 million virtual compositions, the process narrowed the field to just 37 candidates for further evaluation.

The researchers then refined component ratios within the most promising compositional family and selected two formulations for experimental testing. Both samples contained small substitutions of tin (Sn), with one containing 1 mol% and the other 2 mol%.

What Did the Experimental Results Show?

The two tested materials demonstrated impressive performance metrics. At room temperature, they achieved dielectric constants of 3,422 and 3,307, respectively, which are among the highest reported for materials in their compositional family. More importantly, both samples met the international X5R, X6R, and X7R standards for temperature stability, meaning their dielectric constants remained within acceptable ranges across temperature spans from minus 55 degrees Celsius to 125 degrees Celsius.

Adding small amounts of tin produced a critical balance: temperature stability improved without substantial loss in dielectric constant. Compared with barium titanate (BaTiO3), a material widely used in current multilayer ceramic capacitors, the new compositions maintained high dielectric constants more consistently across a broader temperature range.

"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," said Ho Won Jang, Professor of Materials Science and Engineering at Seoul National University.

Ho Won Jang, Professor of Materials Science and Engineering at Seoul National University

To understand why tin substitution worked so well, the researchers used piezoresponse force microscopy, Raman spectroscopy, and atomic resolution electron microscopy to compare what the machine learning models identified as important with actual material measurements. Their analysis revealed that limited tin substitution expands the crystal framework and increases electrical heterogeneity at the atomic scale, changes that enhance temperature stability.

What's the Broader Impact Beyond These Two Materials?

While the two validated materials represent immediate practical applications for high-temperature multilayer ceramic capacitors and electronic components in electric vehicles and aerospace systems, the real significance lies in the methodology itself. The research demonstrates a scalable approach to turning information dispersed throughout scientific literature into actionable datasets that can guide material design.

"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," noted Ho Won Jang.

Ho Won Jang, Professor of Materials Science and Engineering at Seoul National University

The findings were published in Nature Communications and represent a significant step forward in how artificial intelligence can accelerate materials science research. Rather than relying solely on computational predictions, the team combined machine learning with rigorous experimental validation, ensuring that AI-identified candidates actually perform as expected in real-world conditions. This hybrid approach addresses a persistent challenge in AI-driven materials discovery: the gap between theoretical predictions and practical performance.

As demand for heat-resistant materials grows across industries, this methodology offers a template for researchers worldwide to systematically mine existing scientific literature and leverage AI to identify promising candidates faster and more efficiently than traditional approaches would allow.