AI Mines 448 Research Papers to Discover New Heat-Stable, Lead-Free Materials

A research team at Seoul National University's College of Engineering, led by professor Ho Won Jang, has developed a way to discover new lead-free dielectric materials — insulating materials that store electric charge — by combining data mined from scientific papers with physics-informed machine learning. Dielectrics are key components of multilayer ceramic capacitors, which are used in smartphones, electric vehicles and other electronics; the higher a material's dielectric constant, the more electrical energy it can store in the same size, but that performance must also stay stable at high temperatures. The findings are published in the journal Nature Communications.

The team, led by first author Kwanwoo Song along with Youngmin Kim and postdoctoral researcher Jaehyun Kim, used large language models to automatically pull composition, processing conditions and performance data from the text, tables and graphs of 448 research papers, building a dataset of 1,202 records. Because measurement conditions vary between studies, the researchers added 22 physical descriptors and combined 30 separately trained machine-learning models to make the scattered data usable for training.

Applying this framework to a virtual space of about 150 million possible lead-free compositions, the team narrowed the search to 37 promising candidates. They then synthesized two of these compositions, containing small amounts of tin, and confirmed experimentally that both had high dielectric constants and stayed stable across a broad temperature range — unlike barium titanate, a widely used dielectric whose performance changes sharply around 125°C.

As more electronics — including electric vehicles, power systems and aerospace equipment — operate at higher temperatures, the number of possible lead-free element combinations is vast, making trial-and-error searches costly and slow. The researchers said their inverse-design approach, which predicts promising compositions before they are made rather than testing materials one by one, could also apply to designing materials for hydrogen energy, carbon capture and clean-ammonia systems.