A team from Seoul National University combines automatic literature extraction and physics-guided learning to design lead-free dielectrics that are stable at high temperatures.
The data that materials science needs already exists, but it's scattered across thousands of papers, in the form of tables, curves, and legends that no one has ever consolidated. Professor Ho Won Jang's team at Seoul National University's Department of Materials Science took the problem by the end: extracting this information rather than redoing the experiments.
The pipeline combines a multimodal digging of the literature — text and figures — with physics-constrained machine learning. From 448 publications, the researchers reconstructed a set of 1,202 measurements of dielectric properties, and then explored about 150 million possible compositions. The screening reduced this space to 37 candidates, two of which were effectively synthesized and characterized: they have a high dielectric constant and, above all, retain it over a wide temperature range.
The industrial challenge is concrete. Lead-free dielectrics capable of holding heat condition the capacitors of power electronics, from electrified automobiles to computing infrastructures. The work is published in Nature Communications.
Read the full article on the original source
Go to the source →