By digging through the texts, tables and graphs of 448 articles, a team put together a dataset and then designed temperature-stable lead-free dielectrics by inversion.
Published in Nature Communications, the study combines two rarely associated ingredients: a multimodal extraction of information from the scientific literature — the text, but also the tables and curves of figures — and machine learning constrained by physics. The search of 448 articles produced 1,202 readings of dielectric properties, the basis from which the team conducted an inverse design, i.e. a direct search for compositions likely to achieve the targeted performance.
The space explored has about 150 million virtual compositions, reduced to 37 candidates. Two of them, tin substituted titanates, were effectively synthesized and characterized: permittivity greater than 3,300 at room temperature and thermal stability in accordance with the X5R, X6R and X7R standards, the combination that the specifications were aiming for.
The industrial challenge is the replacement of lead dielectrics in multilayer ceramic capacitors, components that are ubiquitous in electronics. The remarkable point here is not so much the model as the raw material: data that has already been published for years, but scattered in figures that no one had aggregated.
Read the full article on the original source
Go to the source →