Scientific Discoveries With Word2vec

How cool is this?! Scientists trained an AI to find promising discoveries from “reading” materials science research papers using word2vec. By itself, the method captured knowledge of chemistry like the structure of the periodic table and how materials are related. When they backtested the method, it successfully predicted later discoveries of thermoelectric materials. They hope to use its recommendations to focus research into new materials.

“Unsupervised word embeddings capture latent knowledge from materials science literature” by Tshitoyan, Dagdelen, Weston, et al in Nature 4 July 2019.

Link to paper

Link to code