Machine-learning models that predict material properties and screen candidate compositions orders of magnitude faster than experimental synthesis.