We may only need a *not-so-deep* neural network for materials science