Deep learning has the potential to accelerate atomistic simulations, but existing models suffer from a lack of robustness, sample efficiency, and accuracy. Simon Batzner, Albert Musaelian, and Boris ...