Abstract: Learning Atomistic Physics Using Deep Neural Network
Computational design of new materials and molecules requires accurately simulating atomistic systems at scale. While quantum mechanical (QM) physics gives precise answers, the simulation of QM is prohibitively expensive — even approximate methods often scale as O(N^3), where N is the number of atoms in a system. For decades, researchers have replaced QM with simple models, such as block-and-spring type approximations, which are fast, but questionably accurate and not transferable between different physical systems.
The next generation of these models is powered by machine learning — a large dataset of QM calculations is created, and ML physics models learn to accurately reproduce the results of these calculations using O(N) models. We describe our methods for learning Deep Neural Network models that describe energy, force, charge, and other physical quantities to enable computational studies of atomistic systems that are accurate, transferable, and scalable.