Advances in machine learned potentials for molecular dynamics simulation

Recent machine learning techniques allow emulation of quantum chemistry with stunning fidelity. For example, deep neural networks can now predict molecular properties with accuracy approaching that of coupled cluster theory, at a tiny fraction of the computational cost. We present methods for building machine learned potentials based on the following key ideas: (1) encoding physical symmetries, (2) active learning to dynamically grow the training dataset, and (3) transfer learning to incorporate data from varying sources. The aim is to enable large-scale and highly accurate molecular dynamics simulations, e.g., for chemistry, materials science, and biophysics applications.