Building Transferable Potential Energy Surfaces with Machine Learning

Abstract

Discovering and characterizing novel compounds to serve society is the overarching goal of chemical sciences. Nowadays, experimental and computational methods provide complementary approaches for navigating the vast diversity of chemical space to identify suitable compounds for a given task. A key component of many computational methods is accurate and efficient mapping of the potential energy surface (PES), which relates the geometry to the energy of a chemical system.1 From a quantum mechanical perspective, the PES can be accurately computed; however, existing methods are infeasible for large systems (e.g., proteins) because their computational cost grows explosively with the size of the system. For this reason, machine learning (ML) approaches are used to circumvent direct computation of the PES of various chemical systems.2 To obtain ML potentials, a supervised algorithm is trained to establish a relationship between the structure of a chemical system, such as a molecule or crystal, and an output, like the system’s total energy and the forces acting on each atom. Even though existing ML potentials can reach excellent accuracy when applied to structures within the training domain (i.e., interpolation), they typically have poor performance for structures outside their applicability domain (i.e., extrapolation). To overcome this, we train NewtonNet,3 a message-passing network for deep learning of interatomic potentials, with more chemically-inspired data to the ML algorithm than has been traditionally used. Our goal is to allow this ML potential to learn transferable chemical properties and test its performance when applied to systems beyond the scope of the training set.