High-Dimensional Neural Network Potentials for Complex Systems

In recent years, machine learning potentials have become a promising new approach for the representation of high-dimensional reactive potential-energy surfaces [1]. After training to a set of reference energies and forces obtained from electronic structure calculations they allow to perform large-scale simulations with first-principles accuracy at a fraction of the computational costs. As they do not contain any system-specific terms, they are applicable to a wide range of problems in chemistry, physics and materials science.
In this talk the methodology of high-dimensional neural network potentials (HDNNPs) [2,3] will be briefly reviewed, and its current scope and limitations will be discussed. The applicability of HDNNPs will be illustrated by recent studies of solid-liquid interfaces using copper [4] and zinc oxide [5,6] as prototypical examples. Since HDNNPs allow to describe the making and breaking of bonds, they enable the detailed investigation of the dissociation and recombination of water molecules as well as proton transport processes at interfaces. While water molecules do not spontaneously dissociate at ideal low-index copper surfaces, the water - zinc oxide system is very dynamic with properties strongly depending on the specific surface termination.

[1] J. Behler, J. Chem. Phys. 145 (2016) 170901.
[2] J. Behler and M. Parrinello, Phys. Rev. Lett. 98 (2007) 146401.
[3] J. Behler, Angew. Chem. Int. Ed. 56 (2017) 12828.
[4] S. Kondati Natarajan and J. Behler, Phys. Chem. Chem. Phys. 18 (2016) 28704.
[5] V. Quaranta, M. Hellström, and J. Behler, J. Phys. Chem. Lett. 8 (2017) 1476.
[6] M. Hellström, V. Quaranta, and J. Behler, Chem. Sci., in press (2018), DOI 10.1039/C8SC03033B.