Enantioselective Processes at Surfaces Studied by High-Dimensional Neural Network Potentials
Final Report Abstract
The major outcome of the Emmy Noether project is the development and implementation of a novel approach to construct high-dimensional potential energy surfaces based on artificial neural networks (NN). In this approach the total energy of the system is constructed as a sum of environment-dependent atomic or pair energy contributions. Further, long-range electrostatic interactions can be included based on environment-dependent charges. The atomic environments are provided in form of many-body symmetry functions serving as structural fingerprints. The resulting neural network potentials (NNPs) combine the reliability of electronic structure calculations with the efficiency of empirical potentials and force fields and thus enable to perform large-scale molecular dynamics simulations of systems containing tens of thousands of atoms with the accuracy of first principles methods. An important advantage of this approach is the ability to describe all types of bonding and atomic interactions, from covalent via ionic to metallic bonding, on an equal footing without any changes in the functional form or other system-specific modifications. NNPs can in principle be constructed employing all types of electronic structure methods. A remaining drawback of the approach is the large number of electronic structure reference calculations required to determine the parameters of the NNPs, but this large number of parameters and the unbiased “non-physical” form of NNPs are the fundamental reason why a high accuracy can be reached. Therefore, the construction of NNPs requires more effort than that of simpler empirical potentials, but this pays off quickly in atomistic simulations, which then can be carried out on significantly extended length and time scales than ab initio MD. To date, the method has been applied to study a variety of systems in my group and in several collaborations. For example, NNPs have been employed to study the high-pressure phase diagram of silicon, the melting anomaly of sodium, the properties of the phase change material GeTe, the structure of copper clusters supported at zinc oxide, and the atomistic origins of the density anomaly and the melting point of water.
Publications
- "A Density-Functional Theory Based Neural Network Potential for Water Clusters Including van der Waals Corrections", J. Phys. Chem. A 117 (2013) 7356
T. Morawietz and J. Behler
(See online at https://doi.org/10.1021/jp401225b) - "Neural Network Potentials for Metals and Oxides - First Applications to Copper Clusters at Zinc Oxide", Phys. Stat. Sol. B 250 (2013) 1191
N. Artrith, B. Hiller, and J. Behler
(See online at https://doi.org/10.1002/pssb.201248370) - "Representing the potential-energy surface of protonated water clusters by high-dimensional neural network potentials", Phys. Chem. Chem. Phys. 17 (2015) 8356
S. Kondati Natarajan, T. Morawietz, and J. Behler
- "How Van der Waals Interactions Determine the Unique Properties of Water". PNAS July 26, 2016 113 (30) 8368-8373
T. Morawietz, A. Singraber, C. Dellago, and J. Behler
(See online at https://doi.org/10.1073/pnas.1602375113)
