Our main research topic is the development and application of efficient interatomic potentials based on artificial Neural Networks. In contrast to conventional empirical potentials, which have a fixed functional form based on a number of assumptions and approximations, Neural Networks represent a machine learning potential providing a very high numerical accuracy. The correct physical description of the systems under investigation is introduced via a large number of accurate electronic structure reference calculations. Based on a large number of these ab initio energies and forces for different atomic configurations, a continuous representation of the potential-energy surface is obtained by a Neural Network interpolation, i.e. the Neural Network 'learns' the shape of the potential-energy surface. Neural Networks are very flexible and, once properly trained, provide an accurate analytic expression for the potential-energy surface.
Because of their advantages and disadvantages, Neural Network potentials are most useful to speed up molecular dynamics simulations, if complex bonding situations or chemical reactions are present which cannot be described by conventional potentials. We believe that in such cases Neural Network potentials are the method of choice to perform long molecular dynamics simulations of large systems, which are not directly accessible by ab initio MD.
We are constantly working on the further development of the Neural Network methodology to improve the accuracy and performance of these potentials. In order to construct Neural Network potentials for a variety of systems, we have developed our own Neural Network package RuNNer.