About RuNNer

RuNNer is a stand-alone Fortran program for the construction of high-dimensional neural network potentials written mainly by Jörg Behler.

The author first became familiar with the use of neural networks for the construction of potential-energy surfaces during his PhD thesis with Matthias Scheffler and Karsten Reuter at the Fritz-Haber Institut der Max-Planck Gesellschaft. He employed neural network potentials for studying the dissociation of oxygen molecules at the Al(111) surface. For this purpose he used and extended a neural network code written by Sönke Lorenz in the same group for another system, and he is very grateful to Sönke for introducing him to the NN method. After having obtained his PhD, the author moved as a postdoctoral fellow to the group of Michele Parrinello at the ETH Zurich. Here, he developed the approach for the construction of high-dimensional neural network potentials and the required atom-centered symmetry function descriptors, which are the heart of the methodology underlying RuNNer. An early experimental code called NeuralCryst was used at that time which was a strongly modified version of the original unnamed NN code by Sönke Lorenz. This code has been the first implementation of high-dimensional NNPs and has been applied e.g. to the high-pressure phase diagram of Silicon.

As the underlying structure of this code, which has originally been designed for low-dimensional systems, did not allow for an easy and efficient implementation of high-dimensional NN potentials, the author started to develop a completely new NN code for high-dimensional systems from scratch, when he established his own group at the Ruhr-Universität Bochum in 2007. This new code was named RuNNer and is still under very active development at its new home at the Georg-August-Universität Göttingen.

RuNNer is primarily a code for the parameterization and validation of high-dimensional NN potentials, but it is not a molecular dynamics code. Therefore it does not aim to achieve the highest performance although a lot of work has been invested in an efficient implementation. Instead, its goal is to offer a convenient way and a lot of functionality for the construction of NN potentials.


For efficient molecular dynamics, a highly optimized and compatible implementation in LAMMPS is availble, which has been written in collaboration primarily by Andreas Singraber in the group of Christoph Dellago (University of Vienna). The files generated by RuNNer can thus directly be used for large-scale MD simulations in LAMMPS.