The Ruhr-university Neural Network energy representation - RuNNer - suite is a high-performance implementation of high-dimensional neural network potentials (HDNNP), a method introduced by Behler and Parrinello in 2007 establishing the first machine learning potential applicable to high-dimensional condensed systems containing large numbers of atoms [1]. Since the introduction of the original method, HDNNPs have been continuously further developed and allow to perform atomistic simulations of a wide range of systems at much reduced computational costs compared to electronic structure calculation, but with essentially the same accuracy.  Several generations of HDNNP are available in RuNNer [2], e.g. third-generation (3G) HDNNPs including long-range electrostatic interactions [3] and fourth-generation (4G) HDNNPs [4] including non-local effects like long-range charge transfer requiring a global description of the system. The code is developed and maintained by the group of Professor Jörg Behler, Chair of Theoretical Chemistry II at Ruhr-Universität Bochum and the Research Center Chemical Sciences and Sustainability, Research Alliance Ruhr. 

RuNNer is written in modern Fortran and has been designed to run on a wide range of systems. The code is freely available under the GPL3 license. As a CPU-based implementation it does not require expensive GPUs but can be run on different platforms, from laptops to powerful HPC compute nodes with a large number of cores. RuNNer can make use of the available RAM in a very flexible way, storing intermediate quantities to speed up calculations significantly if enough memory is available. The parallelization strategy includes both OpenMP and MPI. 

Quick links:
RuNNer repository at gitlab.com

RuNNer online documentation

RuNNer ASE interface repository at gitlab.com

RuNNer ASE online documentation

The methodology of HDNNPs is described in detail in several reviews [2,5,6] and the details of atom-centered symmetry function (ACSF) descriptors can be found in Ref. [7]. Tutorial-style introductions are provided in Refs. [8] and [9].

RuNNer has the following key features:

• implementation of 2G, 3G and 4G HDNNPs
• efficient implementation of long-range electrostatics
• global phenomena like long-range charge transfer
• OpenMP and MPI parallelization, highly-efficient CPU-based implementation
• applicable to very large systems 
• training data can be obtained from arbitrary electronic structure methods and codes
• ASE interface for convenient workflows, e.g. in jupyter notebooks
• LAMMPS interface for efficient atomistic simulations

Contributors:

RuNNer 2.0 has been written by the Behler group in Bochum with contributions by some friends. The main contributors are:
Alexander L. Knoll, Moritz R. Schäfer, K. Nikolas Lausch, Gunnar Schmitz, Redouan El Haouari, Moritz Gubler, Jonas A. Finkler, Alea M. Tokita, Henry Wang, J. Richard Springborn, Marco Eckhoff, and Jörg Behler.

References:

[1] J. Behler, and M. Parrinello, Phys. Rev. Lett. 98, 146401 (2007).

[2] J. Behler, Chem. Rev. 121 (2021) 10037.

[3] N. Artrith, T. Morawietz, J. Behler, Phys. Rev. B 83 (2011) 153101.

[4] T. W. Ko, J. A. Finkler, S. Goedecker, J. Behler, Nat. Commun. 19 (2023) 3567.

[5] J. Behler, J. Phys.: Condens. Matter 26, 183001 (2014).

[6] J. Behler, Angew. Chem. Int. Ed. 56 (2017) 12828.

[7] J. Behler, J. Chem. Phys. 134 (2011) 074106.

[8] J. Behler, Int. J. Quantum Chem. 115, (2015) 1032.

[9] A. M. Tokita, J. Behler, J. Chem. Phys. 159 (2023) 121501.

 

If you are interested in learning more about RuNNer, please feel free to contact us.