Overview

The Ruhr university Neural Network energy represetation (RuNNer) is a program for training and evaluating machine-learning potentials, developed since 2007 in the group of Prof. Dr. Jörg Behler. It underwent a full rewrite and was released again in 2026.

RuNNer focuses on an efficient implementation of the different generations of high-dimensional neural network potentials (HDNNPs).

RuNNer is a Fortran package and provides a standalone binary and the underlying library.

• The RuNNer binary allows training different types of machine-learning potentials or predicting trained potentials for a dataset of structures.
• The RuNNer library exposes the API of RuNNer. It is needed by the binary itself, the ASE and the LAMMPS interface.

We also provide helpful interfaces to other programs:

• a generic Python interface for predicting energies, forces, and stress from given atomic positions.
• interfaces to the Atomic Simulation Environment (ASE).
• an interface to LAMMPS.

Overview¶

Typical use cases of RuNNer are:

• mode train: Training of the potential. Splitting the data in train and test sets, computation of features and their derivatives and determination of the NN parameters in an iterative optimization process.

• mode predict: Application of the potential to predict energies, forces, stress tensors, or atomic properties. Currently RuNNer supports calculations for multiple given structures. RuNNer is not a molecular dynamics code. For MD, please refer to the LAMMPS or ASE interfaces.