Quickstart Guide

Overview

The typical use of RuNNer consists of three steps called modes:

• mode 1: Transformation of the Cartesian coordinates to symmetry functions, splitting of the reference data into a training and a test set, and the conversion of total energies in energies per atom (and removing the free atom energies, if requested).

• mode 2: Analysis of the training and test sets and determination of the NN parameters in an iterative optimization process.

• mode 3: Application of the potential to predict energies, forces, stress tensors and charges. Currently RuNNer is mainly used for calculating these properties for multiple given structures. RuNNer is not a molecular dynamics code. For MD, a compatible implementation in LAMMPS is available, which can use NN potentials constructed with RuNNer.

Mode 1: Generation of the Symmetry Functions

In mode 1 the symmetry functions are generated from the structural information provided in the input.data file. The generation of the symmetry functions is not parallelized as even for different fits the generation has to be done only once with reasonable CPU time requirements. It needs to be repeated only if the symmetry functions or the splitting into the training and test sets (cf. keyword test_fraction) changes.

Relevant files to get you started

Only two files are needed by RuNNer in order to start calculating symmetry functions:

• input.nn
• input.data

Mode 2: Fitting

In mode 2 the weight parameters of the atomic neural networks are determined.

Mode 3: Application of the Neural Network Potential

In mode 3 the finalized NNP can be used to predict the energies and forces of a single structure given in the input.data file. If there is more than one structure included in the input.data file, all structures in input.data are used.