Kai Leonhard, Jadran Vrabec, Johannes Lenhard, Gabriele Raabe, Andreas Mecklenfeld, Hans Hasse, Peter Klein, Gabor Rutkai, Edder J. García, Vitalie Botan, Michael Schappals, Simon Stephan, Colin W. Glass, Andreas Köster, Leif Kröger
Thermodynamic properties are often modeled by classical force fields which describe the interactions on the atomistic scale. Molecular simulations are used for retrieving thermodynamic data from such models, and many simulation techniques and computer codes are available for that purpose. In the present round robin study, the following fundamental question is addressed: Will different user groups working with different simulation codes obtain coinciding results within the statistical uncertainty of their data? A set of 24 simple simulation tasks is defined and solved by five user groups working with eight molecular simulation codes: DL_POLY, GROMACS, IMC, LAMMPS, ms2, NAMD, Tinker, and TOWHEE. Each task consists of the definition of (1) a pure fluid that is described by a force field and (2) the conditions under which that property is to be determined. The fluids are four simple alkanes: ethane, propane, n-butane, and iso-butane. All force fields consider internal degrees of freedom: OPLS, TraPPE, and a modified OPLS version with bond stretching vibrations. Density and potential energy are determined as a function of temperature and pressure on a grid which is specified such that all states are liquid. The user groups worked independently and reported their results to a central instance. The full set of results was disclosed to all user groups only at the end of the study. During the study, the central instance gave only qualitative feedback. The results reveal the challenges of carrying out molecular simulations. Several iterations were needed to eliminate gross errors. For most simulation tasks, the remaining deviations between the results of the different groups are acceptable from a practical standpoint, but they are often outside of the statistical errors of the individual simulation data. However, there are also cases where the deviations are unacceptable. This study highlights similarities between computer experiments and laboratory experiments, which are both subject not only to statistical error but also to systematic error.
