VSC School Seminar: A parallel neural network potential software suite for HPC


    With our VSC School Project ”Parallel software suite for neural network potentials for materials simulations” coming to an end, we are proud to present the first release of our ”Neural Network Potential Package” (N2P2). Over the last three and a half years we developed this software to provide the scientific community with a user-friendly and efficient code for neural network potential training and its application in molecular dynamics (MD) simulations.

    It has been almost 10 years since Behler and Parrinello [1] presented their promising approach extending the applicability of neural network potentials to large-scale simulations. The method is based on training neural networks to reproduce reference potential energy surfaces. Using an ab initio data set as input during training it is possible to later predict energies and forces of atomic configurations with high accuracy but at a fraction of the computational cost. Thus, it is a valuable tool to access long timescales and large system sizes in MD simulations.

    In our presentation we will cover the three cornerstones of our software package: A parallel training algorithm, a fast LAMMPS [2] implementation and a set of user-friendly tools to generate new neural network potentials. All of them depend on the same common code basis, the NNP library, whose structure and capabilities we will describe. In addition we showcase the recent application [3] of the neural network potential method to water and ice which allowed us to determine previously unknown properties of the corresponding ab initio models.

    This talk is based on work supported by the VSC Research Center funded by the Austrian Federal Ministry of Education, Science and Research (BMBWF) and reports about the VSC School Project "Parallel software suite for neural network potentials for materials simulations".

    [1]   J. Behler and M. Parrinello, “Generalized Neural-Network Representation of High-Dimensional Potential-Energy Surfaces,” Phys. Rev. Lett., vol. 98, p. 146401, Apr. 2007.
    [2]   S. Plimpton, “Fast Parallel Algorithms for Short-Range Molecular Dynamics,” J. Comput. Phys., vol. 117, pp. 1–19, Mar. 1995.
    [3]   T. Morawietz, A. Singraber, C. Dellago, and J. Behler, “How van der Waals interactions determine the unique properties of water,” Proc. Natl. Acad. Sci. U.S.A., vol. 113, pp. 8368–8373, July 2016.


    Andreas Singraber
    Computational Physics, University of Vienna



    Date, Time, and Location:

    19.03.2018, 15:00 - 17:00, FH Hörsaal 2 (TU Wien, Wiedner Hauptstraße 8-10, 2nd floor, yellow area)


    If you would like to join us for this event, please email to: vsc-seminar@list.tuwien.ac.at
    (so we will know how many persons will come)


    The course material will be available for registered attendees only.


    Upcoming and past events of the VSC School Seminar:

    Opens external link in new windowvsc.ac.at/training