Dynamically polarizable water potential based on multipole moments trained by machine learning

Chris M. Handley, Paul L A Popelier

    Research output: Contribution to journalArticlepeer-review


    It is widely accepted that correctly accounting for polarization within simulations involving water is critical if the structural, dynamic, and thermodynamic properties of such systems are to be accurately reproduced. We propose a novel potential for the water dimer, trimer, tetramer, pentamer, and hexamer that includes polarization explicitly, for use in molecular dynamics simulations. Using thousands of dimer, trimer, tetramer, pentamer, and hexamer clusters sampled from a molecular dynamics simulation lacking polarization, we train (artificial) neural networks (NNs) to predict the atomic multipole moments of a central water molecule. The input of the neural nets consists solely of the coordinates of the water molecules surrounding the central water. The multipole moments are calculated by the atomic partitioning defined by quantum chemical topology (QCT). This method gives a dynamic multipolar representation of the water electron density without explicit polarizabilities. Instead, the required knowledge is stored in the neural net. Furthermore, there is no need to perform iterative calculations to self- consistency during the simulation nor is there a need include damping terms in order to avoid a polarization catastrophe. © 2009 American Chemical Society.
    Original languageEnglish
    Pages (from-to)1474-1489
    Number of pages15
    JournalJournal of Chemical Theory and Computation
    Issue number6
    Publication statusPublished - 9 Jun 2009


    Dive into the research topics of 'Dynamically polarizable water potential based on multipole moments trained by machine learning'. Together they form a unique fingerprint.

    Cite this