Optimal construction of a fast and accurate polarisable water potential based on multipole moments trained by machine learning

Chris M. Handley, Glenn I. Hawe, Douglas B. Kell, Paul L A Popelier

    Research output: Contribution to journalArticlepeer-review

    Abstract

    To model liquid water correctly and to reproduce its structural, dynamic and thermodynamic properties warrants models that account accurately for electronic polarisation. We have previously demonstrated that polarisation can be represented by fluctuating multipole moments (derived by quantum chemical topology) predicted by multilayer perceptrons (MLPs) in response to the local structure of the cluster. Here we further develop this methodology of modeling polarisation enabling control of the balance between accuracy, in terms of errors in Coulomb energy and computing time. First, the predictive ability and speed of two additional machine learning methods, radial basis function neural networks (RBFNN) and Kriging, are assessed with respect to our previous MLP based polarisable water models, for water dimer, trimer, tetramer, pentamer and hexamer clusters. Compared to MLPs, we find that RBFNNs achieve a 14-26% decrease in median Coulomb energy error, with a factor 2.5-3 slowdown in speed, whilst Kriging achieves a 40-67% decrease in median energy error with a 6.5-8.5 factor slowdown in speed. Then, these compromises between accuracy and speed are improved upon through a simple multi-objective optimisation to identify Pareto-optimal combinations. Compared to the Kriging results, combinations are found that are no less accurate (at the 90th energy error percentile), yet are 58% faster for the dimer, and 26% faster for the pentamer. © 2009 the Owner Societies.
    Original languageEnglish
    Pages (from-to)6365-6376
    Number of pages11
    JournalPhysical Chemistry Chemical Physics
    Volume11
    Issue number30
    DOIs
    Publication statusPublished - 2009

    Fingerprint

    Dive into the research topics of 'Optimal construction of a fast and accurate polarisable water potential based on multipole moments trained by machine learning'. Together they form a unique fingerprint.

    Cite this