Neural networks to approach potential energy surfaces: Application to a molecular dynamics simulation

Authors

  • Diogo A. R. S. Latino,

    1. Grupo de Simulação Molecular, CCMM, Departamento de Química e Bioquímica, Faculdade de Ciências, Universidade de Lisboa, Campo Grande, 1749–016 Lisboa, Portugal
    2. REQUIMTE, CQFB, Departamento de Química, Faculdade de Ciências e Tecnologia, Universidade Nova de Lisboa, 2829–516 Caparica, Portugal
    Search for more papers by this author
  • Filomena F. M. Freitas,

    1. Grupo de Simulação Molecular, CCMM, Departamento de Química e Bioquímica, Faculdade de Ciências, Universidade de Lisboa, Campo Grande, 1749–016 Lisboa, Portugal
    Search for more papers by this author
  • João Aires-De-Sousa,

    1. REQUIMTE, CQFB, Departamento de Química, Faculdade de Ciências e Tecnologia, Universidade Nova de Lisboa, 2829–516 Caparica, Portugal
    Search for more papers by this author
  • Fernando M. S. Silva Fernandes

    Corresponding author
    1. Grupo de Simulação Molecular, CCMM, Departamento de Química e Bioquímica, Faculdade de Ciências, Universidade de Lisboa, Campo Grande, 1749–016 Lisboa, Portugal
    • Grupo de Simulação Molecular, CCMM, Departamento de Química e Bioquímica, Faculdade de Ciências, Universidade de Lisboa, Campo Grande, 1749–016 Lisboa, Portugal
    Search for more papers by this author

Abstract

Potential energy surfaces (PES) are crucial to the study of reactive and nonreactive chemical systems by Monte Carlo (MC) or molecular dynamics (MD) simulations. Ideally, PES should have the accuracy provided by ab initio calculations and be set up as fast as possible. Recently, neural networks (NNs) turned out as a suitable approach for estimating PES from ab initio/DFT energy datasets. However, the accuracy of the properties determined by MC and MD simulation methods from NNs surfaces has not yet, to our knowledge, been systematically analyzed in terms of the minimum number of energy points required for training and the usage of different NN-types. The goal of this work is to train NNs for reproducing PES represented by well-known analytical potential functions, and then to assess the accuracy of the method by comparing the simulation results obtained from NNs and analytical PES. Ensembles of feed-forward neural networks (EnsFFNNs) and associative neural networks (ASNNs) are used to estimate the full energy surface. Training sets with different number of points, from 15 differently parameterized Lennard–Jones (LJ) potentials, are used and argon is taken to test the network. MD simulations have been performed using the tabular potential energies, predicted by NNs, for working out thermal, structural, and dynamic properties which are compared with the values obtained from the analytical function. Our results show that, at least for LJ-type potentials, NNs can be trained to generate accurate PES to be used in molecular simulations. © 2007 Wiley Periodicals, Inc. Int J Quantum Chem, 2007

Ancillary