POLARISABLE ELECTROSTATICS BASED ON HIGH RANK MULTIPOLE MOMENTS TRAINED BY MACHINE LEARNING

Abstract

Quantum Chemical Topology (QCT) can provide chemical insight by rigorously defining atoms and bonds from quantum mechanical wave functions. This theory has been applied towards the construction of a QCT topological multipolar force field (QCTFF) in an attempt to approach quantum mechanical (QM) accuracy, especially in the area of electrostatic interactions. This is achieved by providing a detailed representation of the electrostatic term, which replaces the typical electrostatic terms used in other popular force fields.A machine learning method called Kriging accounts for polarisation by capturing the way atomic multipole moments vary upon conformational change. The Kriging models predict the multipole moments that are used in the calculation of atom-atom electrostatic interaction energies. These predicted energies are then compared to their true values to assess the reliability of the Kriging models.The hydrogen-bonded complexes in the publicly available S22 dataset are chosen to investigate polarisation with Kriging models. Atom-atom contributions to the intramolecular and intermolecular Coulomb energy are investigated using high rank multipole moments (up to hexadecapole). The Coulomb interaction energy between two molecules in a hydrogen-bonded complex is computed by summing the additive atom-atom contributions between the molecules.The generation of training sets to populate multipole moments of each atom has also been investigated by considering the coordinate system employed to generate distorted structures. This work shows that a greater region of conformational space is spanned when generating distorted structures using redundant internal coordinates, thereby making it the superior method when randomly sampling different geometries.

Date of Award	29 Oct 2012
Original language	English
Awarding Institution	The University of Manchester