Developments in machine learning and bonding indicators for the Quantum Chemical Topology Force Field

Abstract

The Quantum Chemical Topology Force Field (FFLUX) takes a novel approach to molecular simulation. In its modern formulation, classic fixed-form valence potentials are replaced by non-parametric regression models that mimic the total quantum mechanical energy. Likewise the electrostatics are represented by high-order multipolar interactions that converge to the quantum electrostatic potential. Altogether, this approach aims to overcome the obstacles of classical force fields in accurately modelling the energetics of complex chemical phenomena, such as ligand-receptor interactions. This line of research is still in its infancy, and many problems remain unsolved. We discuss some of these hurdles, including the neglect of polarisation, the problems of non-parametric regression when extrapolating beyond the training set, and the curse of dimensionality, which imposes an exponential growth of the models as the system size increases. We then introduce two major contributions to the Quantum Chemical Topology Force Field. Our first contribution sprouts from the need to find an inexpensive exchange indicator, and evolved as a qualitative study of the delocalization index, a magnitude from Quantum Chemical Topology that accounts for the amount of electrons delocalized within two atoms. We showed that this index can account for the covalent interaction between both atoms, and that represents a useful indicator of bond order. The findings for well-known bonding patterns were found to be extensible to uncommon compounds, allowing to explain some previously disputed situations. Our second contribution is the introduction of a novel formulation of kriging that implicitly incorporates permutational symmetry, allowing to reduce computational cost and avoid some nonphysical descriptions. The form of this predictor is reminiscent of the monomial symmetrization employed in the theory of invariant polynomials, and together with a heuristic algorithm allows to greatly reduce the number of training points. We implement this procedure in the DL_POLY 4 molecular dynamics package, and show its capabilities in optimisations and simulations of some small molecules.

Date of Award	1 Aug 2019
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Paul Popelier (Supervisor) & Samuel De Visser (Supervisor)