A flexible and polarizable water model built on interpolated multipoles

Abstract: Water has a rich phase diagram with regions of meta-stability, the prime examples being super-cooled and super-heated liquid water.Below ca 50 °C, water exhibits unusual and not yet fully explained anomalous thermodynamic properties upon cooling, for example a minimum in compressibility at 46 °C, a minimum in heat capacity at 35 °C, a density maximum at 4 °C, and a compressibility maximum at -44 °C in the supercooled liquid, at ambient pressure.It is hypothesized that the phase transition (PT) between high- and low-density amorphous ices, below ~120 K, extends toward ambient conditions as a PT between two liquid phases, and ends at a critical point in the deeply supercooled liquid, giving rise to critical fluctuations in density that explain the anomalies.The phase diagram can be examined with molecular dynamics (MD) simulations, where the molecular structure is a direct observable.The realism of a simulation improves with the accuracy of model forces, the number of molecules, and simulated time.Accurate forces can be obtained from wave function (WF) calculations, but the computational demands are high.In computationally efficient models, the accuracy is generally compromised,but with the advent of machine learning, this situation has improved.In this thesis we report on the development of a water model that aims for accuracy similar to the best WF methods, while being efficient enough for predictive MD simulations.Multipole and polarizability tensors have been computed with WF methods for a training set of perturbed water monomers, and fitted with Gaussian process regression as functions of the molecular geometry, in order to create an accurate model of the electrostatic and many-body polarization forces.To obtain accurate short-range forces, due to intermolecular WF overlap, a machine-learning (Gaussian approximation) potential is applied, which is trained on the difference (in forces and energies) between the electrostatic model and an accurate WF method, for a large training-set of water dimers and trimers.To facilitate periodic calculations at constant pressure, the Ewald method and virial stress tensor have been implemented for multipolar forces between flexible molecules.The model shows good agreement with WF calculations of the energies of different test-structures, as well as promising results for thermodynamic properties from MD simulations.