Flexible Polarizable Water Model Parameterized via Gaussian Process Regression.

Water is one of the most common components in molecular dynamics (MD) simulations. Using Gaussian process regression for predicting the properties of a water model without the need of running a simulation whenever the parameters are changed, we obtained a flexible polarizable water model, named SWM4/Fw, that is able to reproduce many reference water properties. The added flexibility is critical for modeling chemical reactions in which chemical bonds can be stretched or even broken and for directly calculating vibrational spectra. In addition to being one of the few water models that are both flexible and polarizable, SWM4/Fw is also efficient thanks to the extended Lagrangian scheme with Drude oscillators. The overall accuracy is on par with or better than the related SWM4-NDP model.