How well can polarization models of pairwise nonadditive forces describe liquid water?

Properties of liquid water have been computed using a near-exact rigid-monomer two-body (pairwise-additive) force field and the same field supplemented by a simple, non-empirical polarization model of pairwise nonadditive many-body forces. The inclusion of nonadditive polarization forces leads to a dramatic decrease, sometimes by an order of magnitude, of the deviations of water properties computed using classical molecular dynamics from experiment results. The remaining deviations are typically of the order of 10%. The model correctly predicts the temperature dependence of the properties except for the density of supercooled water. This good performance is achieved despite the known failure of the polarization model in reproducing trimer nonadditive interaction energies, confirmed here by showing that for a random set of trimers with all O-O separations smaller than 3.4 Å, selected from simulation snapshots, the average error of the model relative to accurate ab initio values is 71%. However, the errors gradually decrease for larger trimers, more abundant in liquid, and one can estimate that the polarization model should reproduce the exact liquid interaction energy to within about 6%. Although this accuracy is consistent with the observed performance of the polarization model, it does not explain the dramatic improvements over the two-body model. These improvements are due to the restructuring of liquid into tetrahedral arrangements instigated by the nonadditive polarization forces. The deviations of our predictions from experiments are generally also consistent with the estimated contributions from leading neglected effects other than the exchange nonadditive forces: the monomer flexibility and quantum nuclear motion effects.