A water potential based on multipole moments trained by machine learning - Reducing maximum energy errors

A potential that strives to represent the Coulomb interaction realistically must include polarization. In our approach, three decisions were made to accomplish this: (i) define an atom according to quantum chemical topology (QCT), (ii) express the interaction between atoms via their multipole moments, and (iii) use machine learning to capture the response of an atomic multipole moment to a change in this atom's environment. This approach avoids explicit (distributed) polarizabilities and eliminates the problem of polarization catastrophe. Previously, we showed (Phys. Chem. Chem. Phys. 2009, 11, 6365) that a machine learning method called kriging predicted atomic multipole moments more accurately than competing machine learning methods. This was established for the atoms of a central water molecule in water clusters, from the dimer to the hexamer. The prediction errors in all multipole moments were collectively assessed by errors in total interaction energy, for thousands of clusters configurations. Here, we target the maximum errors, with an eye on reducing the worst predictions that the potential may return. We demonstrate proof-of-principle for the water dimer using local kriging.