ABSTRACT
Traditionally, most condensed-phase calculations have been performed by approximating the interactions between particles with empirical potentials. These have the advantage of being very computationally efficient as, depending on the sophistication of the potential, the evaluation of each interaction requires only a few arithmetic operations. Potentials vary in their sophistication from the very simple Lennard-Jones potentials [2], which crudely describe a short-range repulsive potential with a longer-range attractive potential, to much more sophisticated potentials such as those of Bukowski and co-workers [3,4] in which interaction potentials are derived from symmetry-adapted perturbation theory [5] and fitted to highly accurate ab initio results. There are myriad potentials (Guillot lists more than 40 for water alone [6]) whose functional forms have been tailored, with parameters fitted to specific systems or to reproduce certain properties, and although tremendously popular and often reasonably successful, they suffer from the drawback that the potentials must often be refitted when a different system is considered or when the conditions of interest are far from those sampled in the parameterization [7]. If a more accurate model is desired, it is not always clear how potentials can be improved and the development of a new potential can be a labor-intensive task and often relies on empiricism as opposed to rigorous justification.
