Systematic improvement of approximations with smooth models of the Coulomb potential

Abstract

Orbital-based methods for electronic-structure calculations are limited to atoms or molecules with up to about 50 electrons. This limitation comes from the requirement of a long expansion in basis functions to approximate correctly the wave function. Replacing the Coulomb interaction with a smooth model potential has two main consequences: first, the wave function becomes cuspless and the expansion in basis functions converges more rapidly, and second, the smooth potential describes a weaker interaction at the electronic coalescence point, which leads to the loss of accuracy. This work explores whether one can construct models with smooth, non-singular, potentials, but without compromising accuracy. The key idea is to use extrapolation procedures to predict the energy for the Coulomb interaction from a sequence of (cheaper) calculations for smooth potentials. By replacing the Coulomb electron-electron interaction with a smooth potential, using the semi-stochastic heat-bath configuration interaction method (SHCI) to select key configurations, and extrapolating to the limiting (non-smoothed) Coulomb potential, we were able to retain the accuracy of full configuration interaction (FCI) calculations, at reduced computational cost.