NEW DEVELOPMENTS IN LOCAL CONFIGURATION INTERACTION THEORY

Abstract

Many chemical phenomena, from the freezing point of water to the strength of a chemical bond, are determined by the distribution of electrons which make up matter. We can model and predict chemical phenomena by using quantum mechanics to determine the electrons' distribution (wavefunction). Multireference configuration interaction (MRCI) provides a flexible wavefunction for capturing crucial electron correlation effects. Unfortunately, MRCI's computational cost grows rapidly, O(N^6 ), limiting its application to small molecules. Over the last three decades, researchers have exploited the spatial locality of electron correlation to reduce the costs of correlated quantum chemistry methods like MRCI. The local electron correlation approximation removes insignificant long range correlations thereby reducing MRCI's cost. By so doing, local MRCI methods can be applied to much large molecules than canonical MRCI. In this thesis, previous efforts by Carter and co-workers applying the local correlation approximation to MRCI are expanded upon: both computational speedups and improved accuracy are considered. The state-of-the art local MRCI algorithm scales as O(N^3 ) which, while cheaper than conventional MRCI, scales rapidly with system size. Converting the previously serial local MRCI code to parallel code allows exploitation of multicore architectures common in modern CPUs. Replacing the Cholesky-decomposed two-electron integrals with cheaper density-fitted two-electron integrals reduces local MRCI's cost. These two advances don't effect the O(N^3 ) scaling, but rather reduce the scaling prefactor, thereby allowing simulation of larger molecules. MRCI's accuracy is hampered by the well-known size extensivity error, which grows with molecular size. We introduce previously proposed MRCI size- extensivity corrections to the O(N^3 ) local MRCI. Both a priori and a posteriori size extensivity corrections can be applied. However, a priori corrections can cause numerical instabilities in both canonical and local MRCI. We show that these instabilities are avoided crossings with spurious low energy states. This analysis suggests two different approaches to maintain a stable size extensivity correction. Finally, we improve the accuracy of local MRCI by optimizing the parameters controlling the local electron correlation. The combination of these developments provides a faster, more accurate method for modeling larger scale chemical phenomena than previously possible.