| One of the fundamental challenges in modern electronic structure theory is accurate modeling of large molecules that display intrinsically quantum-mechanical phenomena such as bond-breaking. While existing ab initio methods have essentially solved the electronic structure problem for small molecules, they are too computationally expensive to apply to large molecules. Semiempirical methods represent a possible solution to this problem. These are simplified versions of ab initio methods, with embedded parameters that are fitted to reproduce known results for model systems.; This thesis presents work towards new, functional-group-based semiempirical methods that are constructed by "mining" large databases of small-molecule calculations. The methods are based on the observation that the electronic structures of molecular functional groups (methyl, benzyl, etc.) are largely transferable between different molecules. This observation implies that ab initio calculations on molecules spend a great deal of time recalculating conserved properties. Semiempirical models of such conserved properties have the potential to be much more computationally efficient than full ab initio calculations. The transferability of electronic structure also implies that a sufficiently rich data set on a functional group in small molecules contains all information needed to describe the group in molecules of arbitrary size. Our goal is to use data-mining methods to build semiempirical models from such data sets.; This thesis presents results from several implementations of this approach: models of acceptor-donor substitution, functional-group-specific DFT exchange-correlation functionals, functional-group-specific basis sets, and large-basis NDDO semiempirical methods. |
