Molecular systems have two properties that provide the basis for approximations that lower computational costs: nearsightedness and molecular similarity. Nearsightedness implies that interactions become simpler at long-range, and is the basis for linear scaling methods. The other main aspect of molecular systems that can reduce computational cost is molecular similarity, whereby molecular fragments behave similarly in similar environments.  Early semiempirical methods such as MNDO and AM1 took advantage of similarity by building experimental information about molecular fragments into quantum chemical models. This presentation explores a new approach that utilizes large volumes of data from ab initio computations. Approaches adapted from machine learning are then used to discover the simplest model that is able to reproduce the ab initio data to chemical accuracy. The approach embeds parameters into a low-?level (LL) ab initio theory and adjusts these to obtain agreement with a higher-level (HL) ab initio theory.  The parameters are fit to obtain agreement with the HL models on data that spans a range of chemical systems over which molecular similarity can be plausibly assumed to hold. The most studied model to date is based on organic molecules consisting of saturated and unsaturated carbon backbones with NH2, OH, and F substituents. Studied properties include total energy, dipoles, charge distributions, and HOMO/LUMO energies. The parameters are embedded by multiplying LL Hamiltonian matrix elements with scaling factors that depend on atom type and operator. Making the scaling factors functions of the geometry, atomic charges, and bond orders also enhances the performance. Performance is measured by testing the model on systems not included in the fitting of the parameters. Both the logistics of developing the models and the performance on test systems will be discussed.

New approaches to the development of semiempirical quantum chemical methods

248th National Meeting (2014)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.