The high computational cost of electronic structure calculations remains a challenge for computational design of materials. Molecular similarity, the tendency of molecular fragments to behave similarly in similar contexts, provides a means to lower this cost. We will compare results from methods that take advantage of molecular similarity in different ways, ranging from pure informatics predictions to methods that blend quantum chemical and informatics models. Two sets of target properties will be considered. The data used to train and test the models is generated from density functional theory (DFT).<br/>The first set of targets are the excitation energy and redox properties of organic semiconductors. Time-dependent DFT results were generated for 5000 oligomers of conjugated polymers that vary both the backbone and attached substituents, and that use three different functionals. Both informatics and semiempirical quantum chemical models can predict these electronic properties to an accuracy that is equivalent to the agreement between functionals. The informatics model uses, as features, strings that encode the makeup of the oligomer. The quantum chemical models use a minimal basis, Intermediate Neglect of Differential Overlap (INDO) form with parameters trained on a portion of the data set. More computationally-efficient coarse-grained quantum models, in which each repeat unit of the oligomer is treated as a single site, were developed for a more restricted set of polymer structures. These site models provide accurate predictions as a function of monomer sequence and rotamer structure, allowing for efficient averaging over disordered ensembles.<br/>The second set of targets are total energy and atomic forces of a diverse sets of organic molecules. Pure informatics models, such as kernel ridge regression, perform well on total energy but extension to atomic forces is challenging. Two approaches to melding semiempirical quantum chemistry with informatics will be compared. The first uses informatics to improve only the classical, atom-atom, repulsive potential. The second extends this by using informatics to also set the matrix elements of the electronic Hamiltonian, such that can these vary with local geometry.<br/>

Predicting properties of organic materials with models that blend informatics with quantum chemistry

253rd National Meeting (2017)

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