Date of Award

12-1992

Degree Name

Doctor of Philosophy

Department

Chemistry

First Advisor

Dr. Robert E. Harmon

Second Advisor

Dr. Robert Trenary

Third Advisor

Dr. William Kelly

Fourth Advisor

Dr. Michael McCarville

Abstract

Computational neural networks (CNNs) are a computational paradigm inspired by the brain’s massively parallel network of highly interconnected neurons. The power of computational neural networks derives not so much from their ability to model the brain as from their ability to learn by example and to map highly complex, nonlinear functions, without the need to explicitly specify the functional relationship. Two central questions about CNNs were investigated in the context of predicting chemical reactions: (1) the mapping properties of neural networks and (2) the representation of chemical information for use in CNNs.

Chemical reactivity is here considered an example of a complex, nonlinear function of molecular structure. CNN’s were trained using modifications of the backpropagation learning rule to map a three dimensional response surface similar to those typically observed in quantitative structure-activity and structure-property relationships. The computational neural network’s mapping of the response surface was found to be robust to the effects of training sample size, noisy data and intercorrelated input variables.

The investigation of chemical structure representation led to the development of a molecular structure-based connection-table representation suitable for neural network training. An extension of this work led to a BE-matrix structure representation th at was found to be general for several classes of reactions. The CNN prediction of chemical reactivity and regiochemistry was investigated for electrophilic aromatic substitution reactions, Markovnikov addition to alkenes, Saytzeff elimination from haloalkanes, Diels-Alder cycloaddition, and retro Diels-Alder ring opening reactions using these connectivity-matrix derived representations. The reaction predictions made by the CNNs were more accurate than those of an expert system and were comparable to predictions made by chemists.

Computational neural networks were shown to have robust mapping properties and were capable of giving excellent predictions of chemical reactivity when trained with suitable molecular structure representations. The CNN methodology developed here may be useful for extracting reactivity rules from databases of chemical reactions.

Comments

Fifth Advisor: Dr. John Kapenga

Access Setting

Dissertation-Open Access

Download

Included in

Chemistry Commons

Share

COinS