Physical properties are fundamental to all chemical, biochemical and environmental industries. One of these properties is the normal boiling point of a compound. However, experimental values in literature are quite limited and measurements are expensive and time consuming. For this reason, group contribution estimation methods are generally used. Group contribution is the simplest form of estimation requiring only the molecular structure as input. Consequently, the aim of this project was the development of a reliable group contribution method for the estimation of normal boiling points of non-electrolytes applicable for a broad range of components. A literature review of the available methods for the prediction of the normal boiling points from molecular structure only, was initially undertaken. From the review, the Cordes and Rarey (2002) method suggested the best scientific approach to group contribution. This involved defining the structural first-order groups according to its neighbouring atoms. This definition also provided knowledge of the neighbourhood and the electronic structure of the group. The method also yielded the lowest average absolute deviation and probability of prediction failure. Consequently, the proposed group contribution method was then developed using the Cordes and Rarey method as a starting point. The data set included experimental data for approximately 3000 components, 2700 of which were stored in the Dortmund Data Bank (DDB) and about 300 stored in Beilstein. The mathematical formalism was modified to allow for separate examination and regression of individual contributions using a meta-language filter program developed specifically for this purpose. The results of this separate examination lead to the detection of unreliable data, the re-classification of structural groups, and introduction of new structural groups to extend the range of the method. The method was extended using steric parameters, additional corrections and group interaction parameters. Steric parameters contain information about the greater neighbourhood of a carbon. The additional corrections were introduced to account for certain electronic and structural effects that the first-order groups could not capture. Group interactions were introduced to allow for the estimation of complex multifunctional compounds, for which previous methods gave extraordinary large deviations from experimental findings. Several approaches to find an improved linearization function did not lead to an improvement of the Cordes and Rarey method. The results of the new method are extensively compared to the work of Cordes and Rarey and currently-used methods and are shown to be far more accurate and reliable. Overall, the proposed method yielded an average absolute deviation of 6.50K (1.52%) for a set of 2820 components. For the available methods, Joback and Reid produced an average absolute deviation of 21.37K (4.67%) for a set of 2514 components, 14.46K (3.53%) for 2578 components for Stein and Brown, 13.22K (3.15%) for 2267 components for Constantinou and Gani, 10.23 (2.33%) for 1675 components for Marrero and Pardillo and 8.18K (1.90%) for 2766 components for Cordes and Rarey. This implies that the proposed method yielded the lowest average deviation with the broadest range of applicability. Also, on an analysis of the probability of prediction failure, only 3% of the data was greater than 20K for the proposed method. This detailed comparison serves as a very valuable tool for the estimation of prediction reliability and probable error. Structural groups were defined in a standardized form and the fragmentation of the molecular structures was performed by an automatic procedure to eliminate any arbitrary assumptions.

Thesis (M.Sc.Eng.)-University of Natal, Durban, 2004.