Quantitative structure-performance relationship (QSPR) and neural network models have been designed to correlate and predict physical properties of pure components and a mixture parameter for a simple equation of state. The key step was to generate and select those structure-related parameters (descriptors) that best described the experimental physical property data by a multilinear regression or a neural network analysis. The descriptors found show theoretical significance and allow insights in the theoretical background of the physical properties investigated. The correlations and neural network models enable us to predict physical properties of compounds related to but not present in the training set of compounds used for the development of the QSPR and neural network models. Examples are presented for the prediction of the normal boiling point of chlorosilanes, the cloud points of surfactants, and the combining rule parameter k(ij) in a modified Peng-Robinson equation of state applied to vapor-liquid equilibria of binary systems containing carbon dioxide.