QUANTITATIVE STRUCTURE-SUBLIMATION ENTHALPY RELATIONSHIP STUDIED BY NEURAL NETWORKS, THEORETICAL CRYSTAL PACKING CALCULATIONS AND MULTILINEAR REGRESSION-ANALYSIS

Three different techniques have been used to analyse the relationship between the structure of 62 organic compounds and their sublimation en thalpies. Using a neural network based on molecular structure descript ors (molecular formula, hydrogen bonding and pi-characteristics), subl imation enthalpies can be modelled, The best of the neural network mod els yielded an average error of 2.5 kcal mol(-1) in a series of 'leave -one-out experiments'. The same sublimation enthalpy data have been st udied using theoretical techniques based upon crystal packing calculat ions, and also with a simple three parameter multilinear regression mo del. The latter two methods produced results that were superior to the neural network in this particular study (mean errors of 1.4 and 1.8 k cal mol(-1), respectively), although in the case of MLRA, this is the result of the model fitting exercise, and not a predictive run. It was surprising to find such a simple linear relationship between characte ristics describing the molecular formula and the sublimation enthalpy. Nevertheless, the results here have highlighted the potential of neur al networks and MLRA as useful tools for the approximate prediction of physical properties, as demonstrated for a series of compounds not in cluded in the training set.