Nonlinear statistical models based on artificial neural networks have been constructed to predict the thickening times of class G, class H, and class A oil well cement slurries from physicochemical data on cement powders. These data include chemical composition, particlesize distributions, and Fourier transform infrared (FTIR) spectra. Thickening times can be predicted with mean residual errors of less than +-10% from measurements of major oxide composition, loss on ignition, insoluble residue, the concentrations of the minerals gypsum, bassanite, anhydrite, calcium hydroxide, and calcium carbonate, and particle-size distribution. This error rises to +-13% when particle-size distribution is omitted from the input data. Similar conclusions apply to models that predict thickening times of cement slurries retarded with 0.2 weight percent calcium lignosulfonate. Slurry thickening times can also be predicted from cement powder FTIR spectra alone, with errors in predictions similar to those for models using the full chemical composition and particle-size distribution data. This implies that the FTIR spectrum of a cement is a signature of its slurry performance properties. Additionally, neural networks can be trained to predict the particle-size distributions and mean particle diameters from FTIR spectra. The occupancies of all particle-size fractions can be predicted to within an absolute error of +1%, approximately twice the expected experimental error on the measurement. However, the model is able to predict general trends in particle-size distributions between cements effectively. AD-