This paper explores the effects of the correlation model, the trend model, and the number of training points on the accuracy of Kriging metamodels. Gaussian correlation models are found to be superior to exponential and linear correlation models. No particular trend model is found to be better than the other models. The number of training points used in constructing the Kriging metamodels is observed to change the relative performances of the trend and the correlation functions. The leave-one-out cross-validation error is found to become a better surrogate for the actual error, as the number of training points is increased. Finally, the use of an ensemble of metamodels is discussed and it is found that using an ensemble may improve the accuracy.