A systematic methodology for improving the efficiency of fermentation process development is presented. It consists of the "smallest composite" experimental design and subsequent feature correlation and analyses. The methodology is applied to a simulation of tylosin fermentation. By looking at the whole parameter space of the simulation, the interesting behavior of a neural net changing from lower order to higher order during training is examined. The results show that, different sizes of neural nets within a certain range give an equally good prediction by using the "stopping training" technique, while quadratic regressions are sensitive to the size of the data sets.