Fermentations are highly non-linear time-varying and inaccurately mode
lled processes. Our work towards the generation of a neural network mo
del of an industrial secondary metabolic fermentation process is prese
nted in this paper. The construction methodology and the various issue
s arising during the design, training and testing stages are also disc
ussed. Such issues addressed in this paper include the treatment of dy
namics and the determination of the number of hidden nodes. The useful
ness of the neural network model with regard to the on-line estimation
of fundamental measurements is demonstrated and its part in a control
system is highlighted.