A fed-batch fermentation for beta-galactosidase production by a recombinant Escherichia coli strain has been simulated with interruptions in the inflow rates of a concentrated substrate and a diluent. Data covering a 12 h fermentation period were simulated by a 6-12-4 Elman neural network with two extra- and two intracellular variables as outputs. Three types of inflow failure were considered : either of the two feed streams separately or both together. Despite the fermentation performance being quite different for each type of failure, a network trained with data pertaining to one kind of-failure was able to mimic the performance adequately for the other two data sets. The largest error and root mean square error were 14 and 10%, respectively, among the plasmid DNA and the intracellular protein concentrations, and 9 and 6% for the concentrations and mass fractions of recombinant cells. The accuracy is better than that reported for back-propagation networks.