A hybrid neural modelling procedure which enables the implementation of an adaptive control scheme for the optimization of fed-batch fermentations is presented. Simulations for the processes of cell mass production and ethanol fermentation by Sacharomyces cerevisae show that, in the presence of modelling errors, the adaptive control leads to nearly optimal results, while open-loop control leads to bad results. Experimental studies show that, for the process of ethanol fermentation by Zymomonas mobilis, a hybrid neural model can be-developed with relatively few experimental data and the use of an approximate mathematical model.