A hybrid neural network algorithm for on-line state inference that accounts for differences in inoculum of Cephalosporium acremonium in fed-batch fermentors.

One serious difficulty in modeling a fermentative process is the forecasting of the duration of the lag phase. The usual approach to model biochemical reactors relies on first-principles, unstructured mathematical models. These models are not able to take into account changes in the process response caused by different incubation times or by repeated I. To overcome this problem, we have proposed a hybrid neural network algorithm. Feedforward neural networks were used to estimate rates of cell growth, substrate consumption, and product formation from on-line measurements during cephalosporin C production. These rates were included in the mass balance equations to estimate key process variables: concentrations of cells, substrate, and product. Data from fed-batch fermentation runs in a stirred aerated bioreactor employing the microorganism Cephalosporium acremonium ATCC 48272 were used. On-line measurements strongly related to the mass and activity of the cells used. They include carbon dioxide and oxygen concentrations in the exhausted gas. Good results were obtained using this approach.