Search

The cultivation of Bacillus licheniformis and the production of subtilisin Carlsberg serine protease were investigated in complex medium with starch and glucose, respectively, and Na-caseinate as substrates to maximize the protease concentration. The turbidity and culture fluorescence were monitored in situ, the optical density on-line and the (dry) sediment and (dry) cell mass concentrations as well as the cell count and the DNA content were monitored off-line. These values are closely interrelated and were quantified by particular relationships. By means of the six-channel flow injection analyser (FIA) system, the following medium components were monitored on-line: glucose, maltose, starch, ammonium, urea, phosphate and protease activity. The same components as well as protein, intracellular phosphate and-α-amylase activity were evaluated off-line. The off-gas composition was analysed on-line as well. Various control strategies were tested in order to maximize the protease concentration: On one hand, starch in various concentrations was used as substrate. These runs were performed at non-controlled starch decomposition, at controlled and non-controlled pH-values, respectively, and non-controlled Po2-values. On the other hand, glucose was used as substrate in fed-batch mode. These runs were performed with closed loop controlled pH- and Po2-values and open-loop and closed-loop controlled glucose concentrations, respectively. The latter strategy yielded a higher protease concentration than the former. With complex medium and closed-loop controlled process, extremely high protease activities (24 000 EPE ml −1) were obtained.

Artificial neural networks are powerful in representing nonlinear processes, but they may not be reliable when they are applied beyond their range of experience. In order to avoid suspect predictions, appropriate extensions must be installed which monitor their reliability. When their performance becomes unsatisfactory, a sufficient safety measure must be provided. A software tool is described, in which these aspects of improving artificial neural nets are implemented. Application examples are presented from a production scale beer brewery fermenter.

An improved technique of hybrid modelling biochemical production processes is described, composed of a set of dynamical differential equations, an artificial neural network and a fuzzy expert system. The applicability of a hybrid model for state estimation, prediction, feed rate optimization, and process control is explained. Its performance is demonstrated by means of an the example of a fed-batch baker's yeast production process. The identification procedure is described in detail. Emphasis is placed on the most difficult part of it, the training of the neural network component of the hybrid model. New approaches to process control are developed based on neural networks.

Penicillin production with a high-producing strain Penicillium chrysogenum was investigated under well-controlled conditions in a stirred tank reactor with complex media containing lard oil and lactose on the one hand, and lactose on the other hand. With lard oil, cell growth and product formation rates were higher, and the production time was shorter by 40 h than without lard oil. On account of the longer production time without lard oil, the amount of β-lactam compounds was higher (29.93 g 1−1), but the mole fraction of the decomposed products (penicilloic acid and penilloic acid) was larger (0.282) than the amount of penicillin V (23.25 g 1−1 and the decomposed mole fraction (0.0747) with lard oil. The final product concentrations were about the same (20.86 g 1−1 or 35 462 IU ml−1 with lard oil, and 20.43 g 1−1 or 34510 IU ml −1 without lard oil). The mole fractions of the by-product (p-OH-penicillin V) were 0.0365 and 0.066. The substitution of lard oil with lactose is possible without a considerable reduction of process performance.