We used a self-organizing feature maps neural network (SOFMNN) to model batch culture of a recombinant yeast, a lipocortin producer. The major concern was modeling of the relationships of various specific rates to three key state variables : glucose, ethanol and cell mass concentrations. The ultimate goal was to develop a model to be used for the on-line control of ethanol in fed-batch cultures.
The key step in the present modeling approach is training of the neural network, so called, learning process. In this process, specific rates data determined from state variables vs. time curves. The data used in the learning process was called learning data. In the determination of the specific rates data, a cubic spline method was employed to smooth out the curves.
In simulation study, we investigated the effect of iteration on learning efficiency of the neural network to find that there existed an optimal number of iteration. The specific rates data and values of the state variables regenerated by the neural network after learning agreed quite well with the learning data. The neural network also showed a good interpolation capability. Although the neural network had been trained with data sets obtained with 12 g/L and 20 g/L of initial glucose concentrations, with the aid of mass balance equation, it accurately predicted the time profiles of the state variables for the case of 15 g/L of initial glucose glucose concentrations.
In experimental study, the neural network model again showed a good accuracy and interpolation capability, which was much superior to that of a mathematical model developed earlier.
Results of preliminary experiment on the effects of the ethanol concentration on cell growth and lipocortin production have been presented. A on-line control algorithm for ethanol concentration regulation has been proposed.
