Glucoamylase (E.C. 3.2.1.3, 1, 4-a-D Glucan Glucohydrolase) is an industrially important enzyme with major use in the starch industry. Online prediction of cell mass is of great value in real time process control during extracellular glucoamylase production process. Experimental data on the effect of carbon sources, pH, temperature, agitation and aeration on cell mass production during batch and continuous glucoamylase fermentation, were used and applied to an artificial neural network model programmed in microsoft visual basic for windows. Based on the Levenberg-Marquardt algorithm with back propagation, the network so developed was used to predict the cell mass. The optimum network configuration and activation function for both feed forward and recurrent neural networks were found. The conventional Monod and Contois models mostly described the fungal cell growth kinetics during batch fermentation. The overall results for all sets of data from the feed forward and recurrent neural network were compared with the conventional Monod and Contois model for cell growth kinetics. It was found that the feed forward neural network was giving better results than the recurrent neural network with coefficient of determination (R²) of 0.9950.

Biological systems do not easily lend themselves to quantitative description due to common characteristics such as nonlinearity, complexity and lack of understanding. Simple linear models with few parameters do not represent the biological system efficiently. These lead us to the Black-box model which is purely based on the experimental data. Online readjustment of operation through a good controlling system helps in the maintenance of productivity. Unfortunately such adjustments are hampered by the lack of online sensors for indication of bioreactor performance (Bhat and McAvoy, 1990). An alternative to the use of sensors is the development of online estimation of algorithms to provide necessary performance measures. A neural network, which has the ability to learn sophisticated nonlinear relationships provides an ideal means of modeling such complicated systems. Feed Forward Neural Networks (FFNN) and Recurrent Neural Networks (RNN), which have excellent characteristics, such as dealing with nonlinear systems with noisy and approximate data, learning from the past data to changing environment, relatively simple in structure and executing action quickly once the network has been trained, offer a great potential in biosensor data processing and in variable prediction, in optimization and in advanced control of dynamic bioprocesses (Karim and Rivera, 1992).

In the last few years, many researchers from diverse disciplines have concentrated in the field of Artificial Neural Networks (ANN), including modeling, system identification and control (Willis et al., 1990) and studied the capability of ANN to characterize nonlinear functional relationships. Di Massimo et al. (1992) investigated the ability of ANN to learn process nonlinearities from plant data. They developed an artificial network-based biomass and penicillin estimator for online application to an industrial fermentation.

Chemical Engineering Journal, Glucoamylase, Neural Network, Cell Mass Estimation, Aspergillus Awamori, Recurrent Neural Networks, RNN, Artificial Neural Networks, ANN, Zymomonas Mobilis, Feed Forward Neural Networks, FFNN, Recurrent Neural Networks, RNN, Glucoamylase Production, Biological Systems.