| Evaluated were the use of black box and hybrid neural networks in modeling and optimizing fed-batch fermentations. A neural network model, being an empirical fit of the data, may appear to model the data well, but perform poorly when applied more rigorously in optimization, a more accurate and useful criterion than the conventional root mean squared (RMS) error calculated from training.; Black box neural network models developed for fed-batch fermentations exhibiting substrate-inhibition kinetics performed poorly in optimization, and hence were poor models. Studies included a cycle-to-cycle optimization scheme in which the training data was modified for each cycle. Node saturation resulting from large weight values prevented the neural network from learning the proper behavior, and was resolved by reinitializing the weights. Successful optimizations resulted from including optimal values in the training data such that the optimal feed rate was a consequence of the neural networks interpolating, which they do well without a priori information.; Building upon this good interpolation ability, a hybrid scheme was employed in which neural networks modeled the complex kinetics, assuming only inputs of substrate, viable cell, and penicillin concentrations, within mass balances for a fed-batch penicillin fermentation. How the hybrid neural network scheme performed in optimization discriminated among the different models for the rate expressions. The failure of a hybrid neural network scheme to converge in optimization was attributed to the neural networks approximating functionally incorrect forms for the rate expressions, and justified elimination of those inputs for further process improvements. For example, when the penicillin production rate was modeled with inputs of substrate and penicillin concentrations, the hybrid neural network scheme optimized with a Sequential Quadratic Programming (SQP) algorithm produced optimal feeding policies comparable to those reported, with maximum penicillin amounts up to 13% larger than the 32.7 g from optimizing with SQP the reported model, which had postulated that penicillin production rate was also dependent on viable cell concentration. |
