| A significant number of pharmaceuticals and other biological products are produced by batch or fed-batch fermentations. Over the course of these fermentations, the concentrations of the key components and their time derivatives may change by several orders of magnitude. The kinetic equations governing these changes are generally nonlinear, representing the complex interplay between cellular physiology and environment. Modeling and prediction of the time course of a fermentation is therefore a challenging task. Accurate models would allow improvements in scheduling of process equipment and personnel, development of predictive control algorithms, and the investigation of new operating conditions.; In order to accurately portray the dynamic behavior of fermentation, this work proposes a hybrid modeling strategy that is a combination of classical fermentation models and artificial neural networks. Classical models are developed using chemical engineering first principles to provide mass balances and knowledge of the biochemistry of microorganisms to develop simplified rate models. Based on experimental data, values for the model parameters are found using regression. However, in the hybrid approach, one or more of these are allowed to become variables. Artificial neural networks are used to represent the variation in these "parameters", based on the fermentation state variables.; Initial test cases for modeling were simple mathematical functions: a circular paraboloid and the substrate-inhibition and cell mass-inhibition growth rate expressions. Modeling of a circular paraboloid showed it is important that the first-principles portion of the model exhibits dependencies on the independent variables similar to the underlying behavior, as provided by theory and shown in experimental data. Growth rate function modeling indicated the advantages of the hybrid modeling approach become more apparent as the complexity of the problem increases. Hybrid and black box neural network models were equivalent for substrate-inhibition models, but hybrid models were superior for cell mass-inhibition models.; Subsequent studies involved simulated experimental fermentation data from a batch fermentation and a fed-batch penicillin fermentation. In both cases, models developed using the hybrid modeling technique presented in this work (neural networks/mass balances/rate expressions) were superior to black box neural network models, classical unstructured models, and the previously proposed hybrid neural networks/mass balances models. |
