Hybrid Models for Biological Reactors: Performance and Possibilities.

The complexity of microbial processes, coupled with the effects of incomplete mixing and the influx of noise, makes it difficult to describe quantitatively the behavior of bioreactors under realistic conditions. By circumventing the need for phenomenological equations, neural networks offer viable descriptions of such systems. However, such networks also require large amounts of carefully screened data and long training schedules. They also have limited extrapolation capability and poor physiological fidelity. Since phenomenological models are derived from physical descriptions of cellular processes, combinations of such models with neural networks, and sometimes also fuzzy logic, have provided good portrayals of nonideal bioreactor behavior while reducing the weaknesses of both approaches. However, like neural networks themselves, different architectures are possible for hybrid neural models. The rationale, architectures and illustrative applications of hybrid models for biological reactors are discussed here, with the possibility of combining them with cybernetic models, exploiting cellular intelligence, to develop self-evolving intelligent systems for optimization and control of microbial processes. [ABSTRACT FROM AUTHOR]