Design Considerations in Hybrid Neural Optimization of Fed-Batch Fermentation for PHB Production by Ralstonia eutropha

Abstract

Mechanistic models are often inadequate for non-ideal fermentation processes. This has been a major limitation in the commercial success of fed-batch fermentations for poly-β-hydroxybutyrate (PHB), a biopolymer that is potentially better than many synthetic polymers. A previous study with Ralstonia eutropha has shown that optimization through neural networks in place of mechanistic models can help to enhance PHB production substantially. Hybrid models combine the strengths of mechanistic and neural models and minimize their weaknesses. However, there is no unique hybrid neural model for a given application. Therefore, three fundamental hybrid designs have been compared with the neural and mechanistic optimizations done earlier for a fed-batch bioreactor for PHB production. Non-ideal features were introduced by adding Gaussian noise in the two feed streams—glucose and NH4Cl—and incorporating the optimum finite dispersion determined previously. All three hybrid designs were superior to the neural and mechanistic approaches. The best hybrid system, using a weighted combination of mechanistic and neural kinetic rates, generated 140% more of biomass and 330% more of PHB than conventional mechanistic optimization. This was achieved with reduced consumption of the two primary substrates. The supremacy of this hybrid system underlines the complexity of the PHB synthesis network and the merit in combining mechanistic and neural kinetics, and the performance enhancement achieved indicates the feasibility of such a system for an economically viable large-scale PHB fermentation.