8.1 General Conclusions
In this book, a number of results related to monitoring, modelling and optimization of fed-batch fermentation processes are presented. The study focuses on AI approaches, in particular, RNNs and GAs. These two techniques can be used either separately or together to fulfill various goals in the research. The great advantages that are offered by these approaches are the flexible implementation, fast prototype development and high benefit/cost ratio. Their applications to biotechnology process control provide a new inexpensive, yet effective way to improve the production yield and reduce the environmental impact.
A comparison of GAs and DP has demonstrated that GAs are superior to DP for optimization fed-batch fermentation processes. An on-line identification and optimization method based on a series of real-valued GAs was successfully applied to estimate the parameters of a seventh order system and to maximize the final concentration of hybridoma cells in a fed-batch culture. In the first two days of the fermentation, the system parameters were found using the GA based on the measured data. Then the optimal feed rate control profiles were determined using the predicted model. In the last eight days of fermentation, the bioreactor was driven under the control of optimal feed flow rates and reached a final MAb concentration of 193.1 mg/L and a final volume of 2L at the end of the fermentation. This result is only 2% less than the best result (196.27 mg/L) obtained in the case which all the parameters are assumed to be known.
The suitability of using a RNN model for on-line biomass estimation in fermentation processes has been investigated. Through simulations, an appropriate neural network topology is selected. This selected neural network topology is further tested experimentally. From the experimental results, the proposed softsensor has shown itself be able to predict the biomass concentration with an RMSP error of 10.3%. The proposed softsensor provides a powerful tool for measuring the biomass on-line.
L. Z. Chen et al.: Modelling and Optimization of Biotechnological Processes, Studies in Computational Intelligence (SCI) 15, 109-110 (2006)
www.springerlink.com © Springer-Verlag Berlin Heidelberg 2006
A cascade RNN model proposed in this work has proved capable of capturing the dynamic nonlinear underlying phenomena contained in the training data set and can be used as the model of the bioprocess for optimization purpose. The structure of the neural network model is selected using validation and testing methods. A modified GA is presented for solving the optimization problem with a strong capability of producing smooth feed rate profiles. The results of optimal feeding trajectories obtained based both on the mechanistic model and the neural network model have demonstrated that the cascade recurrent neural model is competent in finding the optimal feed rate profiles. The proposed approach can partly eliminate the difficulties of having to specify completely the structure and parameters of a bioprocess model.
Finally, the design and implementation of optimal control of bench-scale fed-batch fermentation processes using cascade RNNs and GAs are presented. The neural network that is proposed in the work has a strong capability of capturing the nonlinear dynamic relationships between input-output data pairs, provided that sufficient data that are measured at appropriate sampling intervals are available. It has also shown that proper data processing and zero-appending methods can further improve the prediction accuracy. GAs have been used for solving the dynamic constraint optimization problem. The fast convergence as well as global solution are achieved by the novel constraint handling technique and the incremental feed subdivision strategy. Among all 12 experiments, the one controlled by the optimal feed rate profile based on the DO neural model yields the highest product. The main advantage of the approach is that the optimization can be accomplished without a priori knowledge or detailed kinetic models of the processes. Owing to the data-driven nature of neural networks and the stochastic search mechanism of GAs, the approach can be readily adopted for other dynamic optimization problems such as determining optimal initial conditions or temperature trajectories for batch or fed-batch reactors.
8.2 Suggestions for Future Research
Investigations presented in this book have opened several key areas that the author would like to suggest for future studies.
• Combination of problem-specific process knowledge and RNNs can be considered to enhance the robustness and extrapolability of the fed-batch fermentation model. However, the development cost may increase.
• Combination of conventional mathematical optimization schemes with the GA should further improve the optimality of the optimal feed rate profiles.
• Online adaptation or tuning of the models and the optimal feed rate profiles are required to produce more reliable and repeatable results, especially when the process time is increased.
• Optimal experimental design can be used to increase the span of the space that is covered by the experimental database.
Was this article helpful?