Why is Artificial Intelligence Attractive for Fermentation Control

The last decade or so, has seen a rapid transition from conventional monitoring and control based on mathematical analysis to soft sensing and control based on AI. In an article on the historical perspective of systems and control, Zadeh considers this decade as the era of intelligent systems and urges for some tuning [70]:

"I believe the system analysis and controls should embrace soft computing and assign a higher priority to the development of methods that can cope with imprecision, uncertainties and partial truth".

Fermentation processes, as mentioned in Section 1.1, are exceedingly complex in their physiology and performance. To propose mathematical models that are sufficiently accurate, robust and simple is a time-consuming and costly work, especially in the noisy interactive environment. AI, particularly neural networks, provides a powerful tool to handle such problems. An illustration of a neural network-based biomass and penicillin predictor has been given by Di Massimo et al. [71]. The neural network of relatively modest scale was demonstrated to be able to capture the complex bioprocess dynamics with a reasonable accuracy. The ability to infer some important state variables (eg. biomass) from other measurements makes neural networks very attractive in the applications of fermentation monitoring and modelling [72,73,74], because it can reduce the burden of having to completely construct the mathematical models and to specify all the parameters.

The dynamic optimization problems of such complex, time-variant and highly nonlinear systems are difficult to solve. The conventional analytical methods, such as Green's theorem and the maximum (or minimum) principle of Pontryagin, are unable to provide a complete solution due to singular control problems [75]. Meanwhile, conventional numerical methods, such as DP, suffer from a large computational burden and may lead to suboptimal solutions [21]. An example of a comparison between GA and DP is given in [76]. Both methods are used for determining the optimal feed rate profile of a fed-batch culture. The result shows that the final production of monoclonal antibodies (MAb) produced by using a GA is about 24% higher than that produced by using the DP. In addition to the advantage of global solution, GAs can be applied to both "white box" and "black box" models (eg. neural network models) [45,77]. This offers a great opportunity to combine GAs with neural networks for optimization of fermentation processes.

Finally, AI approaches provide the benefit of rapid prototype development and cost-effective solutions. Due to less a priori knowledge being required in AI methods, monitoring, modelling and optimization of fermentation processes can be achieved using a much shorter time as compared to conventional approaches. This can lead to a significant saving in the amount of investment in process development.

Was this article helpful?

0 0

Post a comment