On the Performance of DoDE in a class of in silico Fermentation Processes and the Impact of the Input Domain

Abstract

In a recent publication (Georgakis, 2013), it was shown that a data-driven model obtained through the proposed Design of Dynamic Experiments (DoDE) was able to accurately optimize a penicillin fermentation process without the use a knowledge-driven model. The resulting optimal operation, just after a set of experiments, is almost identical to the one obtained in (Riascos and Pinto, 2004) using the detailed model of the process by B&R (Bajpai and Reuss, 1980). Here we examine in silico whether a similar number of DoDE experiments will result in an equally accurate estimation of the optimal process operation of 32 other fermentation processes. This set of fermentation processes is defined by significantly varying the values of 10 key parameters of the initial penicillin model. Only between 3 and 7 of the 32 fermentations require additional experiments to obtain a satisfactory process optimization through a more accurate data-driven model. Furthermore, we examine two different time-evolving domains, A and B, within which the substrate in-flow is varied. One of them, domain A, forces the substrate inflow to be zero at the end of the batch. Domain B removes this constraint but requires more experiments. The obtained optimal operation in domain B is always better than that in domain A; sometimes by as much as 280%. This implies that limiting the number of experiments might also limit optimization gains