Monitoring and advanced control of batch processes poses a desire for modelling; however first principles modelling of bio-processes is still in its infancy due to the lacking knowledge of microbial regulatory mechanisms. Thus a significant challenge remains for development of Process Analytical Technology.
An promising candidate methodology is described for
modelling batch and fedbatch processes, based upon data obtained from standard operating processes. The paper also illustrates how designed experiments may be used to improve the model quality. The paper describes how the developed models may be used for process monitoring, for ensuring process reproducibility and for optimizing process performance by enforcing learning from previous
batch runs through Learning Model Predictive Control (L-MPC).
A penicillin cultivation is used as a benchmark based upon a simple model. Data sets are simulated for seven batches each with two perturbed input and four output variables. A validated multivariable predictive model for the non-linear fed-batch process is obtained. Learning Model Predictive Control (L-MPC) is developed for this benchmark simulation. The controller is tested under scenarios with inter-batch disturbances and with intra-batch disturbances where good control performance and robustness is demonstrated in both cases.
A model is developed for an industrial alpha-amylase production in a Fungal cultivation carried out at the Novozymes Pilot Plant. For that purpose, two additional identification experiments are designed with two perturbed input variables to generate data with sufficient information for model estimation. Combining these new data with data from a few previous batches then a set of batch data for model estimation is selected depending on the data quality. The estimated model is in validation shown
to very well predict the process trajectory and the dynamics around it. Thus a candidate model for monitoring fed-batch reactors and for implementing L-MPC is available.