T. Heine

Quality-by-Design Using a Gaussian Mixture Density Approximation of Biological Uncertainties

Abstract In this contribution the uncertainties of a biological process model are taken into account explicitly to calculate optimal process trajectories. For this purpose, the initial condition and the uncertainties of the model parameters are described by a weighted sum of normal distributions. Such a so-called Gaussian mixture density (GMD) approximation is propagated through the nonlinear process model to calculate a second order approximation of the statistical properties of the planed process trajectory. A Value@Risk primary objective is used to obtain an optimal process design procedure in presence of uncertainties. In an extensive simulation study a descriptive fermentation process model is used to compare the classical trajectory planning with the robust design approaches. Here, different degrees of approximation complexity and the influence of the weighting factor in the Value@Risk dual objective criterion is investigated.

EXPERIMENTAL COMPARISON OF MODEL PREDICTIVE CONTROL STRATEGIES FOR THE PRODUCTION OF ANTIBIOTICS IN FED-BATCH FERMENTATIONS

Abstract The main objective of the presented work is to compare two model predictive control strategies by applying them to an antibiotic production in a fed-batch fermentation process. The reactor is modelled as a nonlinear biological compartment system with 13 states. The well-known nonlinear model predictive control (NMPC) is compared with a control strategy based on online optimization of the full trajectory. The control strategies are applied to real fermentation processes, which are strongly disturbed by a temperature shift. To extend the comparison, two state estimators, the Extended-Kalman-Filter (EKF) and the Constrained-EKF (CEKF), are used.

A software supported compartmental modeling approach for Paenibacillus polymyxa

Abstract This paper shows the derivation, selection and validation of a six-compartment model for P. polymyxa fed-batch fermentations on minimal media. P. polymyxa is known to produce a highly potent antibiotic, namely macrolactin. While biological insight is exploited to set up a basic model structure, a modeling tool is used to automatically select between possible reaction rates to describe individual steps. Experiments are used to identify the model parameters, and the most promising model candidate is validated with another experiment.

An automated approach to build process models by detecting biological phenomena in (fed-)batch experiments

Abstract This paper presents a method to detect appropriate process models from an automated analysis of (fed-)batch fermentations. After a data reconciliation the influence of feeding and sampling on the measurement trends is numerically compensated, in order to prepare the data sets for an automated detection of biological phenomena. At this, a probabilistic method is used to divide the compensated curves into several sections, called episodes, according to their qualitative behavior, i.e. increasing, decreasing, constant, or zero. This framework allows to calculate the probability of biological phenomena that reveal crucial information about the underlying reaction network. According to the probability level of the detected reactions a number of structured models is automatically proposed. An experimental validation of the described approach is shown using real fermentation data from fed-batches of Paenibacillus polymyxa , resulting in a simple but structured process model.

Robust Trajectory Planning and Robust NMPC of Processes with Model Uncertainties

Abstract This paper presents an approach for robust open-loop and robust closed-loop control of biological production processes described by models which comprise uncertain initial conditions and parameter uncertainties. The application of these robust techniques dramatically decreases the variability of the realized process trajectory and hence, the fluctuation of the product amount. The algorithm utilizes a second order approximation to calculate the first two statistical moments of the systems output as a function of the stochastic systems state and uncertain model parameters. Extensive simulation studies of both open- and closed-loop controlled fermentations based on a simple unstructured production model as well as on a structured compartment model consisting of 7 biotic states are shown.

A Comparison of Three Approaches for Control of Fed-Batch Process for Antibiotic Production by Streptomyces Strains

Abstract The main objective of the work is to investigate three different algorithms for control of antibiotic production during a fed-batch fermentation of Streptomyces strain and to compare the results from a user’s point of view. The reactor is modeled as a nonlinear biological compartment system with 13 states and 51 parameters. A control strategy based on online optimization of the full trajectory is compared separately with both other strategies (algorithms). The first one is a linearizing adaptive control and the second is an adaptive control designed by using the MATLAB toolbox DESIGNER. The main control task is to obtain maximum antibiotic concentration in the reactor controlling feed rates of three limiting substrates (or the most important among them) maintaining of all state restrictions. The verification of each control algorithm is realized by computer simulation of a control system that includes the structured process (reactor) model. The advantages and disadvantages of the investigated approaches are discussed.