A new training method for hybrid models of bioprocesses

Abstract

Modeling of bioprocesses for engineering applications is a very dif®cult and time consuming task, due to their complex nonlinear dynamic behavior. In the last years several propositions for hybrid models, and especially serial approaches, were published and discussed, in order to combine analytical prior knowledge with the learning capabilities of Arti®cial Neural Networks (ANN). These approaches often require synchronous and equidistant sampled training data. However, in practice concentrations are mostly off-line measured, rare, and asynchronous. In this paper a new training method especially suited for very few asynchronously sampled data is presented and applied for modeling animal cell cultures. The achieved model is able to predict the concentrations of the reaction components inside a stirred tank bioreactor. IntroductionBiotechnology is among the most modern technologies. Its fundamentals are the subject of scienti®c research and its various applications are explored in engineering science. One problem in process engineering is the lack of accurate process models at the macroscopic level. Those are not only necessary for the design of control, but also for the development and understanding of a process. Computer simulations based on a process model are useful to reduce the number of experiments with the real bioprocesses, which are expensive, and particularly time consuming.ANN is one of the most commonly used approach from a large variety of data driven modeling techniques [1]. The disadvantage of neural modeling is the lack of transparency. In ANN models process knowledge is represented in an unstructured``black box''. Therefore, it is extremely dif®cult to make use of existing process knowledge.Numerous authors have shown that so called hybrid approaches offer certain advantages compared to pure ANN approaches. Prior knowledge, which is usually available in form of balance equations, is used as basis for the model and only the unknown parts are learned by an ANN. Thompson and Kramer [2] distinguish between two types of hybrid models. In the ®rst type of approach, the weights of classical ANNs are trained with respect to certain constraints which limit the search space. These constraints are de®ned using the existing knowledge, e.g. in order to insure the stability of the model [3]. In the second type classical``white box'' models are combined with neural``black box'' components. Among several possible architectures the parallel and serial structures, as shown in Fig. 1, are the most common ones.In the parallel structure (Fig. 1, left) a complete classical model is connected in parallel with a neural network. The classical model provides an estimation of the output, while the neural network is trained to compensate remaining errors between the model and the observed process behaviour. In the serial structure (Fig. 1, right) an incomplete classical model is used. The unknown or hardly known terms ± e.g. the process kinetics ± are represented by an ANN. The latter approach is used in the present work.In ...

Macroscopic modeling of mammalian cell growth and metabolism

“…By using this hybrid model in a fed-batch process, they were able to increase the final productivity of IgG1-IL2 by 10 % (Teixeira et al 2007 ). Graefe et al ( 1999 ) applied a serial hybrid model to CHO-K1 by combining mass balances and neural network kinetics. A convincing prediction of components concentrations in the stirred tank bioreactor was achieved.…”

“…LMA was also used by Vande Wouwer et al ( 2004 ) for batch CHO cell culture. LMA was applied in the training process for hybrid models of bioprocesses (Graefe et al 1999 ). Dorka et al ( 2009 ) successfully identified Monod-type parameters of hybridoma cell culture during exponential phase by using quadratic programming (QP).…”

Modeling of growth and sporulation of Bacillus thuringiensis in an intermittent fed batch culture with total cell retention

“…Additionally, computer simulations based on a process model are useful to reduce the number of experiments with the real biological processes, which are expensive, and particularly time consuming [5].…”

Parameter identification for state estimation—application to bioprocess software sensors

“…Rare and asynchronous measurements of biomass, glutamine, glucose and lactate concentrations were performed in batch CHO (Chinese Hamster Ovaries) cell cultures (for details about the experimental conditions, see (Graefe et al, 1999)). In order to describe the macroscopic phenomena occurring in these cultures, a simple reaction scheme is proposed:…”