ABSTRACT
The modeling of chemical and biotechnological processes suffers from two conflicting challenges: (1) the quantitative mechanistic understanding of essential subprocesses, such as reaction kinetics or cellular activity, is insufficiently available; (2) the availability of retrospective data for the training of black-box models is limited, and even worse, retrospective data rarely cover the data space sufficiently to inform black-box modeling tools. Hence, the predictivity often remains insufficient for applications such as model-based optimization. In this chapter it will be shown that the combination of knowledge of the model structure, eventually in combination with mechanistic models for a subset of subprocesses, significantly reduces the data demand for model identification and enables extrapolation. In this way, integration of partial knowledge of the process and data in a hybrid modeling platform provides a tool for systematic balancing of the investments into knowledge and data with respect to the required predictivity. This approach will be demonstrated through several examples from chemical industry.
