In chemical process engineering, data are collected in experiments in order to calibrate physically motivated models. These experiments are always time- and cost-intensive. Therefore, their planning is about deriving as reliable models as possible from as few experiments as possible. In a cooperative project with BASF, we develop and implement methods that support this.
Conflict for a Reliable Model
The reliability of model calibrations is influenced in two ways: On the one hand, the error bars of the estimated parameters, but also the prediction errors of the model are directly proportional to the measurement accuracy in the experiments. In other words, the more accurate the sensors, the more reliable the model prediction. On the other hand, in order to calibrate successfully, it is crucial to consider correlations in the sensitivity of the models - especially with regard to the model parameters at the measurement points. This is illustrated in the following example.