Advanced techniques for model-based experiment design under uncertainty

The use of first-principles mathematical models to describe bio-chemical phenomena is paramount to the effective development of microalgae processes. Model-based design of experiments (MBDoE) represents a valuable tool providing several methodologies that assist the model-building strategy of such complex mathematical models. The objective is to design experiments (or campaigns of experiments) that will provide the highest information content possible, thus reducing the time and resources needed to develop and validate models.

Model-based design of experiments can be mainly used towards two distinct but highly related goals: model discrimination, which aims to select the most appropriate model whenever multiple competing models have been proposed to describe a certain system, and improving parameter precision, which corresponds to improving the reliability of the estimates of the adjustable parameters once a specific model has been identified.

The initial focus of this research project has been directed to MBDoE for improving parameter precision. Several criteria are available in the scientific literature, but the ones that are most commonly used are not free from drawbacks. Local design criteria involve a linearisation of the model around the currently available estimates of the model parameters to predict the information content of the experiments being designed. This can lead to the design of suboptimal (non-informative) experiments, especially at the early stage of the model development procedure when parameter estimates may be highly uncertain or not available at all. Several criteria have been proposed that are more robust to the initial guesses of the model parameters, but they are often too computationally demanding.

The current research work involves attempting to combine local criteria with other approaches to the experiment design problem, with the aim of providing a methodology that can be used to design the most informative experiments possible. The design of campaigns of parallel experiments, in place of single experimental runs, is also currently under investigation.

by Marco Sandrin