(351a) Model-Driven Design of Experiments and Optimization

Numerous scientific phenomena are investigated using complex computer models. Despite the surging computation power over the past few decades, certain models are still computationally prohibitive for systems tasks, such as process design and optimization, when a large number of evaluations is required. To alleviate the computational cost, one practical approach is to approximate expensive models with surrogate models. In this work, we investigate the development of surrogate models using information-theoretic tools and introduce an active learning model-driven design of experiments approach. In this approach, optimized Latin hypercube design is used as the initial sample plan and the next-best sample points are supplied iteratively by calling the information-based algorithm and evaluating the objective model. To benchmark the efficiency and effectiveness of our framework, we test it on various kinetic models, such as fructose to 5-hydroxymethyl furfural (HMF) conversion, as well as in approximating computational fluid dynamics (CFD) reactor models. By comparing the corresponding surrogates with the surrogates from one-stage factorial design, we demonstrate that the proposed method possesses superior efficiency in exploration and global optimal search. The surrogate models are fairly simple and can easily be applied in large scale systems tasks and used to bridge different scales in multiscale modeling. Finally, we demonstrate how to design actual experiments using this approach and demonstrate it experimentally using the HMF chemistry mentioned above.