(183f) An Optimization Framework to Combine Design Space Maximization with Design of Experiments | AIChE

(183f) An Optimization Framework to Combine Design Space Maximization with Design of Experiments


Chen, Q. - Presenter, Imperial College London
Adjiman, C. S. - Presenter, Imperial College London
Paulavicius, R. - Presenter, Imperial College London, Center for Process Systems Engineering
Garcia-Munoz, S. - Presenter, Eli Lilly and Company

While pharmaceutical quality by design enables the specification of a design space in terms of process inputs, the space of process outputs may be more relevant to safety and product quality. Thus, while design of experiments techniques are often used to explore the effect of process parameters, it is often desirable during development, and for the purpose of process understanding, to maximize the explored space in terms of process outputs or quality attributes. However, timeline and resource constraints limit the number of experimental trials available to properly characterize both spaces, prompting the need for systematic design of experiments. An optimization framework based on the sometimes conflicting objectives of maximizing the size of the output or quality attribute design space and maximizing the D-optimality of the input parameters is proposed and applied to a spray-coating process. This example shows that optimizing only one of the two objectives yields unacceptable results in the neglected objective. Two strategies are explored. First, a multi-objective optimization approach is presented, enabling visualization of possible tradeoffs as a Pareto-optimal front—a set of solutions where no objective can be further improved without degrading another objective value. This allows the decision maker to choose among a set of efficient designs. It also provides insights into the potential benefits of increasing (or decreasing) the number of experiments. Second, a bilevel formulation is also presented which enables process engineers to quickly identify good compromise designs without the need to map out multiple points along the Pareto-optimal front.