Bayesian statistical methods provide a sound mathematical framework to combine prior knowledge about controllable variables of importance in a process, if available, with the actual data, and has proved useful in several data analytic tasks in the process industry. We present a review and some extensions of bayesian predictive methods for process optimization based on experimental design data, an area that is critical in Quality by Design activities and where the bayesian perspective has received limited attention from the Chemometrics and process analytics communities. The goal of the methods is to maximize the probability of conformance of the predicted responses to their specification limits by varying the process operating conditions. Optimization of multiple response systems and of systems where the performance is given by a curve or “profile” are considered, as they are more challenging to model and optimize, yet are increasingly common in practice. We discuss the particular case of Robust Parameter Design, a technique due to G. Taguchi and popular in discrete manufacturing systems, and its implementation within a bayesian optimization framework. The usefulness of the models and methods is illustrated with three real-life chemical process examples. MATLAB code that implements all methods and reproduces all examples is made available.
All Science Journal Classification (ASJC) codes
- Analytical Chemistry
- Process Chemistry and Technology
- Computer Science Applications