Sequential Bayesian Experimental Design for Process Optimization with Stochastic Binary Outcomes

Abstract

For innovative products, the issue of reproducibly obtaining their desired end-use properties at industrial scale is the main problem to be addressed and solved in process development. Lacking a reliable first-principles process model, a Bayesian optimization algorithm is proposed. On this basis, a short of sequence of experimental runs for pinpointing operating conditions that maximize the probability of successfully complying with end-use product properties is defined. Bayesian optimization is able to take advantage of the full information provided by the sequence of experiments made using a probabilistic model (Gaussian process) of the probability of success based on a one-class classification method. The metric which is maximized to decide the conditions for the next experiment is designed around the expected improvement for a binary response. The proposed algorithm's performance is demonstrated using simulation data from a fed-batch reactor for emulsion polymerization of styrene.