Chemical engineers routinely wish to optimize the inputs to and operating conditions of a chemical reactor to achieve maximal conversion, yield, profit, etc.
Often, the complex reaction kinetics and fluid contact patterns in the reactor are poorly characterized, precluding [physical-]model-based optimization of the reactor.
Bayesian optimization (BayesOpt) combines supervised machine learning and automated decision-making for the adaptive, data-driven design of a sequence of experiments to find the optimal inputs to and/or operating conditions of a chemical reactor–using the fewest experiments. BayesOpt treats the reactor as a black-box function and constitutes a closed feedback loop:
(i) run the reactor with a given input and observe the output, (ii) use all data from reactor runs thus far to construct a surrogate model of the reactor, (iii) employ the surrogate model to choose the input for the next experiment, while balancing exploration and exploitation.
Intriguingly, BayesOpt may orchestrate autonomous chemical reactor optimization via automated instrumentation for manipulating inputs to and measuring outputs of the reactor.
In this talk, I will explain BayesOpt of chemical reactors and demonstrate, within a computer simulation, BayesOpt of the input feed to a semi-batch reactor for maximizing yield of a desired product in the face of a competing side reaction.
Cory Simon is an associate professor of chemical engineering at Oregon State University. He earned his PhD in Chemical Engineering from the University of California, Berkeley. His research group develops mathematical models, trains machine learning models, and conducts computer simulations to tackle or deliver insights into problems in chemistry and materials science.