Categorical-Continuous Bayesian Optimization Applied to Chemical Reactions

Abstract

Chemical reaction optimization is a challenging task for the industry. Its purpose is to experimentally find reaction parameters (e.g. temperature, concentration, pressure) that maximize or minimize a set of objectives (e.g. yield or selectivity of the chemical reaction). These experiments are often expensive and long (up to several days), making the use of modern optimization methods more and more attractive for chemistry scientists.

Recently, Bayesian optimization has been shown to outperform human decision-making for the optimization of chemical reactions [16]. It is well-suited for chemical reaction optimization problems, for which the evaluation is expensive and noisy.

In this paper we address the problem of chemical reaction optimization with continuous and categorical variables.

We propose a Bayesian optimization method that uses a covariance function specifically designed for categorical and continuous variables and initially proposed by Ru et al. in the COCABO method [14].

We also experimentally compare different methods to optimize the acquisition function. We measure their performances in the optimization of multiple chemical reaction (or formulation) simulators.

We find that a brute-force approach for the optimization of the acquisition function offers the best results but is too slow when there are many categorical variables or categories. However we show that an ant colony optimization technique for the optimization of the acquisition function is a well-suited alternative when the brute-force approach cannot be (reasonably) used.

We show that the proposed Bayesian optimization algorithm finds optimal reaction parameters in fewer experiments than state of the art algorithms on our simulators.