Adaptive Boundary Constraint in Bayesian Optimization: A General Strategy to Prevent Futile Experiments in Complex Reaction Optimization

Abstract

Efficiently identifying optimal reaction conditions with minimal experimental effort is a fundamental challenge in chemical research, given the high cost and time involved in performing experiments. Recently, Bayesian Optimization (BO) has gained popularity for this purpose. However, we identify that for some common objective functions (e.g., throughput), some experimental conditions suggested by the algorithm are futile to perform. These experiments can be identified by determining whether the given experimental conditions can improve the existing best objective, even when assuming a 100% yield. We propose a strategy that incorporates knowledge of the objective function into BO, termed Adaptive Boundary Constraint Bayesian Optimization (ABC-BO). The proposed algorithm was tested in three in silico experiments using two different optimization solvers with various acquisition functions. ABC-BO effectively avoided futile experiments, increasing the likelihood of finding the best objective value. The effectiveness of ABC-BO was further demonstrated in a experimental case study of real-world complex reaction optimization involving multiple categorical, continuous, and discrete numeric variables. In the optimization performed using standard BO, 50% of the experiments were futile. In contrast, ABC-BO avoided futile experiments and identified a superior objective value compared to BO in a relatively smaller number of experiments. We show that the number of promising experimental conditions in the overall search space reduces as the optimization process progresses. Identifying and focusing on these conditions is more beneficial for optimizing the complex reaction space, especially when working with a limited experimental budget.