Abstract:Multi-objective Bayesian optimization (MOBO) has shown to be a promising tool for reaction development. However, noise is usually unavoidable during experiments and makes it challenging to find reliable solutions. In this study, we focus on finding a set of optimal reaction conditions using multi-objective Euclidian expected quantile improvement (MO-E-EQI) under noisy settings. First, the performance of MO-E-EQI is evaluated by comparing with some recent MOBO algorithms \textit{in silico} with linear and log-linear heteroscedastic noise structures and different magnitudes. It is noticed that high noise can degrade the performance of MOBO algorithms. MO-E-EQI shows robust performance in terms of hypervolume-based metric, coverage metric and number of solutions on the Pareto front. Finally, MO-E-EQI is implemented in a real case to optimize an esterification reaction to achieve maximum space-time-yield and minimal E-factor. The algorithm identifies a clear trade-off between the two objectives.

Multi-objective reaction optimization under uncertainties using expected quantile improvement