Every chemical reaction must be optimized before industrialization. The goal of this optimization is to determine, through experiments, the reaction parame- ters (e.g., temperature, concentration, pressure) that minimize or maximize an objective (e.g., yield, selectivity). Conducting these experiments requires a sub- stantial amount of expert knowledge. Nevertheless, it is not reasonably possible to establish an analytical form of the relationship between the reaction para- meters and the objectives. Therefore, we refer to this as the optimization of a black box.

Experiments can be expensive, requiring substantial resources, time, and financial investment. Motivated by the ecological and economic challenges in the industry, this manuscript focuses on a specific paradigm : optimizing a black-box model with the minimum number of evaluations needed.

Recently, Bayesian optimization has emerged as the benchmark method for optimizing difficult-to-evaluate black-box functions. However, there are many variants of the classical reaction optimization problem. The work presented in this manuscript addresses a specific case where the reaction parameters are of different types. We focus on continuous variables, such as the choice of reaction temperature, and categorical variables, such as the choice of base or solvent. These latter variables, being discrete and unordered, are inherently challenging to handle due to their non-numeric nature.

Several directions have been explored in depth:

• The simulation of reactions and formulations with continuous and categorical parameters.
• The optimization of these simulations.
• The creation of synthetic functions inspired by well-known synthetic functions used by the optimization community.
• The analysis of the balance between exploration and exploitation in our optimization method.
• The improvement of the modeling part of our optimization method.