In chemistry, a reaction yield depends on several parameters (at least 4 or more), such as the reactant concentrations or the temperature. Together, these parameters define a reaction setting. When the number of possible settings is large, it becomes impossible to test them all due to time and material constraints. 

Then, the goal is to identify a setting that leads to a good yield within a limited number of successive experiments. To address this challenge, Bayesian optimization (BO) offers an active learning framework that selects new settings to test. Sensitivity analysis provides more information by measuring each parameter's influence, with indices like the Hilbert–Schmidt Independence Criterion (HSIC). 

To overcome BO's limitations, this work proposes Bayesian Optimization With Sensitivity Analysis (BOWSA), a method that incorporates sensitivity information to reorder the settings proposed by BO. BOWSA was compared to a standard BO on synthetic functions and chemical reactions problems. It consistently reaches productive settings earlier than BO, and reduces the variability across the different runs. BOWSA has also been applied to an electrochemical reaction to accelerate the identification of a productive reaction setting. This work was done in collaboration with the chemistry lab Chimie Physique et Chimie du Vivant (CPCV) of the ENS ULM of Paris.

The presentation will describe BOWSA, with a particular focus on the scoring function used to evaluate each candidate. We will then present our experimental evaluation and the results obtained.