Data Science-Assisted Workflow for Reaction Optimization in Process Chemistry

This is the authors' abstract. We don't add key points for this paper.

We present a newly developed, data-assisted workflow at Bayer that integrates Bayesian optimization (BO) with CIME4R, an open-source data visualization tool with explainable AI features, to facilitate chemical reaction optimization. The workflow leverages the efficiency of BO for navigating high-dimensional reaction spaces while using CIME4R to visualize the algorithm’s decision-making process, exploration of the reaction space, and the influence and interactions of individual variables. These visualizations aid the interpretation of complex data sets and provide a platform for scientists to efficiently develop a deeper understanding of machine-learning-guided optimization campaigns, thereby improving accessibility and user trust, as well as decision-making efficiency. We demonstrate the workflow in a case study involving a new class of oxime amide ligands evaluated in Ullmann-type C–N cross-coupling reactions. This data-science-driven approach enabled the rapid identification of high-yielding conditions by sampling only 0.5% of the full parameter space within 4 days of experimental work. Feature-importance analysis revealed the solvent, propylene glycol methyl ether, as the most influential parameter, followed by K3PO4 as the preferred base.