Bayesian Optimization for Resource-Efficient Hydroformylation

Hydroformylation remains a cornerstone transformation in industrial catalysis, especially for the production of fine chemicals, fragrances, and pharmaceuticals. However, achieving high linear selectivity in the hydroformylation of vinyl arenes remains challenging due to intrinsic substrate preferences and the high cost of rhodium-based catalysts. In this work, we employed Bayesian optimization (BO) to explore a seven-dimensional (7D) space of reaction conditions, spanning catalyst type and loading, ligand identity and loading, solvent, temperature, and pressure for a total of ∼2.9 billion possible combinations, to enhance selectivity toward the linear aldehyde at reduced catalyst usage and reaction time within just 88 experiments. Despite the vast search space, our optimization campaign using the Atinary FalconGPBO algorithm was able to maintain conversion and linear selectivity at a halved reaction time and with a 10–30 times reduction in rhodium catalyst loading (from 0.1 to <0.01 mol %) compared to high-loading prescreening conditions. Specifically, for two promising ligands, we were able to reduce the rhodium cost contribution by 95% (from 102 to 5 €/kg of product) and 97% (from 127 to 4 €/kg of product). This work showcases the potential of using machine learning (ML) tools such as BO to accelerate reaction optimization, improve process efficiency, and reduce resource consumption in fine chemical production.