Machine Learning-Driven Optimization of Continuous-Flow Photoredox Amine Synthesis

Photoredox catalysis plays an important role in the synthesis of pharmaceutically relevant compounds such as C(sp³)-rich tertiary amines. The difficulty of identifying underlying mechanistic models for such novel transformations, coupled with the large reaction space of this reaction class, means that developing a robust process is challenging. In this work, we demonstrate the machine learning-driven optimization of a photoredox tertiary amine synthesis with six continuous variables (e.g., concentration, temperature, residence time) and solvent choice as a discrete variable, in a semiautomated continuous flow setup. Starting with a large library of solvents, the workflow included multiple steps of a priori knowledge generation (e.g., solubility predictions) to narrow the discrete space. A novel Bayesian optimization algorithm, nomadic exploratory multiobjective optimization (NEMO), was then deployed to identify and populate the Pareto front for the two reaction objectives─yield and reaction cost. Permutation feature importance and partial dependence plots identified the most important parameters for high yield, sig3, the asymmetry of the s-profile for the discrete space, and equivalences of alkene and Hantzsch ester for the continuous variables. Catalyst loading and residence time were found to be correlated to absorbed photon equivalence, while catalyst loading was additionally the main parameter to drive cost. Even though productivity was not an optimization objective, the best result achieved in flow was ∼25 times higher than reactions in batch, which equals to ∼12 g per day throughput.