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

Area of Science:

• Organic Chemistry
• Catalysis
• Chemical Engineering

Background:

• Photoredox catalysis is crucial for synthesizing pharmaceutically relevant C-(sp3)-rich tertiary amines.
• Optimizing these reactions is challenging due to complex mechanistic models and vast reaction spaces.

Purpose of the Study:

• To demonstrate machine learning-driven optimization for photoredox tertiary amine synthesis.
• To identify key reaction parameters and improve process robustness in a continuous flow setup.

Main Methods:

• Utilized a semiautomated continuous flow setup with six continuous and one discrete variable.
• Employed a priori knowledge generation (e.g., solubility predictions) and a Bayesian optimization algorithm (NEMO).
• Analyzed results using permutation feature importance and partial dependence plots.

Main Results:

• Identified critical parameters influencing yield and cost, including catalyst loading, residence time, and solvent choice.
• Discovered correlations between catalyst loading, residence time, and absorbed photon equivalence.
• Achieved a throughput approximately 25 times higher than batch reactions, reaching ~12 g/day.

Conclusions:

• Machine learning, particularly NEMO, effectively optimizes complex photoredox reactions.
• Continuous flow synthesis offers significant productivity gains over batch processes for tertiary amine synthesis.
• The developed workflow accelerates the discovery and optimization of valuable pharmaceutical intermediates.