(283c) Bayesian Reaction Optimization of Rac-Lactide Polymerization Catalyzed By Aluminum Complexes

Reaction optimization is experimentally costly due to the tremendous chemical search space spanned by functional materials and condition variables that determine the reaction outcomes. Recent advances in data science provide opportunities for accelerating the process via the identification of promising points in the feature space with only a small number of experiments. However, those studies largely focus on small molecule synthesis, and scarcity remains on polymerization. Herein, Bayesian optimization (BO), an iterative global optimization algorithm, is employed to fill in the gap, in a case study of rac-lactide polymerization catalyzed by aluminum (Al) complexes. BO model performance is tested on around 60 existing experimental data points from the literature within the overall chemical space (576 in size) built upon Al-complex catalysts with different functional groups, targeting towards the highest isotacticity (Pm) and syndiotacticity (Pr) values (defined from 0 to 1), two important quantities that determine polymer physical properties. The model shows a good prediction capability with mean absolute errors around 0.1. The optimization performance comparing BO with the traditional random search indicates a higher search efficiency and stability of BO. By using the model trained on these existing data points, we are currently searching through the chemical space to efficiently identify new optimal Al-complexes with high Pm or Pr values. Synthesizability is considered via incorporating the number of synthesis steps into the searching procedures for efficient synthesis. Two new catalysts with NMR-measured Pm and Pr over 0.9 were suggested by BO and later experimentally validated, indicating the effectiveness of the BO approach in finding polymerization catalysts.