Active Learning as a Tool for Optimizing “Plug‐n‐Play” Electrochemical Atom Transfer Radical Polymerization
Boyu Zhao, Jiahao Cheng, Junlong Gao, David M. Haddleton, Paul Wilson
Abstract
Abstract A recently reported “plug‐n‐play” approach to simplified electrochemical atom transfer radical polymerization (seATRP) is investigated using machine learning. It is shown that Bayesian optimization via an active learning (AL) algorithm accelerates optimization of the polymerization of oligo(ethylene glycol methyl ether acrylate) 480 (OEGA 480 ) in water. Molecular weight distribution ( M w / M n ; dispersity; Ɖ m ) is the output selected for optimization targeting poly(oligo[ethylene glycol methyl ether acrylate]) (POEGA 480 ) with low dispersity ( Ɖ m < 1.30). Input variables included applied potential ( E app ), [M] and [M]/[I], which led to a potential space of 275 possible reaction conditions. From a training data set of seven reactions, selected to yield uncontrolled POEGA 480 with higher dispersities ( Ɖ m > 1.5), ten iteration loops are performed. During each iteration the algorithm suggests the next reaction conditions. The reactions are then performed and the conversion, number average molecular weight ( M n ) and Ɖ m values are recorded and the Ɖ m values fed back into the algorithm. Overall, 80% of the experiments yield POEGA with Ɖ m < 1.30. Conversely, only 30% of experiments performed using reaction conditions selected at random from the possible reaction space yield POEGA with Ɖ m < 1.30. This study suggests that adopting AL methods can improve the efficiency of optimizing a given seATRP reaction.