Tuning the molecular weight distribution from atom transfer radical polymerization using deep reinforcement learning

Abstract

We devise a novel technique to control the shape of polymer molecular weight distributions (MWDs) in atom transfer radical polymerization (ATRP). This technique makes use of recent advances in both simulation-based, modelfree reinforcement learning (RL) and the numerical simulation of ATRP. A simulation of ATRP is built that allows an RL controller to add chemical reagents throughout the course of the reaction. The RL controller incorporates fully-connected and convolutional neural network architectures and bases its decision upon the current status of the ATRP reaction. The initial, untrained, controller leads to ending MWDs with large variability, allowing the RL algorithm to explore a large search space. When trained using an actorcritic algorithm, the RL controller is able to discover and optimize control policies that lead to a variety of target MWDs. The target MWDs include Gaussians of various width, and more diverse shapes such as bimodal distributions. The learned control policies are robust and transfer to similar but not identical ATRP reaction settings, even under the presence of simulated noise. We believe this work is a proof-of-concept for employing modern artificial intelligence techniques in the synthesis of new functional polymer materials.

Deep learning for molecular design—a review of the state of the art

“…While much of the work so far has focused on deep generative modeling for drug molecules, 24 there are many other application domains which are benefiting from the application of deep learning to lead generation and screening, such as organic light emitting diodes, 25 organic solar cells, 26 energetic materials, 10,27 electrochromic devices, 28 polymers, 29 polypeptides, [30][31][32] and metal organic frameworks. 33,34 Our review touches on four major issues we have observed in the field.…”

Deep Learning for Deep Chemistry: Optimizing the Prediction of Chemical Patterns

“…In Chemistry domains, researchers have had access to multidimensional data of unprecedented scale and accuracy, that characterize the systems/processes to be modeled. A collection of different examples of optimization based on ML approaches can be found in Kowalik et al (2012) Specifically, ML contributions have involved a variety of systems including drugs (Griffen et al, 2018), polymers (Li et al, 2018a), polypeptides (Grisoni et al, 2018;Müller et al, 2018), energetic materials (Elton et al, 2018), metal organic frameworks (He et al, 2018;Jørgensen et al, 2018a;Shen et al, 2018), and organic solar cells (Jørgensen et al, 2018a).…”

Molecular Weight Distribution in Ring‐Opening Polymerization of Propylene Oxide Catalyzed by Double Metal Complex: A Model Simulation

“…[1][2][3][4][5][6][7][8][9][10] MWD of produced polymer usually depends on the polymerization mechanism and kinetics. [11][12][13][14][15][16][17][18][19][20] In general, living anionic polymerization produces chains with narrow MWD, while traditional free radical polymerization yields broadly distributed chains. [20] To better understand the connection between kinetics and MWD as well as to clarify which theory is correct or more proper, many efforts have been made to simulate MWD.…”

“…[11,12,16,[18][19][20][21][22][23][24][25][26][27][28][29][30][31][32][33] Three methods, i.e., moments, Monte Carlo, and solving ordinary differential equations directly, have been developed. [14,16] Among them, the most used is the method of moments because it is not only based on clear reaction mechanism and kinetic equation, but also ratio (RR) of chain propagation to chain transfer can give an impact on the degree of uniformity of chain length.…”

“…[ 11,12,16,18–33 ] Three methods, i.e., moments, Monte Carlo, and solving ordinary differential equations directly, have been developed. [ 14,16 ] Among them, the most used is the method of moments because it is not only based on clear reaction mechanism and kinetic equation, but also easy to calculate. Even though this method cannot show out the full MWD curves, it gathers the most important information such as monomer conversion ( x ), molecular weight (MW), and polydispersity index (PDI).…”