(169de) Using Reinforcement Learning to Design Polymers with Specified Properties

Designing polymers with specified properties is crucial in various fields, for example, in the aerospace field, polymers with high tensile stress strength at yield are required to maintain device stability. Currently, the design of polymers with specified properties relies mainly on the experience of chemists and wet experiments, which is often a time-consuming and labor-intensive process. Therefore, a computational method is needed to improve efficiency.

In this work, we constructed a model RLPolyG (Reinforcement Learning for Polymer Generation) for de novo design of polymers with specified properties, which consists of a polymer property prediction model and a polymer generation model combined with a reinforcement learning model. We demonstrated the workflow of RLPolyG by designing polymers with high tensile stress strength at yield. For the polymer property prediction model, we collected a total of 267 homo-polymers and their corresponding tensile stress strength at yield from the PoLyInfo database^[1]. The Morgan fingerprint with frequency of polymer monomers was used to represent the polymer and served as input to the random forest model to predict its tensile stress strength at yield^[2]. The R² of the model on the train and test dataset were 0.92 and 0.94, respectively, demonstrating the accuracy of the model. For the polymer generation model, we constructed an autoregressive language model based on the long short-term memory(LSTM) network. The model iteratively predicted the next character in the PSMILES string based on the previous k characters to generate PSMILES strings with the correct structure. Finally, we jointly trained the polymer generation model and the polymer property prediction model with reinforcement learning to generate polymers with specified properties^[3]. We compared the performance of polymers generated by RLPolyG and PolyG(without reinforcement learning) models and found that the mean tensile stress strength at yield of polymers generated by RLPolyG and PolyG model were 0.05783Gpa and 0.04888Gpa, respectively. The mean tensile stress strength at yield of polymers generated by the RLPolyG model was increased by 18.31%. The computational results demonstrated that the RLPolyG model can efficiently and accurately design and optimize novel polymers with specified properties.

Reference:

[1] Otsuka S, Kuwajima I, Hosoya J, et al. PoLyInfo: Polymer database for polymeric materials design[C]//2011 International Conference on Emerging Intelligent Data and Web Technologies. IEEE, 2011: 22-29.

[2] Yue T, He J, Tao L, et al. High-throughput screening and prediction of high modulus of resilience polymers using explainable machine learning[J]. Journal of Chemical Theory and Computation, 2023, 19(14): 4641-4653.

[3] Popova M, Isayev O, Tropsha A. Deep reinforcement learning for de novo drug design[J]. Science advances, 2018, 4(7): eaap7885.