De Novo Design of Polymers with Specified Properties Using Reinforcement Learning

Designing polymers with specified properties is crucial
for industries
such as aerospace, automotive, and construction, where high yield
strength is necessary for stability and performance. Traditionally,
the design of polymers has relied on decades of trial-and-error experiments
that are time-consuming and inefficient. While recent advancements
in computational methods have emerged as promising tools for polymer
design, they predominantly focus on property prediction or unbiased
polymer generation and do not fully progress toward the tailored design
of novel polymer structures that meet specific performance criteria.
To accelerate the exploration of new high-performance polymers, we
proposed Reinforcement Learning for Polymer Generation (RLPolyG),
an integrated goal-oriented exploration workflow for de novo polymer
design with specified properties. This framework employs a forward
model for predicting polymer properties and an inverse model optimized
via reinforcement learning to generate polymers with specific yield
strength. Our forward model achieved an R2 of 0.84 in predicting yield strength, enabling the inverse model
to generate 4991 novel polymer candidates, resulting in a significant
45.20% improvement in average yield strength. We further screened
these candidates based on synthetic accessibility (SA) scores and
degradability, identifying 3099 polymers with excellent feasibility
for synthesis and degradation performance. Finally, we validated the
nine top-performing polymers through molecular dynamics (MD) simulations,
which showed an average related error of 14.64% between the predicted
and MD-validated values. This work demonstrates the potential of using
reinforcement learning to transform polymer design, providing a systematic
and efficient pathway to explore the vast polymer space and accelerate
the discovery of materials tailored to meet specific industrial needs.