The Global Polymer Summit

Design of Elastomer Monomers with Generative Machine Learning

Speaker:
Croix Lancosay, Senior Scientist I, Schrödinger

Abstract:
The next generation of elastomeric materials for wire coatings and seals requires components that perform reliably under extreme conditions, placing stringent and often competing specifications on the underlying polymer chemistry. The simultaneous optimization of multiple properties remains challenging, as structural motifs that improve one property may be detrimental to another, and the vast polymer chemical space makes exhaustive experimental screening impractical. For this reason, we employ a computational workflow that combines generative AI, machine learning, and physics-based simulation to identify and validate materials that show promise as elastomeric insulators.

In this work, we apply REINVENT, a recurrent neural network-based generative framework with reinforcement learning optimization, to the inverse design of elastomer monomers targeting simultaneously two desired properties. A prior is pre-trained on a polymer database to learn the grammar of polymer chemistry, then focused toward an elastomeric chemical space. Reinforcement learning is used to steer generation toward candidates satisfying multi-objective property criteria, using Schrödinger’s built-in machine learning models as scoring functions. We evaluate generated candidates for chemical validity, diversity, and assess the most promising structures through computational property validation. This study demonstrates how an established generative AI framework can be used for goal-directed polymer discovery, and highlights opportunities for accelerating the design of specialty polymers for sealing and insulation applications.