A Closed-Loop Deep Generative Model for the Inverse Design of Radiation-Resistant Polymers

The discovery of radiation-resistant polymers is vital for aerospace, medical, and energy applications, where ionizing radiation rapidly degrades conventional materials. Inspired by the impact of Google DeepMind’s AlphaFold in structural biology, this study presents a closed-loop generative AI framework for polymer inverse design. The starting dataset contained only SMILES representations of polymer repeat units. To enable property-targeted generation, 17 RDKit molecular descriptors were computed, while available experimental glass transition temperatures (Tg) and mass attenuation coefficients (MAC) were integrated. Since experimental coverage was sparse, random forest surrogate models were trained to predict Tg and MAC, filling in missing values and producing a fully annotated dataset. These predictors achieved high accuracy (R² > 0.99 for MAC, R² > 0.90 for Tg) and were then used in tandem with a property-conditional Transformer that generates chemically valid SMILES conditioned on target properties. Generated candidates were automatically featurized, evaluated by the surrogate models, and selected through a score–diversity scheme that balances performance with novelty. The resulting closed-loop system enables iterative sampling, prediction, and refinement, accelerating the exploration of chemical space beyond what is feasible through experiment-driven discovery. This framework identifies polymers with Tg values near 215 °C and MAC above 0.0569 cm²/g, consistent with desired radiation shielding performance. By uniting surrogate modeling, molecular descriptors, and generative deep learning, this work establishes a foundation for AI-driven inverse design of high-performance radiation-resistant polymers.