Physics-Guided Neural Networks for Transferable Property Prediction in Architecturally Diverse Copolymers

The architectural, compositional, and chemical complexities of polymers are fundamentally important to their properties; however, these same factors obfuscate effective predictions. Machine learning offers a promising approach for predicting polymer properties, but model transferability remains a major challenge, particularly when data is scarce due to high acquisition costs or the growth of the parameter space. Here, we examine whether integration with polymer physics theory effectively enhances the transferability of machine learning models to predict properties of architecturally and compositionally diverse polymers. To do so, we first generate ToPoRg-18k--a dataset reporting the moments of the distribution of squared radius of gyration for 18,450 polymers with diverse architectures, molecular weights, compositions, and chemical patterns. We then systematically assess the performance of several different models on a series of transferability tasks, such as predicting properties of high molecular weight systems from smaller ones or predicting properties of copolymers from homopolymers. We find that a tandem model, GC-GNN, which combines a graph neural network with a fittable model based on ideal Gaussian chain theory, surpasses both standalone polymer-physics and graph neural network models in predictive accuracy and transferability. We also demonstrate that predictive transferability varies with polymer architecture due to deviations from the ideal Gaussian chain assumption. Furthermore, the integration with theory endows GC-GNN with additional interpretability, as its learned coefficients correlate strongly with polymer solvophobicity. Overall, this study illustrates the utility of combining polymer physics with data-driven models to improve predictive transferability for architecturally diverse copolymers, showcasing an extension of physics-informed machine learning for macromolecules.