Extensive Evaluation and Advancement of Extrapolation Performance in Polymer Property Prediction

Polymer Informatics (PI), an emerging field which integrates computational science and data science, has gained significant attention as a powerful approach for accurately predicting intricate polymer properties. The prevalent approach in PI heavily depends on structure-based models, specifically deep learning and nonlinear regression models, which excel at capturing complex relationships within polymer molecular structures. However, these structure-based models face challenges when dealing with small data predictions, particularly in PI where the available database is still insufficient. Despite their effectiveness with interpolation of large datasets, the generalization performance of these models can be compromised in limited data [1]. Moreover, the highly expressive nature of these black-box models poses difficulties in interpreting and explaining their predictions, especially when extrapolating beyond the range of available data [2]. It is crucial to explore the potential of PI in identifying highly functional polymers that surpass the limits of existing data.<br/>In this study, we evaluate the extrapolation performance of baseline ML models used in polymer property prediction. Our evaluation employs experimental and computational datasets encompassing a wide range of data scales, including mechanical, thermal, and electrical properties. Additionally, we propose a novel ML model that effectively extrapolates polymer properties by leveraging a framework previously demonstrated in extrapolating the properties of small organic molecules. In contrast to conventional approaches relying on black-box ML techniques, our constructed ML model incorporates microscopic physical features derived from molecular computations. This integration allows for the explicit extraction of correlations between microscopic physical quantities and macroscopic polymer properties, resulting in an explainable ML model with enhanced extrapolation capabilities. Through extensive evaluations of the extrapolation performance, we demonstrate the limitations of conventional models in extrapolation tasks while highlighting the effectiveness of our proposed approach.<br/>[1] Xu, P., Ji, X., Li, M. et al. Small data machine learning in materials science. npj Comput Mater 9, 42 (2023). https://doi.org/10.1038/s41524-023-01000-z<br/>[2] Li, K., DeCost, B., Choudhary, K. et al. A critical examination of robustness and generalizability of machine learning prediction of materials properties. npj Comput Mater 9, 55 (2023). https://doi.org/10.1038/s41524-023-01012-9