Abstract

Computational methods, from theoretical modeling to numerical simulations, have transformed materials science by providing crucial guidelines for experimental design. The recent integration of deep learning has further revolutionized this field, enabling unprecedented insights into structure-property relationships at the molecular level. However, applying these computational tools to polymer science remains challenging due to complex molecular interactions, vast combinatorial spaces, and multi-scale phenomena. We present HAPPY (Hierarchically Abstracted rePeat unit of PolYmers), a novel string-based representation system that simplifies complex polymer structures by grouping sub-structures and utilizing grammatically complete connectors. When integrated with Recurrent Neural Networks, HAPPY enables accurate property prediction even with limited training data. Our work demonstrates HAPPY's effectiveness not only in property prediction but also in inverse design, where we aim to design polymers with specific target properties.