Year: 2026
Type: preprint
Abstract: Foundation models have emerged as a powerful paradigm for learning transferable knowledge from broad, heterogeneous data and adapting them to diverse downstream tasks, driving a growing interest in AI-accelerated scientific discovery. In polymer science, this paradigm offers a promising route beyond conventional machine learning methods that are often confined to isolated tasks and struggle to generalize under data scarcity and across heterogeneous downstream settings. Polymer foundation models (PFMs), however, are not merely general scientific foundation models applied to polymer datasets. They must address representation, generation, transfer, and evaluation challenges specific to polymer informatics. This review synthesizes recent progress in PFMs for representation learning and generative modeling, spanning models designed for polymer systems and those adapted from general foundation models. We organize this review around two stages: (1) polymer modeling as the foundation, covering representations, datasets, model architectures, and pretraining objectives; and (2) downstream applications, evaluation, and adaptation. By synthesizing recent progress, open challenges, and future directions, this review outlines a roadmap toward robust and generalizable foundation models for polymer informatics.
Source: ChemRxiv
Institution: University of Notre Dame
Language: English
Subfield: Materials Chemistry
Field: Materials Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7220681316