Polymer materials, as an important branch of materials science, are widely applied in some fields such as energy, electronics, and aerospace due to their adjustable structures, diverse properties, and low cost. Although polymer functional materials permeate into numerous technical fields, their multi-scale complex behaviors still pose significant challenges for precise design. In this context, the data-driven research paradigm advocated by the Materials Genome Initiative emerges, marking the entry of materials research into the "fourth paradigm" characterized by data-intensive approach. Machine learning as an important branch of artificial intelligence demonstrates outstanding performance prediction and design capabilities in various systems such as alloys, polymer dielectrics, and photovoltaic materials. Although machine learning makes progress in predicting and designing polymer properties, it still faces several challenges, i.e., the scale and quality of data are limited by the costs of synthesis and characterization; the representation of polymer molecules and feature engineering still require the development of more effective encoding methods; and the experimental-computational loop has not yet been fully connected.

This review first, in conjunction with the changes in the research paradigm, systematically explores the progress of AI in the field of organic polymer functional materials. The fourth paradigm centered on data significantly shortens the material research and development cycle and accelerates the screening and design of new materials. Although several review articles summarize the relevant progress of machine learning in the field of polymer material research and design from different perspectives, there is still insufficient discussion on key aspects such as the construction of polymer-specific databases, data-driven characterization of functional units, and model interpretability. Establishing a new research framework centered on artificial intelligence is expected to achieve groundbreaking progress in exploring the vast feature and property space of polymers.

This review represents the general process of applying artificial intelligence to polymer materials, mainly consisting of four stages, i.e., data generation, feature engineering, model selection, and model interpretation. Data quality and scale remain the weakest links in the existing data-driven approaches. The OPECM database, which contains a total of 3225 polymers, is constructed due to the scarcity of available datasets for organic polymer functional materials (such as semiconductors, photovoltaics, and dielectric materials), and the limited chemical diversity of these materials. To address the challenges faced by the polymer representation methods in quantitative structure-property relationship (QSPR) modeling, the "Polymer-Unit" concept is proposed and two novel structure-encoding schemes (i.e., the Polymer Unit Fingerprint (PUFp) and the Polymer-unit Graph) are developed.

Through the lens of database construction, feature representation, and model interpretability, this review highlights machine learning applications in functional polymer design, with some case studies covering semiconductor, photovoltaic, and dielectric materials. PUFp-based framework enables pre-screening and classification of materials via machine learning, offering an efficient alternative route to discovering high-mobility organic semiconductors. The Polymer-unit Graph representation, implemented with graph neural networks (GNNs), provides a visually interpretable and topology-aware approach to polymer modeling.