Machine-Learning Prediction of Carrier Mobility and Polarity Type of Polymer Semiconductors: from High-Fidelity Data to Experimental Realization

Abstract Machine learning has rapidly advanced the discovery of functional materials. However, its application to polymer-based organic field-effect transistors (OFETs) remains limited by the scarcity of high-fidelity databases and the complex structure-process-property couplings. Here, we introduce an open-access integrated platform for polymer semiconductors in OFETs, termed OFETnow (https://ofetnow.top). This platform combines a high-quality database with automatically extracted and manually verified data from reported works, together with an efficient fingerprint-based machine-learning model. Importantly, this framework requires no quantum-chemically derived input descriptors while implicitly encoding key processing variables that are often inaccessible to a unified database. The practical utility of OFETnow is validated through a closed-loop experimental workflow, where recently reported polymers excluded from the database are first used for an algorithmically retrieved external validation. Building on this, interpretability analysis guided the selection of two representative building blocks, resulting in five designed polymers through a two-stage prediction protocol that employs standardized parameters for initial screening and experimentally determined values for the refined prediction. The experimentally measured device performances are in close agreement with model predictions, enabling reliable discrimination between high-mobility candidates and an intentionally selected low-mobility control as well as accurate identification of carrier transport types. This work establishes a data-driven and experimentally grounded framework for predicting mobility levels and carrier transport types in polymer OFETs, directly bridging machine learning screening with device-level validation.