Advancing organic photovoltaic materials by machine learning-driven design with polymer-unit fingerprints

Abstract

To enhance the power conversion efficiency (PCE) of organic photovoltaic (OPV) cells, the identification of high-performance polymer/macromolecule materials and understanding their relationship with photovoltaic performance before synthesis are critical objectives. In this study, we developed five algorithms using a dataset of 1343 experimentally validated OPV NFA acceptor materials. The random forest (RF) algorithm exhibited the best predictive performance for material design and screening. Additionally, we explored a newly developed polymer/macromolecule structure expression, polymer-unit fingerprint (PUFp), which outperformed the molecular access system (MACCS) across diverse machine learning (ML) algorithms. PUFp facilitated the interpretability of structure-property relationships, enabling PCE predictions of conjugated polymers/macromolecules formed by the combination of donor (D) and acceptor (A) units. Our PUFp-ML model efficiently pre-evaluated and classified numerous acceptor materials, identifying and screening the two most promising NFA candidates. The proposed framework demonstrates the ability to design novel materials based on PUFp-ML-established feature/substructure-property relationships, providing rational design guidelines for developing high-performance OPV acceptors. These methodologies are transferable to donor materials, thereby supporting accelerated material discovery and offering insights for designing innovative OPV materials.