Virtual Characterization via Knowledge-Enhanced Representation Learning: from Organic Conjugated Molecules to Devices

Unlocking the potential of organic functional devices requires elucidating structure–property–performance relationships through comprehensive multiscale, multi property characterization. However, this process is limited by high costs and multidisciplinary complexity. This work introduces OCNet, a domain-knowledge-enhanced representation learning network for organic conjugated systems, pre-trained on self-generated ten-million-level structures. OCNet offers a general microscopic representation with strong expressiveness and transferability, enhancing optoelectronic property prediction accuracy by 20% compared to existing models. Leveraging a million-scale, quantum mechanics(QM) calculated transfer integral database, OCNet establishes the first generalized transfer integral model for thin films, further enabling mesoscale carrier mobility predictions with QM accuracy. At the device level, it improves the accuracy of power conversion efficiency predictions by 12% relative to QM descriptor-based models, aligning well with experimental characterization. Overall, OCNet enables accurate and low-cost multiscale, multi-property virtual characterization, offers an innovative solution to accelerate material design and device optimization for organic functional materials, and demonstrates strong potential in applications like display technologies and photovoltaics.