Accurate and expandable graph deep learning for accelerating predictive identification of high-potential organic photovoltaics

To overcome the prohibitive synthetic costs and the vastness of unexplored chemical space in developing high-performance organic photovoltaics (OPVs), a research team (Prof. Hui-Hsu Tsai) from National Central University, Chemistry has developed a robust, highly accurate graph deep learning framework. The team utilized a physics-informed directed message passing neural network (D-MPNN). By representing molecules as fragment-level graphs enriched with theoretically computable descriptors—such as DFT-derived energies and structural parameters—the model achieves exceptional predictive accuracy without relying on experimental data. Leveraging this expandable framework, the researchers conducted a massive high-throughput virtual screening (HTVS) of approximately 334 million ternary OPV combinations. Remarkably, they successfully identified 4,904 novel candidate systems with predicted power conversion efficiencies (PCEs) surpassing the current experimental dataset ceiling of 19.36%. Furthermore, their statistical analyses unveiled actionable design rules, highlighting the advantages of "No-pi-spacer" architectures and specific non-fused rigid cores. This innovative computational approach provides a powerful platform for accelerating the predictive identification and de novo design of next-generation clean energy materials. It also proved that AI Framework Accelerates the Discovery of High-Efficiency Solar Cells. The results were recently published in the prestigious Journal: Chemical Engineering Journal, 2026.

圖/化學系提供