Artificial intelligence-driven materials development has emerged as a powerful alternative to traditional trial-and-error methods. Despite its promise, these methods often struggle to uncover novel materials or generate actionable insights in emerging fields due to limited data availability. This challenge is particularly pronounced in electronic materials, where the intricate interplay of physical mechanisms and structure-property relationships impede progress. Here, we report a methodology combining Physical-Knowledge-Undergirded Transfer Learning for accurate property prediction with limited data, coupled with physical knowledge and artificial intelligence-driven hypothesis generation to yield scientific insights. Using this approach, we successfully identify low-voltage, high-performance organic electrochemical transistor materials and yield material design knowledge. The approach is experimentally validated through the synthesis of n-type polymers, demonstrating accurate property prediction and revealing critical structure-property relationships. We believe this approach can be applied to other emerging material systems with limited data availability and complex physical mechanisms, and accelerates the development of materials in emerging fields.
Tian et al. (Tue,) studied this question.