Jilong Wang, Yaqi Fan, Zheng Wan, Xiangxiang Shen, Xinyue Zhao, Ziwen Niu, Xintong Li, Xianchen Gong, Jingang Jiang, Zhendong Wang, Yi Luo, Yejun Guan, Xian Wei, Hao Xu, Yanhang Ma, Xiao He, Weimin Yang, Peng Wu
Zeolite crystallization is a metastable process under harsh conditions with poorly understood mechanisms, making the directed synthesis of specific frameworks challenging. Organic structure-directing agents (OSDAs) are key to framework control, but their discovery remains dominated by trial-and-error screening. Here, we develop a domain knowledge-informed machine learning model to predict OSDAs, which enables the successful synthesis of three novel zeolites, namely, ECNU-30, ECNU-34, and ECNU-40 (named after East China Normal University), validating the efficacy of the model. Traditional descriptor-based machine learning models exhibit limited predictive performance in screening OSDAs for unknown zeolite frameworks. Combining an end-to-end architecture with active learning, the ECNU-Zeoformer effectively overcomes this limitation, enabling more accurate prediction of OSDA-zeolite binding energies for selecting suitable OSDAs and superior generalizability to different framework topologies.