CRCU

Chemical Research in Chinese Universities ›› 2026, Vol. 42 ›› Issue (5): 1604-1614.doi: 10.1007/s40242-026-6126-9

• Research Articles • Previous Articles    

Multi-Scale Memory-Enhanced Transformer for Optical Chemical Structure Recognition

WANG Rui1, JI Yujin2, LI Youyong1,2, LEE Shuit-Tong1,2   

  1. 1. Macao Institute of Materials Science and Engineering, Faculty of Innovation Engineering, Macau University of Science and Technology, Taipa, Macau SAR 999078, P. R. China;
    2. State Key Laboratory of Bioinspired Interfacial Materials Science, Institute of Functional Nano and Soft Materials, Soochow University, Suzhou 215123, P. R. China
  • Received:2026-06-04 Accepted:2026-08-01 Online:2026-10-01 Published:2026-09-24
  • Supported by:
    This work was supported by the National Key Research Program of China (Nos. 2025YFE0121800, 2022YFA1503101), the National Natural Science Foundation of China (No. 92580104), the Suzhou Science and Technology Development Project—Basic Research in Medical Applications, China (No. SKY2023040), the Science and Technology Development Fund, Macau SAR (FDCT No. 0087/2026/AGJ), the Project of the Collaborative Innovation Center of Suzhou Nano Science & Technology, the Priority Academic Program Development of Jiangsu Higher Education Institutions (PAPD), the ‘111’ Project, and the Fund of the Joint International Research Laboratory of Carbon-Based Functional Materials and Devices.

Abstract: Optical Chemical Structure Recognition (OCSR) is a key technology at the intersection of chemistry and computer science, aiming to convert molecular structure images into machine-readable sequence representations. However, existing OCSR models still fall short in building long-range dependencies within sequences, making it difficult to accurately generate long sequences with complex structures. To address this, we propose a Multi-Scale Memory-Enhanced Transformer (MSMET) model and design two modules: the Layer-wise Hierarchical Fusion (LHF) module and the Cross-Attention Memory Augmentation (CAMA) module. The LHF module employs adaptive feature recalibration across encoder blocks through learnable gating, effectively alleviating feature forgetting phenomena in deep architectures while maintaining multi-scale visual discriminability. The CAMA module equips the decoder with a learnable memory matrix that dynamically enriches cross-modal attention, significantly enhancing the model's capacity to capture long-distance topological interactions in molecular graph structures. We evaluated the performance of MSMET, and the experimental results demonstrate that the model achieved strong overall state-of-the-art performance.

Key words: Optical chemical structure recognition, Transformer, Attention, Deep learning