Finding Harmony in Chemical Data: Hierarchical and Balanced multimodal Fusion for Reaction Yield Prediction

YuHan Li, Zekun Li, Jian Zhang, Yang Gao, Jing Ma, xin chen, Yinghuan Shi

Keywords: AI for Chemistry, Chemical Reaction Yield Prediction, Hierarchical Multi-Modal Fusion, Balanced Modality Contribution

Abstract: Multimodal yield prediction aims to integrate heterogeneous molecular descriptors across distinct data modalities to predict the conversion efficiency of chemical reactions. However, existing approaches often face limitations in effectively utilizing multimodal information, primarily due to inadequate consideration of both hierarchical relationships and imbalanced contributions across modalities during the fusion process. To address these challenges, we propose a **H**ier**ar**chical and balanced **m**ulti-m**o**dal fusion framework for reactio**n** **y**ield prediction, termed **Harmony**. Specifically, to enhance multimodal information utilization, we design a hierarchical fusion architecture comprising three modality encoders and two feature fusion modules for different levels of granularity. Furthermore, we introduce a novel contribution assessment mechanism that quantitatively evaluates modality-specific impacts, coupled with a prefer-balancing optimization objective. Extensive experimental evaluations demonstrate that Harmony not only consistently outperforms existing methods but also exhibits robust out-of-sample (OOS) generalization. Specifically, it achieves a **22\% improvement** in the $R^2$ metric over the strongest baseline on the most challenging Amide Coupling Reaction dataset. Our code can be found at https://anonymous.4open.science/r/F6BB.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 15482