Dual-Branch Deep Learning with Multimodal-Multiscale Descriptors for Surface Tension Prediction of Binary Liquid Mixtures

Abstract Surface tension originates from collective intermolecular interactions at fluid interfaces and represents a cross-scale mapping from molecular structure to macroscopic interfacial properties. To capture hierarchical interactions beyond single-scale descriptors, we develop a deep learning framework integrating multimodal molecular representations, including semantic embeddings, topological features, and physicochemical descriptors, together with mixture composition and thermodynamic variables. Within this framework, multihead self-attention and gated fusion integrate multimodal-multiscale descriptors and capture nonlinear relationships with mixture surface tension. Trained on 40,834 data points spanning 1583 unique binary pairs with diverse chemical structures and polarity regimes, the model achieves high predictive accuracy (R2 = 0.9798, MAE = 1.060 mN/m, RMSE = 2.635 mN/m) and demonstrates robust predictive performance. This framework provides a scalable approach based on multimodal molecular representations for predicting the surface tension of binary mixtures and linking molecular representations to macroscopic interfacial behavior.