FusNet: unlocking molecular fusion properties through machine learning

Accurate prediction of molecular fusion properties is critical for energy-efficient material design and sustainable process optimization, yet remains challenging due to data scarcity and complex thermodynamic interdependencies. This work introduces machine learning tools to address these gaps by combining expert-curated molecular descriptors with deep learning. By systematically evaluating statistical machine learning algorithms and attention-based architectures, optimized models are identified: a SMILES-augmented Transformer-Convolutional Neural Network for fusion temperature and a graph attention network for fusion enthalpy. Prediction power is further validated experimentally on four structure diverse compounds (γ-butyrolactone, methyl octanoate, N-phenylbenzenesulfonamide, and triethylene glycol dimethyl ether). Interpretability analyses reveal that these models prioritize key structures in molecules: attention in text-based models focuses on key atoms while that in graph models focuses on key chemical bonds, aligning with empirical thermodynamic evidences. By providing rapid, interpretable fusion property predictions, this framework can support the development of low-energy phase-change materials and sustainable solvent systems, advancing data-driven green chemistry.