Multimodal molecular representation with latent space mixing for RON prediction of pure and mixed fuels

Abstract

Accurate prediction of the research octane number (RON) is essential for clean fuel design, engine optimization, and gasoline blending, yet experimental measurements remain expensive, and mixture blending behavior is difficult to describe using conventional linear rules. To address this challenge, we developed an interpretable multimodal molecular representation framework based on latent space mixing. Because the reliability of mixture prediction depends on the quality of the pure-component representations, a multimodal encoder was built for pure components by integrating graph neural network embeddings, molecular access system (MACCS) fingerprints, and molecular descriptors selected through a residual-guided strategy. The trained pure-component encoder was then transferred to mixture prediction. For the pure-component test set, the model achieved an R² value of 0.9373 and a mean absolute error (MAE) of 4.04. For the mixtures, the latent space model achieved an R² of 0.9736 and an MAE of 1.46. The ablation results showed that the graph topology provided the strongest contribution, whereas the MACCS fingerprints and descriptors supplied complementary information. Atom-level attention analysis identified a graph-branch emphasis on the local structural environments associated with the RON. Finally, the mixture model was applied to fuel formulation design, where inverse searching under the target RON constraints generated a feasible candidate formulation. These results demonstrated that composition-weighted molecular embedding is a promising strategy for mixture RON prediction and fuel formulation.