The research octane number (RON) is a critical measure of a fuel's resistance to engine knocking. Fuels with higher RON allow spark-ignition engines to operate at higher compression ratios, improving thermal efficiency and reducing fuel consumption. However, experimental RON measurements are costly and time-consuming, and conventional linear blending rules—which assume that a mixture's RON is simply the mole-fraction-weighted average of its components—often fail to capture the complex, nonlinear blending behaviour observed in real fuel mixtures containing multiple hydrocarbons and oxygenates.
In a study published in ENG. Chem. Eng., researchers at China University of Petroleum (Beijing) and Shandong Kegu Jiequan Technology Co., Ltd. developed an interpretable multimodal molecular representation framework to address this challenge. The framework integrates three complementary types of molecular information: graph neural network embeddings that capture atom-bond topological relationships, MACCS fingerprints that encode predefined substructure fragments, and molecular descriptors selected through a residual-guided strategy.
For pure-component RON prediction, the trimodal model achieved an R² of 0.9373 and an MAE of 4.04 on the test set. Ablation experiments revealed that the graph topology provided the strongest contribution, while the MACCS fingerprints and descriptors supplied complementary information. The atom-level attention mechanism visualised which local structural environments the model emphasised—the model focused on branching sites in alkanes, substitution sites in cycloalkanes, double-bond regions in olefins, and connection points between side chains and aromatic rings. These attention patterns are consistent with known structure-RON relationships, enhancing the model's interpretability.
The trained pure-component encoder was then transferred to mixture RON prediction. For a mixture containing components, the embedding vectors of each component were weighted according to their mole fractions and combined in the latent space before being fed into an XGBoost regressor. This first-order latent-space mixing model achieved an R² of 0.9736 and an MAE of 1.46 on the mixture test set, significantly outperforming the conventional linear blending baseline (R² = 0.7501, MAE = 4.83). The introduction of second-order interaction terms did not result in significant improvement, suggesting that the first-order model already captured the primary composition-dependent variation in the current latent space.
The model was also applied to fuel formulation design. Using a stochastic sampling search method, the researchers identified feasible ternary blending compositions that met target RON constraints across four case studies. In all cases, the known literature formulations fell within the predicted feasible solution space, confirming that the method can support inverse fuel design.
This work demonstrates that molecular representations learned from pure components can be effectively transferred to mixture property prediction, and that latent-space composition weighting offers a promising strategy for mixture RON prediction and fuel formulation. The framework provides a foundation for future multi-objective optimisation incorporating additional fuel properties such as vapour pressure, density, and viscosity.
DOI
10.1007/s11705-026-2702-7