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.

Graphical abstract

Keywords

RON / graph neural network / multimodal fusion / attention mechanism / latent space mixing rule

References

[1] Ali Khan M Z , Ali Khan H , Ravi S S , Turner J W , Aziz M . Potential of clean liquid fuels in decarbonizing transportation–An overlooked net-zero pathway. Renewable and Sustainable Energy Reviews, 2023, 183: 113483

[2] Ouni M , Mrad S , Mraihi R . Pathways to sustainable transportation in G-20 countries: unveiling the role of green technology, green energy, green finance and digital economy using panel data and machine learning analyses. Research in Transportation Business & Management, 2025, 60: 101355

[3] ASTM D2699-12 . Standard Test Method for Research Octane Number of Spark-Ignition Engine Fuel. West Conshohocken, PA: ASTM International, 2012,

[4] Cai G Q , Liu Z F , Zhang L Z . Transformation rule-based molecular evolution for automatic gasoline molecule design. Chemical Engineering Science, 2022, 263: 118119

[5] Katritzky A R , Kuanar M , Slavov S , Hall C D , Karelson M , Kahn I , Dobchev D A . Quantitative correlation of physical and chemical properties with chemical structure: utility for prediction. Chemical Reviews, 2010, 110(10): 5714–5789

[6] Banchero M , Manna L . Comparison between multi-linear- and radial-basis-function-neural-network-based QSPR models for the prediction of the critical temperature, critical pressure and acentric factor of organic compounds. Molecules, 2018, 23(6): 1379

[7] Albahri T A . Structural group contribution method for predicting the octane number of pure hydrocarbon liquids. Industrial & Engineering Chemistry Research, 2003, 42(3): 657–662

[8] SubLaban A , Kessler T J , Van Dam N , Mack J H . Artificial neural network models for octane number and octane sensitivity: a quantitative structure property relationship approach to fuel design. Journal of Energy Resources Technology, 2023, 145(10): 102302

[9] Wu Y S , Liu Y S , Li X L , Huang Z , Han D . Gasoline octane number prediction from near-infrared spectroscopy with an ANN-based model. Fuel, 2022, 318: 123543

[10] Sun X Y , Zhang F , Liu J P , Duan X B . Prediction of gasoline research octane number using multiple feature machine learning models. Fuel, 2023, 333: 126510

[11] Saghir H , Ahmad I , Kano M , Caliskan H , Hong H . Prediction and optimisation of gasoline quality in petroleum refining: the use of machine learning model as a surrogate in optimisation framework. CAAI Transactions on Intelligence Technology, 2024, 9(5): 1185–1198

[12] Kubic W L Jr , Jenkins R W , Moore C M , Semelsberger T A , Sutton A D . Artificial neural network based group contribution method for estimating cetane and octane numbers of hydrocarbons and oxygenated organic compounds. Industrial & Engineering Chemistry Research, 2017, 56(42): 12236–12245

[13] Hong H X , Xie Q , Ge W G , Qian F , Fang H , Shi L M , Su Z Q , Perkins R , Tong W D . Mold², molecular descriptors from 2D structures for chemoinformatics and toxicoinformatics. Journal of Chemical Information and Modeling, 2008, 48(7): 1337–1344

[14] Ben Hicham K K , Rittig J G , Grohe M , Mitsos A . Tabular foundation models for in-context prediction of molecular properties. , 2026,

[15] Mauri A , Consonni V , Pavan M , Todeschini R . DRAGON software: an easy approach to molecular descriptor calculations. MATCH Communications in Mathematical and in Computer Chemistry, 2006, 56(2): 237–248

[16] Gao P, Su H, Miao C, Tang P, Cao Y. Multi-feature optimization of QSPR model for fuel octane number. In: Proceedings of the Second International Conference on Electrical, Electronics, and Information Engineering (EEIE 2023). Bellingham, WA: SPIE, 2024, 186–193

[17] Liu Z F , Zhang L Z , Elkamel A , Liang D , Zhao S Q , Xu C M , Ivanov S Y , Ray A K . Multiobjective feature selection approach to quantitative structure property relationship models for predicting the octane number of compounds found in gasoline. Energy & Fuels, 2017, 31(6): 5828–5839

[18] Chen X S , Wu R L , Lan Y Y , Ma T , Liu Y . MolEvolve: LLM-guided evolutionary search for interpretable molecular optimization. , 2026,

[19] Schweidtmann A M , Rittig J G , König A , Grohe M , Mitsos A , Dahmen M . Graph neural networks for prediction of fuel ignition quality. Energy & Fuels, 2020, 34(9): 11395–11407

[20] Cui C , He S , lv W J , Zhang L Z , Zhou X . Graph convolution neural network for predicting the octane number of pure hydrocarbons in gasoline. Petroleum Processing and Petrochemicals, 2021, 52(7): 82–87

[21] Bounaceur R , Barthélemy N , Delort N , Herbinet O , Battin-Leclerc F . A multimodal learning model based on a QSPR approach for the estimation of RON, MON and CN, for any C, H, O hydrocarbons. Fuel, 2025, 381: 133438

[22] Sun S H , Wang P , He Y , Yang J , Li S J . Multimodal cross-attention molecular property prediction for text, sequence, graph, and geometry. ACS Omega, 2025, 10(46): 56225–56239

[23] Correa Gonzalez S , Kroyan Y , Sarjovaara T , Kiiski U , Karvo A , Toldy A I , Larmi M , Santasalo-Aarnio A . Prediction of gasoline blend ignition characteristics using machine learning models. Energy & Fuels, 2021, 35(11): 9332–9340

[24] Benavides A , Zapata C , Benjumea P , Franco C A , Cortés F B , Ruiz M A . Predicting octane number of petroleum-derived gasoline fuels from MIR spectra, GC-MS, and routine test data. Processes, 2023, 11(5): 1437

[25] Ghosh P , Hickey K J , Jaffe S B . Development of a detailed gasoline composition-based octane model. Industrial & Engineering Chemistry Research, 2006, 45(1): 337–345

[26] Tipler S , D’Alessio G , Van Haute Q , Parente A , Contino F , Coussement A . Predicting octane numbers relying on principal component analysis and artificial neural network. Computers & Chemical Engineering, 2022, 161: 107784

[27] Kelly J J , Barlow C H , Jinguji T M , Callis J B . Prediction of gasoline octane numbers from near-infrared spectral features in the range 660-1215 nm. Analytical Chemistry, 1989, 61(4): 313–320

[28] Li R Z , Herreros J M , Tsolakis A , Yang W Z . Machine learning regression based group contribution method for cetane and octane numbers prediction of pure fuel compounds and mixtures. Fuel, 2020, 280: 118589

[29] Westbrook C K , Sjöberg M , Cernansky N P . A new chemical kinetic method of determining RON and MON values for single component and multicomponent mixtures of engine fuels. Combustion and Flame, 2018, 195: 50–62

[30] Schlichting S , Methling T , Oßwald P , Zinsmeister J , Riedel U , Köhler M . Numerical prediction of research octane numbers via a quasi-dimensional two-zone cylinder model. Applications in Energy and Combustion Science, 2022, 11: 100079

[31] Kuzhagaliyeva N , Horváth S , Williams J , Nicolle A , Sarathy S M . Artificial intelligence-driven design of fuel mixtures. Communications Chemistry, 2022, 5: 111

[32] Yalamanchi K K , Kommalapati S , Pal P , Kuzhagaliyeva N , AlRamadan A S , Mohan B , Pei Y J , Sarathy S M , Cenker E , Badra J . Uncertainty quantification of a deep learning fuel property prediction model. Applications in Energy and Combustion Science, 2023, 16: 100211

[33] Yalamanchi K K , Pal P , Mohan B , AlRamadan A S , Badra J A , Pei Y J . A variational autoencoder model toward molecular structure representation learning of fuels. Journal of Energy Resources Technology, Part A: Sustainable and Renewable Energy, 2025, 1(5): 052301

[34] Guo Z C , Guo K H , Nan B Z , Tian Y J , Iyer R G , Ma Y H , Wiest O , Zhang X L , Wang W , Zhang C X . et al. Graph-based molecular representation learning. , 2022,

[35] Li L L , Zhang Y P , Wang G H , Xia K L . Kolmogorov–Arnold graph neural networks for molecular property prediction. Nature Machine Intelligence, 2025, 7(8): 1346–1354

[36] Coley C W , Barzilay R , Green W H , Jaakkola T S , Jensen K F . Convolutional embedding of attributed molecular graphs for physical property prediction. Journal of Chemical Information and Modeling, 2017, 57(8): 1757–1772

[37] Sandström H , Rinke P . Similarity-based analysis of atmospheric organic compounds for machine learning applications. Geoscientific Model Development, 2025, 18(9): 2701–2724

[38] Hong H , Wu X , Sun H , Xie C , Wang Q , Li Y . A hierarchical interaction message net for accurate molecular property prediction. Communications Chemistry, 2026, 9: 150

[39] Rijal D , Vasilyev V , Yang Y X , Wang F . Evaluating fuel properties of strained polycycloalkanes for high-performance sustainable aviation fuels. Energies, 2025, 18(19): 5253

[40] Ghironi S , Viganò E L , Selvestrel G , Benfenati E . QSAR models for predicting the antioxidant potential of chemical substances. Journal of Xenobiotics, 2025, 15(3): 80

[41] Homafar A , Nasiri H , Chelgani S C . Modeling coking coal indexes by SHAP-XGBoost: explainable artificial intelligence method. Fuel Communications, 2022, 13: 100078

[42] Hu W H , Liu B W , Gomes J , Zitnik M , Liang P , Pande V , Leskovec J . Strategies for pre-training graph neural networks. , 2019,

[43] Xu K , Hu W H , Leskovec J , Jegelka S . How powerful are graph neural networks. , 2018,

[44] Nizami A R , Ali S F , Afzal M Z , Inc M . Novel descriptors for the prediction of molecular properties. Open Chemistry, 2025, 23: 20250194

[45] AlRamadan A S , Sarathy S M , Khurshid M , Badra J . A blending rule for octane numbers of PRFs and TPRFs with ethanol. Fuel, 2016, 180: 175–186

[46] Foong T M , Morganti K J , Brear M J , da Silva G , Yang Y , Dryer F L . The octane numbers of ethanol blended with gasoline and its surrogates. Fuel, 2014, 115: 727–739

[47] Ershov M A , Grigorieva E V , Abdellatief T M M , Kapustin V M , Ali Abdelkareem M , Kamil M , Olabi A G . Hybrid low-carbon high-octane oxygenated gasoline based on low-octane hydrocarbon fractions. Science of the Total Environment, 2021, 756: 142715

Rights & permissions

Higher Education Press