Machine learning-quantitative structure property relationship (ML-QSPR) method for fuel physicochemical properties prediction of multiple fuel types

The workflow of virtual fuel screening by ML-QSPR and chemical kinetics. [Display omitted] •ML-QSPR method enables to predict 15 fuel properties of 23 fuel types.•QSPR-UOB 3.0 system extracts and digitalizes fuel molecular structure features.•ML algorithms describe the dependence of fuel properties on chemical structure.•UOB Fuel Property Database provides a comprehensive dataset for model training.•ML-QSPR identifies molecules with desired properties to enable efficient & clean combustion. A machine learning-quantitative structure property relationship (ML-QSPR) method is proposed to predict 15 fuel physicochemical properties of 23 fuel types. QSPR-UOB 3.0 functional group classification system is developed to extract and digitalize the molecular structure feature. ML algorithms are used to map the molecular structure feature and fuel properties as well as model parameter tuning. UOB Fuel Property Database (1797 pure compounds and 465 mixtures) is established to provide massive properties data for model training. Cross-validation is chosen to examine predictive precision, avoid overfitting and estimate inter/extrapolation capacity. ML-QSPR method has 4 distinct advantages compared to published statistical methods: (1) It applies to 15 properties of CN, RON, MON, Tm, Tb, ΔHvap, surface tension γ, LHV, liquid density ρ, YSI, IT, FP, VP, LFL, UFL. (2) It applies to 23 fuel types of alkanes, cycloalkanes, alkenes, cyclic alkenes, alkadienes, alkynes, alcohols, cycloalcohols, aldehydes, ketones, cyclic ketone, saturated esters, unsaturated esters, acyclic ethers, furans, other cyclic ethers, aromatics, carbonate ester, carboxylic anhydride, peroxide, hydroperoxide, polyfunctionals, carboxylic acids. (3) High predictive accuracy is achieved and the average R2 of 15 fuel properties reaches 0.9816. (4) The regression models display reasonable interpolation and extrapolation capacity to test new molecules. The success is attributed to 2 key factors: (1) QSPR-UOB 3.0 system accounts for the contribution of structural features, functional group interaction and fuel reactivity. (2) ML algorithms accurately capture the dependence of fuel properties on chemical structures.