A big database containing molecular structures and multiple properties of 2899 hydrocarbon molecules （the number of carbon atom is from 1 to 50）， was constructed via data collection， structure optimization and quantum chemistry calculation. Seven properties were focused， including melting point （T_m）， boiling point （T_b）， density （ρ）， internal energy at 0 K （U₀）， internal energy at 298.15 K （U）， enthalpy at 298.15 K （H） and Gibbs free energy at 298.15 K （G）. Four different machine learning models were established， including Decision Tree Regressor， Lasso CV， Ridge CV and XGBoost Regressor， using coulomb matrix representing molecular structures as the input. In comparison， the XGBoost Regressor model is more suitable for regressing experimental melting point， boiling point and density of hydrocarbon molecules； Ridge CV model is more suitable for the prediction of four thermodynamic energy properties. In addition， the optimized machine learning combined model can accurately predict the properties of the hydrocarbon molecules with same carbon numbers， hydrocarbons with different types and hydrocarbon isomers. Furthermore， the densities of 34 high-density hydrocarbon fuels reported experimentally were calculated by the optimized machine learning model. The mean absolute error between the calculated values and the experimental values is only 0.0290 g cm^-3. Next， the fuel properties of 319，893 hydrocarbon molecules in GDB-13C were predicted by the machine learning model to establish a big database containing hydrocarbon structure and fuel properties. Based on high-throughput screening， 37 hydrocarbon fuel molecules with low freezing point and high density have been discovered. Through the proof-of-concept via group contribution method and DFT method， the net heat of combustion and specific impulse of the as-screened new molecules are similar to those of JP-10 and quadricyclane （QC）.