Abstract: To solve the time and safety problems faced by the traditional experimental methods for measuring the auto-ignition temperature of binary mixed liquids,a method optimizing the quantitative structureproperty relationship（QSPR） prediction model by using neural networks is proposed.BP neural network（BPNN） and the one-dimensional convolutional neural network（1DCNN） are used to process the mixed molecular descriptor data,respectively.Then,the molecular structure map data is processed using the convolutional neural network（CNN）,and in this way,two prediction models,BPNN+CNN and IDCNN+CNN,are established.After that,the prediction ability,fitting ability and stability of the experimentally designed BPNN+CNN and 1DCNN+CNN models are verified through cross-validation,residual analysis and application domain analysis.Finally,the effects of different optimizers on the model performance are discussed.And the results of two-dimensional and three-dimensional molecular structure diagrams on the model performance are analyzed.The experimental results show that the coefficients of determination of the two models are 0.989 8 and 0.987 1,respectively.The 10-fold cross-validated complex correlation coefficients are 0.961 1 and 0.963 3,respectively.And the cross-validated coefficients are 0.982 6 and 0.992 5,respectively.The results indicate that both models can predict the self-ignition temperature of most binary mixed liquids.The BPNN+CNN model has better fitting ability,and the 1DCNN+CNN model has better stability.