Abstract: Based on the graph convolution neural network, an improved neural fingerprint method was established by introducing pooling operation. A model for predicting octane number of pure hydrocarbons in gasoline was established by this method as an important part of blending model of gasoline on molecular level. Through the verification on the boiling point and critical temperature data set, the prediction ability of neural fingerprint method was improved obviously by introducing pooling operation. The favorable and unfavorable features could be selected automatically by the improved neural fingerprint method. It was also found that the effect of double bond on motor octane number was greater than that of aromatic ring. The improved neural fingerprint method also had good results in the prediction of both research octane number and motor octane number of pure hydrocarbons in gasoline.

Key words: graph neural network, neural fingerprint, octane number, gasoline