PREDICTION OF OCTANE NUMBERS OF ORGANIC COMPOUNDS USING MACHINE LEARNING METHODS

Currently, petrochemical processes are aimed at producing gasoline fractions with a high octane number. The octane number is an indicator that characterizes the resistance of fuel to detonation (explosion) in internal combustion engines. The higher this indicator is the more energy it is able to produce. Isooctane (2,2,4-trimethylpentane) was used as a reference. Sometimes a mixture of it with normal heptane is used. Anti-knock additives are usually introduced to increase the octane numbers of gasoline fractions. Organic compounds of various structures are used as anti-knock additives. These are usually isoalkanes, olefins, cycloalkanes, alcohols, etc. with high octane numbers. Numerous works have been devoted to the peculiarities of the effect of anti-knock additives on the octane numbers of fuels [1-8]. These studies show the possibility of directional control of the octane numbers of gasoline fractions by varying the nature of modifying additives and their amount. An additive increase in the octane number of gasoline fractions is usually observed in proportion to the content of the anti-knock additive. However, there may be exceptions that are established empirically. The most promising is the establishment of patterns of connection between the chemical structure of hydrocarbons and the detonation of fuels. In this approach, organic compounds are represented as a set of descriptors characterizing the features of the chemical structure of molecules. Flash point data for organic compounds were taken from the PubChem database. Information on the flash point for 219 organic substances was included in the database for this work. To simplify the analysis of the representation of organic compounds, we used MACC keys, as they are among the best descriptors for predicting the properties of chemical compounds. These descriptors are created based on the shared keys of the substructure. In addition, the models were calculated using Morgan molecular prints, also known as circular prints with a radius of 2. Within the framework of this work, ridge regression, the random forest algorithm, the kNN nearest neighbor method, the support vector machine (SVM) method, and artificial neural networks were implemented. For the training sample, the resulting classification model of a random forest showed an error-free classification, the prediction error for it is 0. The statistical characteristics of the constructed XGBoost model for the sample have the following values: R2 =0.83 and RMSE=9.96 °С.