Ensemble learning directed classification and regression of hydrocarbon fuels

Abstract

Predicting the properties of fuels based on their structures is essential to develop next generation fuels for various applications, e.g., aircraft, rocket, and missile. In this study, an ensemble learning assisted method has been developed for accurate and quick prediction of fuel properties based on their structures. Two descriptors of hydrocarbon molecules, continuous operable molecular entry specification (COMES) and Coulomb matrix (CM), are compared as the input of the learning model. The optimized stacking of various base learners, including Boosting, Bagging, NeuralNetwork, Trees, Voting, Linear, Neighbors and Kernel, have been achieved for efficient screening of potential high energy density fuels (HEDFs) and accurate prediction of fuel properties. Hydrocarbon molecules are efficiently classified into HEDFs and non-HEDFs based on their structures by the stacking model of ensemble learning, which exhibits superior generalization performance to the single learner. The properties of these potential HEDFs are accurately predicted by another stacking model, including density, heat value, freezing point, boiling point, flash point, and specific impulse. Both underfitting and overfitting have been effectively restricted by the COMES descriptor and the ensemble-learning model. Based on efficient classification and regression realized by our method, the discovery of new hydrocarbon fuels with specific criteria will be accelerated.