Simple descriptor based machine learning model development for synergy prediction of different metal loadings and solvent swellings on coal pyrolysis

Abstract

The co-effect of solvent swelling and metal loading on coal pyrolysis was investigated through statistical analysis and machine learning. The distributions and properties of pyrolysis products, and the pyrolysis parameters were all considered. 22 targets were screened out by analysis of variance (ANOVA). Both linear and non-linear regression models aiming to predict these values were constructed, in which the swelling ratio and atomic descriptors collected from handbooks were taken as inputs. The symbolic transformation (ST) algorithm was involved to assemble a new feature and the model built on the advanced feature set displays higher prediction accuracy for all targets. The result of leave-one-out cross validation shows acceptable performance(R2 > 0.8) for most targets (19 of 22), and good performance (R2 > 0.9) for half of them (12 of 22). The importance of ST feature was verified, and the contribution of each single feature was clearly reflected in the formula of ST.