Abstarct: In this research, studies on Quantitative Structure-Property Relationship (QSPR) has been applied by using linear and nonlinear modeling in order to predict the self-diffusion coefficient for linear alkanes. The self-diffusion coefficient of cases were classified by the use of principal component analysis (PCA) into four categories; training, validation, test and external test. Descxriptors of molecular structure was calculated. The linear modeling started after removing the additional descxriptors and adding two parameters: temperature and pressure. Two methods, principal component regression (PCR) and partial least squares regression (PLS), were used for linear modeling which in both principal components has been made first and then optimum number of principal components has been calculated. Afterward modeling started by optimized principal components and the final models were applied in order to case prediction related to external test set. The Mean squared error (MSE) related to external test set for PCR and PLS models were obtained respectively 0.6129 and 0.5327. Two methods, stepwise regression (SR) and genetic algorithm (GA), were used for separating optimum descxriptors in nonlinear modeling. Optimum descxriptors entered the neural network and modeling was done for both methods in those descxriptors have been chosen. After training and optimization, built networks were used for predicting related cases to external testing series. The MSE for the external test set in SR-ANN and GNN models was 0.0379 and 0.0219, respectively. These results indicate the superiority of nonlinear methods in describing selection phase and modeling stage.