Abstarct: The main aim of the present work is to development of a quantitative structure-property relationship (QSPR) method using an artificial neural network for prediction of thermal conductivity of n-alkanes over a wide range of temperature and pressure. Improved model for predicting this property were generated using three different methods for descxriptor selection: stepwise multiple linear regression (MLR), combined data splitting-feature selection (CDFS) strategy and genetic (GA) algorithm. Then, the selected descxriptors using each of these methods were used as input nodes for generating three-laxyer network. The data set was divided into three data sets using principle component method: training (798point), validation (201point) and test set (216point). After training and optimization of the ANNs parameters (weights and biases) and architecture, the performance of the optimized models was evaluated by the test set. The obtained mean square error (MSE) for test set using MLR-ANN, CDFS-ANN and GA-ANN were 1.1757, 1.0411 and 1.1079, respectively. The obtained results showed that all three methods produce excellent nonlinear models, but the model obtained by the CDFS-ANN is somewhat better.