Abstarct: Abstract In the present work, a QSPR study was performed to predict the density of saturated and unsaturated aliphatic hydrocarbons containing linear and branched alkanes, cycloalkanes, alkenes and cycloalkenes as well as aromatic hydrocarbons over a wide range of temperatures and pressures. The first selected descxriptors were temperature (T), pressure (p) and molecular mass (M) because the density is related to them. The other descxriptors have been obtained from chemical structure of compounds using two different approaches. In the first approach, the best descxriptors were selected baxsed on group contribution method. After analyzing the chemical structures of 79 studied hydrocarbons, There were found 15 useful functional groups. In the second method, a large number of molecular descxriptors (3224 descxriptors) were calculated using Dragon sofrware from chemical structure of compounds. Then, a subset of calculated descxriptors was selected from 22 classes of Dragon descxriptors with a genetic algorithm baxsed on partial least aquare (GA-PLS) method as a feature selection technique. Only 11 descxriptors were obtained by genetic algorithm (GA) as the most feasible descxriptors. The selected descxriptors with two feature selection methods and 3 variable (T, P and M) were used as input nodes for constructing different artificial neural networks. 4 feed-forward artificial neural networks were designed by combination of Levenberg-Marquardt (LM) and Bayesian regularized (BR) algorithms with two transfer function: logarithm-sigmoid and tangent-sigmoid. Different parameters for designed networks such as number of neurons in input and hidden laxyers and the number of iteration (epoch) were optimized. After optimization of parameters for generated artificial neural networks using two types of descxriptors, the performance of the models was investigated by the external test set. The mean square error (MSE) and determination coefficient (R2) for the external test set using optimized GCM-ANN model are 9.556 and 0.9987, and usimg optimized GA-ANN model are 70.1612 and 0.9982 respectively. The result showed that the Levenberg-Marquardt artificial neural network with logarithm-sigmoid transfer function and GCM baxsed descxriptors, may be simulated the relationship between the structural descxriptors and density of the desired molecules accurately.