Urbanization is one of the most significant global changes. The rapid urban growth has been imposing high pressure to land and their resources. Neural networks are a powerful tool for understanding the spatial processes and patterns. Hence, the neural network model called multi-layer perceptron was applied as a tool for simulating the urban growth in Hashtpar township. The Root Mean Square Error (RMSE) was used as an index in design and stopping the training process of the network in this study. After normalization and removal of the covariate variables, distance to city center, main transportation and hydrographical networks, agriculture, grassland, barren land and slope were chosen as effective variables on the urban growth for study area. Architecture of the network has been designed as 7-14-1, which stands for number of input, hidden and output nodes, respectively. The training process was conducted by implementation of the sigmoid function and extracting the training samples of the urban change (1989-2000) and then simulating the urban growth for 2007. Investigating the performance of the model and analyzing the pattern of the simulated landscape was carried out using the relative operating characteristic and the landscape ecological metrics. The values of the ROC and landscape Ecological metrics indicate an acceptable spatial agreement between the simulated and classified maps. Accordingly, the neural network model has a good reliability in simulation of the class area, Euclidean nearest distance and fractal dimension index of the urban patches, as well. Finally, sensitivity of the model was examined using stepwise independent variable elimination and comparing the results with the full model. The results revealed that the distance to city center and the main transport network can be considered as the most effective variables in simulating the urban growth in the study area.