A supervised feed-forward back-propagation neural network model has been developed and used to estimate tight gas sand permeability from porosity, mean pore size, and mineralogical data. A sensitivity study on different neural network architectures led to the choice of an optimal network. topology consisting of an eight-neuron input layer, two five-neuron hidden layers that use nonlinear sigmoid transfer functions, and a linear single-neuron output layer. The network model has been trained on a data set from a tight gas sand well and tested on some core samples data that were not seen by the network during training. The optimal network architecture was able to estimate back the permeability from the training set within 0.89% average relative error and was able to predict the permeability of the test data set within 3.3% average relative error. This is a remarkable result, since Linear and nonlinear multivariate regression models were unable to predict the intrinsic permeability within less than 40% average relative error.