The complexity of the porous medium in tight gas reservoirs due to the presence of small pore throats, the high overburden pressures and the unusual fluids distribution makes the petrophyisical parameters (measured from samples assays) to achieve misleading values. In addition, they tend to be optimistically and that set an important discussion about accuracy in permeability assessments, limited in the range of 0.001 and 0.01 mD Therefore, we intend to find the way to achieve accurate calculations, integrating probabilistic, statistical and analytical models towards a computational model that best represents the insitu conditions of the reservoir. An artificial neural network is an information-processing system that has certain performance characteristics in common with biological neural networks, and that have been developed as generalizations of mathematical models of human cognition or neural biology [2]. They learn from experience in a way that no conventional computer can and they will shortly solve all of the world's hard computational problems, Neural networks are used automatically for solve recognition, classification problems and forecasting. This paper takes the equation from C.Y. YAO and S.A HOLDITCH [1], and integrates probabilistic models such as bivariate distributions fitted with multiple regression and neural network model [2] to predict permeability from different well logs (GR, ILD, RHOB, NPHI) correlated with data obtained from core samples. The model was applied to the Lajas Formation located in the Neuquén Basin. The data used include previous experience with more than 5 years and a monthly seasonality. The construction of the predictive model was developed following the process of knowledge discovery and model Neural Networks - The Multilayer Perceptron (Back Propagation), with the lowest error, using 80% of the data to train and 20% for validation, a number of competing models with different parameters. This is in aligned to the improvement of the integrated model, expecting to find results with an error less than 10%.