Abstarct: Soil dispersivity coefficient is a key parameter in determining the distribution of pollution concentration in soil, it is also an important parameter of the advection-dispersion equation, that it ̓ s measurement is very crucial. Soil dispersivity coefficient is also typically difficult to model due to the complexity of the phenomenon. With respect to importance of this parameter, this research aims to forecast soil dispersivity coefficient by using Multi laxyer Perceptron (MLP) - Artificial Neural Network (ANN). whereas studies conducted by number of investigators have shown that the dispersivity of soil is dependent on the travel distance, this research also aims to extrapolation or interpolation the soil dispersivity coefficient values in each distance. then comparing capability of ANN model with Multiple Linear Regression method (MLR). So, the experimental data which measured by chavoshinejad and maroofpur, (1389) in the rectangular tank with 1550mm length, 100mm width and 600mm height, were used in this study. The collected data related to sandy soil with five sizes of very coarse, coarse, medium, fine and very fine and five distances of 25, 50, 75, 100 and 125 cm. NaCl also was used as a conservative tracer with five velocities. The measured data such as transport distance (L), bulk density (ρb), porosity (n), hydraulic conductivity (K), average diameter of particles (D50) and the pollutant velocity (Vc) were used as input data and soil dispersivity coefficient (α) was the output of the models. For comparison of results, statistical criteriaes such as coefficient of determination ( R2), Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) were used. The results showed that MLP - ANN with one hidden laxyer and Secant Hyperbulic function (Sech) and R2= 0.987 and RMSE and MAE equal to 0.0004 and 0.0003 respectively in testing stage, gives the best result. so, ANN model can be used as a useful tool for dispersivity coefficient estimation. The ANN model showes satisfactory performance for all extrapolation and interpolation stages. The results also clearly showed that MLP-ANN can estimate more accurate than multiple linear regression method.