Loss of containment of storage tanks can involve atmospheric dispersion of toxic or flammable gases. The accident of Buncefield oil storage depot in 2005 and the Viareggio LPG explosion in 2009 tragically illustrate the potential consequences of dispersion phenomena in complex environments. Gas dispersion modeling in a complex geometry is a tricky task because of the effects of obstacles, which involves for example a high level of turbulence. Indeed, flows around obstacles have specific behavior like boundary layers effects and recirculation zones. Different approaches exist to estimate atmospheric dispersion, depending on the modeling strategy. Gaussian models solve advection-diffusion equation (ADE) assuming several hypotheses, in particular uniform flow and turbulence, but are not very efficient in congested areas. On the other hand, Computational Fluid Dynamics (CFD) computes very likely turbulent flows and dispersion by solving non-linear Navier-Stokes equations for congested areas. Turbulence can be modeled by considering turbulent kineticenergy k and its dissipation rate e (k-e model standard). A full 2D CFD modeling requires solving 5 equations (continuity, momentum on x and y axes, transport of k and e). These numerous calculations induce longcomputing time on large areas with fine meshing. Once the wind field is calculated, dispersion of a pollutantcan be rapidly computed by ADE. As the major part of computing time of CFD models is dedicated to determine the turbulent flow, Artificial Neural Networks (ANN) were thus investigated to calculate x and y velocities and turbulent diffusion coefficient Dt. This machine learning method is a powerful statistical tool as it is able to reproduce accurately any nonlinear and dynamic behavior from a database without any physical assumption. This study focused on turbulent flows around cylindrical storage tanks, with a diameter in therange [10 m – 52 m]. Database is designed by RANS k - e CFD model. Several neural networks solutions are proposed and their efficiency is compared and discussed in terms of quality, real-time applicability and real-life plausibility. Four criteria (coefficient of determination, factor of two, fractional bias and normalized mean squared error) are used to evaluate the model. While the accuracy is kept within satisfying criteria values, computational time is reduced by a factor of 600.