The numerical dispersion which occurs in seismic data due to use of coarse meshes in numerical modeling is reduced using a neural network. The training dataset is a sample of wave fields that have been modeled on a fine grid. The trained NDM-net (Numerical Dispersion Mitigation neural network) is applied to the entire set of seismic data corresponding to all of the source positions. This approach makes it possible to obtain seismic data with suppressed dispersion at a lower cost compared to numerical modeling on a fine mesh. The use of a neural network for test cases confirms the approachs effectiveness. We present several approaches to generate a training dataset. All of them are based on the minimization of the Hausdorff distance, however, they use different element-wise metric. In this research, we provide the global sensitivity analysis of the network accuracy on the choice of the metrics. It is shown that the training dataset formed using the properties of the seismic data itself has the greatest weight for the resulting error.