A neural network is used to approximate the transition operator from seismic data modeled on a large computational grid to data obtained on a small one. Thus, we obtain an effective way of suppressing numerical dispersion in numerically modeled seismic fields. This article discusses a method for constructing an optimal training dataset based on the properties of a velocity model. We build a distance matrix for the parts of the model that correspond to the positions of the sources and build a dataset in such a way that the distance between the training set and all sources is limited.