The main way that numerical error in seismic modeling manifests itself is through numerical dispersion brought on by coarse grids. Refining the mesh has the potential to minimize it, but doing so will result in an unacceptably high computational cost. Numerical Dispersion Mitigation network (NDM-net), a new technique that was recently created, is applied to the full dataset computed using a coarse mesh after the network has been trained using a relatively small number of seismograms that were previously computed using a fine mesh. The creation of the training dataset is the component of the procedure that requires the greatest computation. Therefore, reducing the quantity of precomputed seismograms may help the approach perform better as a whole. In this article, we describe a method for building the training dataset so that the difference between it and the full dataset does not go above the allowed limit.