The neural network was designed to mitigate numerical dispersion in three-dimensional seismic data. The algorithm operates in two stages. In the first stage, seismic modeling is performed using a coarse computational grid to generate seismograms, followed by the recording of a subset of seismograms with greater accuracy using a fine computational grid. This subset serves as the training data for the neural network. The second stage involves training the neural network using the acquired training dataset. Once trained, the network is applied to rough seismograms to achieve enhanced results. A series of numerical experiments were carried out to assess the efficacy of the proposed methodology.