In this study, we present a novel approach for seismic modeling combining conventional finite differences with deep neural networks. The method includes the following steps: First, a training dataset composed of a small number of common-shot gathers is generated. The dataset is computed using a finite-difference scheme with fine spatial and temporal discretization. Second, the entire set of common-shot seismograms is generated using an inaccurate numerical method, such as a finite difference scheme on a coarse mesh. Third, the numerical dispersion mitigation neural network is trained and applied to the entire dataset to suppress the numerical dispersion. We tested the approach on two 2D models, illustrating a significant acceleration of seismic modeling.
