NDM-Net and Model Changes

The paper discusses the problem of suppressing the numerical error, which together represents the numerical dispersion and the error of the roughened velocity model. Such a numerical error is suppressed by using a NDM-net (Numerical Dispersion Mitigation neural network), which was originally proposed to map seismograms with numerical dispersion to accurate solutions calculated on a fine grid. This approach makes it possible to accelerate the seismic modeling by 6 times without significant changes in quality. Also, the paper discusses two-stage training of NDM-net, in which the numerical dispersion is first suppressed, and then the interface error.