Reducing Numerical Dispersion in Pseudo-3D Space and Constructing Training Dataset

The paper discusses the extension of NDM-net (Numerical Dispersion Mitigation neural network) to pseudo-3D cases and the construction of training datasets. NDM-net is initially designed to reduce numerical dispersion in seismic data, which are generated as a result of simulating elastic waves. Previously, seismograms on a coarse grid with numerical dispersion and a certain number of seismograms on a fine grid are calculated to form a training sample. The paper discusses three approaches for building a representative sample to speed up the learning process in case of pseudo-3D. Additionally, the paper discusses the combination of metrics based on statistical analysis.