Training data set construction based on the Hausdorff metric for numerical dispersion mitigation neural network in seismic modelling

The article outlines a strategy for constructing a training data set for a numerical dispersion mitigation network (NDM-net), consisting in the calculation of the full set of seismograms by the finite difference method on a coarse grid and the calculation of the training sample using a fine grid. The training dataset is a small set of seismograms with a certain spatial distribution of wave field sources. After training, the NDM-net allows approximating low-quality coarse-grid seismograms into seismograms with a smaller sampling step. Optimization of the process of constructing a representative training dataset of seismograms is based on minimizing the Hausdorff metric between the training sample and the full set of seismograms. The use of the NDM-net makes it possible to reduce time costs when calculating wave fields on a fine grid.