Adaptive training dataset generation for neural network numerical dispersion mitigation approach in seismic modeling

Abstract

We introduce a novel method for developing the training dataset for the Numerical Dispersion Mitigation network (NDM-net), aimed at diminishing numerical inaccuracies in seismic modeling. Our strategy involves using a limited set of seismograms, produced with coarse and fine grids, to train the network. This training enables the network to transform less accurate coarse-grid data into higher-quality fine-grid data. Subsequently, the network is employed on a more extensive set of seismograms, initially computed with the coarse grid, to lower numerical errors. Creating the training dataset is the most demanding aspect of this method, requiring a balance between the number of seismograms used and maintaining training effectiveness. We propose a method to create the training dataset that maintains a specific Hausdorff distance with the complete dataset. However, this distance can vary based on the seismic-geological model used in simulations. Our work shows that an adaptive approach in setting the Hausdorff distance limit is more advantageous than a fixed limit, as it reduces the training dataset size without compromising accuracy.