Hausdorff-distance-based training dataset construction for numerical dispersion mitigation neural network

Abstract

We present a novel approach for constructing the training dataset for the Numerical Dispersion Mitigation neural network (NDM-net). The NDM-net is a multi-step approach to reduce numerical error in seismic modeling. First, a coarse grid is used to simulate the entire dataset with lower accuracy quickly. A fine grid is then used to simulate selected cases with higher accuracy. Next, the high-accuracy solutions are used together with corresponding low-accuracy solutions as a training dataset for the neural network. Finally, the neural network is used to improve the accuracy of the entire dataset of low-accuracy solutions. Running high-accuracy simulations for the training dataset is the most time-consuming step of the NDM-net approach. Thus, reducing the training dataset may significantly improve the NDM-net performance. In this study, we propose a training dataset construction method that maintains the Hausdorff distance between the training dataset and the complete dataset, allowing the training dataset to be generated from as few as 3%–5% of the total number of sources.