Machine Learning-Based Numerical Dispersion Mitigation in Seismic Modelling

We present an original approach to improving seismic modelling performance by applying deep learning techniques to mitigate numerical error. In seismic modelling, a series of several thousand simulations are required to generate a typical seismic dataset. These simulations are performed for different source positions (equidistantly distributed) at the free surface. Thus, the output wavefields that corresponded to the nearby sources are relatively similar, sharing common peculiarities. Our approach suggests simulating wavefields using finite differences with coarse enough discretization to reduce the computational complexity of seismic modelling. After that, solutions for 1 to 10 percents of source positions are simulated using fine discretizations to obtain the training dataset, which is used to train the deep neural network to remove numerical error (numerical dispersion) from the coarse-grid simulated wavefields. Later the network is applied to the entire dataset. Our experiments illustrate that the suggested algorithm in the 2D case significantly (up to ten times) speeds up seismic modelling.

Automated summary

The proposed method utilizes deep learning techniques to reduce numerical errors in seismic modeling, significantly improving performance. By using coarse discretization for finite difference simulation of wavefields and training a deep neural network to remove numerical errors, seismic modeling is accelerated in 2D cases by up to ten times.