DEEP LEARNING-BASED NUMERICAL DISPERSION MITIGIATION IN SEISMIC MODELLING

Abstract

Seismic modelling is the most computationally intense and time consuming part of seismic processing and imaging algorithms. Indeed, generation of a typical seismic data-set requires approximately 10 core-hours of a standard CPU-based clusters. Such a high demand in the resources is due to the use of fine spatial discretizations to achieve a low level of numerical dispersion (numerical error). This paper presents an original approach to seismic modelling where the wavefields for all sources (right-hand sides) are simulated inaccurately using coarse meshes. A small number of the wavefields are generated with computationally intense fine-meshes and then used as a training dataset for the Deep Learning algorithm - Numerical Dispersion Mitigation network (NDM-net). Being trained, the NDM-net is applied to suppress the numerical dispersion of the entire seismic dataset.