Neural network seismic velocity model building: A frequency-stepping approach

Abstract

Data-driven artificial neural networks (ANNs) demonstrably offer a number of advantages over conventional deterministic methods in a wide range of geophysical problems. For seismic velocity model building, judiciously trained ANNs offer the possibility of estimating subsurface velocity models more efficiently than deterministic full-waveform inversion (FWI) approaches with reduced sensitivity to many inherent FWI non-linearity issues; however, there are substantial challenges with effective and efficient network training. Motivated by the multi-scale approach commonly used to tackle FWI non-linearity challenges, we develop a frequency-stepping approach using a sequence-to-sequence recurrent neural network to address the velocity model building challenge. Input sequences consist of the frequency-domain seismic data ordered by frequency from lowest to highest, while the corresponding output sequences are smoothed velocity models with different frequency-dependent smoothness levels. Qualitative and quantitative analyses of the network testing results show that such a network has the potential to build complex velocity models starting from their smoothest components.