RNN seismic velocity model building: improving generalization using a frequency-stepping approach and hybrid training data

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 lead to the possibility of estimating high-resolution subsurface velocity models at a low computational cost. However, a significant challenge of ANNs is training generalization, which is the ability of an ANN to apply the learning from the training process to evaluate test data not previously encountered during the training process. In the context of velocity model building, this means learning the relationship between velocity models and the corresponding seismic data from a set of training data, and then using acquired seismic data to accurately estimate unknown velocity models. While generalizing to testing models with structures similar to those found in the training data has become a manageable task as evidenced in the recent literature, extending generalization to more realistic scenarios where testing models may exhibit drastically different velocity structures and/or distributions than those in the training data set remains an important and ongoing research challenge. To address this issue, this thesis develops and present the applications of a multi-scale approach inspired by physics-based full-waveform inversion that uses recurrent neural networks to invert frequency-domain seismic data using a frequency-stepping scheme. The input data consist of a sequence of seismic frequency slices that are fed to the network progressively from the lowest available to the highest usable in the data. I combine this approach with a hybrid training approach that merges background velocity gradient models with purely geometrical and geologically realistic model structures. This combination increases the range of spatial wavenumbers as wells as the variability of geological structures present in the training data. I demonstrate the potential for improved generalization by comparing the model estimates results from two trained networks: one using a hybrid set of models, the other with only geological models. I test the two networks using subsets of the community BP2004 benchmark model with complex salt structures fully absent from the models used in the training process. Qualitative analysis shows that models recovered using hybrid training data are significantly more accurate than those recovered using geological training data alone, with arbitrarily shaped salt bodies being accurately delineated by the trained hybrid network. In addition, I demonstrate through a quantitative SSIM metric analysis that the developed RNN extends the range of structures recoverable by the trained ANN. The developed approach illustrates the potential of neural networks to learn the seismic velocity model building problem at a general level from a representative set of training models, and opens the way for more research into improving the design of non-geological training data to further improve network generalization.