Velocity model building by deep learning: From general synthetics to field data application

The promise of fully automatic full-waveform inversion (FWI) — a (seismic) data-driven velocity model building process — has proven elusive in complex geologic settings, with impactful examples using field data unavailable until recently. In 2015, success with FWI at the Atlantis Field in the U.S. Gulf of Mexico demonstrated that semiautomatic velocity model building is possible, but it also raised the question of what more might be possible if seismic data tailor-made for FWI were available (e.g., with increased source-receiver offsets and bespoke low-frequency seismic sources). Motivated by the initial value case for FWI in settings such as the Gulf of Mexico, beginning in 2007 and continuing into 2021 BP designed, built, and field tested Wolfspar, an ultralow-frequency seismic source designed to produce seismic data tailor-made for FWI. A 3D field trial of Wolfspar was conducted over the Mad Dog Field in the Gulf of Mexico in 2017–2018. Low-frequency source (LFS) data were shot on a sparse grid (280 m inline, 2 to 4 km crossline) and recorded into ocean-bottom nodes simultaneously with air gun sources shooting on a conventional dense grid (50 m inline, 50 m crossline). Using the LFS data with FWI to improve the velocity model for imaging produced only incremental uplift in the subsalt image of the reservoir, albeit with image improvements at depths greater than 25,000 ft (approximately 7620 m). To better understand this, reprocessing and further analyses were conducted. We found that (1) the LFS achieved its design signal-to-noise ratio (S/N) goals over its frequency range; (2) the wave-extrapolation and imaging operators built into FWI and migration are very effective at suppressing low-frequency noise, so that densely sampled air gun data with a low S/N can still produce useable model updates with low frequencies; and (3) data density becomes less important at wider offsets. These results may have significant implications for future acquisition designs with low-frequency seismic sources going forward.

Download Full-text

Deep-learning inversion: A next-generation seismic velocity model building method

Geophysics ◽

10.1190/geo2018-0249.1 ◽

2019 ◽

Vol 84 (4) ◽

pp. R583-R599 ◽

Cited By ~ 38

Author(s):

Fangshu Yang ◽

Jianwei Ma

Keyword(s):

Deep Learning ◽

Seismic Data ◽

Model Building ◽

Seismic Velocity ◽

Velocity Model ◽

Computational Time ◽

Learning Methods ◽

Velocity Models ◽

Velocity Model Building ◽

Training Stage

Seismic velocity is one of the most important parameters used in seismic exploration. Accurate velocity models are the key prerequisites for reverse time migration and other high-resolution seismic imaging techniques. Such velocity information has traditionally been derived by tomography or full-waveform inversion (FWI), which are time consuming and computationally expensive, and they rely heavily on human interaction and quality control. We have investigated a novel method based on the supervised deep fully convolutional neural network for velocity-model building directly from raw seismograms. Unlike the conventional inversion method based on physical models, supervised deep-learning methods are based on big-data training rather than prior-knowledge assumptions. During the training stage, the network establishes a nonlinear projection from the multishot seismic data to the corresponding velocity models. During the prediction stage, the trained network can be used to estimate the velocity models from the new input seismic data. One key characteristic of the deep-learning method is that it can automatically extract multilayer useful features without the need for human-curated activities and an initial velocity setup. The data-driven method usually requires more time during the training stage, and actual predictions take less time, with only seconds needed. Therefore, the computational time of geophysical inversions, including real-time inversions, can be dramatically reduced once a good generalized network is built. By using numerical experiments on synthetic models, the promising performance of our proposed method is shown in comparison with conventional FWI even when the input data are in more realistic scenarios. We have also evaluated deep-learning methods, the training data set, the lack of low frequencies, and the advantages and disadvantages of our method.

Download Full-text