Seismic attributes (derived quantities) such as P-wave and S-wave impedances and P-wave to S-wave velocity ratios may be used to classify subsurface volume of rock into geologic facies (distinct lithology-fluid classes) using pattern recognition methods. Seismic attributes may also be used to estimate subsurface petrophysical rock properties such as porosity, mineral composition, and pore-fluid saturations. Both of these estimation processes are conventionally carried out independent of each other and involve considerable uncertainties, which may be reduced significantly by a joint estimation process. We have developed an efficient probabilistic inversion method for joint estimation of geologic facies and petrophysical rock properties. Seismic attributes and petrophysical properties are jointly modeled using a Gaussian mixture distribution whose parameters are initialized by unsupervised learning using well-log data. Rock-physics models may be used in our method to augment the training data if the existing well data are limited; however, this is not required if sufficient well data are available. The inverse problem is solved using the Bayesian paradigm that models uncertainties in the form of probability distributions. Probabilistic inference is performed using variational optimization, which is a computationally efficient deterministic alternative to the commonly used sampling-based stochastic inference methods. With the help of a real data application from the North Sea, we find that our method is computationally efficient, honors expected spatial correlations of geologic facies, allows reliable detection of convergence, and provides full probabilistic results without stochastic sampling of the posterior distribution.
Three digital earth models were designed and constructed during SEAM Phase II to study exploration challenges at the scale of modern land seismic surveys. Although built as generic models, each was based on one or more related geologic type areas. The Barrett model represents the seismic anisotropy of complex laminated and fractured shale reservoirs, based on the Woodford and Eagle Ford formations and set below a stratigraphic overburden and near surface of a North American midcontinent basin. The Arid model features the extreme property contrasts of desert terrains in a 500 m thick near surface that juxtaposes hard carbonate bedrock and soft sediments filling karsts, typical of the Saudi Arabian Peninsula. The Foothills model contains sharp surface topography and alluvial fan-like sediments above complex fold-and-thrust structures based on the compressive tectonics of the Llanos Foothills of South America. All three models were built in workflows that combined automated steps with a large measure of manual model building, which represents the current state of the art in geologic modeling for large-scale geophysical simulations. The Barrett and Arid models each contain about 1.5 billion grid cells representing regions 10 × 10 × 3.75 km in physical size. The Foothills model has about 2 billion cells representing a region about 14.5 × 12.5 × 11 km. Full elastic-wave simulations with these models were run for a combined total of about 170,000 shots, usually with millions of recorded channels per shot, generating several petabytes of seismic data in standard and novel shot-receiver geometries. Selected shots from these simulations show that large, detailed earth models can reproduce features of land seismic surveys that continue to challenge the best modern seismic data processing and imaging techniques.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.