Seismic Inversion In Fluvial Reservoirs: Building A Geologic Model-based Inversion For The Stratton Field 3D Survey

Seismic surveys, in particular 3-D surveys, are an important tool for imaging geologic structure in the subsurface; they are routinely applied to map structural trends and geometry to define the reservoir structure. Many of the world's oil and gas reservoirs are sands deposited by river systems but the scales of fluvial architectural elements challenge the resolution of the seismic method. One such challenging field is the Stratton field in the FR-4 Gas Play of south Texas; the reservoir is Frio formation, a thick collection (greater than 3000 feet) of stacked fluvial deposits consisting of thin channels and splays of reservoir sand that are separated by low-permeability floodplain deposits. Of further interest, the Texas Bureau of Economic Geology (BEG) makes available a 3-D seismic survey over a portion of this field; this creates a common data base for researchers to test inversion techniques to transform the seismic data into reservoir parameters. This data set has dense well coverage with 19 wells in the two-mile by one-mile seismic survey, an important feature for testing inversion techniques,Several techniques for mapping reservoir elements using the seismic data have been proposed, they generally correlate a specific sand to a specific seismic time-horizon. The most straightforward processes track a seismic reflector or take a horizontal slice through the seismic volume to create a map of attributes to image subsurface features. In the Frio formation a river is expressed as an amalgamated ribbon of channel sands, typically 30 feet thick. To find this target in the seismic data requires a very accurate time-to-depth correlation. Techniques such as mapping the amplitude of a horizon can produce interesting patterns, but in the BEG data-set the correlation to well data is poor. This research challenges the notion that the seismic inversion, as a one-to-one mapping for the BEG survey, can produce a usable geologic-model. In other words, seismic attributes at a specific time may represent the convolution of the desired reservoir element with neighboring layers (e.g. tuning, multiples, etc.) and interfering parameters (e.g. gas, thin beds, and tight sands). Conversely, information on a fluvial reservoir element at a specific depth may be spread over time.To analyze the depth-to-time mapping, a limited section (650 ft) of the upper middle Frio was chosen that has simple post-depositional deformation. The BEG data comes with well-logs for 19 wells that cover the study area. The well-logs are used to construct a detailed stratigraphic column; this strata is dissected by several floodplains that can be correlated for chronostratigraphic control. The well-log stratigraphy is compared to the seismic traces near the well-bores; from the analysis the information on a specific depth target appears spread over a broad time-window (greater than 40 ms). Given the information spread noted in the seismic data, a technique for tracking the fluvial reservoir elements using a neural network (NN) classifier is designed and tested. Similar to fingerprint tracking, the character of the seismic trace in a broad search window is tracked away from the well-bores. Three different fluvial parasequences are the tracking targets for the NN classifier, the first and third parasequences have river channels as the target; the target in the second parasequence is a splay that deposited thin (~ 15 ft. thick) sheet sand. The wells are divided into 12 model wells to train the network, 7 testing wells to test the resulting facies maps. The performance of the NN classifier is compared to the simple inversion technique of making an amplitude map at a specific time-slice.Given a successful neural network design, the seismic features used for training are analyzed by a clustering and sub-setting algorithm. The NN feature vectors are examples of the seismic trace near the well-bores that have been classified; for a 44 ms search window (with 2 ms resolution) the feature vector would be 22 points along the trace (time, amplitude). The sub-setting of the features ranks which time segments in the search window contain information on the target class. The most potent features, when viewed as amplitude time-slices, often show plausible shapes for fluvial elements, in particular we often note a "phantom-thalweg" as a trough in the seismic and high amplitude patterns that correlate to splays. Fusing the images from the "best" time-slices and the NN classifier, a fluvial model is mapped that has good correlation to the well-bore stratigraphy.