Stochastic inversion of seismic data by implementing image quilting to build a litho-facies model for reservoir characterization of Delhi field, LA

Abstract

In this research a new stochastic inversion approach along with an image reconstruction method is implemented to build a litho-facies model with a focus on Delhi Field, LA. The field is under CO2 injection as an enhanced oil recovery (EOR) method. This makes it critical to define the CO2 flow paths and flow baffles in higher resolution to plan for the EOR project. The algorithm starts at the well location by defining the litho-facies, using well logs, K-mean clustering method and core studies, and updated by the elastic properties distribution. The whole inversion approach is performed including multiple point statistics (MPS). The key element in the MPS algorithms is the training image. It is a conceptual model from the reservoir, which is built based on information from the reservoir regarding the depositional environment, structure, and any other information from the reservoir. A 3D training image is built for the reservoir, but the inversion is performed on a 2D seismic line, therefore the training image is sub-sampled in the direction parallel to the direction of the 2D inline of interest. Then a square template is chosen of sizes of 5×5 and 7×7 are chosen and all the 2D planes are scanned with this template and the pattern database is constructed. The pattern database includes all the possible configurations of the litho-facies from the training image. At the next step, the search algorithm begins and searches for all the patterns from the database that have similar configuration to the litho- facies at the well location. A distance function is defined (here Manhattan distance) and the patterns providing the smallest distance with the patterns at the well location are stored. Multiple realizations of litho-facies from the stored patterns are generated. The next step is to choose the realization, which provides the highest correlation or simi- larity to the subsurface. At this step, seismic forward modeling is implemented. Pseudo-logs of density and P-wave velocity are generated from the joint distribution of the properties at the well location that are conditioned to each litho-facies. Multiple realizations of pseudo-logs are generated (15 in this case) and synthetic seismic traces are created, having extracted the wavelet from the seismic volume. The realization that has the highest cross correlation is chosen as the litho-facies at the well location. To continue the algorithm away from the well location, an image reconstruction method that is called image quilting is implemented. This algorithm searches for similar patterns that have some overlapping criteria with the previously accepted pattern. The distance function is defined in a way to search for the overlapping grid nodes. The algorithm continues and the seismic forward modeling is im- plemented in a stochastic approach to find the best elastic properties and the corresponding litho-facies realization. Multiple realizations of litho-facies for the whole 2D inline is gener- ated and the maximum probability map of multiple realizations (ten in this case) is obtained as a representative of the litho-facies in the reservoir. The structural and depositional complexity of Delhi Field, presents a heterogeneous reser- voir in the vertical and horizontal directions. Due to the fact that the field is under an EOR process, obtaining a detailed definition of litho-facies and flow paths distributions is of great importance. The method conducted in this research incorporates stochastic inversion and image reconstruction and provides a new methodology for constructing a detailed and high resolution litho-facies model by integrating multi-scale and multiple data types for complex and heterogeneous reservoirs like Delhi Field. Because of the stochastic characteristics of this methodology, equi-probable scenarios are generated and the most probable one is calculated.