Seismic driven reservoir classification using advanced machine learning algorithms: A case study from the Lower Ranikot/Khadro sandstone gas reservoir, Kirthar Fold Belt, Lower Indus Basin, Pakistan

Abstract

Reservoir characterization of thin sand bed reservoirs has been a challenge for petroleum explorers across the globe. In this study, we have studied the heterogeneous Paleocene sandstone gas reservoir of Lower Ranikot/Khadro, Kirthar Fold Belt, Lower Indus Basin, Pakistan. The studied reservoir is par below seismic resolution, with an average thickness that varies from 4 to 7 m at places and has provided a good amount of prosecution in some producing wells. The solution to the above challenges has been put forth where an integrated approach of seismic attributes, petrophysical properties, well logs, and facies data sets are combined with advanced machine learning algorithms to get better resolution for the studied thin-bedded sandstone resource. In addition, the thin heterogeneous sands of the Ranikot/Khadro Formation were characterized by predicting the elastic properties of the reservoir and classifying the facies distribution of hydrocarbon-bearing sands within the Zamzama Gas Field. Several machine learning (ML) algorithms were chosen and implemented on pre-stack seismic data from seven wells, two of which were produced from the Paleocene reservoir. The Gradient Boosting Regressor (GBR) produced promising results with the demarcation of thin sands with high accuracy levels. Based on a successful petro-elastic relationship, thin gas sands were highlighted. The low VCL of 30% and Sw of 45% cut-off reflect gas sands with P-impedances of 7500–9500 m/s × g/cc and a Vp/Vs ratio of 1.4–1.6 as decisive ML characteristics. Based on the applied workflow aided by ML, three reservoir units have been assigned for the Paleocene hydrocarbon-bearing sands, and gas sand probability distribution maps have been developed for each reservoir unit. Aligning the seismic inversion technique with machine learning advanced algorithms not only speeds up the prediction of elastic properties for facies classification but also increases accuracy to over 90% while drastically reducing uncertainty and associated costs in the process. Advanced machine learning algorithms have allowed us to cater to thin sand beds by eliminating the need for injecting frequencies. Instead, it recognizes features from the well data, trains itself, and predicts the corresponding features on the seismic trace, thus effectively allowing us to characterize the beds that are below seismic resolution in any given complex geology. The generated probability distribution maps of the three recognized thin-bedded sand units can be the basis for petroleum exploration and development in the study area. The discussed technique is successfully applied to complex folds and thrust belts, and hence can be implemented successfully in other heterogeneous thin-bed reservoirs for the successful classification of hydrocarbon-bearing sand distribution. It would optimize the results in a short period of time and with fewer economics.