An integrated comprehensive approach describing structural features and comparative petrophysical analysis between conventional and machine learning tools to characterize carbonate reservoir: A case study from Upper Indus Basin, Pakistan

Abstract

More than 70% of the global hydrocarbon reserves are present in carbonated rocks. Evaluating prospects in carbonate reservoirs is a complicated task because of their unique depositional features. The Eocene carbonates in the Joyamair oil field are heterogeneous and present challenges defining the entrapment and sealing mechanism by applying traditional methods. Although structural interpretation revealed a positive triangular geometry, estimating accurate reservoir properties requires an effective model for assessing hydrocarbon presence. Therefore, an optimized machine learning (ML) approach has been deployed to address reservoir challenges and delineate the potential with a high success rate after drawing a comparison with the conventional approach. Two wells were utilized for petrophysical evaluation in the conventional method, while one well (Joyamair-04) was kept blind in a supervised ML approach. Extra Tree Regressor (ETR) produced a low volume of shale and effective porosity (PHIE) high results with more than 99% R² and least mean square error score. Random Forest Regressor (RFR) showed water saturation (S_w) results with about 100% accuracy compared to conventional interpretation at a blind well. Volumetric reserve estimation also proved economical hydrocarbon reserves present in the reservoir formation. The study revealed that integrating conventional and ML techniques along with structural geometry aided better reservoir characterization and reserve estimation. The study proved that ML algorithms outperformed traditional petrophysical methods in accuracy and efficiency.