Extracting useful information from sparsely logged wellbores for improved rock typing of heterogeneous reservoir characterization using well-log attributes, feature influence and optimization

Abstract

The information from sparsely logged wellbores is currently under-utilized in reservoir simulation models and their proxies using deep and machine learning (DL/ML). This is particularly problematic for large heterogeneous gas/oil reservoirs being considered for repurposing as gas storage reservoirs for CH₄, CO₂ or H₂ and/or enhanced oil recovery technologies. Lack of well-log data leads to inadequate spatial definition of complex models due to the large uncertainties associated with the extrapolation of petrophysical rock types (PRT) calibrated with limited core data across heterogeneous and/or anisotropic reservoirs. Extracting well-log attributes from the few well logs available in many wells and tying PRT predictions based on them to seismic data has the potential to substantially improve the confidence in PRT 3D-mapping across such reservoirs. That process becomes more efficient when coupled with DL/ML models incorporating feature importance and optimized, dual-objective feature selection techniques.