Petrophysical assessment and permeability modeling utilizing core data and machine learning approaches – A study from the Badr El Din-1 field, Egypt

Abstract

Reservoir petrophysical characterization involving permeability estimation in the laboratory using core samples is time-consuming and that too accords limited coverage of the sub-surface reservoir. Machine learning (ML) algorithms have been utilized to overcome such difficulties, which offers better and more accurate prediction models. This work integrates core data and ML approached and reports a comprehensive petrophysical evaluation of the Cenomanian Bahariya and Albian Kharita reservoirs from the Badr El Din-1 field, Western Desert of Egypt. Three distinct reservoir rock types are identified from the detailed petrographic investigation: i) glauconitic sandstones (RRT1) and ii) subfeldspathic quartz arenites (RRT2) within Bahariya Formation and iii) quartz arenites (RRT3) within Kharita Formation. Reservoir quality index (RQI), flow zone indicator (FZI), and hydraulic flow units (HFU) are interpreted from the routine core analysis, which infers that the RRT1 has poor permeability (KH < 1 mD) and impervious (RQI < 0.25 μm, FZI <1 μm). Transgressive shoreface deposits of RRT2 display a wide range of porosity (5.9–18.5%) and permeability (0.82–76 mD) distribution and offer poor to fair reservoir qualities (1 μm < FZI < 5 μm) with moderate to high permeability anisotropy. Well sorted medium grained quartz arenites (RRT3) were deposited in a braided fluvial channel environment and contributed to the excellent reservoir quality with an average KH of 468 mD, FZI > 5 μm, and lesser permeability anisotropy. Finally, two-hybrid ML-based algorithms such as particle swarm optimization (PSO) trained neural network (NN) and least squares support vector machines (LS-SVM) have been employed to estimate the permeability. Based on the statistical performance metrics, we conclude that both the ML algorithms performed exceptionally well as compared to the conventional method, which is attributed to their enhanced competence to model the connectivity within the pore microstructures of the reservoir rocks with proper recognition of the pattern in the training data. Our results indicate that amongst the PSO-NN and LS-SVM algorithms, the latter is more superior since it is comparatively more accurate and reliable in permeability prediction.