Bayesian optimization of capillary pressure data in hydraulic flow units using NMR.

Area of Science:

• Petrophysics
• Geoscience
• Machine Learning Applications

Background:

• Nuclear magnetic resonance (NMR) data is crucial for reservoir petrophysical characterization.
• Estimating capillary pressure (Pc) from NMR T2 data is challenging due to fluid saturation variations and the need for hydrocarbon correction.
• Existing methods often lack effective hydrocarbon correction, hindering accurate reservoir characterization.

Purpose of the Study:

• To develop a novel methodology for estimating capillary pressure (Pc) in reservoirs at hydraulic flow units (HFUs).
• To incorporate NMR hydrocarbon correction and an ensemble-committee machine model (ECMM) for improved Pc estimation.
• To analyze capillary pressure variability at HFUs and its impact on model performance.

Main Methods:

• Utilized NMR T2 data and cumulative desaturation rate (∑(dSnw/dT2)) features.
• Developed a new workflow incorporating NMR hydrocarbon correction.
• Employed an ensemble-committee machine model (ECMM) with Bayesian-optimized machine, ensemble, and deep learning algorithms.

Main Results:

• The ECMM workflow significantly improved mean squared error (MSE) compared to individual intelligent models for Pc prediction.
• Analysis revealed that higher capillary pressure variability among HFUs negatively impacts model MSE.
• The methodology provides a robust and cost-effective approach for continuous Pc estimation.

Conclusions:

• The proposed ECMM-based workflow offers a robust and cost-effective solution for estimating continuous capillary pressure in reservoirs.
• Accurate hydrocarbon correction and advanced machine learning models are key to overcoming challenges in NMR-based Pc estimation.
• Understanding capillary pressure variability at HFUs is essential for optimizing reservoir characterization models.