Machine-Learning Approach Optimizes Formation-Pressure Testing in Complex Reservoirs

_ This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper SPE 224365, “Optimizing Depth Selection for Formation-Pressure Testing in Complex Reservoirs: A Machine-Learning Approach,” by Mohammad K. Aljishi, SPE, Muhammad Abdullah, and Deepak Devegowda, SPE, University of Oklahoma, et al. The paper has not been peer reviewed. _ Traditionally, formation-pressure-test (FPT) depth selection relies heavily on the expertise of engineers and geoscientists in analyzing well-logging data to determine test locations. A manual approach can be inconsistent, time-consuming, and prone to human bias. To address these challenges, this study introduces a machine-learning (ML) framework to enhance FPT depth selection, systematically improving decision-making based on well-log data. The proposed framework aims for consistent and reliable test placement, minimizing invalid tests, enhancing safety, and reducing operational costs. Its scalability and adaptability across various well types and reservoir conditions make it an important step toward efficient, data‑driven FPT operations. Methodology This study used well logs and formation-testing data from 16 wells in the Mangahewa Field, New Zealand. The data set included approximately 1,000 formation-testing points, along with corresponding well logs capturing borehole conditions, rock properties, and mudcake characteristics critical to FPT outcomes. Selected logs included caliper, gamma ray (GR), density, neutron porosity, resistivity, acoustic properties, and nuclear magnetic resonance (NMR)-derived permeability and bound fluid volume (BFV). Data Preprocessing. Data preprocessing ensured quality and consistency of inputs for ML models. Key steps included: 1. Log resampling—Logs were resampled to a consistent scale for compatibility across data. 2. Outlier removal—Anomalies such as caliper-log deviations and negative total neutron porosity values were removed to enhance reliability. 3. Well-compatibility adjustments—Wells with significantly different bit sizes or geometries were excluded to ensure uniformity. 4. Feature transformation—Logarithmic transformations were applied to features with wide dynamic ranges (e.g., resistivity and permeability) to normalize data and reduce skewness. 5. Feature selection—Highly correlated features, such as shear slowness and compressional slowness, were analyzed, and shear slowness was retained to minimize redundancy. Feature Engineering. To better understand the relationships between well-log parameters and FPT outcomes, statistical analysis was conducted using correlation heatmaps and pair plots. These visualizations revealed key dependencies between features and their effect on test validity. Key insights from feature analysis include: - BFV—Derived from NMR logs, BFV represents the volume of bound fluids, which influences pressure stabilization. Zones with high BFV are less likely to yield valid pressure tests because of restricted fluid mobility. While there is some overlap in BFV values, this parameter provides critical insights into fluid distribution and pressure behavior. - Caliper log—This measures borehole size and shape, identifying washouts and irregularities. Washouts can lead to invalid tests by preventing proper tool contact or causing fluid leakage. Although caliper log primarily reflects wellbore conditions, its influence on tool performance indirectly affects pressure test validity. - GR—Statistical distributions revealed that valid tests were strongly associated with GR values below 50 API, indicating clean reservoir sands. - Density-neutron crossovers—Higher-density neutron porosity crossovers correlated with valid test depths, suggesting favorable porosity and reservoir quality.