This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper SPE 201698, “Finding a Trend Out of Chaos: A Machine-Learning Approach for Well-Spacing Optimization,” by Zheren Ma, Ehsan Davani, SPE, and Xiaodan Ma, SPE, Quantum Reservoir Impact, et al., prepared for the 2020 SPE Annual Technical Conference and Exhibition, originally scheduled to be held in Denver, Colorado, 5–7 October. The paper has not been peer reviewed. Data-driven decisions powered by machine-learning (ML) methods are increasing in popularity when optimizing field development in unconventional reservoirs. However, because well performance is affected by many factors, the challenge is to uncover trends within all the noise. By leveraging basin-level knowledge captured by big data sculpting, integrating private and public data with the use of uncertainty quantification, a process the authors describe as augmented artificial intelligence (AI) can provide quick, science-based answers for well spacing and fracturing optimization and can assess the full potential of an asset in unconventional reservoirs. A case study in the Midland Basin is detailed in the complete paper. Introduction Augmented AI is a process wherein ML and human expertise are coupled to improve solutions. The augmented AI work flow (Fig. 1) starts with data sculpting, which includes information retrieval; data cleaning and standardization; and smart, deep, and systematic data quality control (QC). Feature engineering generates all relevant parameters entering the ML model. More than 50 features have been generated for this work and categorized. The final step is to perform model tuning and ensemble, evaluating model robustness and generating model explanation and uncertainty quantification. Geology The complete paper provides a detailed geological background of the Permian Basin and its Wolfcamp unconventional layer, an organic-rich shale formation with tight reservoir properties. To find a solution for the multidimensional well-spacing problem in the Permian Basin, multiple sources and types of data were gathered using publicly available sources. The detailed geological attributes, including structure, petrophysics, geochemistry, basin-level features, and cultural information (such as counties or lease boundaries) have been combined in an integrated database to extract and generate features for the ML algorithm. Most attributes are available either in a limited number of wells, mostly vertical, or through the low number of available cored wells across the basin. Therefore, a significant amount of data imputation has been processed with mapping exercises using geostatistical modeling techniques. The mapping process augmented the ML attribute-generation step because these features were distributed in both vertical and lateral dimensions. All horizontal wells within the area of interest across the Permian Basin have been resampled with the logged and mapped information. The geological features also are reengineered into multiple indices to reduce the number of labeled features to include in the ML process. This feature-reduction process also has helped in ranking and selecting the most-important parameters relevant to the well-spacing problem. Here, a key attribute called the shale-oil index was introduced, which is generated for the ML-driven process and is used in understanding the level of contribution of geological sweet spots to well-spacing optimization. In addition, the initial well, reservoir, or laboratory data, including logs, have been normalized before mapping and modeling to eliminate potential bias. This study has focused on Wolfcamp layers; however, both geological and engineering attribute generation work flows used for this practical ML methodology to find optimization solutions for common problems are highly applicable to other unconventional layers, such as Bone Spring or Spraberry.
Read full abstract