This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper OTC 31419, “Bringing Huge Core-Analysis Legacy Data to Life Using Machine Learning,” by Siti N.F. Zulkipli, SPE, Benard Ralphie, and Jamari M. Shah, SPE, Petronas, et al. The paper has not been peer reviewed. Copyright 2022 Offshore Technology Conference. Reproduced by permission.
Among the sources of subsurface data, rock and fluid analyses stand out as the best means of directly measuring subsurface properties. The implication of modeling this data into an organized data store means better assessment of economic viability and producibility in frontier basins and the capability to identify bypassed pay in old wells that may not have rock material. The complete paper presents agile technologies that integrate data management, data-quality assessment, and predictive machine learning (ML) to maximize company asset value with legacy core data.
The complete paper integrates data gathering, data filtering, the connecting of scattered data, and the building of useful knowledge models using legacy core data from various operating assets. This integration is achieved by a data quality check (QC) work flow and ML to improve the definition of reservoir rock properties that affect field development and asset management. ML and deep-learning capabilities have been implemented to predict critical reservoir properties based on qualified legacy data.
The work flow is aimed at bridging gaps in areas with limited core data coverage, reducing subsurface uncertainty and risk, and accelerating project delivery. In addition, data-prediction accuracy can be improved with ML applications as new data become available.
In this stage, more than 60,000 data points are gathered from 327 wells and grouped into different categories such as routine core analysis, special core analysis, geology, and rock mechanics. Leveraging on best practices in laboratory core analysis and subject-matter expert (SME) expertise, 88 QC rules are generated to delineate the data, identify outliers, and classify the data into good quality (labeled as 0), data requiring further QC (1), and rejected data (2). The generated QC rules target various rock properties such as porosity, permeability, capillary pressure, electrical properties, relative permeability, and rock mechanics. QC-rule implementation is an iterative process in which, after the initial screening, the rules are further optimized and enhanced to assist data classification based on practical experiences. These QC rules are needed because some observed data errors are related to data tabulation and incorrect naming.
A total of 1,826 relevant variables, generated by different vendors, was extracted during QC-rule creation. These variables had to be merged in the QC process, which reduced the number of variables to 240.