Analysis of Machine Learning Models for Prediction of Petrophysical Data
P. Palash,D. List,G. Maxwell,B. J. McPherson,C. Vega-Ortiz,E. Edelman
DOI: https://doi.org/10.56952/arma-2024-0627
2024-06-23
Abstract:The growing need for better reservoir characterization methods in unconventional hydrocarbon extraction, geothermal energy production, and carbon sequestration has led to a demand for more accurate and reliable petrophysical data prediction techniques. However, data standardization across multiple wells in an area of interest is a challenging task, but essential for effective reservoir characterization. Petrophysical data is often statistically erratic due to the use of multiple service providers and logging tools, as well as the natural variability of the reservoir lithology. The article underscores the potential of the RTM, ANN and GPR ML models to generate accurate predictions of petrophysical properties, such as density and total organic carbon (TOC), even in the face of incomplete data sets. Such predictions are not merely academic but have practical implications, with ML-derived pseudo-curves aiding in the fine-tuning of reservoir models. By providing detailed case studies, the article offers insights into the successful application of ML models for pseudo-curves generation, which not only enhances the accuracy of reservoir simulations but also mitigates the risks associated with reservoir development and management. In conclusion, it encourages a shift from traditional empirical modeling to data-driven, algorithmic approaches, which promise to redefine the standards for accuracy and reliability in reservoir analysis and development.
The historical approach in the oil and gas industry applies conventional models to establish correlations for generating pseudo-data. However, these models often fall short due to their reliance on region-specific petrophysical assumptions and the problematic generalization of constants or variables, leading to inaccurate estimations. To circumvent these limitations, ML regression models are increasingly employed. By leveraging existing petrophysical measurements—such as gamma ray, density, porosity, acoustic, and resistivity logs—as training datasets, ML models can infer the complex relationships between these measurements and reservoir characteristics. This training data encapsulates critical lithological information reflecting the interactions of minerals, fluids, and specific environmental conditions within the reservoir, providing a nuanced understanding that underpins the prediction process.
Engineering,Environmental Science