Title: Advanced Permeability Prediction Through Two-Dimensional Geological Feature Image Extraction with CNN Regression from Well Logs Data
Source: Mathematical Geosciences
Authors: Wakeel Hussain, Miao Luo, Muhammad Ali, Syed Naheel Raza Rizvi, Harith F. Al-Khafaji, Nafees Ali & Salah Alshareef Alkfakey Ahmed
Published: 14 January 2025
Link: https://link.springer.com/article/10.1007/s11004-024-10171-4
Abstract:
The evaluation of permeability plays an essential role in understanding subsurface fluid behavior, optimizing hydrocarbon recovery, managing reservoir performance, and facilitating the sequestration of carbon dioxide. Conventional methods for its computation, which depend on well tests and core samples, are time-consuming, costly, and limited. There is a need for more efficient and adaptable approaches that better support decision-making in the petroleum industry. Machine learning, particularly convolutional neural networks (CNNs), provides a quick and cost-effective solution for permeability prediction. This study builds a CNN regression model to predict permeability in the Sawan gas field in Pakistan, a prospective field for hydrocarbon storage and enhanced oil recovery. The geological feature image, derived from geophysical logs, includes variables like spectral gamma-ray log (SGR), density log (RHOB), neutron porosity log (NPHI), volume of shale (VSH), and effective porosity (PHIE). The significant contribution of this research is in illustrating how the integration of CNNs with geological images leads to a more accurate and efficient approach for predicting permeability, enhancing the performance of conventional neural network models. The model demonstrates exceptional efficiency, with a processing time of just 1.14 s. The training performance metrics reveal R-squared values of 0.9821 and adjusted R-squared values of 0.9818, indicating strong predictive ability. The root mean square error (RMSE) is recorded at 0.0288, the mean squared error (MSE) at 0.0008, and the mean absolute error (MAE) at 0.0213. In the cross-validation of the entire dataset, the CNN model maintains a high R-squared of 0.9812 and adjusted R-squared of 0.9810, with RMSE at 0.0507, MSE at 0.0026, and MAE at 0.0345. Furthermore, the subset testing results show that the CNN model achieves an R-squared value of 0.9869 and adjusted R-squared of 0.9865, with RMSE at 0.0297, MSE at 0.0009, and MAE at 0.0249. These results demonstrate a notable improvement over other methodologies, such as GMDH (group method of data handling)-based modified Levenberg–Marquardt (LM) and backpropagation neural networks (BPNN), confirming the CNN’s enhanced performance in permeability prediction. Through the successful application of CNNs for permeability prediction, this study not only enhances the methodological framework in reservoir characterization but also enhances decision-making processes related to subsurface resource management in the petroleum industry. The implications of our findings extend to various essential areas, including hydrocarbon management and carbon dioxide sequestration, thereby enriching the existing body of knowledge in reservoir studies.
