Machine learning - a novel approach to predict the porosity curve using geophysical logs data: An example from the Lower Goru sand reservoir in the Southern Indus Basin, Pakistan

• Published in: Journal of applied geophysics Vol. 214; p. 105067
• Main Authors: Hussain, Wakeel, Luo, Miao, Ali, Muhammad, Hussain, Syed Mumtaz, Ali, Sajid, Hussain, Sartaj, Naz, Asim Falak, Hussain, Saddam
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.07.2023
• Subjects: Borehole geophysics; Fuzzy logic; Goru Reservoir; Indus Basin, Pakistan; Neural networks; PHIT curve; Fuzzy logic; Goru Reservoir; PHIT curve; Neural networks; Indus Basin, Pakistan; Borehole geophysics
• DOI: 10.1016/j.jappgeo.2023.105067
• ISSN: 0926-9851

Porosity estimation is one of the essential issues in oil and natural gas industries to evaluate the reservoir characteristics properly. Therefore, it is imperative to predict porosity with the optimum way to reduce logging tests. In this paper, fuzzy logic (FL) and neural networks (NNs) are considered effective approaches to predict the Lower Goru sand reservoir's porosity curve. The input dataset for the study contained four known logs, gamma ray (GR), neutron porosity (NPHI), density (RHOB), and sonic (DT) of five wells drilled. For the fuzzy logic model, ten bins were used. The closeness of fit (Cfit) curves were computed using the most likely and second most likely curves. The weighted average final probability Pi, or the most likely solution, was also calculated. The curve histogram distribution and set of curve bin distribution cross plots were built using a fuzzy model. In the fuzzy logic model, the Gaussian membership function provided the optimum match for the examined geophysical log data. Fuzzy logic models indicate Cfit values ranging from 94 to 100% for Sawan-01, Sawan-02, Sawan-03, Sawan-07, and Sawan-08, with standard deviations of 1.248, 1.241, 1.254, 1.336, and 1.374, respectively. The neural networks model was trained using the backpropagation (BP) algorithm. The neural networks model has a Cfit_nn of 88%–100% across five wells with standard deviations of 0.016, 0.014, 0.015, 0.017, and 0.018. The results show that the predicted modeling evaluations using fuzzy logic and neural networks techniques fit the geophysical log data quite well. The multiple linear regression (MLR) assessments were conducted using the same geophysical log datasets of five studied boreholes for comparison. The coefficients of determination (R2) for the fuzzy logic (PHIT_ml) and neural networks (PHIT_nn) models were 0.960127, and 0.973039, respectively, whereas the values of the PHIT curve for multiple linear regression (PHIT_mlr) 0.926329. The high R2 values show that fuzzy logic and neural networks are more effective methods for PHIT curve prediction than the multiple linear regression approach. The relevant correlation was derived by comparing synthetic log values to actual log values. The evaluations between recorded and predicted values applying the two distinct approaches fuzzy logic and neural networks revealed that both are effective at synthesizing PHIT logs. The confirmation of this efficiency was further verified by the low values obtained in the root mean square error (RMSE) analysis. The study conducted on the Sawan Gas Field wells revealed that both fuzzy logic and neural networks are reliable approaches for predicting the PHIT curve. By using a composite of GR, RHOB, NPHI, and DT logs, these techniques can provide a realistic fit for both actual and synthesized PHIT curves. The findings of this study suggest that the implementation of these methods can contribute to the improvement of hydrocarbon exploration and production in the region by reducing uncertainty in predicting the PHIT curve. Moreover, the methods used in this study have the potential for wider application beyond the Sawan Gas Field. These methods can be applied globally to predict the PHIT curve and evaluate the reservoir prospects. The successful application of fuzzy logic and neural networks in this paper provides a solid foundation for future research on using machine learning techniques in reservoir characterization and modeling. •FL and NNs showed potential in assessing complex conditions for porosity prediction in the Lower Goru sand reservoir.•FL and NNs provide better results than MLR approach.•In the FL model, the Gaussian membership function provided the optimum match for the examined geophysical log data.•FL and NNs effectively synthesize PHIT logs based on evaluations of real and predicted values.