Practical content of the elements of petrophysical model of terrigenous sandstones — oil and gas reservoirs in neural networks, deep learning and regression methods

• Date: 2025
• Main Authors: Vyzhva, Serhiy, Gozhyk, Andriy, Shabatura, Oleksandr, Onyshchuk, Viktor, Onyshchuk, Dmytro, Onyshchuk, Ivan
• Format: Article
• Language: Ukrainian
• Published: S. Subbotin Institute of Geophysics of the NAS of Ukraine 2025
• Subjects: petrophysics; terrigenous oil and gas reservoirs; multiple regression; machine learning; neural networks
• Online Access: https://journals.uran.ua/geofizicheskiy/article/view/309312
• Journal Title: Geofizicheskiy Zhurnal

An important problem in the search for oil and gas fields is the ability to predict key petrophysical properties such as porosity, permeability, etc. Along with traditional regression analysis, neural network methods and deep learning technologies are becoming increasingly common. All of them require verification of the efficiency of the petrophysical model, i.e. the ability to correctly predict the desired value with the least errors based on selected sets of independent petrophysical data. The object of study was samples of Lower Carboniferous sandstones of deep horizons (interval 4931—5879 m) from 14 wells of promising formations of the northwestern part of the Dnipro-Donetsk Basin (Bakumivska, Zorkivska, Voloshkivska, Komyshnyanska, Chervonozavodska, Lutsenkivska, Piskivska and Chervonolutska areas). Nine families of geological and petrophysical characteristics were used as independent search features, for which effective approximations of multiple regression were obtained, and their informative weight was determined. In total, 38 empirical regression equations were obtained that can be used to predict the key reservoir characteristics of terrigenous reservoir rocks (effective porosity, permeability, residual water saturation, etc.). The residual water saturation ratio and effective porosity are traditionally effectively predicted by a simple linear regression model using petrodensity, petroelectric, petro-velocity and geochemical attributes. Carbonate and structural features can only be used in a piecewise linear regression model for predicting the residual water saturation factor. All these regression equations usually have small errors. Practical analysis of the behavior of the composition features indicates the importance of using Na2O, TiO2 and Fe2O3 oxides, which most likely convey the influence of mineralized solutions, chemical composition of the cement aggregate and films on the surface of mica minerals — siderite, iron oxides-hydroxides and ore mineral. Three families of features in the predictive model of the residual water saturation coefficient: petrodensity, carbonate content, and structural, give a close value of the critical point of the predictive response (kc.v=0.39), which is probably critical for these terrigenous reservoirs, since this value indicates the limit of impact of conditional bound water. The authors have formed regression relationships in all selected models between all families of traits and the permeability coefficient. It was found that the use of a nonlinear regression model significantly increases the level of its reliability compared to the reliability of traditional linear models. For example, the geochemical predictors in the linear regression model for predicting effective porosity have a low value of the explained variance (75 %), which is on the verge of statistical reliability. At the same time, the SVM provides a reliable correlation of geochemical characteristics and permeability at the level of 90 %; the active role of SiO2, sodium and chlorine was established. All implementations of linear regression models of the porosity parameter in reservoir conditions and the proportion of supercapillary pores in the total volume of voids have increased and/or high modeling errors, except for the family of petroelectric and geochemical features. The preliminary results showed that neural network and deep learning methods outperform traditional regression analysis in terms of prediction accuracy and can effectively cope with uncertainty in test results, making it one of the most effective tools for petrophysical modeling and prediction.