Abstract:Methods for predicting mineralogy from logging-tool measurements have been an active area of research for several decades. In spite of these efforts, methods for predicting quantitative mineralogy including clay types from well-logging data were not fully achieved. The introduction of geochemical logging tools in the 1980s offered promise; however, early versions of geochemical logging tools did not measure elemental chemistry with enough accuracy and precision to enable reliable and quantitative determination… Show more
“…One method in the literature to overcome overfitting problems is by using regularization (Orr, 1996;Freedman, 2006;Freedman et al, 2014;Trevizan et al, 2014). In this paper, a forward selection method (Orr, 1996) is used to address overfitting problems.…”
Section: Brief Description Of Rbf Methodsmentioning
Characterizing heavy oil viscosity by nuclear magnetic resonance (NMR) relaxation time (T 1 and T 2 ) measurements is much more challenging than characterizing light oil viscosities. Crude oils contain a wide range of hydrocarbons, resulting in broad T 1 and T 2 distributions that vary with the oil composition. Most often, a single geometric mean value T 1;gm or T 2;gm is correlated with the crude oil viscosity, which cannot accurately account for the inherent complexity of the oil constituent information. Furthermore, as the viscosity increases, some of the protons in the oil relax too quickly to be observable by logging or laboratory NMR instruments. This results in deficiencies of relaxation time and signal amplitude that give rise to apparent T 1 and T 2 distributions (T 1;app and T 2;app ) and apparent hydrogen index (HI app ). Using T 1;app and T 2;app distributions in NMR viscosity models could produce erroneous heavy oil viscosity estimations. Several attempts have been made to overcome these challenges by taking into account HI app at a fixed interecho time (TE), or a TE-dependent HI app . We have developed a new radial-basis-function-based heavy oil viscosity model using the entire T 2;app distribution, rather than T 2;gm , with an option of including the NMR-derived HI app . Because both of these quantities are TE dependent, it is desirable to include multiple TE data to develop the model. In addition, the principal component analysis (PCA) method was applied to extract major variations of features embedded in the T 2;app distributions, while discarding distribution features that are derived from random noise. The coefficients of the RBFs were derived using laboratory NMR T 2 measurements at ambient and elevated temperatures between 23.5°C and 39.5°C and corresponding viscosity measurements on 50 oil samples. These oil samples were collected from different parts of a shallow viscous oil reservoir in Kuwait. It was observed that the use of this newly developed RBF method showed significant improvement in terms of the reliability of the viscosity prediction compared to some recently published heavy oil viscosity correlations.
“…One method in the literature to overcome overfitting problems is by using regularization (Orr, 1996;Freedman, 2006;Freedman et al, 2014;Trevizan et al, 2014). In this paper, a forward selection method (Orr, 1996) is used to address overfitting problems.…”
Section: Brief Description Of Rbf Methodsmentioning
Characterizing heavy oil viscosity by nuclear magnetic resonance (NMR) relaxation time (T 1 and T 2 ) measurements is much more challenging than characterizing light oil viscosities. Crude oils contain a wide range of hydrocarbons, resulting in broad T 1 and T 2 distributions that vary with the oil composition. Most often, a single geometric mean value T 1;gm or T 2;gm is correlated with the crude oil viscosity, which cannot accurately account for the inherent complexity of the oil constituent information. Furthermore, as the viscosity increases, some of the protons in the oil relax too quickly to be observable by logging or laboratory NMR instruments. This results in deficiencies of relaxation time and signal amplitude that give rise to apparent T 1 and T 2 distributions (T 1;app and T 2;app ) and apparent hydrogen index (HI app ). Using T 1;app and T 2;app distributions in NMR viscosity models could produce erroneous heavy oil viscosity estimations. Several attempts have been made to overcome these challenges by taking into account HI app at a fixed interecho time (TE), or a TE-dependent HI app . We have developed a new radial-basis-function-based heavy oil viscosity model using the entire T 2;app distribution, rather than T 2;gm , with an option of including the NMR-derived HI app . Because both of these quantities are TE dependent, it is desirable to include multiple TE data to develop the model. In addition, the principal component analysis (PCA) method was applied to extract major variations of features embedded in the T 2;app distributions, while discarding distribution features that are derived from random noise. The coefficients of the RBFs were derived using laboratory NMR T 2 measurements at ambient and elevated temperatures between 23.5°C and 39.5°C and corresponding viscosity measurements on 50 oil samples. These oil samples were collected from different parts of a shallow viscous oil reservoir in Kuwait. It was observed that the use of this newly developed RBF method showed significant improvement in terms of the reliability of the viscosity prediction compared to some recently published heavy oil viscosity correlations.
“…1957년 처음 소개되어 케이싱 내에서 포화도를 추정하는 검층 법으로 또는 지화학검층(Geochemical log)법으로 최근에는 비 전통저류층 평가에 중요한 검층법으로 적용되고 있다 (Baker, 1957;Muench and Osoba, 1957;Lock and Hoyer, 1974;Freeman et al, 2014 (Ku et al, 2012a(Ku et al, , 2012bWon et al, 2013Won et al, , 2014 …”
Section: 이다 이 검층법은 지층의 구성 원소 평가가 가능한 검층으로unclassified
For efficiently designing neutron induced gamma spectroscopy sonde, Monte Carlo simulation is employed to understand a dominant location of thermal neutron and classify the formation elements from the energy peak of capture gamma spectrum. A pulsed neutron generator emitting 14 MeV neutron particles was used as a source, and flux of thermal neutron was calculated from the twelve detectors arranged at each 10 cm intervals from the source. Design for reducing borehole effects using shielding materials was also applied to numerical sonde model. Moreover, principal elements and quantities of numerical earth models were verified through the energy spectrum analysis of capture gamma detected from a gamma detector. These results can help to enhance the signal-to-noise ratio, and determine an optimal placement of capture gamma detectors of neutron induced gamma spectroscopy sonde.
“…In hydrocarbon reservoir characterization, the correct estimation of the rock matrix's mineral fractions is crucial in quantifying hydrocarbons. In studies related to reservoirs, in addition to directly impacting the calculation of porosity (Freedman et al., 2015), the knowledge of mineral composition can significantly improve the water saturation calculation (Clavier et al., 1984; Garcia et al., 2017; Poupon & Leveaux, 1971; Waxman & Smits, 1968) and estimate the cation exchange capacity (Herron, 1986) and total organic carbon (Gonzalez et al., 2013). Additionally, reliable mineralogy estimates can aid in monitoring a well's hydrocarbon/water contact during its productive life (North, 1987; Ulloa et al., 2016; Westaway et al., 1983) and aid in acid fracturing operations (Jin et al., 2019).…”
Well mineralogy can be estimated from probabilistic, direct and machine learning models; however, all these models have limitations. The maximum number of components in probabilistic models is restricted to the number of logs plus one. Direct models require the precise composition of minerals. Machine learning models demand unbiased databases, a challenge as the samples are collected in reservoir intervals. These limitations impact the evaluation for the Santos Basin pre‐salt rocks due to the complexity of facies and magnesian clays. This work proposes creating a hybrid model through the combination of probabilistic and machine learning models. First, mineral fractions of calcite, dolomite, quartz, k‐feldspar, detrital clay, plagioclase and pyroxene are estimated by the algorithm XGBoost trained using rock samples. Then, a probabilistic model reconstructs the well logs and machine learning estimations through the seven minerals mentioned plus magnesian clays, pyrite, barite and fluids. The difference between the real and reconstructed responses is minimized, weighted by the curves’ uncertainties. The hybrid model is used to estimate the mineralogy of three wells drilled in the Santos Basin, honouring the mineralogy of the rock samples collected in these wells and improving the quantification of dolomite, pyroxene and magnesian clay. Among the advances introduced, the following stand out: The use of machine learning estimates and well logs improved the quantification of magnesian clay; the machine learning estimates regularized the probabilistic model, generating more coherent results; the uncertainties of the machine learning algorithms dealt with database bias. The hybrid model mitigated limitations related to database bias without the costs associated with collecting more samples.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.