Goto

Layer | Fill | Outline |
---|

Theme | Visible | Selectable | Appearance | Zoom Range (now: 0) |
---|

Fill | Stroke |
---|---|

We present a novel semi-supervised algorithm for well-log facies classification. This method considers both the guidance of domain experts and the distribution characteristics of well-log properties. The semi-supervised algorithm aims to obtain well-log facies that are geologically and seismically meaningful. We impose guidance from domain experts as pairwise constraints (must-link and cannot-link). We incorporate the constraints into facies classification in two ways: modification of the objective function and optimization of the classification subspace. We applied the method to a set of well logs from the Glitne field, North Sea. We compared the semi-supervised approach with some popular classification methods (quadratic determinant analysis and expectation-maximization with a Gaussian mixture model) to evaluate classification results. Results demonstrated that the semi-supervised approach produced facies that were more consistent with expert intention and more geologically interpretable. Presentation Date: Tuesday, October 13, 2020 Session Start Time: 1:50 PM Presentation Time: 1:50 PM Location: 360D Presentation Type: Oral

SEG-2020-3426030

algorithm, Artificial Intelligence, classification, clean sandstone 1, clean sandstone 2, constraint, exploration geophysicist 10, facies, facies classification, facies profile, log analysis, machine learning, objective function, pairwise constraint, sandstone, seg international exposition, semi-supervised algorithm, shale, shaley sandstone, silty shale, Upstream Oil & Gas, well logging, well-log facies classification

Oilfield Places:

- North America > United States > West Virginia > Appalachian Basin > Marcellus Shale (0.99)
- North America > United States > Virginia > Appalachian Basin > Marcellus Shale (0.99)
- North America > United States > Pennsylvania > Appalachian Basin > Marcellus Shale (0.99)
- (6 more...)

SPE Disciplines:

Technology: Information Technology > Artificial Intelligence > Machine Learning > Statistical Learning > Clustering (0.47)

Abstract Distal turbidites consist of thin laminations (inch-scale) usually ranging from fine sand to clay-rich deposits and may represent major hydrocarbon reservoirs: conventionally, they are studied by means of a log-based binary modeling that discriminates productive and non-productive layers. Nevertheless, the binary model represents a major drawback when dealing with laminations in the silt grain-size range, as their allotment to either end-member can be extremely problematic. This paper deals with a novel lithological facies classification approach that integrates core data and a high-resolution dielectric dispersion wireline log: its 1-inch vertical resolution and a related fit-for-purpose petrophysical model make the log tool's response suitable to describe the lithological heterogeneity of these reservoirs. The approach is presented by means of a study performed on the cored section of a well drilled into a laminated gas-bearing Pleistocene reservoir in the Adriatic Basin. A core-based classification was first carried out using sedimentological descriptions, mineralogical analyses, cation exchange capacity measurements, routine and special core analyses, and a statistical investigation of grain-size distributions: this allowed the identification of four litho-facies ranging from hemipelagite to coarse silt. Next, a log-based classification was carried out with a multivariate statistical numerical technique (cluster analysis) run on the dielectric dispersion model output curves along the cored section of the well. In the end, a four-facies log-based classification was obtained that matches the core-based classification with an overall agreement in excess of 93%. When compared to the conventional methodology, the presented approach shows the added value of identifying intermediate lithologies, thus leading to a more accurate quantification of the thickness of the potentially hydrocarbon-bearing net reservoir.

Artificial Intelligence, classification, clay content, core analysis, core-based classification, core-facies classification, dispersion, drilling operation, entropy, grain-size weight fraction 2, information, log analysis, log-facies classification, machine learning, modeling, permittivity, probability, Reservoir Characterization, salinity, Scenario, silt, structural geology, textural parameter, turbidite reservoir, Upstream Oil & Gas, water-filled porosity, well logging

Oilfield Places:

- Europe > Spain > Mediterranean Sea > Mediterranean Basin (0.99)
- Europe > Italy > Adriatic Sea > Adriatic Basin (0.99)

SPE Disciplines:

- Well Drilling > Drilling Operations > Coring, fishing (1.00)
- Reservoir Description and Dynamics > Reservoir Characterization > Exploration, development, structural geology (1.00)
- Reservoir Description and Dynamics > Formation Evaluation & Management > Open hole/cased hole log analysis (1.00)
- Reservoir Description and Dynamics > Formation Evaluation & Management > Core analysis (1.00)

Technology:

Aleardi, M. (University of Pisa) | Calabrò, R. A. (Edison R,D&I) | Ciabarri, F. (Edison R,D&I) | Garcea, B. (Edison E&P) | Giussani, M. (Edison E&P) | Peruzzo, F. (Edison E&P) | Terdich, P. (Edison E&P) | Mazzotti, A. (University of Pisa)

**ABSTRACT **

Recent advances in seismic-constrained reservoir characterization combine statistical rock-physics and amplitude versus offset/angle (AVO/AVA) inversion in order to directly estimate petrophysical properties such as porosity, shaliness and water saturation from pre-stack seismic data. By exploiting the Bayesian inversion formalism, it is possible to propagate uncertainty from seismic to petrophysical properties, including the effect of noise on seismic data and the approximation of physical models. The results of such petrophysical-seismic inversion are spatial probability density distributions of rock and fluid properties that can be effectively integrated in the reservoir modeling workflows. This paper discusses two target-oriented Bayesian petrophysical-AVA inversion techniques: a *two-stage *approach and a *single-stage *approach, developed as part of a collaborative research project between Edison and the Earth Sciences Department of the University of Pisa. The two approaches are evaluated on the gas-bearing sands of the Pliocene interval in the Northern area of the offshore Abu Qir field where a 3D seismic survey was acquired using long-offset cables and well-control is available to validate the inversion results. The *two-stage *approach, is performed over the whole target-interval and is based on two cascade steps: first, seismic angle-gathers are inverted into acoustic and shear impedances using the convolutional model and a narrow-angle, time-continuous approximation of the Zoeppritz equations; then, a rock-physics model is used to transform the elastic parameters into petrophysical properties. Differently, the *single-stage *approach uses the rock-physics model to re-parameterize the exact Zoeppritz equations in terms of petrophysical variables; the derived equations are used to directly estimate the petrophysical property along the top-horizon of target interval by taking into account wide-angle seismic reflections. Independently from the inversion approach considered, the rock-physics model plays a crucial role in petrophysical-AVA inversion as it provides the link between elastic and petrophysical properties. In the Abu Qir field, borehole data acquired at the target-depths were exploited to derive a single rock-physics model, valid for different lithologies and for the full-ranges of shaliness and water saturation values. Despite the differences in the forward-model parameterization, the results of the two inversions are comparable and consistent with borehole data. In particular, the described inversion approaches were both able to identify the increase of porosity and the decreases of shaliness and water saturation in the target sands. It results that porosity is well resolved by both *two-stage *(narrow-angle) and *single-stage *(wide-angle) inversions. The water saturation remains poorly resolvable in both inversions due to its limited influence in determining the AVA response. Finally, wide-angle reflection inversion has demonstrated to be particularly effective in better constrain the shaliness estimations.

a-posteriori pdf, a-priori pdf, Artificial Intelligence, coefficient, equation, estimation, information, inversion, inversion technique, machine learning, offshore abu qir field, petrophysical property, petrophysical-seismic inversion, petrophysical-seismic inversion technique, porosity, Reservoir Characterization, RPM, seismic data, shaliness, Upstream Oil & Gas, water saturation

Oilfield Places: Africa > Middle East > Egypt > Nile Delta > Nile Delta Basin > West Mediterranean Concession > Abu Qir Field (0.99)

SPE Disciplines: Reservoir Description and Dynamics > Reservoir Characterization > Seismic processing and interpretation (1.00)

Technology: Information Technology > Artificial Intelligence > Machine Learning > Statistical Learning (1.00)

ABSTRACT Cement bond evaluation is a critical step in the early-life stages of newly drilled wells since it rules the way for obtaining useful information about wellbore integrity. Conventionally, this is carried out by means of a detailed interpretation of cased-hole sonic and ultrasonic log data. However, this standard approach can be highly time-consuming and challenging in long completion sections and when complex scenarios have to be handled in operative time. In this respect, oil companies have stored huge datasets for their wells, with quality-checked cased-hole acoustic logs and associated interpretations in terms of wellbore integrity. This paper deals with a novel, probabilistic data analytics approach aimed at obtaining a fast and robust cement bond facies classification. The latter is deemed able to automatically provide an exhaustive quantitative cement placement evaluation, hence avoiding time-consuming processes and possible subjectivity issues. The implemented methodology takes advantage of the Multi-Resolution Graph-based Clustering (MRGC) algorithm that gathers its knowledge by recognizing patterns in sonic and ultrasonic logs/maps from dozens of wells, including more than 500K meters of logged intervals. This allows the system to learn through experience how the log measurements are related to the common cement bond scenarios (e.g. good, partial, poor cementation, dry or wet microannulus, free pipe). The MRGC is then integrated in a Bayesian framework to obtain the probability of the cement bond facies, the most probable scenarios, and the associated uncertainty by means of entropy computation. In detail, an automated screening can be performed in newly drilled wells to detect possible problems of hydraulic sealing. The potentialities of the discussed method are demonstrated by real case applications consisting of cement log data collected from several blind-test wells. First, the probabilistic approach is used to predict the cement bond scenarios together with the uncertainties of their classification. Then, an unbiased evaluation of the results is performed. The successful outcomes coming from the final step of the workflow show how, with a statistically representative and good quality dataset, data analytics can efficiently mimic high-skill expert work in harsh circumstances and within a time-efficient template. In fact, this data-driven methodology takes few seconds to provide an exhaustive interpretation against, at least, one day with the conventional one.

acquisition, Artificial Intelligence, Bayesian Inference, casing and cementing, CBL amplitude, cement and bond evaluation, cement job, classification, facies, facies classification, fraction, free pipe, interpretation, machine learning, multi-resolution graph-based clustering, probability, Scenario, spwla 61, spwla-5060, st annual logging symposium, University, Upstream Oil & Gas

SPE Disciplines:

Technology:

- Information Technology > Data Science (1.00)
- Information Technology > Artificial Intelligence > Machine Learning > Statistical Learning (0.68)
- Information Technology > Artificial Intelligence > Representation & Reasoning > Uncertainty > Bayesian Inference (0.49)
- Information Technology > Artificial Intelligence > Machine Learning > Learning Graphical Models > Directed Networks > Bayesian Learning (0.49)

**Summary**

Seismic reservoir characterization aims to provide an accurate reservoir description of rock and fluid properties estimated from seismic data. However, in several applications, seismic data only, cannot accurately discriminate the fluid effect, and the integration of other geophysical measurements, such as electromagnetic data, is required to improve the reservoir description. In this work, we propose a joint rock physics inversion to estimate porosity and fluid saturations from seismic velocity and electrical resistivity. The method is based on a Bayesian approach to inverse modeling and combines inverse theory and statistical rock physics relations. The advantages of this approach are the joint estimation of rock properties, achieved by a coupled rock physics model, and the estimation of the uncertainty associated to the predicted model, achieved through the Bayesian approach. The method has been applied to a real dataset, the Rock Spring Uplift field in Wyoming, a CO_{2} sequestration study.

**Introduction**

The goal of seismic reservoir characterization is to provide a reliable model of the reservoir, in terms of rock properties, such as porosity and lithology, and fluid saturations. In rock physics models, when rock properties are known, we can predict the effect of fluid saturations on P-wave and S-wave velocity and density (Mavko et al., 2009; and Dvorkin et al., 2014). However, the solution of the inverse problem, i.e. the estimation of rock and fluid properties from velocities and density, is generally a challenging task (Avseth et al., 2005; and Doyen, 2007). Indeed, the solution of the inverse problem is not necessarily unique: two different rocks could have different porosities, lithologies and fluids, and the same elastic response. Furthermore, when the inverse problem is solved using seismic data instead of well log data, the low resolution and low signal-to-noise ratio of the data often increase the uncertainty in the estimation of seismic velocities and density, which makes the rock-fluid property estimation more challenging. To improve the reservoir description and reduce the associated uncertainty, we propose to integrate electromagnetic (EM) data, together with seismic attributes, in the reservoir modeling workflow (Du and MacGregor, 2010; MacGregor, 2012).

Artificial Intelligence, Bayesian Inference, bayesian rock physics inversion, classification, electrical property, estimation, facies, facies classification, inversion, log data, machine learning, madison formation, plot show, porosity, posterior distribution, posteriori, probability, Reservoir Characterization, resistivity, rock physics inversion, rock physics model, Upstream Oil & Gas

Oilfield Places:

- North America > United States > West Virginia > Appalachian Basin > Marcellus Shale (0.99)
- North America > United States > Virginia > Appalachian Basin > Marcellus Shale (0.99)
- North America > United States > Pennsylvania > Appalachian Basin > Marcellus Shale (0.99)
- (7 more...)

SPE Disciplines: Reservoir Description and Dynamics > Reservoir Characterization > Seismic processing and interpretation (1.00)

Technology:

- Information Technology > Artificial Intelligence > Representation & Reasoning > Uncertainty > Bayesian Inference (0.69)
- Information Technology > Artificial Intelligence > Machine Learning > Learning Graphical Models > Directed Networks > Bayesian Learning (0.69)
- Information Technology > Artificial Intelligence > Machine Learning > Statistical Learning (0.68)

Thank you!