Strategic Geosteeering Workow with Uncertainty Quantication and Deep Learning A Case Study on the Goliat Field

2025-05-02 0 0 2.37MB 40 页 10玖币

侵权投诉

Strategic Geosteeering Workﬂow with Uncertainty Quantiﬁcation

and Deep Learning:

A Case Study on the Goliat Field

Muzammil Hussain Rammay1, Sergey Alyaev2, David Selv˚ag Larsen3, Reidar Brumer

Bratvold1, and Craig Saint4

1University of Stavanger, Stavanger, Norway

2NORCE Norwegian Research Centre, Bergen, Norway

3V˚ar Energi, Stavanger, Norway

4Baker Hughes, Bergen, Norway

October 28, 2022

Abstract

The real-time interpretation of the logging-while-drilling data allows us to estimate the po-

sitions and properties of the geological layers in an anisotropic subsurface environment. Ro-

bust real-time estimations capturing uncertainty can be very useful for eﬃcient geosteering

operations. However, the model errors in the prior conceptual geological models and forward

simulation of the measurements can be signiﬁcant factors in the unreliable estimations of the

proﬁles of the geological layers. The model errors are speciﬁcally pronounced when using a deep-

neural-network (DNN) approximation which we use to accelerate and parallelize the simulation

of the measurements. This paper presents a practical workﬂow consisting of oﬄine and online

phases. The oﬄine phase includes DNN training and building of an uncertain prior near-well

*Corresponding authors: Muzammil Hussain Rammay (muzammil.h.rammay@uis.no) and Sergey Alyaev

(saly@norceresearch.no)

arXiv:2210.15548v1 [physics.geo-ph] 27 Oct 2022

geo-model. The online phase uses the ﬂexible iterative ensemble smoother (FlexIES) to perform

real-time assimilation of extra-deep electromagnetic data accounting for the model errors in the

approximate DNN model. We demonstrate the proposed workﬂow on a case study for a historic

well in the Goliat Field (Barents Sea). The median of our probabilistic estimation is on-par

with proprietary inversion despite the approximate DNN model and regardless of the number

of layers in the chosen prior. By estimating the model errors, FlexIES automatically quantiﬁes

the uncertainty in the layers’ boundaries and resistivities, which is not standard for proprietary

inversion.

Keywords: Model error; Deep neural networks; Real-time interpretation; Flexible iterative

ensemble smoother; Historical ﬁeld case; Strategic Geosteering; Uncertainty quantiﬁcation;

1 Introduction

Recent advances in logging-while-drilling (LWD) technology allow us to sense the subsurface envi-

ronment tens of meters away from the well using a suit of extra-deep azimuthal resistivity (EDAR)

logs. These logs can be inverted in real-time to estimate the geo-layer proﬁle during drilling [Sey-

doux et al., 2014, Sviridov et al., 2014]. However, achieving ”strategic” geosteering [Arata et al.,

2016] requires the integration of the measurements into a subsurface model, which includes the

relevant geological uncertainties.

Ensemble-based methods such as the Ensemble Kalman ﬁlter and iterative ensemble smoothers

present a statistically-consistent Bayesian framework for assimilating data to update an uncertain

model [Alyaev et al., 2019]. However, their performance relies on thousands of parallel executions

of forward models of the obtained measurements [Jahani et al., 2022, Rammay et al., 2022]. The

thousands of function evaluations using expensive high-ﬁdelity forward models are computationally

expensive in real-time workﬂows or require special infrastructure [Dupuis and Denichou, 2015]. Re-

cently introduced deep-neural-network (DNN) proxies showed robust approximation for modelling

azimuthal EM measurements [Kushnir et al., 2018, Shahriari et al., 2020], including the deepest

sensing EDAR measurements [Alyaev et al., 2021a, Noordin et al., 2022]. Most of the computational

cost for the DNN can be oﬄoaded to oﬄine training, thus giving superior performance compared

to Maxwell’s equations solvers during operational use. real-time statistical data-assimilation work-

ﬂows with EDAR data required for strategic geosteering [Alyaev et al., 2021b, Fossum et al., 2022,

Jahani et al., 2021, Noordin et al., 2022, Rammay et al., 2022].

Although DNNs bring sub-second model performance, they come with additional unknown

model errors. In realistic scenarios, the negligence of the uncertainties related to the model er-

rors during modelling will result in an unreliable estimation of model parameters and uncertainties

[Oliver and Alfonzo, 2018, Rammay and Alyaev, 2022, Rammay et al., 2021]. Achieving practically

usable results for the real-time data assimilation with proxy models requires algorithms that can

automatically account for model errors and detect possible multi-modality. Recently, ﬂexible itera-

tive ensemble smoother (FlexIES) introduced in Rammay et al. [2021] showed good performance in

the synthetic tests for assimilating EDAR data modelled by a DNN where model errors were coming

from the inaccuracies in the DNN approximation of the forward model [Rammay et al., 2022]. In a

realistic setting, additional model errors will come from inaccuracies in the physical simulation and

the mismatch between the chosen geomodel and real, complex geology. Moreover, data mismatch

can be due to the local minima (local or multiple modes), which might be misinterpreted as data

errors by IES-type algorithms [Rammay et al., 2022].

This paper describes a complete FlexIES-based workﬂow for strategic Bayesian geosteering with

steps to account for model errors and multi-modality. It consists of oﬄine and online phases. The

oﬄine phase takes care of the determination of one or several geological priors consistent with an

ensemble method and training a DNN proxy model. The online phase combines the FlexIES with

the DNN model to assimilate the real-time EDAR data. The workﬂow is veriﬁed on data from a

historical operation in the Goliat ﬁeld and compared to a deterministic inversion delivered by the

service company post-job [Larsen et al., 2015]. The objectives of this case study are:

1. Demonstrate that the DNN from Alyaev et al. [2021a] can be retrained to handle anisotropic

ﬁeld data and then be used for probabilistic real-time interpretation;

2. Use FlexIES to estimate probabilistic layered geomodel from ﬁeld data automatically account-

ing for model errors coming from geomodel and the DNN;

3. Compare the FlexIES results with the proprietary deterministic inversion.

The rest of the paper is organised as follows: Section (2) introduces the workﬂow and describes

the steps in its oﬄine and online phases in more detail. The workﬂow is applied to a reservoir

section of a well in the Goliat ﬁeld using the available historical data, as described in Section (3).

The conclusions of this paper are summarised in Section (4).

2 Workﬂow

In this paper, we propose an ensemble-based workﬂow for assimilating borehole electromagnetic

measurements to estimate proﬁles of geo-layers along with associated uncertainties in real-time

to support Geosteering operation. This workﬂow consists of oﬄine pre-job stage and an online

real-time inversion stage as shown in Figure 1.

The oﬄine phase enables the real-time data assimilation. We start the pre-job phase by con-

structing a geologically-relevant ensemble of prior geomodel realizations, see Figure 1.1. The sam-

pled 1.5D layering conﬁgurations from the prior (Figure 1.3), are used for training a DNN (Figure

1.2), that can model the suite of EDAR measurements [Sviridov et al., 2014] in milliseconds. A

possibly constrained set of realizations (Figure 1.5), is used as the prior for the real-time data

assimilation loop.

For the online phase we use an Ensemble Kalman Filter (EnKF) type method, namely Flexible

Iterative Ensemble Smoother (FlexIES), see Figure 1.4. FlexIES compares the synthetic mea-

surements modelled for realizations of the prior ensemble of geomodels with the real-time EDAR

measurements in the data assimilation update loop. The loop integrates the measurements and

reduces the uncertainty in the ensemble yielding the posterior. The posterior realizations can be

further used for real-time decision support [Alyaev et al., 2019]. In case of a ﬁlter-type sequential

data assimilation the posterior serves as the prior for assimilation of future data [Chen et al., 2015].

Ensemble of

prior

realizations of

geomodels

DNN model for

measurements

Modelled

measurements 4. FlexIES Realtime LWD

measurements

Posterior

realizations of

geomodels

with reduced

uncertainty

Constructing

geologically-

relevant

prior

5. Solving possible

inversion problems due

to multi-modality by

constraining the prior

3. Creating

relevant

training

dataset

2. Training deep neural

network (DNN)

Real-time

decision

support

Offline pre-job stage Online real-time data assimilation stage

Data Assimilation Loop

Setting new prior for EnKF-type methods

Not implemented in this paper

N. Workflow element described in Section 2.N.

Figure 1: The full workﬂow for preparation and real-time data assimilation during drilling.

We describe how the workﬂow is applied to a part of a historical drilling operation, and point

out possible modiﬁcations needed to apply it for a real operation. The rest of the section describe

the major building blocks of the workﬂow with numbering that follows Figure 1.

2.1 Constructing geologically-relevant prior

The ﬁrst step of the workﬂow is related to the prior description of the geo-model. This can be

done by utilizing the prior knowledge or experience of the geologists or the interpretation of the

data set from the oﬀset wells. In the considered historical case the geology can be represented by a

layer-cake model with continuous layers Larsen et al. [2015]. Thus, we describe the geo-model by a

number of layers, the 2D proﬁle of each layer, and the anisotropic layer’s resistivities. For example,

three layers geo-model and four layers geo-model are shown in Figure 6.

The prior uncertainties of a geo-model are described by prior realizations of the geo-layer proﬁles

(estimated as boundary positions or layers thicknesses) and layers’ resistivities. We model the

prior realizations of the thickness proﬁles of the geo-layers as multivariate Gaussian with mean

(µ= [35,30,20]) and an exponential covariance function:

c=σ2exp (−3h

l) (1)

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

StrategicGeosteeeringWorkowwithUncertaintyQuanticationandDeepLearning:ACaseStudyontheGoliatFieldMuzammilHussainRammay1,SergeyAlyaev2,DavidSelvagLarsen3,ReidarBrumerBratvold1,andCraigSaint41UniversityofStavanger,Stavanger,Norway2NORCENorwegianResearchCentre,Bergen,Norway3VarEnergi,Stavanger,Norway...

收起<<

Strategic Geosteeering Workow with Uncertainty Quantication and Deep Learning A Case Study on the Goliat Field.pdf

共40页,预览5页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Strategic Geosteeering Workow with Uncertainty Quantication and Deep Learning A Case Study on the Goliat Field

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: