Download as pdf or txt
Download as pdf or txt
You are on page 1of 15

Exploration Geophysics

ISSN: 0812-3985 (Print) 1834-7533 (Online) Journal homepage: https://www.tandfonline.com/loi/texg20

Probabilistic reservoir characterisation using 3D


pdf of stochastic forward modelling results in
Vincent oil field

Junhwan Choi, Soyoung Kim, Bona Kim & Joongmoo Byun

To cite this article: Junhwan Choi, Soyoung Kim, Bona Kim & Joongmoo Byun (2020) Probabilistic
reservoir characterisation using 3D pdf of stochastic forward modelling results in Vincent oil field,
Exploration Geophysics, 51:3, 341-354, DOI: 10.1080/08123985.2019.1696151

To link to this article: https://doi.org/10.1080/08123985.2019.1696151

Published online: 02 Dec 2019.

Submit your article to this journal

Article views: 45

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at


https://www.tandfonline.com/action/journalInformation?journalCode=texg20
EXPLORATION GEOPHYSICS
2020, VOL. 51, NO. 3, 341–354
https://doi.org/10.1080/08123985.2019.1696151

Probabilistic reservoir characterisation using 3D pdf of stochastic forward


modelling results in Vincent oil field

Junhwan Choi a , Soyoung Kimb , Bona Kimc and Joongmoo Byuna


a RISE.ML (Reservoir Imaging with Seismic & EM technology. Machine Learning), Hanyang University, Seoul, Korea; b KOGAS (Korea Gas
Corporation), Daegu, Korea; c KIGAM (Korea Institute of Geoscience and Mineral Resources), Daejeon, Korea

ABSTRACT ARTICLE HISTORY


Reservoir characterisation using a crossplot of elastic properties can be used to determine fluid Received 16 November 2019
and lithology in a seismic survey area. P-impedance and P-wave/S-wave ratios are commonly Accepted 16 November 2019
used as axis parameter for a 2D crossplot. To achieve this goal, the fluid and lithology must first
be identified using well log data. However, when the well log data are too sparse, they cannot
encompass the full suite of reservoir properties. Stochastic forward modelling (SFM) methods
have been developed to overcome the sparseness problem. However, when the results of SFM
are plotted on the 2D crossplot, the augmented data for different facies frequently overlap. To
overcome this problem, we propose a probabilistic reservoir characterisation using 3D crossplot-
ting of the SFM results. Axis-parameters of the 3D crossplot consist of the seismic attributes by
which the facies are distinguished. The acoustic impedance (Ip ), pseudo gamma ray (GR) log,
and pseudo water saturation (Sw ) log were used as the axis parameters of the 3D crossplot.
To perform SFM, pseudo GR and pseudo Sw log data must be expressed mathematically with
well log data. Linear multi-regression analysis was used to derive the mathematical relation-
ships of the different parameters. The probability distributions of the pseudo GR and pseudo
Sw logs were extracted using these relationships. Using the probability distributions of the Ip ,
pseudo GR log, and pseudo Sw log, the data were augmented by Monte Carlo simulation. The
trivariate probability density function (3D PDF) of each facies was determined by the mean
and covariance of the augmented data. The pseudo GR log and pseudo Sw log volumes were
extracted using a probabilistic neural network. Finally, a Bayesian inference was applied to calcu-
late the facies probabilities using the 3D PDFs. We confirmed that the proposed method is more
effective than the conventional reservoir characterisation method using 2D crossplot of SFM
results.

Introduction to determine the distribution of petrophysical facies.


Two-dimensional (2D) crossplotting of elastic proper- Goodway, Chen, and Downton (1997) introduced λρ,
ties, such as P-impedance (Ip ) and the P-wave/S-wave µρ, and λ/µ from Ip and S-impedance (Is ) to improve
ratio (V p /V s ), is commonly used to analyse the dis- the discrimination of lithology and pore fluid distri-
tribution of lithology and pore fluid around wells to bution and to extract quantitative information. Close,
increase the possibility of success in the exploration Taylor, and Nixon (2015) and Reine (2014) predicted
of hydrocarbon reservoirs, as it shows the geologi- lithology and pore fluid distribution using a rock physics
cal structure and indicates the presence of hydrocar- template depicted on a λρ–µρ crossplot. However, the
bon. Various crossplotting methods including ampli- crossplotting method has an overlap problem of facies
tude variation with offsets analysis (Ross 2000; Chen clusters in the oil reservoir. To address the limitation,
et al. 2001; Whitcombe and Fletcher 2001; Avseth, Muk- the statistical methods of crossplot of seismic attributes
erji, and Mavko 2005), have been developed. Chi and have been proposed to estimate the facies probabil-
Han (2009) predicted lithology and pore fluid distribu- ity and quantify the uncertainty (Gallop 2006; Ng et al.
tion based on a rock physics template designed using 2008; Michelena, Godbey, and Rodrigues 2010). Further-
the 2D crossplot of Ip and V p /V s , which links the elas- more, to improve the separability of facies clusters, Ross
tic properties to petrophysical parameters such as water and Sparlin (2000) and Chopra, Alexeev, and Xu (2003)
saturation (Sw ), porosity (φ), and shale volume (V sh ). expanded the 2D crossplot to a three-dimensional (3D)
Furthermore, Goodway, Chen, and Downton (1997) crossplot, making the anomaly more distinguishable.
and Close, Taylor, and Nixon (2015) used lambda–rho However, when well log data are too sparse on the
(λρ) and mu–rho (µρ) as crossplot axis parameters crossplot, they cannot encompass the variation in the

CONTACT Joongmoo Byun jbyun@hanyang.ac.kr RISE.ML (Reservoir Imaging with Seismic & EM technology. Machine Learning), Hanyang
University, Seoul, Korea

© 2019 Australian Society of Exploration Geophysicists


342 J. CHOI ET AL.

reservoir properties; thus, the accuracy of calibration


with seismic data is decreased, causing uncertainty in
the interpretation of the data.
SFM (i.e. data augmentation) methods have been
developed to address the problem of sparse well log
data (Lamont, Thompson, and Bevilacqua 2008). The
bivariate normal probability density function (2D PDF),
which provides probability information for each facies,
can be built from the augmented data. When 2D PDFs
and seismic inversion results, such as Ip and V p /V s ,
are applied to a Bayesian inference method (Doyen
2007), the facies probability can be calculated. Lam-
ont, Thompson, and Bevilacqua (2008) used stochas-
tically modelled multidimensional crossplots to build
2D PDFs (Ip versus V p /V s ) and quantify the uncertainty
in a gas reservoir field based on the Bayesian infer-
ence method. They applied the SFM method and used
a depth-dependent rock physics model to analyse the Figure 1. (a) Workflow of the conventional probabilistic reser-
depositional environment and burial history. They suc- voir characterisation using the two-dimensional crossplot of
SFM results and (b) the original data (upper panel); the aug-
cessfully predicted the lithology and pore fluid distri- mented data, where colours indicate the facies (middle panel);
bution away from well control in a way that effectively and the bivariate probability density functions of the facies
quantifies risk. Nieto, Batlai, and Delbecq (2013) pre- (lower panel). Ip , P-impedance; V p /V s , P-wave/S-wave ratio.
dicted facies in an unconventional reservoir probabilis-
tically using the Bayesian inference method. They cal- Conventional approach of SFM
culated the proportion of each facies in the well log In areas where petrophysical analysis is difficult because
data to use as a priori probabilities. They built the bivari- of sparse well log data, the conventional SFM method
ate probability mass functions of the facies using the using a rock physics model and Monte Carlo simulation
well log data and adopted the confusion matrix, which augments the known well log data (Lamont, Thompson,
is a key tool for ensuring the quality of the prediction. and Bevilacqua 2008). This method is an extension of
Pendrel, Schouten, and Bornard (2017) also performed rock physics modelling that predicts the V p , V s , and den-
Bayesian facies estimation, using 3D priors defined from sity (ρ) using petrophysical properties, such as φ, V sh ,
the facies probability curves for wells for the Bayesian and Sw . More specifically, as shown in Figure 1(a), the
method of the inversion properties. Grana (2018) used probability distributions of the petrophysical parame-
Bayesian inference using 2D PDFs for facies and reser- ters are used as input parameters for the rock physics
voir properties inversion. The 2D non-parametric PDFs model. Monte Carlo sampling is then performed based
defined in elastic and petrophysical domains are used on the rock physics model to simultaneously augment
as prior information. the multi-dimensional datasets such as V p , V s , ρ, Ip and
However, the 2D PDFs of facies overlap on 2D cross- V p /V s (Figure 1(b), middle panel). When performing the
plots, rendering a clear distinction of facies difficult Monte Carlo simulation, the Iman–Conover technique
because of similar 2D PDF values in Vincent oil reser- (Iman and Conover 1982), based on rank correlations,
voirs. This problem increases uncertainty in the inter- simulates the correlation between Ip and V p /V s . The
pretation of data. To overcome this problem, in this Iman–Conover method provides a way of determining
study, we apply the probabilistic reservoir characteri- reasonable associations and, thus, inducing correlation
sation method using 3D crossplots of SFM results and between samples of variables (Mildenhall 2005). Assum-
compare the results with those of the conventional 2D ing that the data are normally distributed, the 2D PDF of
crossplotting method. A field data example from the each facies can be built based on the mean and covari-
Vincent oil field is used to demonstrate the effective- ance of the augmented data (Figure 1(b), lower panel).
ness of the proposed method. Three parameters, Ip , The equation of the normal 2D PDF is written as follows
pseudo gamma ray (GR) and pseudo Sw logs are used (Rose and Smith 2002):
as axis-parameters for 3D crossplot in this example. To  −1/2
perform the SFM, the axis-parameters should be math- 1 
p(x |ci ) =  
ematically established with petrophysical and elastic 2π  
i
property logs. We use linear multi-regression analysis to  
derive mathematical relationships between properties 1  −1
 i )T
exp − (x − μ (x − μ
 i) ,
and pseudo GR and Sw . And then, the probabilistic neu- 2
i
ral network (PNN) is applied to estimate the pseudo GR
and Sw volumes. i = 1,2, · · · N, (1)
EXPLORATION GEOPHYSICS 343

where N is the number of facies, ci is the facies (e.g. Effective reservoir characterisation using the

sand), x = (Ip , Vp /Vs ) is a seismic attribute vector, and 3d crossplot of stochastic forward modelling
i and μ  i are the variance–covariance matrix (2 × 2) (SFM) results
and mean vector (1 × 2) of the seismic attributes in the
To overcome the limitation of conventional probabilis-
ith class, respectively. When the 2D PDF and seismic
tic reservoir characterisation using 2D crossplots of SFM
inversion results are applied to the Bayesian inference
results, we propose a new approach using 3D crossplots
method, the probability of each facies can be calculated.
of SFM results. The workflow of the proposed method
Bayesian inference (Doyen 2007) is used to calculate the
is shown in Figure 2. The main points of the proposed
probability of a particular facies, given a particular set of
method are as follows:
seismic attributes, as follows:

(1) Instead of using the 2D crossplot, we use the 3D


p(x |ci ) · p(ci ) crossplot composed of three parameters. Thus, a 3D
p(ci |x ) = , (2)
p(x ) PDF is built instead of 2D PDF.
(2) The proposed method uses parameters that can
where p(ci |x ) is the posterior probability, p(ci ) is the prior separate the facies more clearly. The parameters
probability for ci (e.g. the probability of obtaining sand, used in this study are Ip , pseudo GR log, and pseudo
in general), and p(x ) is the probability of the seismic Sw log because these are sensitive to φ, lithology,
attribute vector. In Bayes’ theorem, the probability, p(x ), and pore fluid, respectively.
is a normalisation factor that is computed so that the
conditional probabilities of all facies to sum to 1: To perform SFM based on the probability distributions
of the Ip , pseudo GR log, and pseudo Sw log, which
are the axis parameters of the 3D crossplot, mathemat-

N ical relationships between the elastic or petrophysical
p(x ) = p(ci ) · p(x |ci ). (3) logs and pseudo GR or pseudo Sw logs must first be
i=1
built. There are many methods for estimating GR and
Sw logs like petrophysical inversion (Bosch, Zamora,
The posterior probability, p(ci |x ) is what we are trying and Utama 2002; Grana 2018) and extended elastic
to calculate (see Equation (2)). Finally, the calculation impedance (Whitcombe, Connolly, and Reagan 2002). In
of the most probable facies is performed by determin- this study, the linear multi-regression method was used
ing the highest probability in each seismic grid. How- to establish mathematical relationships due to simplic-
ever, the effectiveness of this method is limited when ity and efficiency. Based on these derived mathemat-
the 2D PDFs calculated from the augmented data are ical relationships, SFM should be performed on a 3D
not clearly distinguishable on the 2D crossplot, espe- crossplot. Thus, the data are augmented based on the
cially for oil reservoirs. For example, the distinction of probability distributions of the Ip , pseudo GR log, and
the facies at the location marked by a black star on the pseudo Sw log, and the 3D PDF of each facies has to be
lower panel of Figure 1(b) is difficult because the 2D built.
PDFs overlap. These similar facies probabilities result in From seismic data, the Ip , Is , and ρ volumes are
increased interpretation uncertainty. calculated using seismic inversion. Using these values,

Figure 2. Workflow of the new probabilistic reservoir characterisation method using the three-dimensional crossplot of the SFM
results (modified from Choi et al. 2017). GR, gamma ray; Sw , water saturation; 3D PDF, trivariate probability function; Ip , P-impedance;
Is , S-impedance; V p /V s , P-wave/S-wave ratio; ρ, density.
344 J. CHOI ET AL.

Figure 3. Location map of Exmouth Sub-basin and Vincent field (modified from Loro et al. 2015).

elastic property volumes, such as λρ, µρ, Young’s mod- between well logs and pseudo GR and pseudo Sw logs
ulus, and Poisson’s ratio, are computed. Using these had to be established. Linear multi-regression analysis
volumes, pseudo GR and pseudo Sw log volumes are was adopted to estimate these target logs using the
extracted using a PNN (Hampson, Schuelke, and Quirein mathematical combination of elastic and petrophysi-
2001). The reservoir characterisation can then be per- cal logs (Hampson, Schuelke, and Quirein 2001). Linear
formed probabilistically based on the Bayesian infer- multi-regression models a target log as a linear combi-
ence method using 3D PDFs generated from well log nation of several elastic and petrophysical logs using
data and Ip , pseudo GR log, and pseudo Sw log volumes the following formulas, which can be solved using a
extracted from seismic data. standard least squares approach:

Establishment of mathematical relationships for n


GRpseudo = wGR,0 + wGR,i A(t), (4)
SFM i=1
n
To perform SFM, the mathematical relationships bet Swpseudo = wSw ,0 + wSw ,i B(t), (5)
i=1
ween target logs (in this study, GR and Sw logs) and
elastic or petrophysical logs must be established. In
the conventional SFM method, Ip and V p /V s are gen- where w0 is a constant value of each equation, wi is a
erally used for the 2D crossplot, and probability distri- weighting value of the ith well logs, A(t) are the nonlin-
butions of these parameters are easily calculated from ear transformed well logs for estimation of the pseudo
the probability distributions of V p , V s , and ρ that are GR log (GRpseudo ). B(t) are the nonlinear transformed
derived from the rock physics model. However, in this well logs for estimation of the pseudo Sw log (Swpseudo ).
study, pseudo GR and pseudo Sw logs were used for the The weighting values can be derived by minimising
axes of the 3D crossplot with Ip , and the relationships the mean squared prediction error (Hampson, Schuelke,
EXPLORATION GEOPHYSICS 345

Figure 4. Inversion results section, which includes the Vincent 1 and Vincent 3 wells: (a) P-impedance; (b) S-impedance; and (c)
density.

and Quirein 2001), as follows: data on the 3D crossplot, it must be expressed as 3D


PDF. The 3D PDF is defined as follows (Rose and Smith
1 
N M
2
EGR = (GRlog − wGR,0 − wGR,j Aj (t))2 , (6) 2002):
N  −1/2
i=1 j=1

1  
1
N  M P(x |ci ) = 3/2  
2
ESw = (Swlog − wSw,0 − wSw,j Bj (t))2 , (7) (2π ) 
i
N
i=1 j=1  
1 T
−1
exp − (x − μ  i) (x − μ
 i) ,
where GRlog and Swlog are real GR and Sw logs at the 2
i
well, respectively.
Therefore, the weights can be expressed in matrix i = 1, 2, · · · N, (10)
form as shown below:

wGR = [AT A]-1 AT GRlog , (8) where x =(Ip , GRpseudo , and Swpseudo ), and  i and μ
i
are the components of the variance–covariance matrix
wsw = [BT B]-1 BT Swlog . (9) 
(3 × 3) and mean vector (1 × 3) of x in the ith class,
respectively. The shape of the 3D PDF on the 3D cross-
3D crossplotting (3D PDF) of SFM results plot is an ellipsoid and its projection onto any plane
made by two axis parameters is an ellipse, which is the
The advantage of 3D crossplotting over the 2D method shape of the 2D PDF on the 2D crossplot. Therefore, by
is that SFM results can be separated more clearly. To using more axes on the crossplot, the PDFs can avoid
define the probability distribution of the augmented overlap.
346 J. CHOI ET AL.

Table 1. Weighting of the values of the nonlinear transformed


dependent variables for gamma ray log prediction.
Variables Nonlinear transform Weight
Porosity (φ) Square root −225.636
µρ Square −40.1271
P-wave (V p ) Square 228.7914
S-wave (V s ) 1/X 83.65727
Resistivity None −12.5427
P-impedance (Ip ) Square −143.303
λρ 1/X 1.147085
S-impedance (Is ) Square 102.7628
V p /V s Square 162.0432
Density (ρ) 1/X −36.031
Constant −138.738

Table 2. Weighting of the values of the nonlinear transformed


dependent variables for water saturation log prediction.
Variables Nonlinear transform Weight
Resistivity Log −0.25443
Porosity (φ) Square −1.30283
λρ 1/X −0.13016
µρ Square −0.28723
P-impedance (Ip ) Square 0.622253
S-impedance (Is ) None 0.620334
V p /V s 1/X −0.11866
Density (ρ) Square −0.54554
S-wave (V s ) Square −0.48685
P-wave (V p ) 1/X 0.170768
Constant 0.158722

Figure 5. Linear multi-regression application using 10 nonlin-


ear transformed well logs. The original gamma ray logs are
Application shown in black and the predicted logs are shown in red.
Vincent oil field is adjacent to the North West Cape of
Western Australia. The Vincent field is located in the
Exmouth Sub-basin (Figure 3), which is the southern-
most rift basin within the Carnarvon Basin, containing
Jurassic to Early Cretaceous sediments (Blazak et al.
2015). The Vincent structure is confirmed as a three-
way dip/fault closure at the base of the Muderong Shale.
The claystone of the Muderong Shale serves as the top
seal for the Lower Barrow Group, which is the reservoir
of the Vincent field. The API of the oil is 17.3, meaning
that it is classified as heavy oil. Four partial-angle stack
seismic data sets (8°–19°, 19°–30°, 30°–41°, and 41°–52°)
and three wells (Vincent 1, Vincent 2, and Vincent 3)
were used. The seismic survey area containing the wells
is 6.25 by 5 km. A simultaneous inversion (Hampson,
Russell, and Bankhead 2005) was performed with angle-
dependent wavelets to yield the elastic property vol-
umes (Figure 4). The Vsh and Sw log curves were used
to define the facies: shale, water sand, and hydrocarbon
sand.
Prior to linear multi-regression analysis to establish
the mathematical relationships for the pseudo GR and
pseudo Sw logs, well log data were normalised to deter-
mine the stable weighting values. The types of well
log used as dependent variables in the linear multi-
regression analysis were Ip , Is , V p , V s , ρ, λρ, µρ, V p /V s , φ,
Figure 6. Linear multi-regression application using 10 nonlin-
and resistivity. Nonlinear (e.g. non, root, square, log, and ear transformed well logs. The original water saturation logs are
inverse) transformations of well logs were performed shown in black and the predicted logs are shown in red.
first. The nonlinear transformed dependent variables
EXPLORATION GEOPHYSICS 347

Table 3. Correlation coefficients and average relative errors between original logs and predicted logs.
Correlation coefficients Average relative errors

Vincent 1 Vincent 2 Vincent 3 Vincent 1 (%) Vincent 2 (%) Vincent 3 (%)


GR 0.84 0.93 0.97 GR 14.24 9.69 6.94
Sw 0.98 0.94 0.93 Sw 5.77 5.04 6.56
Note: GR, gamma ray; Sw , water saturation.

Figure 7. Augmented data (left) and trivariate probability density functions (right) of each facies on the three-dimensional crossplot
of P-impedance (Ip ), pseudo gamma ray (GR), and pseudo water saturation (Sw ). Shale data are shown in grey, water sand data are
shown in yellow, and oil sand data are shown in green.

correlate better with the target log when the relation- the target logs. The weighting value of each dependent
ship between the target log and the dependent variable variable (Tables 1 and 2) was derived by minimising the
is nonlinear. Stepwise regression was performed to find mean squared prediction error. Figures 5 and 6 show the
the best single dependent variable, then the best pair results of applying linear multi-regression to the well log
of dependent variables, then the best trio, etc. The best data. As shown in these Figures, the predicted pseudo

dependent variable set for the pseudo GR log was { φ, GR and pseudo Sw logs generally match the original
µρ 2 , Vp 2 , 1/V s , resistivity, Ip 2 , 1/λρ, Is 2 , (V p /V s )2 , 1/ρ}, well logs, except for the high-frequency components at
and the best variable set for the pseudo Sw log was some depth intervals. The correlation coefficients and
{log(resistivity), φ 2 , 1/λρ, µρ 2 , Ip 2 , Is , 1/(V p /V s ), ρ 2 , Vs 2 , average relative errors are shown in Table 3. Note that
1/V p }. The components of the variable sets are arranged the correlation coefficients were high and the average
in decreasing order of the strength of correlation with relative errors were low in all cases.

Figure 8. Augmented data (left) and bivariate probability density functions (right) of each facies on the two-dimensional crossplot
of P-impedance (Ip) and the P-wave/S-wave ratio (V p /V s ). Shale data are shown in grey, water sand data are shown in yellow, and oil
sand data are shown in green.
348 J. CHOI ET AL.

Figure 9. PNN application using eight nonlinear transformed Figure 10. PNN application using eight nonlinear transformed
elastic properties at the well locations. The scaled gamma ray elastic properties at the well locations. The scaled water satura-
(GR) log is shown in black and the predicted GR is shown in tion (Sw ) log is shown in black and the predicted Sw is shown in
red. red.

A rock physics model based on Gassmann’s equation were Vsh , quartz volume, φ, Sw , shear modulus, and
(Smith, Sondergeld, and Rai 2003) was built to pre- dry rock bulk modulus. The probability distributions of
dict the probability distributions of V p , V s , and ρ the input parameters were then calculated from the
using the distributions of petrophysical parameters. well data. To estimate the properties of the pore fluid
The input parameters used in this rock physics model and the matrix, Wood’s equation (Wood 1955) and the

Figure 11. Pseudo gamma ray section, which includes the Vincent 1 and Vincent 3 wells, extracted using the PNN.
EXPLORATION GEOPHYSICS 349

Figure 12. Pseudo water saturation (Sw ) section, which is the same inline section as that in Figure 10, extracted using the PNN.

Figure 13. Probability volumes of (a) shale, (b) water sand, and (c) oil sand facies obtained using bivariate probability density
functions in the same section as that in Figure 10.
350 J. CHOI ET AL.

Figure 14. Probability volumes of (a) shale, (b) water sand, and (c) oil sand facies obtained using trivariate probability density
functions in the same section as that in Figure 10.

Voigt–Reuss–Han equation (Wang and Nur 1989) were the data; however, this problem is resolved on the 3D
applied. Based on the rock physics model, the proba- crossplot.
bility distribution of the Ip can be estimated using the Elastic property volumes, such as V p , V s , λρ, and
distributions of V p and ρ. From the mathematical rela- µρ, were calculated from the Ip , Is , V p /V s , and ρ vol-
tionships determined using linear multi-regression as umes. In total, eight elastic property volumes were
described above, the probability distributions of the used for the PNN input data. Before applying the
pseudo GR and pseudo Sw logs were estimated using PNN, nonlinear transformations of the elastic prop-
the distributions of the elastic and petrophysical prop- erty volumes were performed, and the best variable
erties. Finally, using the probability distribution of the set of nonlinear transformed elastic property volumes
Ip , pseudo GR log, and pseudo Sw log, Monte Carlo sim- was calculated by stepwise regression, as applied in
ulation was performed to augment the data (Figure 7). the linear multi-regression analysis. Furthermore, the
The 3D PDFs were determined by the mean and the convolutional operator was used to more correctly
covariance–variance matrix of each facies, calculated estimate the target (GR and Sw ) logs, which typically
from the augmented well data (Figure 7). To verify the have higher-frequency contents than those of the elas-
effectiveness of the proposed method, the PDFs of the tic properties (Hampson, Schuelke, and Quirein 2001).
facies on the 3D crossplot were compared with those The convolutional operator is a process that predicts
on the general 2D crossplot (Ip versus V p /V s ). As shown each sample of the target logs using a group of neigh-
in Figure 7, the distributions of each of the facies in bouring samples on the elastic property logs, rather
the 3D crossplot are clearly distinct compared with than using only one sample. PNN was performed to
those on the 2D crossplot (Ip versus V p /V s ; Figure 8). derive the pseudo GR and pseudo Sw log volumes. Fig-
Facies are difficult to distinguish in the area of the ures 9 and 10 show the predicted pseudo GR and Sw logs
2D crossplot where the facies probabilities are simi- obtained using the PNN with scaled well logs at the well
lar, which increases uncertainty in the interpretation of locations. The cross-correlation coefficient between the
EXPLORATION GEOPHYSICS 351

Figure 15. The most probable facies, identified by determining the highest probability of each seismic grid among the probabilities
> 0.7 using the 3D crossplotting method. Shale data are shown in grey, water sand data are shown in yellow, and oil sand data are
shown in green.

Figure 16. The most probable facies volumes, identified by determining the highest probability of each seismic grid among the
probabilities > 0.7 using the 2D crossplotting method. Shale data are shown in grey, water sand data are shown in yellow, and oil
sand data are shown in green.

scaled GR log and predicted pseudo GR log for all wells predicted pseudo Sw log for all wells was 0.95 and the
was 0.92 and the training error was 7.74. The cross- error was 0.049. Figures 11 and 12 show the pseudo GR
correlation coefficient between the scaled Sw log and and pseudo Sw log sections in the same inline section,
352 J. CHOI ET AL.

Figure 17. 3D view of the facies cube generated using the 3D crossplotting method in the Vincent oil field. Shale data are shown in
grey, water sand data are shown in yellow, and oil sand data are shown in green.

which includes the Vincent 1 and Vincent 3 wells. Com- Conclusions


paring the superimposed well logs at the Vincent 1 and
A 2D crossplotting method using well log and seismic
Vincent 3 well locations, the pseudo GR and pseudo Sw
data has commonly been used to identify hydrocar-
logs generally match the scaled well logs, except for the
bon reservoirs. However, when well log data are too
GR log at depth intervals corresponding to interbedded
sparse on the 2D crossplot, the data cannot fully rep-
shale in the Vincent 3 well.
resent reservoir variation, reducing the accuracy of cal-
The Bayesian inference method was applied to the
ibration with seismic data. SFM (data augmentation)
full 3D seismic volume to calculate the probability
methods have been applied to overcome the prob-
volume for each facies and the most probable facies vol-
lem of well log data sparseness. From the mean and
ume using the facies probability distribution. Figure 13
covariance of the augmented data, we can build 2D
shows the probability volumes from the conventional
PDFs and probabilistically estimate the distribution of
2D PDFs (2D crossplot of Ip and V p /V s ), and Figure 14
facies. However, in many cases, the overlap in the dis-
shows the probability volumes from the 3D PDFs (3D
tributions of facies on the 2D crossplot makes the iden-
crossplot of Ip , pseudo GR log, and pseudo Sw log).
tification of the facies for a given seismic data set dif-
The probability volumes calculated using the 2D PDFs
ficult and leads to increased uncertainty in interpre-
have intermediate values because the 2D PDFs over-
tation. To solve this overlap problem, we propose a
lapped on the 2D crossplot. Compared with the results
3D crossplotting method using SFM results to separate
obtained using the conventional 2D PDFs, the proba-
the PDFs. In the proposed solution, expansion of the
bility volume of each facies using the 3D PDFs is either
crossplot dimensions from 2D to 3D and the attributes
high or low, meaning that the 3D probability distribu-
selected clearly separate the distribution of each facies.
tions of the facies are well separated. To make a more
In this study, we used Ip , pseudo GR log, and pseudo Sw
effective comparison, the most probable facies volumes
log, which are sensitive to φ, lithology, and pore fluid,
were calculated using 3D PDFs and 2D PDFs, and proba-
respectively, as the axis parameters of the 3D crossplot.
bility values less than 0.7 were replaced by 0. The results
To perform SFM based on the probability distributions
are shown in Figures 15 and 16, respectively. The white
of the Ip , pseudo GR log, and pseudo Sw log (indepen-
areas indicate the regions in which all facies probabili-
dent variables), the mathematical derivation equations
ties are less than 0.7. The white area in the 2D PDF result
for these variables must be defined using the well logs
is larger than that in the 3D PDF result, demonstrating
(dependent variables). The probability distribution of
greater uncertainty in the interpretation of the results
the Ip was extracted from the probabilities of V p and
obtained using 2D PDFs than of those obtained using
ρ, which were calculated using a rock physics model.
3D PDFs. A 3D facies cube with facies logs at the well
To establish the derivation equations of the pseudo
locations is shown in Figure 17. Note that the oil sands
GR and pseudo Sw logs, a linear multi-regression anal-
are distributed below the shale, which is the seal rock of
ysis was adopted. Using the probability distributions
the reservoir, and that the sand is interbedded with the
of the Ip , pseudo GR log, and pseudo Sw log, the data
shale in other areas.
EXPLORATION GEOPHYSICS 353

were augmented by Monte Carlo simulation and 3D Choi, J., B. Kim, S. Kim, and J. Byun. 2017. Probabilistic facies
PDFs were determined by the mean and covariance analysis using 3D crossplot of stochastic forward-modeling
of the augmented data. A PNN was used to compute results. 87th Annual International Meeting, SEG, Expanded
Abstracts, 3077–81.
the pseudo GR and pseudo Sw logs from seismic data Chopra, S., V. Alexeev, and Y. Xu. 2003. 3D AVO crossplotting?
using the elastic property attributes of the seismic data. An effective visualization technique. The Leading Edge 22
Finally, using the probability distributions of the facies no. 11: 1078–89.
and inverted seismic volumes, Bayesian classification Close, D., R. Taylor, and S. Nixon. 2015. Rock physics and quanti-
was applied to calculate the facies probabilities. The tative interpretation using lambda–mu–rho in the Shipwreck
Trough, Otway Basin: ASEG, Extended Abstracts, 1–4.
effectiveness of our method was demonstrated using
Doyen, P.M. 2007. Seismic reservoir characterization: an earth
a field data example with shale, water sand, and oil modelling perspective, EAGE.
sand facies. The facies are clearly distinguishable on Gallop, J. 2006. Facies probability from mixture distributions
a 3D crossplot from 3D PDFs compared with conven- with non-stationary impedance errors. 76th Annual Interna-
tional 2D PDFs on a 2D crossplot, because of the overlap tional Meeting, SEG, Expanded Abstracts, 1801–4.
in the probability distributions of the facies in the 2D Goodway, B., T. Chen, and J. Downton. 1997. Improved AVO
fluid detection and lithology discrimination using Lamé petro-
crossplot. However, with the 3D crossplotting method, physical parameters: “λρ”, “µρ”, & “λ/µ fluid stack”: from P
because of the clearly distinguishable 3D PDFs on the and S inversions. 67th Annual International Meeting, SEG,
3D crossplot, the facies probabilities were close to 0 or Expanded Abstracts, 183–6.
1, decreasing the interpretation uncertainty. Grana, D. 2018. Joint facies and reservoir properties inversion.
Geophysics 83 no. 3: M15–24.
Hampson, D.P., B.H. Russell, and B. Bankhead. 2005. Simultane-
Acknowledgements ous inversion of prestack seismic data. 75th Annual Interna-
We thank Schlumberger for providing Petrel and Techlog soft- tional Meeting, SEG, Expanded Abstracts, 1633–7.
ware, and CGG for providing Hampson–Russell software. We Hampson, D.P., J.S. Schuelke, and J.A. Quirein. 2001. Use of
are also grateful to SK innovation for providing the seismic multiattribute transforms to predict log properties from
data. seismic data. Geophysics 66 no. 1: 220–36.
Iman, R.L., and W.J. Conover. 1982. A distribution-free app
roach to inducing rank correlation among input variables.
Disclosure statement Communications in Statistics-Simulation and Computation
11 no. 3: 311–34.
No potential conflict of interest was reported by the authors. Lamont, M., T. Thompson, and C. Bevilacqua. 2008. Drilling
success as a result of probabilistic lithology and fluid pre-
diction: a case study in the Carnarvon Basin, WA. The APPEA
Funding
Journal 48 no. 1: 31–42.
This work was supported by the Human Resources Program Loro, R., R. Hill, M. Jackson, and T. Slate. 2015. Technologies
in Energy Technology of the Korea Institute of Energy Tech- that have transformed the Exmouth into Australia’s premier
nology Evaluation and Planning, granted financial resources oil producing basin. The APPEA Journal 55 no. 1: 233–46.
from the Ministry of Trade Industry and Energy Korea [grant Michelena, R.J., K.S. Godbey, and P.E. Rodrigues. 2010.
numbers 20164010201120 and 20182510102470]. Facies probabilities from multidimensional crossplots of seis-
mic attributes. 80th Annual International Meeting, SEG,
ORCID Expanded Abstracts, 1377–81.
Mildenhall, S.J. 2005. Correlation and aggregate loss distri-
Junhwan Choi http://orcid.org/0000-0003-4987-9469 butions with an emphasis on the Iman–Conover Method.
Report of the research working party on correlations and
dependencies among all risk sources, Casualty Actuarial
References Society.
Avseth, P., T. Mukerji, and G. Mavko. 2005. Quantitative seismic Ng, S., P. Dahle, R. Hauge, and O. Kolbjørsen. 2008. Estimation
interpretation: applying rock physics tools to reduce interpre- of facies probabilities on the Snorre Field using geostatisti-
tation risk. Cambridge: Cambridge University Press. cal AVA inversion. 78th Annual International Meeting, SEG,
Blazak, D., G. Smith, D. Baker, C. Falconer, J. DePledge, C. McNa- Expanded Abstracts, 1971–1974.
mara, B. Jury, G. Johnston, and R. Gupta. 2015. Vincent Nieto, J., B. Batlai, and F. Delbecq. 2013. Seismic lithology pre-
multi-lateral drilling: use of 4D seismic, scenario analysis diction: a Montney shale gas case study. CSEG Recorder 38:
and deep directional resistivity for geo-steering, comple- 34–42.
tions and anti-collision analysis to re-develop a thin oil Pendrel, J., H. Schouten, and R. Bornard. 2017. Bayesian estima-
field, offshore Carnarvon sub-basin, Australia: International tion of petrophysical facies and their applications to reservoir
Conference and Exhibition, Melbourne, Australia, 510. characterization. 87th Annual International Meeting, SEG,
Bosch, M., Zamora, M., and Utama, W. 2002. Lithology dis- Expanded Abstracts, 3082–6.
crimination from physical rock properties. Geophysics 67(2): Reine, C. 2014. Quantitative interpretation guided by rock–
573–81. physics templates: examples from unconventional reservoirs.
Chen, H., J.P. Castagna, R.L. Brown, and A.C. Ramos. 2001. 83rd Annual International Meeting, SEG, Expanded Abs
Three-parameter AVO crossplotting in anisotropic media. tracts, 2397–402.
Geophysics 66 no. 5: 1359–63. Rose, C., and M.D. Smith. 2002. Mathematical statistics with
Chi, X.G., and D.H. Han. 2009. Lithology and fluid differentia- mathematica. Berlin: Springer.
tion using a rock physics template. The Leading Edge 28 no. Ross, C.P. 2000. Effective AVO crossplot modeling: a tutorial.
1: 60–65. Geophysics 65 no. 3: 700–11.
354 J. CHOI ET AL.

Ross, C.P., and M.A. Sparlin. 2000. Improved crossplot analysis Whitcombe, D.N., P.A. Connolly, and R.L. Reagan. 2002.
using visualization techniques. The Leading Edge 19 no. 11: Extended elastic impedance for fluid and lithology predic-
1188–99. tion. Geophysics 67 no. 1: 63–67.
Smith, T.M., C.H. Sondergeld, and C.S. Rai. 2003. Gassmann Whitcombe, D.N., and J.G. Fletcher. 2001. The AIGI crossplot as
fluid substitutions: a tutorial. Geophysics 68 no. 2: 430–40. an aid to AVO analysis and calibration. 71st Annual Interna-
Wang, Z., and A.M. Nur. 1989. Seismic and acoustic velocities tional Meeting, SEG, Expanded Abstracts, 219–22.
in reservoir rock: recent development. Society of Exploration Wood, A.W. 1955. A textbook of sound. New York: The MacMil-
Geophysicists 10. lan Co.

You might also like