Sei sulla pagina 1di 58

3D Seismic Attributes for Prospect

Identification and Reservoir Characterization

Kurt J. Marfurt (The University of Oklahoma)

Reservoir Characterization Workflows

15-1
Course Outline
Introduction
Complex Trace, Horizon, and Formation Attributes
Multiattribute Display
Spectral Decomposition

Geometric Attributes
Attribute Expression of Geology
Tectonic deformation
Clastic depositional environments
Carbonate deposition environments
Shallow stratigraphy and drilling hazards
Igneous and intrusive reservoirs and seals
Attribute expression of shale reservoirs and correlation to hydraulic fracturing

Impact of Acquisition and Processing on Attributes


Attribute Prediction of Fractures and Stress
Data Conditioning
Inversion for Acoustic and Elastic Impedance
Image Enhancement and Object Extraction

Unsupervised Multiattribute Classification


Supervised Multiattribute Classification
Statistical Attribute Analysis
Reservoir Characterization Workflows
3D Texture Analysis

15-2
Objectives
• Describe value of pseudowells
• Discuss alternative multiattribute workflows:
• workflow #1: Constraining ant-tracking to map faults consistent
with image log data
• workflow #2 : Impedance inversion followed by geostatistics
estimate of facies
• workflow #3 : Multiattribute linear stepwise regression followed by
geostatistics to predict porosity
• workflow # 4: Neural net with a nonlinear search for best attribute
combination
• workflow #5 : Seismic waveform classification/clustering followed
by forward modeling calibration
• workflow #6 : Matching seismic attributes to attributes of well log
synthetics
• Recognize potential risks when using a small well
15-3
population
The value of pseudowells

tight facies
ss – low porosity
ss –hi porosity
ls –low porosity
ls –hi porosity
dol –low porosity
dol –hi porosity

15-4 (Joseph et al., 1999)


The value of pseudowells

15-5 (Joseph et al., 1999)


The value of pseudowells

Pseudo-well synthetics projected on


first two planes of the seismic
attribute space.

15-6 (Joseph et al., 1999)


The value of pseudowells

Isolated anomalies
(Can be fixed by
geostatistics)

Seismic facies map


15-7 (Joseph et al., 1999)
Workflow #2: Facies constrained geostatistics (RC2-IFP)

1. establish a correlation between facies with a certain seismic


attribute.
2. construct facies geostatistical models using:
a) facies logs as hard data (honored at 100%),
b) 3-D seismic attribute model as soft data (honored at less
than 100%), and
c) other geologic data such as preferred channel direction and
global facies proportions as additional soft data.
3. model geologically constrained seismic facies by:
a) stochastic seismic inversion,
b) facies modeling by sequential indicator simulation with
collocated cokriging, and
c) porosity modeling by facies dependent collocated cokriging.
15-8 (Lo and Bashore, 1999)
A) Petrophysically-constrained geostatistics
S impedance,IS

P impedance, IP

Cross plot of P impedance vs. S impedance.


Shale volume in color.

15-9 (Barens and Biver, 2004)


1) Construct background model using
wells and seismic velocities
Background

15-10 (Barens et al., 2003)


2) Classical seismic inversion for the ‘most
likely’ impedance model at seismic resolution

Seismic bandwidth
Background

Trace Inversion

15-11 (Barens et al., 2003)


3) Geostatistical Inversion at the geologic
model resolution using
– detailed well control
– seismic data Uncertainties
on Inversion
– variograms
Seismic bandwidth
Background

Geostatistical Modelling

15-12 (Barens et al., 2003)


Given a P impedance, IP, use the Prior Joint Distribution and Bayes’
rule to constrain the shear impedance, IS: P(IS|d)IP ~ L(d|IS) · P(IS)IP.

= P(IS)

(IS Kriging)
= P(data|IS)

= P(IS|data)

15-13 (Barens and Biver, 2004)


For each (IP,IS) realization,
forward model near and far AVO effects
IP Seismic Synthetic Near

Aki & Richards approximation

1
  IP
R( )   1  tan 2  
2  IP
 IS 2  IS Seismic Synthetic Far
IS  4 2 sin2  
 IP  IS

Pre-Stack Geostatistical Inversion


Constrained by:
- Seismic data
- Geostatistical Modelling

15-14 Probability Density Functions (Barens and Biver, 2004)


Result 1: mean impedances

The mean impedances


for a stratigraphic layer
in the grid.
P impedance

S impedance

15-15 (Barens et al., 2003)


Result 2: variability of impedance estimates

The standard deviation


of impedances for a
stratigraphic layer in the
P impedance grid.

S impedance

15-16 (Barens et al., 2003)


Result 3: multiple realizations of P- and S-impedances
that honor both well and seismic data.

15-17 (Barens and Biver, 2004)


B) Facies-constrained inversion
1) Data preparation
– Pick surfaces
– Estimate variograms
• Lateral (seismic)
• Vertical (wells)

15-18 (Barens and Biver, 2004)


1) Define Geologic Facies by relating geologic
interpretation and inversion results

Vsh IP IS Facies
Geologic Facies
50 ms Shale
Debris Flows
Lags
Laminated Sands
Coarse Sands

15-19 (Barens and Biver, 2004)


2) Results of cross plotting - Facies estimation
Shale

Shale
IP/IS Ratio

Debris Flows
Lags
Laminated Sands
Coarse Sands

Laminated Coarse sand


sand

IP (with shale detrend)

15-20 (Barens and Biver, 2004)


3) Generate a seismic facies realization at each
trace

For a given IP and IS pair, each voxel has a probability of being a


given seismic facies.

15-21 (Barens and Biver, 2004)


3) Realisation of Seismic Facies

26
Realisation 12
34

SF Region 1: Shale
SF Region 2: Laminated Sands
SF Region 3: Coarse Sands

15-22 (Barens and Biver, 2004)


Workflow #3: Choosing the optimum attributes for
inversion using supervised learning:
Multiattribute linear stepwise regression
1) Correlate candidate attributes with a subset of the available well data.
Russell et al (2000) and Cooke et al (1999) find that seismic inversion
is much better correlated with porosity than original seismic
amplitudes.

2) Add additional attributes to enhance the final well-seismic tie. This


technique is called thereby obtaining a multiattribute transform (this
vector of weights vector will be ‘cross correlated’ with the
multiattribute vector at each trace) 

3) Avoid over fitting the data by validating the transform with wells left
outside the training step (cross validation step).

4) Apply geostatistics to honor both well data and the mulitattribute


transform, with the well data being a hard constraint.
15-23 (Russell et al. 2001)
Multiattribute linear stepwise regression
75 CDPs

125 CDPs

Well distribution map


seismic sonic porosity

15-24 (Russell et al. 2001)


Multiattribute linear stepwise regression

Inst. freq.
impedance

inst. phase

Integ. trace
Average error

ANN system trained


without target well
(Validation error)

ANN system trained


using all wells

Number of attributes
Average error for curve of the form (x,y)=w 0+w1A1(x,y)+…+wmAm(x,y).
Using 5 attributes ‘overtrains’ the neural net system to fit the control
data.
15-25 (Russell et al. 2001)
Multiattribute linear stepwise regression

Porosity prediction from Porosity prediction from linear


acoustic impedance only: combination of 3 attributes
Correlation =-0.65 Correlation = +0.82

15-26 (Russell et al. 2001)


Multiattribute linear stepwise regression

Porosity prediction after Kriging with external drift.


Porosity at wells uses the same color scale as the
seismic prediction

15-27 (Russell et al. 2001)


Workflow # 5: Stratigraphic Interpretation via Neural Net
Driven Seismic Trace Shape Classification (a workflow
used at Paradigm Geophysical)
 Generate a seismic facies (similarity) map by correlating the modeled
wave forms with the actual traces
 Correlate synthetic seismic response to field seismic data.
 Perform forward modelling at the well bore.
 Perturb reservoir properties and interactively observe changes in
synthetic seismic response.
 Predict reservoir properties away from well location.
 Use synthetic seismic response to improve understanding of neural
network facies traces.
 Use neural networks to re-classify data based on modeling to
produce property maps.

15-28 (Poupon et al., 1999)


What’s inside Stratimagic (or what Marfurt has pieced
together)
1. Window a suite of traces of length N samples using picked and/or
phantom horizons and obtain djk(t)
2. Increase the frequency content by taking a time difference:

u jk (nt )  d jk (nt )  d jk ((n  1)t )


3. Normalize each windowed trace to have energy=1.0:

 u 
N
jk ( nt )  1
2

n1

4. Select a subset of the data volume (e.g. every 10th line and 10th cdp) and
cluster in N-dimensional ‘attribute’ space, where each time sample
below the picked horizon is a separate attribute. Call the mean of each
cluster the ‘model trace’ for the cluster.
5. Crosscorrelate every trace in the extracted volume with each and every
15-29 modeled trace. Assign it to the cluster whose mean is closest.
Seismic Facies Classification
• Unsupervised Regional Seismic Facies Analysis
Actual seismic traces are crosscorrelated with the the model traces obtained
using a Kahonen Self Organized Map. A color is assigned corresponding to
the nearest model trace.
Interval of interest

Data courtesy of CGG-USA

Model Traces
(cluster means)
15-30 (Poupon et al., 1999)
Seismic Facies Classification
• Supervised Seismic Facies Analysis
The set of model traces can be updated by inserting a seismic trace at
well location. The classification process is repeated.
Seismic Facies Map
Model Traces

Seismic response
at Well location
Data courtesy of CGG-USA

15-31 (Poupon et al., 1999)


Seismic Facies Classification
The correlation of seismic trace shape to a particular seismic response
(usually a producing well) allows one to identify prospect based on
waveform.

Correlation Map (in %)


Between trace shape at
well and seismic traces
over the studied
Data courtesy of CGG-USA
interval
15-32 (Poupon et al., 1999)
Petro-acoustic Modeling
Optimizing Seismic Interval & Interpreting
Classification maps

Seismic Seismic
line classes

Facies
Correlation to
Synthetic
Models
Blocked
Acoustic
Impedance

Well curves Petro-acoustic model


15-33 (Depth domain) (Time domain) (Poupon et al., 1999)
Facies Modeling – Austral Basin, Argentina

Max flooding surface

S- N strat section

Max flooding surface

W-E strat section

15-34 (Silva-Telles et al., 2003)


Facies Modeling – Austral Basin, Argentina

40 ms analysis window
along the maximum
flooding surface

S- N seismic section

40 ms analysis window
along the maximum
flooding surface

W-E seismic section

15-35 (Silva-Telles et al., 2003)


Mix dip/isochron Average abs amp Shape class

don’t drill
drill

don’t drill

Only wells D and E were drilled when map was made


15-36 (Silva-Telles et al., 2003)
15-37 (Silva-Telles et al., 2003)
Channel-fan complex (Winters
Sands, California). Wells
indicated by red dots. Note that
Coh the SOM differentiates the main
gas producer (70 ft of net pay in
Winters Sands classified as blue
trace shape) from the shaled-out
well to the east classified as
yellow trace shape. These two
wells could not be separated
based on coherence and/or
conventional amplitude analysis

SOM
facies

15-38 (Coleou et al, 2003)


Comparison between an average absolute
amplitude map (top figure) and a SOM facies
map (central) over a meandering channel
(Wolfcamp Sands, Permian, West Texas).
Green dots correspond to producing
wells. Note that the seismic facies map
identifies a meandering channel (brown
classes) not expressed in the interval
amplitude map. Trace shapes and a section
through the channel are shown at the bottom.

15-39 (Coleou et al, 2003)


Unsupervised classification of a fluvial channel

Without PCA With PCA


Natuna Sea, Indonesia. Note that the seismic facies map without PCA does not differentiate the
channel facies from the flood plain deposits (same red class). It also tends to bias the
interpretation toward a channel system not affected by the W-E strike-slip fault. On the other
hand, the PCA seismic facies map differentiates the channel facies (blue class) and clearly
highlights the effect of the W-E strike-slip fault in a relatively noisy area. Also, note the
development of the overbank deposits to the east (yellow facies outlined by black dotted line).
15-40 (Coleou et al, 2003)
Clustering using AVO and impedance attributes

Plots showing the classification results. (a) Amplitude versus fluid factor
and (b) amplitude versus acoustic impedance. The square dots connected
by the solid black lines show cluster nodes.
15-41 (Linari et al, 2003)
Clustering using AVO and impedance attributes
Comparison between a
horizon slice through the
facies classification
volume (a) and a
conventional amplitude
map (b). Both maps are
overlaid by depth
contours. Note that most
producing wells are
outside amplitude
anomalies. Also note that
the NE-SW trend
penetrated by producing
wells is not visible on the
amplitude map.

15-42 (Linari et al, 2003)


San Luis Pass weather prediction exercise
August 24, 2005 – sunny Exercise: flip 6 coins:
August 25, 2005 - storms Heads=sunny
August 26, 2005 - sunny Tails=stormy
August 27, 2005 - sunny
August 28, 2005 - sunny Read out your correlation rate:
August 29, 2005 - storms 0/6 = -1.00 3/6 = -0.00
1/6 = -0.67 4/6 = +0.33
2/6 = -0.33 5/6=+0.67
6/6 = 1.00

heads tails

15-43
San Luis Pass weather prediction exercise

Which coins best


predict the weather in
San Luis Pass?
Should Marfurt go
fishing?

15-44
Potential risks when using seismic
attributes as predictors of reservoir properties

When the sample size is small, the uncertainty about the


value of the true correlation can be large.

• given 10 wells with a correlation of r=0.8, the 95%


confidence level is [0.34,0.95]

• given only 5 wells with a correlation of r=0.8, the 95%


confidence level is [-0.28,0.99] !

15-45 (Kalkomey, 1997)


Spurious Correlations
A spurious correlation is a sample correlation that is large
in absolute value purely by chance.

15-46 (Kalkomey, 1997)


The more attributes, the more
spurious correlations!

15-47 (Kalkomey, 1997)


Risk = expected
loss due to our
uncertainty
about the truth *
cost of making a
bad decision

Cost of a Type I Error (using a seismic attribute to predict a reservoir


property which is actually uncorrelated) is:
• Inaccurate prediction biased by the attribute.
• Inflated confidence in the inaccurate prediction — apparent prediction
errors are small.

Cost of a Type II Error (rejecting a seismic attribute for use in predicting a


reservoir property when in fact they are truly correlated) is:
• Less accurate prediction than if we’d used the seismic attribute.
• Larger prediction errors than if we’d used the attribute.
15-48 (Kalkomey, 1997)
Validation of Attribute Anomalies
1. Basic QC
• is the well tie good?
• are the interpreted horizons consistent and accurate?
• are the correlations statistically meaningful?
• is there a physical or well-documented reason for an attribute to
correlate with the reservoir property to be predicted?

2. Validation
• does the prediction correlate to control not used in training?
• does the prediction make geologic sense?
• does the prediction fit production data?
• can you validate the correlation through forward modeling?

15-49 (Hart, 2002)


Validation of Attribute Anomalies
(Porosity prediction in lower Brushy Canyon)

From multivariate linear regression From probabilistic neural network.

Right map has higher statistical significance and is geologically more realistic

15-50 (Hart, 2002)


Validation of Attribute Anomalies
(Through modeling the Smackover formation)

Seismic

Instantaneous
frequency

Envelope

Field data Model data


Seismic Attribute Correlations: “Trust, but verify!”
15-51 (Hart, 2002)
Validation of Attribute Anomalies
(Through engineering and geology)
Neural Net. R=0.96

Multivariate Linear Regression. R=0.89

Dip map. Engineering and geologic analyses


indicate fractures, associated with high dip
areas, play an important role in enhancing gas
production from these tight carbonates. Stars
indicate locations of wells drilled in 1999

15-52 (Hart, 2002)


Stochastic Modeling
• compute variograms of length, width, height, and spacing of tidal bars
from satellite, outcrop, and shallow seismic data

• compute seismic attributes that differentiate sands from shales

• using well control, compute a probability map of sands and shales from
seismic attribute maps

• stochastically generate a suite of geocellular lithofacies models

• identify infill drilling locations that have a high probability of tidal sands
along with more conventional deeper targets

15-53 (Chowdary et al., 2008)


Ridge and
runnel

Modern tidal ridge


Length = 21 km

Satellite image of tidal ridges


in Gulf of Khambhat
Seismic
images along
water bottom

15-54 (Chowdhury et al., 2008)


Cumulative Probability Curves
(From modern analogues)
99 99

95 95
P90=3.23 km
Probability

Probability
P90=18.6 km
70 70
40 P50=12.3 km 40 P50=0.47 km

10 P10=0.21 km 10
P10=0.11 km
0 0
0 2 4 6 0 2 4 6
Length (km) Width (km)

99
99
95 P90=8.14 km
95 P90=21.5 m

Probability
Probability

70
70 P50=0.32 km
P50=1. m 40
40
10 P10=0.13 km
10 P10=0.3 m
0
0 0 2 4 6
0 10 20 Spacing (km)
Height (m)
15-55 (Chowdhury et al., 2008)
Use lambda-rho and porosity
inversion slices to generate
Lambda-rho probability map for tidal sands
deposits

Porosity
inversion
15-56 (Chowdhury et al., 2008)
Generate facies realizations using
well interpreted facies as hard
A data, cummulative probability
curves to constrain the geometric
B ranges , and attribute-based
probability maps to further guide
Tidal bar the geometry.

A One realization of tidal facies B

15-57 (Chowdhury et al., 2008)


Reservoir Characterization Work Flows

In Summary:

• Pseudowells use well-established methods of rock physics to


expand our statistics

• Always keep some wells out of your calibration to avoid overtraining


and to validate the reliability of your prediction algorithm

• The probability of observing spurious sample correlations between


attributes and well data can be large for small amounts of wells or
large numbers of independent attributes

• If there is no rock physics underpinning, the progressive correlation


approach of attribute selection can then be quite dangerous!

15-58 (Kalkomey, 1997)

Potrebbero piacerti anche