Professional Documents
Culture Documents
ARTICLE
deterministic and probabilistic approaches
Dennis Cooke* and John Cant**
* Santos, Ltd. Adelaide, SA, Australia; ** Makaira Geotechnical, Perth WA, Australia
The target audience of this article is the geoscientist who Forward model refers to the algorithm used to generate
desires to learn more about seismic inversion and its limita- synthetic seismic data. As geophysicists, we have some very
tions. The content is especially relevant to those using seismic complex forward models like a full elastic wave-equation and
inversion to help build static reservoir models for field devel- simple models like the convolutional model. A common
opment and/or estimate probabilities for presence of hydro- forward model used in post-stack inversion consists of calcu-
carbons. The viewpoint and experience base of the authors (at lation of reflection coefficients from an impedance model and
least the first author) is that of a consumer of seismic inver- convolution of those reflection coefficients with a source
sion as opposed to a provider of inversion software or serv- wavelet. For pre-stack inversion, the forward model needs to
ices or a researcher in inversion. We have assumed that the be expanded to include AVO effects – which is usually done
reader has some knowledge of AVO and statistics. Our treat- by use of Shuey’s equation.
ment of the probabilistic versus deterministic topic is not
detailed – we intentionally focus on concepts instead of Model-based inversion
theory and equations. If we are successful in engaging the
A model-based inversion uses a forward model (see above) to
reader in this topic, the reader will have many follow-up
calculate synthetic seismic data as part of the inversion algo-
questions.
rithm. All of the inversion techniques discussed here – deter-
In the title of this article, we have used the terms deterministic ministic inversion, probabilistic inversion, stochastic inversion
and probabilistic to describe two different approaches to and GLI inversion – here are model-based inversions.
seismic inversion. While the reader may not be familiar with
these terms used in the context of inversion, the geoscientists Property models: impedance model, Vp model, Vs model,
that interpret seismic data, generate structure maps and esti- rhob model
mate hydrocarbon volumes of prospects and fields will be
familiar with deterministic and probabilistic reserve estimates. Property models are the input to a forward model. They are
The input variables in those reserve estimates – porosity, also the output of a seismic inversion. For a post-stack inver-
water saturation, gross rock volume and recovery factor – sion the property model is an impedance model. Pre-stack
frequently have a range of values and probabilistic reserve inversions require input (Vp, Vs, and rhob) property models.
estimates attempt to deal with these uncertainties by repre-
senting field reserves as a distribution characterized by the Relative and absolute inversion/ Relative and absolute
P10, P90 and mean reserve size. Contrasted with the proba- property models
bilistic approach is the deterministic field reserve method
Most inversion algorithms solve for absolute property models
where a geophysicist estimates a single reserve volume. The
(ie absolute impedance for a post-stack inversion) which are
deterministic method will commonly be used to give a ‘best
comparable to impedances from well log data. Creating rela-
estimate’ or ‘most likely’ case, but it can also be used to
tive impedance is one method for dealing with non-unique
generate ‘high-side’ and ‘low-side’ hydrocarbon volumes.
inversion results. Relative impedance models show – in a
Probabilistic and deterministic reserve estimates – and the
relative sense – high and low impedances that correspond to
associated concepts about uncertainty – are a very good
different lithologies, but those relative impedances are not are
analogue for dealing with the uncertainties inherent in
not comparable to the absolute impedances of well logs. The
seismic inversion. This article shows how seismic inversion is
key difference between relative and absolute impedance is
non-unique. A deterministic seismic inversion is any inver-
that a relative impedance model is missing the low frequency
sion that outputs just ONE of many different acceptable
component (approximately 0-15 Hz) found in an absolute
inversion solutions. An alternative inversion approach is to
impedances model.
search for ALL of the acceptable inversions (or more realisti-
cally, the statistics that characterize those inversions) which
GLI inversion
we refer to here as a probabilistic inversion.
Generalized linear inversion (GLI) has been and probably
Inversion terminology still is the most commonly used seismic inversion technique.
For each post-stack trace or pre-stack gather inverted, GLI
Before getting into a discussion of inversion and non-unique- inversion requires that the user supply a single initial estimate
ness, it is helpful to define some of the terms we will use here. of the earth’s property model(s) which is iteratively refined
There will be a more detailed technical discussion of these so as to give a synthetic (via the forward model) to match the
terms later. seismic data being inverted. GLI will output a single imped-
ance model for each trace or gather inverted.
Continued on Page 29
Deterministic inversions will output just one earth property Post-stack Deterministic Deterministic Deterministic/ Probabilistic
model. GLI inversion is always a deterministic inversion. If inversion for AI Probabilistic
stochastic inversion is outputting only the best fit property
model it finds, then it is a deterministic inversion. Pre-stack Deterministic Deterministic/ Probabilistic Probabilistic
inversion for Probabilistic
Probabilistic Inversion AI & PR
Probabilistic inversion is a term we use here to denote those Table 1: Recommended applications for deterministic and probabilistic
inversion algorithms that combine stochastic inversion with inversion.
Bayes’ Theorem1 to give rigorous probabilistic estimates of reser- Figure 1 shows a line from a 3D seismic survey and the associ-
voir properties and pore fluid (brine vs water vs gas). ated GLI/ deterministic inversion results. This GLI inversion
Probabilistic inversion is expensive and difficult to use, but it can was used to interpret geology/ stratigraphy and to initially high-
solve a long-standing challenge in our industry; how to use light the presence of hydrocarbons. One discovery and two
seismic attributes in a quantitative manner for risking exploration delineation wells were drilled based on this inversion. These
and development prospects. Probabilistic inversion is an deterministic results were followed up with probabilistic inver-
evolving technology, both in a technical and commercial sense. sions used to risk the presence of gas in near-by fault blocks, esti-
mate volumes of gas present and to help build static and
Geostatistical inversion dynamic reservoir models. We will use this same data set to
Geostatistical inversions are stochastic inversions that use vari- calculate a probabilistic inversion later in this article. But before
ograms to prepare the input property models. The purpose of we discuss the deterministic and probabilistic inversion of this
those variograms is to ensure that the property models fit dataset, we first need to explain how GLI inversion can be non-
expected spatial patterns (i.e. facies models). We will not discuss unique. That requires a bit of background on GLI inversion.
geostatistical inversion here.
Deterministic/ GLI Inversion
Probabilistic and deterministic inversions have different but
overlapping uses. Table 1 attempts to organize a high-level view How GLI inversion works
of where one might best apply deterministic and probabilistic
inversions. On the top row of Table 1 are some different applica- For this explanation of GLI inversion, we will simplify things by
tions for seismic inversion. These are: to facilitate interpretation assuming the forward model and input data are post-stack and
of geology and stratigraphy, as an initial scan for hydrocarbons, the property model being inverted for is thus an impedance
calculating the probability of hydrocarbon presence and using model2. The GLI impedance model is ‘parameterized’ using
Continued on Page 30
Figure 1. Deterministic-GLI inversion for AI and PR. Two separate gas reservoirs are indicated by the arrows.
Figure 2. Iterative loop used in model-based inversion. See text for a complete expla-
nation of each part of this algorithm. Note that the sensitivity matrix and the model
updates are calculated on band-pass filtered data (that is, data that has been filtered Figure 3. Convergence of impedance models with successive iterations in GLI algo-
with the seismic wavelet). Those updates thus do not have reliable information at rithm. User-supplied initial estimate (in blue) and successive iterations converging
higher-than-seismic and lower-than-seismic frequencies. on the final answer (red).
Continued on Page 31
Continued on Page 32
Figure 6. Two very similar synthetic seismograms (on the right) made from two
very different impedance models (on the left). Both impedance models describe a
Figure 4. Frequency spectra of input seismic data being inverted (in black) and reservoir thinner than the tuning thickness. Both synthetics have the same RMS
output impedance model (in blue). amplitude.
Figure 5. Two different impedance models. The model on the right can be described Figure 7. Two different impedance models are shown in black and blue on the left.
with blocky layers, while the model on the left can not. The model on the left They differ only in their low frequency content (lower than the seismic band-pass),
contains two low-impedance reservoirs: the shallow one is a ‘coarsening-upwards’ which causes different absolute impedance values. Synthetic seismograms made
reservoir and the deeper one is a ‘fining-upwards’ reservoir. from these two models are shown on the right – and are essentially identical.
Continued on Page 33
Figure 8. The dots represent porosity-impedance data from well logs. The heavy
solid line represents an impedance-to-porosity transform developed from the well
log data. This solid heavy line is frequently used to transform absolute impedance
inversion results into porosity. When the impedance inversion has low frequency Figure 9. The absolute impedance model on the right is filtered back to the seismic
errors (as in Figure 7), the absolute impedance results are non-unique. The thin band-pass to give the relative impedance model on the left. The x-axis scale refers
vertical and horizontal lines show how non-unique absolute impedances translate only the absolute impedances. Relative impedances have no low frequencies and will
into porosity errors. oscillate about zero impedance.
Continued on Page 34
Figure 10. Example of well data trend analysis. Dots are shale P-wave velocities from 9 different well logs. Green solid line is the average shale velocity as a function of depth.
Green dashed lines indicated the trend +/- two standard deviations. The yellow solid line indicates the sand P-wave velocity trend. Sand data points are not shown on this
plot. The sand trend is shown here to indicate how the velocity difference between sand and shale decreases with depth. Overpressure at deeper depths cause velocities to slow
down. Sand and shale trends are for normally pressured section only.
Continued on Page 35
Probabilistic/stochastic inversion trend analysis for sand and shale. (This figure is from a location
different from Figures 1, 11, 12 and 13. This well trend data set
The relative inversion solution discussed above is one way of was chosen to illustrate how sand and shale trends converge
dealing with non-unique inversion results. Relative inversion with depth and how pressure impacts velocity.) Note that the
essentially filters out all of the non-unique – and potentially average Vp trend and its standard deviation are defined in this
misleading – information. This filtering may be a bit drastic as the analysis. These trends are defined as a function of depth and
blocky layer assumption for thin reservoirs is frequently correct, apply to an inversion over any depth interval. Once the depth
as is the lower-than-seismic portion of the input initial model. An coordinate of the inversion is specified, a distribution of Vp
alternate way of dealing these non-unique solutions is stochastic (mean plus standard deviation) can be extracted from this trend.
or probabilistic inversion which attempts to find ALL of the A similar process is done for Vs and rhob, however, Vs and rhob
acceptable earth impedance models (or the statistics that repre- trends are developed as a function of Vp, not as a function of
sents all acceptable impedance models).
The motivation for doing many
different inversions on the same input
trace is the recognition that a single
model-based inversion result has non-
unique high and low frequency content
(discussed above). The high/ low
frequency content of an inversion solu-
tion is influenced by the high/ low
frequency of the initial impedance
model – so the input models for proba-
bilistic/stochastic inversion are chosen
to sample the range of range of possible
input models.
Probabilistic/stochastic inversion
needs information about the distribu-
tion of properties (Vp, Vs, rhob) in the
input models. These distributions are
generated prior to the inversion using
local well log data. This is a somewhat
interpretive step called trend analysis.
Well data trend analysis starts with
selecting relevant wells (we usually
need a minimum of three to five wells)
and manually interpreting/ picking a
lithology flag. The lithology flag just
indicates which intervals should be in
the sand trend analysis and which
intervals should be in the shale trend.
Figure 10 shows an example output Vp
Continued on Page 36
Figure 11a. PDFs for fluid factor AVO attribute from 400 stochastically simulated layer models. Red PDF is for gas only models. Blue PDF is for brine only models. Figure
11b shows the same PDFs, but scaled to reflection the prior probability of gas = 0.33. Note probability values of ‘b’ and ‘g’ for fluid factor = 0.25. Figure 11c shows calcu-
lated posterior probability of gas and probability of brine as a function of fluid factor. For each fluid factor, probability = g/(b+g) where b and g are taken from Figure 11b.
Continued on Page 37
Figure 13.
Continued on Page 38
Continued on Page 39