Professional Documents
Culture Documents
Transmissibility for 2D
Reservoir Studies
Alexander J. Desbarats, Geological Survey of Canada, and
Roussos Dlmltrakopoulos, SPE, McGill U.
Summary. A geostatistical approach is used to characterize reservoir transmissibility with the aim of assigning sirnulator parameters
in 2D models. Transmissibility is represented as a spatial random function where heterogeneity is described by the probability distribu-
tion and the variogram of sample values. The key element of the geostatistical model is the definition of block transmissibilities as
spatial geometric averages. Published analytical results have shown that the effective transmissibility of infinite, statistically isotropic
flow fields is equal to the ensemble geometric mean. Numerical results presented here show that a spatial geometric average is an excel-
lent approximation of effective transmissibility in such finite fields as simulator gridblocks. The geostatistical model for transmissibility
is used to show that the mean and variance ofblock-averaged values depend on the averaging area. As the averaging area is increased,
mean block transmissibility decreases toward the ensemble geometric mean while the block variance decreases to zero. The geostatisti-
cal model is also used to investigate the kriging of block transmissibilities from well data. The current method of correcting bias in
kriged values is found to cause artifacts of gridblock size in flow simulation results. The simpler, uncorrected kriging estimator is shown
to preserve overall flow-field transmissibility, regardless of gridblock size.
/////,//////// ///////////////////
"1. . T46
Clnrular ve Frequency
lr
-s3
E(Yt1:- lElY(x.lld-x:o ......(9a)
a 1..
T45 and var(Y5):-J \ ollr-sl;drds:o1S,S), ......(9b)
J'ss
where o(S,S) represents the average value of dft) when the extremi-
ties of the lag h independently describe the area,S. Values of o]S,S)
for a specific covariance or variogram model are determined with
numerical integration charts.3 Fig.4 shows the decrease of olS,D
with increasing dimensionless length Z/tr of an averaging area
S=L2, for the variogram model of the H sandstone data.7
Consider the spatial geometric average of (.r), T5, oVor a block
^S:
Data Model
Logarithmic mean 3.735 3.735 o-o
5oo 1000 1500 2000 2500
Logarithmic variance 1.324 1.324 Distance Lag h (m)
Mean, md.m/Pa.s 83.560 81 .194
Geometric mean, md.m/Pa's 41 .88 41.88 Fig. 3-Experimental varlogram of log tlansmlsslbility.
Variance, (md.m/Pa's)2 22420.07 18182.66
FO
o!
o.o 1 I t.o ro-o
L/)\'
o.o
Flg. A-Varlance of ys vs. S of slde L/tr. 0.o 2.O 3.O
Ts/Tc
The early work of Cardwell and Parsonsll and Warren and
Pricel has traditionally provided an empiricaljustification for ge- Fig. S-Normallzed f" vs. normallzed f5 lor a 500.m block.
ometric averaging in random, uncorrelated media. Later theoreti-
cal workl2-14 in the geostatistics and hydrology literature has
shown that effective transmissibility in infinite, correlated, 2D het- mated by ?5 in such fields because 15 tends toward 16 and
erogeneous media is approximately equal to 26. In particular, therefore toward {, as the field size becomes large. 1l This con-
Matheron12 showed that T":fo if transmissibilities are distrib- jecture is evaluated here with the H sandstone transmissibility model
uted log-normally. These analytical results are valid only when for- as an example.
mations are statistically isotropic in the (-r,y) plane. Fortunately, A Monte Carlo method was used to generate spatially correlated
this is often the case in thin formations of large areal extent where point support-scale transmissibilities at the nodes of a grid discretiz-
2D studies are applicable. The case of infinite, statistically anisotrop- ing a 2D flow freld. The simulated values respected the log-normal
ic fields is considered by Gelhar and Axness.15 distribution model and the variogram nnodel obtained above for the
These results also assume that transmissibilities are distributed H sandstone. Tg and T" were calculated for each realization of the
log-normally, or at least have a small variance. Numerical workl random freld. T, was calculated numerically by solving the steady-
has shown that the results are relatively insensitive to moderate state-flow equation and determining the total flux through the sys-
departures from these conditions. Further numerical workl6 ex- tem. Two different flow-field sizes were considered: 500x5fi) m
amined the case of sand/shale sequences where high-permeability (l \ x I X) discretized by a 40 x 4O-block grid and 2500 x 2500 m
contrasts and anisotropic-spatial-correlation structures create (5I x 5^) discretized by a 50 x 50-block grid. Figs. 5 and 6 show
preferential flow channels. Z6 was found to be a poor estimate of Z" plotted vs. Ig for the two block sizes. Values are normalized
true effective permeability in this case. by ?6. The results for both field sizes show that T5 is indeed an
The work cited above does not address the present problem of excellent approximation to T" under the present geostatistical
determining T, in realistic hnite correlated frelds, such as simula- model. For the small freld, the average of T"lT6 is equal to the
tor gridblocks. It does suggest, however, that T" may be approxi- average of TslT6 and is found to be 1.51. For the larger field, the
average of normalized transmissibilities is 1.07. Thus, while
TS=7", the average of these values appears to depend on the size
of the averaging area. This important observation is investigated
next.
2-O
F9
-FO 1-5 ,(.,
l.
7ro
't.o uJ
I
,
o.5 tl
o.o
o.o o.5 1,O 1.5 2.O 2.5 3.O o.o1 1 0.o
Ts/Tc L/\
Fig. 6-Normalized f" vs. normalized 75 for a 2500-m block. Fig. 7-Normalized mean of 7s vs. S of slde L/tr.
5 lo 20 ao 00 ao 90 95 sa99 so-6
Support.Scale Eflects bilities from buildups is 0.0128 md2, compared with 1.440 md2
Support scale refers to the physical area of spatial averaging, ,S, for cores. Geometric mean permeabilities from core (48 md) and
which appears in the definition ofblock-averaged quantities such buildups (37.5 md) agree reasonably well, as predicted by Eq. 13.
as Is. 17 This section examines the important effects of support This need not always be the case, however, because permeabilities
scale on the mean and variance of block transmissibilities. from buildups are not truly spatial geometric averages as defined
T5 is a function of the point-support random variables I(.r) and in Eq. 10.
therefore is also a random variable. Using a Taylor series expan-
sion of ers truncated at order two, the mean and variance of 15 Krlglng Block Transmlsslbllltles
can be approximated by The 15 gridblock values required by flow simulators can be ob-
tained by conditional simulations or interpolation from data at sur-
E(Tr7=B1rrt',-loso(s,S)t2 ..(l2a) rounding wells. This section examines the use of kriging for the
and var(T5)=var(eYs)=f,1792[e;(s's)-l]. . . .. .(l2b) estimation of gridblock transmissibilities. Kriging is a least-squares
technique for spatial interpolation that uses the lanowledge of a vari-
If the random function I(.r) is multivariate normal, then it is easily
able's spatial covariance structure to obtain a best linear unbiased
shown with the properties of the log-normal distribution 18 that the
estimator of the variable at some unsampled location.S
relations in Eqs. l2a and l2b are exact. In particular, when the Following the same approach as previous authors,4'G8 the first
support size S is reduced to a point, Eqs. l2a and l2b become the
step is to determine a kriging estimate, I$, of the true spatial aver-
classical formulas linking the mean and variance of (r) with a age of log-transmissibility, 15. IS is a weighted average of data
and o2, the mean and variance of Y(x). values Z(.r1) at n well locations:
These two relations show that both the mean and the variance
n
of Z5 depend on the area of averaging S through the term o(S,S).
Figs. 7 and 8 show the relationships for the H sandstone transmis- r$: D oiY(x). ....(14)
sibility model. Normalized meanlE(TjlT6l and normalized var- i: I
iance [var(25) lT62l are plotted vs. the dimensionless length Z/tr The n weights, c,ri, are determined so as to minimize the expected
of a field S:22. As the averaging area becomes very large, the squared error or kriging variance ofr;
mean of T5 tends toward T6, while the variance tends toward zero.
As the averaging area tends toward a point, the mean and variance ok:EIVs-y$21, .. ........(15)
of T5 tend toward the mean and variance of Z(x). Eq. l2a explains
subject to the unbiasedness constraint
the results of the previous simulation experiment, where the aver-
age transmissibility of the 500x500-m (l x l-X) blocks was great- n
er than that of the larger 2500 x 2500-m (5 x 5-I) blocks. The results D ,i:1. .........(16)
obtained (1.07 and l.5l for the large and small blocks, respective- i= I
ly) agree well with the theoretical values of l.l0 and 1.49 predicted
The constrained least-squares optim2ation leads to a system of n * I
from Eq. l2a.
The ensemble geometric mean of block transmissibilities is
equations in z*l
unknowns: the <d; and a Lagrange multiplier, 4,
introduced by the constraint. The optimum kriging variance can
another quantity of interest, given by
be written in terms of the solution of the svstem:
exp[E(ln Ts)]:explE(Is)l=e":Tc. ...(13) n
estimated values, however, is always less than the variance of true 2Oa can be corrected with the estimator 7F$* 19'
values because (ofr+2n1 is a positive quantity. This variance
reduction is common to all least-squares methods and is known as
T{*:f{e<"rtr2)+n. . .........(22)
the "s.moothing effect" in the geostatistics literature. Maps of I$ The mean of T$* then is
and o'x for the rectangular region outlined in Fig. I are shown in
Figs. 10 and 11, respectively. Areas where f$ shows little^spa-
E(r$*):E(T$1e@2xtz)+n:E(Ti. ......(23)
tial variability are seen to correspond with areas of high of . In similar bias correction5,T'8 4 has erroneously been neglected.
Kriged estimates of block transmissibilities, ?$, are then ob- Although the corrected estimator resolves the bias problem, the en-
tained from f$ by exponentiation: semble geometric mean of ?$* is now:
Tt:rry. ....(1e)
"E(hT!+):eE(vpe@2K12)+tt:TGe@2K12)+4.
.. ... (24)
Following the same method used to derive Eqs. 12, it can be shown In other words, T$* causes an overestimation of the ensemble ge-
that the mean and variance of T$ are ometric mean, and therefore, of the effective transmissibility of the
entire discretized field.
E(T$)=611r;s-(o2dz')-'t .....(2Oa) A kriging study was performed on the H sandstone transmissi-
and var Q$=E(Til21"its,s)-o2*-z,r-ll. .. ....(20b) bility data to illustrate the discussion of estimators T$ and ?$*.
The rectangular region (Fig. l) was subdivided into a regular grid
As before, these relations are exact if I(.r) is multivariate normal. ofsquare blocks, which then were kriged one by one using all 23
Also, the ensemble geometric mean of T$ is given by samples. The array of kriged block transmissibilities was entered
sE(Y$):sa:Tc. . ....(21) into a single-phase, steady-state-flow simulator. Constant poten-
tial boundaries were applied on the north and south sides of the
According to Eq. 20a, T$ is a biased estimator of T5, where the region, while no-flow boundaries were applied to the east and west
biasing factor, e-k'xl2)-tt, results from the variance reduction or sides. The effective transmissibility of the region was then calcu-
"smoothing" effect. However, T$ does preserve the ensemble ge- lated from the numerical solution to the flow equation. The krig-
ometric mean, which is equal to the effective transmissibility of ing and flow simulation were repeated for blocks of 250,500, and
the entire field under ergodic conditions. The bias problem in Eq. 1000 m. Tables 2 and 3 show the results for T$ and I$*, respec-
tively, including block size, the arithmetic aveftge of kriged block 2. Geometric spatial averaging is used to determhe block-support
values, T,a [i.e., an estimate of the theoretical expected transmis- transmissibilities from point support-scale values. Numerical re-
sibility E(T5), which is also givenl, the geometric average of sults presented here show that spatial geometric averaging provides
kriged block transmissibilities, T6, the effective transmissibility, excellent estimates of effective transmissibility in the case of finite
T, of the field calculated from the flow simulation, and the en- correlated fields, such as simulator gridblocks.
semble geometric mean, T6, which is equal to the effective trans- 3. The geostatistical model for transmissibility explains the de-
missibility of the region under the present conditions. pendence of the mean and variance of block-averaged quantities
A comparison of I,a and E(T5) values in Table 2 shows the un- on the averaging area and on the spatial-correlation structure of
derestimation of E(I5) predicted in Eq. 20a. The relative underes- transmissibility. A consequence ofthis support-scale effect is that
timation worsens with smaller block sizes because of the higher the mean of gridblock or well-test transmissibilities is lower than
kriging variances. The E(?"5) values also reveal their significant the mean of core values.
dependence on support scale with a20% re&rction in mean trans- 4. The geostatistical technique known as kriging is used to de-
missibility as block size increases from 250 to 1000 m. 76 and T, termine optimum estimates of block transmissibilities by spatial in-
values agree remarkably well, and results show no dependence on terpolation of well daa. It is shown how the bias correction of kriged
block size. This further confirms that geometric spatial averaging estimates described in the literature leads to a spurious overesti-
does indeed yield an excellent estimate of effective transmissibili- mation of the overall effective transmissibility of the flow field when
ty. The T6 and Z" values also are very close to the ensemble geo- kriged block values are entered into a simulator. The original un-
metric mean, 16. Thus, while I$ underestimates T5, it preserves corrected kriging estimator does not introduce artifacts ofgridblock
the geometric mean and effective transmissibility of the region as size in simulation mass flux results and therefore should be used.
a whole. 5. The analytical resuls presented here are valid only for the case
The T1 and E(25) values in Table 3 are in much better, but not of 2D flow in a statistically isotropic, low-variance transmissibili-
perfect, agre€ment. The discrepancies between estirnates and model ty field without sources or sinks.
values are caused by the nonuniform distribution of sample loca-
tions over the kriged region. The undersampled, lower- J{omenclature
transmissibility areas in the northeast and southwest areas of the E( ) : expected mean value
region are more proportionately represented in the calculation of
TA thar. in the calculation of the average sample value used in the
i : distance lag in variogram or autocovariance
function, m
geostatistical model. Also, because ofthe nonuniform sample spac-
ing, kriging variances are quite variable (Fig. I l). Eq. 23 is strict- I: permeability, md
ly verifred only if the kriging pattern is identical for each block ,Z : length, m
throughout the region. Under these circumstances, the agreement n: number of samples in kriging estimator
between Ta and E(Tj values in Table 3 is acceptable. As in Ta- .l : averaging area or gridblock, ha
ble 2, T6 and T" are in excellent agreement. These results, how- I, : effective transmissibility, md'm/Pa's
ever, show a disturbing dependence on block size and significantly Tc :'ensemble geometric mean transmissibility,
overestimate T6. md'm/Pa's
Thus, ?$* is an unbiased estimator of 75, but it introduces a spu- Is : spatial geometric average transmissibility,
rious support-scale artifact in the overall effective transmissibility md'm/Pa's
ofthe field. Ifkriging is to be used to determine gridblock trans- (x) : transmissibility at location.r, md'm/Pa's
missibilities for flow simulators, then the estimates should ensure f5 : spatial average of Y(.r) over S
that conservation of mass flux through the held is not affected by
an arbitrary choice of block size. For this reason, T$ should be
(x) : *1ot.t logarithm of transmissibility at location .r
used as an estimator ofT5 rather than I$* as is the current prac-
a : logarithmic mean transmissibility
tice. Although the problems created by the smoothing effect of krig- 7(h) = variogram function of distance lag ft
ing can be corrected in the present case, no similar correction is r1 : Lagrange multiplier in kriging system of equations
known for 3D reservoir models and conditional simulation is there- ), : integral range of correlation in variogram or
fore the preferred spatial interpolation method. autocovariance function, m
:
1^r viscosity, cp
Concluslons o(ft): uolo.ovariance function of Y(x) at distance lag ft
1. Geostatistical theory provides a rigorous mathematical frame- o? : logarithmic variance' of transmissibility
work for quantitative reservoir description and the numerical model- of : kriging variance
ing of transmissibility in heterogeneous formations. co; : kriging weight of rlh sample