Computers & GeosciencesVol. 20, No. 7/8, pp.

1161-1168, 1994
Pergamon Copyright© 1994 ElsevierScienceLtd
009~-3004(94)F_,002~R Printed in Great Britain.All rights reserved
0098-3004/94 $7.00+ 0.00

IMPROVING VARIOGRAM REPRODUCTION ON DENSE

SIMULATION GRIDS
THOMAST. TRAN
Department of Petroleum Engineering, Stanford University, Stanford, CA 94305, U.S.A.

(Received 18 October 1993; accepted 28 December 1993)

Abstract--In any sequential simulation approach, limited computer resources call for retaining only the
conditioning data closest to the point being simulated. As a consequence, reproduction of large range
variograms is poor for dense simulation grids. This paper describes a multiple-grid implementation of the
sequential simulation algorithm that improves variogram reproduction significantly with minimal
additional computer and memory costs.

Key Words: Geostatistics, Simulation, Variogram.

INTRODUCTION Variogram reproduction as a function of size of the

In the sequential simulation approach, attribute
values at grid nodes are simulated sequentially ac-
cording to a random path visiting all nodes. The
attribute value at each grid node is drawn from a
local probability distribution (cdf) which is modeled
conditional to all original data and all previously
simulated node values (Deutsch and Journel, 1992,
p. 123-125). In actual implementation, because of
limited computer resources, only nearby conditioning
data within a limited search neighborhood are re-
tained. These nearby conditioning data are deter-
mined by visiting the nodes closest to the current
node first, and spiraling away until either enough
previously simulated nodes are located or the bound-
ary of the search neighborhood is reached. As
simulation progresses, the distances between the con-
ditioning data retained and the node to be simulated
become smaller relative to the variogram range;
hence, large range features of the variogram model
may be reproduced inadequately.
Variogram reproduction depends on a number of
interrelated factors: grid spacing, range of the vario-
gram model, size of the field to be simulated, size of
the search neighborhood, and the number of con-
ditioning data retained. The following experiment
was performed to better understand these relation-
ships. Using a spherical variogram model with an
isotropic range of 10 units and a 20% relative nugget
effect, unconditional Gaussian simulations were gen-
erated for various combinations of grid sizes, field
sizes, and numbers of nearest nodes retained. Each
realization was standardized to have mean 0 and
variance 1, and the corresponding exhaustive vario-
gram was computed. To attenuate the effect of er-
godic fluctuations, 100 realizations were generated
for each case, and 100 corresponding variograms
were averaged.
field to be simulated, number of data retained, and
size of the search neighborhood (Figs. 1-3) improves
as the field becomes large with regard to the range
(Fig. 1), the number of data retained increases
(Fig. 2), and these data span a larger range of
distances (Fig. 3).
In the sequential simulation approach, searching
for the nearby previously simulated node values is a
recurrent process which needs to be optimized. We
simply can compute and sort distances between each
data point and the node to be simulated and retain
the closest ones. If the data are on a regular grid,
however, a more efficient "spiral search" can be
utilized. Data are searched in the order of increasing
variogram distance from the node to be simulated.
This ordering is established by a process that also
computes and stores the corresponding covariance
values in a table. Ideally, this table should be large
enough to allow a "global" spiral search, that is a
search that could include all nodes of the simulation
grid. Because of computer memory and speed con-
siderations, the size of the covariance table, and
hence, the spiral search capacity, is limited. The
denser the simulation grid, the larger is the covariance
table for the same maximum variogram distance.
This entails that variogram reproduction deteriorates
as the grid becomes denser (Fig. 4); this situation can
be corrected by increasing the number of condition-
ing data retained and size of the search neighbor-
hood. Deutsch and Journel (1992, p. 30) pointed out,
however, that doubling the number of data retained
leads to an eight-fold increase in the CPU time
required to solve the kriging systems. Increasing the
size of the search neighborhood requires additional
computer time and memory for the spiral search and
the table of covariance values.
This paper presents an implementation of the
multiple-grid approach (G6mez-Hern~indez, 1991;

CAGEO 20/7-8~G 1 161

1162 T, T, TRAN

1.20 Field size: lOxlO units MULTIPLE-GRID UNCONDITIONAL SIMULATION

An unconditional multiple-grid simulation consists

f"
1.00
of several stages of simulation (Fig. 5). In the first
0.80 stage, an unconditional simulation is generated on a
coarse grid. In each of the subsequent stages, simu-
0.60 lation is performed on a finer grid, covering the same
5, area (or volume) and conditioned to values simulated
> 0.40 /
at the previous stage. This process is repeated until
/
the final grid is complete&
0.20
Within each grid, the nodes are visited according to
0.00
a random path. A spiral search is performed at each
0,0 4.0 8.0 12.0 16.0 20.0 grid node to gather the nearest conditioning data, as
Distance in the single-grid approach. The long-range com-
1.20 Field size: 25x25 units ponents of the variogram model are honored in the
early stages of the simulation, whereas short-range
1.00 components are reproduced in the later stages of the
simulation. This allows the size of the covariance
0.80 table to be kept small even for dense simulation
grids~
o 0.60
Variogram reproducuon of the multiple-grid ap-
>" 0 . 4 0
proach is compared against that of traditional single-
grid approach (Figs. 6 and 7). The same comparison
0.20 is made for a variogram model with a geometric
anisotropy ratio of 5:1 (Figs. 8 and 9) and a vario-
0.00 gram model with both zonal and geometric an-
0.0' ' '4:0' ' '8:0' ' "d.o' ' 'lg.O' ' '26.0
isotropy (Figs. 10 and ! I). The following conclusions
Distance
can be drawn from the comparisons:
t .20 F i e l d size: 50x50 units
• As expected, zonal anisotropy is difficult to re-
1.00
produce, particularly if the field size is small
(Fig. 10).
0.80
• Improvement by the multigrid approach in-
o 0.60
creases with the ratio of variogram range to field
size.
> 0.40
Average CPU time (in seconds) required to generate
0.20 one unconditional realization on a DECstation
5000/240 (Table 1) indicates that, with the muttiplc
0.00 grid approach, CPU time increases linearly with grid
0.0 4.0 8.0 12.0 16.0 20.0
size, whereas this is not the situation for the single-
Distance grid approach with global spiral search. This is
Figure 1. Varying size of field to be simulated. Parameters because early in the simulation, most of the grid
fixed are: grid size, 500 × 500; 16 nearest nodes retained. nodes have not been simulated, so the single-grid
Thin lines represent variogram model; thick lines represent
average of 100 variograms from 100 realizations. approach takes longer to search for the specified
number of conditioning data. With the size of the
covariance table fixed at 79 × 79 grid cells, the multi-
Deutsch and Journel, 1992, p. 124-125, p. 187; Verly, grid approach is slightly faster and yields better
1993) that improves variogram reproduction con- variogram reproductions.
siderably at minimal additional CPU and memory The given results were obtained using GSLIB's
cost. program sgsim (Deutsch and Journel, 1992,

Table 1. CPU time (in seconds) comparison

Method lOOxlO0 250x250 500x500,[ lO00xlO00
Single-grid
w/500x500 grid cells global spiral search 15 109 ,t60 1825
Single-grid
w/79x79 grid cells spiral search i
15 91 37:~ 1544
[ Multi-grid
Lw/ 79x79 grid ceils spiral search [ 14 85 337 il339
 Improving variogram reproduction 1163

1.20 8 nearest nodes retained 1.20 16 nearest nodes retained

1.00 1.00

0.80

0.60 o 0.60

>" 0.40 0.40

0.20 0.20

0 . 0 0 . • . , • • • ~ . • . , • • . , . . . ,
0.00 • ' ' ~ • ' " , " " " ~ " " • , "

0.0 4.0 8.0 12.0 16.0 20.0 0.0 4.0 8.0 12.0 16.0 20.0

Distance Distance

1.20 32 nearest nodes retained 1.20 64 nearest nodes retained

1.00

~
o
0.80

0.60 o 0.60
0.80

>" 0 . 4 0 >" 0 . 4 0

0.20 0.20

0.00 .... • . . , • • . , • • • , • • • , 0.00

0.0 4.0 8.0 12.0 16.0 20.0 0.0 4.0 8.0 12.0 16.0 20.0
Distance Distance
Figure 2. Varying number of nearest nodes (conditioning data) retained. Parameters fixed are: grid size,
500 × 500; field size, 25 × 25 units. Thin lines represent variogram model; thick lines represent average
of 100 variograms from 100 realizations.

p. 164-167) modified to incorporate the multiple-grid grid spacing (Ax)] and grid size NxL are computed as
approach. The number of intermediate simulation
(Ax)I = (ax)n × 2"- 1
stages is calculated by the program so that the initial
grid size is 2 x 2 for a 2-D grid. Grids are constructed and
so that each is refinement of the previous one: all
nodes of a current grid are a subset of the nodes of N x , l ~-

the subsequent grids. Int(...{Int[Int(N,,,/2 + 0.5)/2 + 0.5/2 + 0.5}...)

The number of simulation stages is
f [Nx"'~
. : max~log2t-~),
where N<.. and Ny,. are final grid sizes in the x and y
direction, respectively. For the x direction, the initial
n - 1 times
i { N"~"'~ } + 1,
og2t -7- ) where (Ax), is the final grid spacing.
Grid size and spacing for the y direction are
computed similarly. Note that the area covered by the
final grid.

Table 2. Grid specification for a nine-stage simulation

Stage Grid size x grid spacing y grid spacing Grid area
1 2x2 12.8 12.8 25.6x25.6
2 4x4 6.4 6.4 25.6x25.6
3 8x8 3.2 3.2 25.6x25.6
4 16x16 1.6 1.6 25.6x25.6
5 32x32 0.8 0.8 25.6x25.6
6 63x63 O.4 0.4 25.2x25.2
7 125x125 0.2 0.2 25.0x25.0
8 250x250 0.1 O.1 25.0x25.0
9 500x500 0.05 0.05 25.0x25.0
1164 T . T . TRAN

1,20 Search radius = 2 units i ,20 Grids&e: lOOxlO0

#
1.00 1.00

0.80 0.80

0.60
o 0.60
q >
:> 0.40 (140
f

0.20 0.20

0.00 ' r , , , , . . . . . . . . . 0.00 • , , t , , . n , , , I , • • , , , , ,

0,0 ' 4,0 8.0 12.0 16.0 26.0 0.0 4.0 8.0 12.0 16.0 20.0
Distance Distance
1.20 Search radius = 5 units 1,20 I Gridsize: 250x250

1.00 1.00

0.80 0.80

o 0.60 o 0.60
>
> 0.40 0,40

0.20 0.20

0.00 • , , , , , , , , , • , , . , , • . , , 0.00
0.0 4.0 8.0 12.0 16.0 20.0 00 4.0 8o 12.0 16.0 20.0
Distance Distance
1,20 Search radius = I 0 units 1.20 Gridsize: 5 0 0 x 5 0 0

1.00 1.00

fi~ 0.80 0, 80

e~ 0.60 0,60

> 0.40 >" 0.40

I).20 0,20

0.00 - . . , . , - , . , , , • , • , • , • , 0,00
0.0 4.0 8.0 12.0 16.0 20.0 00 410 *do* ' ,Lo ' t60 200
Distance Distance
Figure 3. Varying size of search n e i g h b o r h o o d . P a r a m e t e r s Figure 4, Varying grid spacing. P a r a m e t e r s fixed are: size of
fixed are: field size, 25 x 25 units; grid size, 500 x 500; 16 c o v a r i a n c e l o o k u p table; 79 × 79 grid cells; field size 25 × 25
nearest nodes retained. Thin lines represent v a r i o g r a m units: 16 nearest nodes retained. Thin lines represent
model; thick lines represent average of 100 v a r i o g r a m s from v a r i o g r a m model; thick lines represent average o f 100
100 realizations. v a r i o g r a m s from 100 realizations.

9×9
3x3 5x5
• 0 • 0 •

• - 0 O O •

• • • ' O •

• - O ~ •" O •

• e • e • • o •

Figure 5. Multiple-grid: u n c o n d i t i o n a l simulation, Black dots are c o n d i t i o n i n g data-nodes.

 Improving variogram reproduction 1165

1.20 A. Single grid 1.20 B. Multigrid

1.00

0.80

~z 0.60
o~

>" 0.40

0.20
F
1.00

0.80

o~ 0.60

>" 0.40

0.20
/
0.00 . • . , • . . , . • . , - . . , • . . , 0.00 • • • i . • • i • • . , • - . i . . • i

0.0 4.0 8.0 12.0 16.0 20.0 0.0 4.0 8.0 12.0 16.0 20.0
Distance Distance

Figure 6. Comparison of single vs multigrid approach for isotropic variogram model. Field size is 25 x 25
units; grid size is 500 × 500; number o f nearest nodes retained is 16; size o f covariance lookup table is
79 x 79 grid cells. Thin lines represent variogram model; thick lines represent average of 100 variograms
from 100 realizations.

F o r t h e e x a m p l e s s h o w n (Figs. 6, 8, a n d 10), t h e W e c a n use the m o r e e x p e n s i v e t w o - p a r t search at

final grid size was 500 x 500 with grid s p a c i n g s o f 0.5.
E a c h s i m u l a t i o n was p e r f o r m e d in nine stages w i t h
the specifications given in T a b l e 2.
MULTIPLE-GRID CONDITIONAL SIMULATION
F o r a c o n d i t i o n a l s i m u l a t i o n , at e a c h stage, the
original c o n d i t i o n i n g d a t a s h o u l d be used in a d d i t i o n
to t h e s i m u l a t e d d a t a f r o m the p r e v i o u s stage. A t
e a c h stage, care m u s t be t a k e n in d e c i d i n g w h e t h e r to
use a t w o - p a r t search o r a n e a r e s t - n o d e search. In a
t w o - p a r t search, t h e original c o n d i t i o n i n g d a t a are
s e a r c h e d for s e p a r a t e l y f r o m t h e newly s i m u l a t e d
n o d e values. In a n e a r e s t - n o d e ( o r o n e - p a r t ) search,
the original c o n d i t i o n i n g d a t a are assigned to t h e
n e a r e s t grid n o d e s ; all d a t a are t h u s o n a regular grid
w h i c h allows for fast spiral s e a r c h a n d r e a d i n g f r o m
a l o o k u p table the c o v a r i a n c e values n e e d e d for the
kriging systems.
each stage. Alternatively, we can save s u b s t a n t i a l
C P U time by r e l o c a t i n g the original d a t a to t h e
n e a r e s t n o d e s a n d e m p l o y a o n e - p a r t search. In t h e
first few stages, the grids are c o a r s e a n d a t w o - p a r t
search s h o u l d be used. In the later stages, t h e original
d a t a c a n be r e l o c a t e d to t h e n e a r e s t n o d e s . I f the
original d a t a are n o t clustered, r e l o c a t i n g t h e d a t a
w h e n t h e grid s p a c i n g b e c o m e s less t h a n o n e - h a l f the
smallest d i s t a n c e b e t w e e n t w o d a t a is a rule o f t h u m b
t h a t has yielded g o o d results.
CONCLUSION
T h e m u l t i p l e - g r i d a p p r o a c h to sequential simu-
lation allows a b e t t e r r e p r o d u c t i o n o f v a r i o g r a m
m o d e l s at essentially the s a m e C P U c o s t w h e n
c o m p a r e d to the t r a d i t i o n a l single-grid a p p r o a c h .
T h e a l g o r i t h m easily can be i n c o r p o r a t e d into
any simulation program using the
sequentialapproach.

1.20 A. Single grid 1.20 B. Multigrid

j /
1.00 1.00

~ 0.80

o 0.60
0.80

o 0.60

0.40 > 0.40

0.20 0.20

0.00 . . . , . . , . . . , . . . , . . . , 0.00 • • . i . • • , • • . i • • • i . . . J

0.0 4.0 8.0 12.0 16.0 20.0 0.0 4.0 8.0 12.0 16.0 20.0
Distance Distance
Fisure 7. Comparison of single vs multigrid approach for isotropic variogram model. Field size is 50 x 50
umts; grid size is 500 x 500; number o f nearest nodes retained is 16; size of covariance lookup table is
79 x 79 grid cells. Thin lines represent variogram model; thick lines represent average of 100 variograms
from 100 realizations.
 1.20 t Single grid: EWdirection / t.20 Single grid: NS direction

1.00 ~ 1.00

~
o
0.80

0.60 ~0.60
0.80

> 0.40 0.40 /

/
/
0.20 (1 21}

0.00 1 , , . , , . . , , , , . . . , . . . . OX)O
0.0 4.0 8.0 12.0 16.0 20.0 0.0 4.0 8{) 12.0 16,0 20.0
Distance Distance

Muitigrid: EWdirection
1.20 1.20 Multigrid: NS direction

1.00 "l 1,00

10,80 ~0.80

.:- 0.60 .~ 0.60

>" 0,40 > 0.40

0.20 (t.20

0,00 0.00 ' ' ' i '

0.0 4,0 8.0 12.0 16.0 20.0 0.0 4.0 8.0 12.0 16.0 20.0
Distance Distance
Figure 8. Comparison of single vs multigrid approach for a variogram model with a geometric anisotropy.
Field size is 25 × 25 units; grid size is 500 x 500; number of nearest nodes retained is 16: size of covariance
lookup table is 79 x 79 grid cells. Thin lines represent variogram model; thick lines represent average of
100 variograms from 100 realizations,

1.20 7 Single grid: E W direction

,00 1.20 Single grid: NS direction

/
1.00

~
o 0.60
0.80

o
0.80

0.60
~,
> 0 . 4 0 > 0.40

0.211 {).20

0.00 £ . . , , , , , . , • , . , , , , • , , 0.00

0.0 4.0 8.0 12.0 16.0 20.0 0.0 4.0 80 120 16.0 200
Distance Distance

1.20 Multigrid: E W direction 1.20 Muitigrid: NS direction

?
1.00 I .00

E 0.80 (}.80

0.60 o (1,60

> 0.40 > 0.40

0.20 0.20

0.00 0.00

0,0 4.0 8.0 12.0 16.0 20.0 0.0 4.0 8.0 l 2.0 16.0 20.0
Distance Distance
Figure 9. Comparison of single vs multigrid approach for a variogram model with geometric anisotropy
Field size is 50 × 50 units; grid size is 500 x 500; number of nearest nodes retained is 16: size of covariance
lookup table is 79 × 79 grid cells. Thin lines represent variogram model; thick lines represent average of
100 variograms from 100 realizations.
 Single grid: EW direction 1.20 Single grid: NS direction
1.20

1.00 1.00

0.80
~ 0.80

o 0.60
0.60

> 0.40 0.40

0.20 0.20

0.00 0.00 • . , i . • • i " • • i " " " i • " " i

0.0 4.0 8.0 12.0 16.0 20.0 0.0 4.0 8.0 12.0 16.0 20.0
Distance Distance

Multigrid: NS direction
1.20 Multigrid: EW direction 1.20

1.00 1.00

0.80
~ 0.80

o 0.60
... 0.60
¢¢

;> 0.40 > 0.40

0.20

0.00
f • • • i . . . i • - . i . • . i • . . i
0.20

0.00 ' ' • i - • . . . • . i • • . i . . . i

0.0 4.0 8.0 12.0 16.0 20.0 0.0 4.0 8.0 12.0 16.0 20.0
Distance Distance
Figure 10. Comparison of single vs multigrid approach for a variogram model with zonal anisotropy. Field
size is 25 x 25 units; grid size is 500 x 500; number of nearest nodes retained is 16; size of covariance
lookup table is 79 x 79 grid cells. Thin lines represent variogram model; thick lines represent average of
100 variograms from 100 realizations.

1.20 Single grid: EW direction 1.20 Single grid: NS direction

1.00 1.00

~ 0.80

o 0.60 o 0.60

> 0.40 0.40

0.20 0.20

0.00 • • • , • • , i , , • i • • • J . . . ,
0.00 • . , i . . . i . . . i • • . i " " " i

0.0 4.0 8.0 12.0 16.0 20.0 0.0 4.0 8.0 12.0 16.0 20.0
Distance Distance

1.20 Multigrid: EW direction 1.20 Multigrid: NS direction

1.00
1.00

~ 0.80

o 0.60
~ 0.80

o 0.60

0,40 ;> 0.40

0.20 0.20

0.00 . . . , . . . , . . . . . . . , . . . , 0 . 0 0 • • • , • . . , . • • t • • • i • • . ,

0.0 4.0 8.0 12.0 16.0 20.0 0.0 4.0 8.0 12.0 16.0 20.0
Distance Distance
Figure 11. Comparison of single vs multigrid approach for a variogram model with zonal anisotropy. Field
size is 50 x 50 units; grid size is 500 x 500; number of nearest nodes retained is 16; size of covariance
lookup table is 79 x 79 grid cells. Thin lines represent variogram model; thick lines represent average of
100 variograms from 100 realizations.

1167
1168 T.T. TkA~

REFERENCES data measured at a smaller scale: unpubl, doctoral

Deutsch, C. V., and Journel, A. G., 1992, GSL1B: Geostatis-
tical Software Library and User's Guide: Oxford Uni-
versity Press, Oxford, 340 p.
G6mez-Hern~ndez, J., 1991, A stochastic approach to the
simulation of block conductivity fields conditioned upon
dissertation, Stanford Univ., 351 p.
Verly, G. W., 1993, Sequential Gaussian cosimulation
a simulation method integrating several types of
information, m Soares, A., ed., Geostatistics troia
'92: Kluwer Academic, Dordretcht, Netherlands,
p. 543 554.