You are on page 1of 23

Publications of the University of Miskolc, Series A, Mining, Vol. 59.

Geosciences, (2001),

pp.115-138

T H E INVERSION OF WELL LOG DATA USING SIMULATED ANNEALING


METHOD

M. Dobrka1-2 and P. N. Szab2

Abstract
In the paper we present a global optimization method called Simulated Annealing is used to solve the nonlinear geophysical well logging inverse problem.
There are two possible ways to solve this problem. The first one is a conventional
inversion method, which estimates the unknowns in different depth points separately. The other way is to get these parameters is the so called interval inversion
procedure, which uses the data set along the whole interval in a joint inversion process. We made numerical investigations by synthetic and in-situ well log data to
compare these two algorithms so as to decide which is more powerful and yields
more accurate information about the downhole geological situation.

1. Introduction

The first stage of petrophysical interpretation for hydrocarbon exploration


is the evaluation of data sets measured in bore-holes. This operational problem is
generally solved by modern inversion methods with the application of up-to-date
informatics. The main object of this procedure is the detection of reservoir rocks in
the succession of strata, and determination of their layer-thicknesses and some petrophysical parameters. Consequently, the unknowns of the inverse problem are

1
2

MTA-ME Research Group of Geophysical Inversion and Tomography 3515 Miskolc-Egyetemvros


Department of Geophysics, University of Miskolc, 3515 Miskolc-Egyetemvros

115

M. Dobrka and P. N. Szab


geometrical parameters and several petrophysical characteristic values with relatively lot of zoneparameters of the formations, where the petrophysical parameters
have constant values in a layer, but the zoneparameters are invariable in a longer
depth-interval. These physical and geometrical parameters are also used in hydrocarbon forecasting and exploitation planning, therefore it is essential to determine
them as accurate as possible. In the industrial practice there are several softwares
for this goal, e.g. the CLASS, SAND, CRA which are deterministic procedures, but
the OPTIMA, OPTIMA 10 are probabilistic methods making complex reservoir
analyses by processing field data based on linear optimization methods. It is wellknown, that the linearized inversion methods are relatively quick, but the procedure
can often be trapped in local minima. Because of this fact the use of global optimization methods is reasonable.

2. The forward problem

In formulating the forward problem let us introduce the transposed vector


of the petrophysical model parameters as
m ={ POR,SXO,SW ,VCL,VSD,VLM }T
where
POR - porosity,
SXO - water saturation in flushed zone,
SW - water saturation beyond invaded zone,
VCL - specific volume of clay,
VSD - specific volume of sandstone,
VLM-specific

116

volume of limestone.

(1)

The inversion of geophysical well logging data...


To determine these parameters we utilise logs measured in the drill-hole, that record different parameters of natural and/or induced physical fields in the function of
depth, yielding continuos geological information on the formations of the drilled
hole. Let us take the following transposed vector, which contains the observed data
of a possible combination of well logs:

d(oil) = {SP, GR, PORN, DEN, A T, RMLL, RLLD}7

(2)

where
SP - spontaneous potential data,
GR - natural gamma-ray data,
PORN-

neutron porosity data,

DEN- density (scattered gamma-ray) data,


AT- acoustic traveltime data,
RMLL - microresistivity data,
RLLD - macroresistivity data.

The measured data reflect the immediate vicinity of the downhole, the SP and GR
logs are mainly sensitive to Iithology, the PORN-DEN-AT

logs indicate porosity,

and RMLL-RLLD data are influenced by water saturation. Since the measurements
are carried out in a relatively complicated borehole surroundings both in physical
and geometrical sense (high pressure and temperature, vertically and horizontally
inhomogeneous layers, anisotropy, changing hole calliper, different penetrationdepth and vertical resolution capability of tools etc.), therefore at first we need to
correct these data then match these logs in pursuance of depth. In the first step of
forward modelling we create a petrophysical model for the formation of interest on
the basis of the observed data set and a priori information (e.g. lab-examination of
cores), characterising quantitatively each layer. In the next step we calculate physical data from the parameters of this model by certain petrophysical relationships.

117

M. Dobrka and P. N. Szab


These relations are called response functions and they associate the model parameters with log data. The direct problem is solved by these multivariable functions, where there is a g connection between the predicted model parameter vector
and the approximated data vector by calculation

dtmk)=g(m,c),

(3)

where c- vector of set values like DENSD, ATCL etc. Let us survey the detailed
response functions of (2), which is applied also in a well logging data interpretation
system named ULTRA by Halliburton-Gearhart Co. (TH means theoretical value):
SPTH = SPSD - VCL (SPSD - SPSH),
DENTH

= POR[DENMF

"
+ VMAj
i=l

-C(l-SXO)]

+ VCL

DENCL
(5)

DEN MA,

GRTH - {POR[GRMF

SXO DENMF + GRCH{l -

+ VCL GRCL DENCL + VMA,


i=i
/ DENTH
PORNTH

(4)

= POR[PORNMF

SX0)DENCH]

DENMA, GRMA,}

- BCOR( 1 - SXO ) + BC] +


n

(7)

VCL PORNCL + VMAi PORNMA,

ATTH

= PORfSXO

+ YjVMAi

ATMF + ATCH (1-SXO)]

/ r x o t h

118

+ VCL

yjRCL

ATCL

(8)

ATMAjt

VCL

(6)

'

POR
SXO

siBA

RMF

f nN_~\
2 i,
2

(9)

The inversion of geophysical well logging data...


1
JRTTH

(10)

where n - the number of mineral components of rocks. On the left side of Eq. (4)(10) we can find the calculated data in bold letter, and on the right the model parameters stand in the same type, and there are also a lot of constants. Obviously,
the set of equations is nonlinear referring to the model parameters, so it is better to
solve the inverse problem by the help of a global optimalization method. During
the forward modelling we substitute the initial (and later the estimated) values of
(1) model parameters into the Eq. (4)-(10) equations, then after having data received by this procedure we compare them with in-situ data and make a prediction
for the real petrophysical model by an inversion method.

3. Inversion algorithms

The nonlinear well logging inverse problem is conventionally solved by


separated inversion which estimates the petrophysical model parameters in a depthpoint by the use of all available data in the point measured by different physical
principles. We allow for that adjacent depth-points independent, this way we can
not determine the layer thicknesses and we must fix the zoneparameters during the
procedure as the inverse problem is marginally overdetermined. The theoretical
data in the point is calculated by means of a local set of response equations (see Eq.
(4)-(10)). The calculated data vector of the y-th log can be written in a general
forms as

d^=gJ{ml,...,mu),

(11)

where M- the number of model parameters in the depth-point.

119

M. Dobrka and P. N. Szab


Since the number of observed data is only just more than the unknown
model parameters in the point, this narrow overdetermination causes that the accuracy and reliability of the estimation of model parameters is relatively limited. For
example there are six petrophysical parameters in (1) against seven well log data in
(2). Therefore it is worth inverting data of a greater interval jointly in one procedure. The so called interval inversion algorithm is based on the series expansion of
the petrophysical parameters. It develops depth dependent layer characteristic par a m e t e r s DOBRKA (1991):

(12)

where m - the z'-th model parameter, z - the depth, B^1 - series expansion coefficients,

- basis functions. For instance for layerwise homogeneous model we

can describe the model with the combination of unit step functions adopting the
least unknown parameters:

0, if z < zq_
1 , if z . < z < z
'

q-1

0 , if z > z

(If the petrophysical parameter shows vertical variation in a layer, we can choose
power or other appropriate basis function types). With this development in a series
the (11) connection modifies, it is now a response function interpreted in a depthinterval. The synthetic data calculated from they'-th log is

120

The inversion of geophysical well logging data...


(13)

where the unknown parameters are the B coefficients. By optimalization we can


determine B-s that approach the petrophysical model parameters along the overall
observed interval after substituting them into (12) formula.

3.1. Linearized inversion methods

These conventional search methods reduce the solution of the nonlinear inverse problem to sequences of linear problems. Before using these procedures, the
problem should be linearized. Let m0 be a point in the parameter space close to the
solution of the nonlinear inverse problem
m-m0+m

(14)

where 8 m denotes the parameter correction vector. Solving the linearized inverse
problem means that in a method of successive approximation m is chosen m0 in
the subsequent iteration step, and the process will be continued until a stop criterion is satisfied.

3.1.1.

Linearized separated (point by point) inversion

Let us assume that the M-dimensional model parameter column vector and
the Z-dimensioned transposed calculated data vector are:

in

={ml,m2,...,mM} T

={dt,d2,...,dL}T.

121

M. Dobrka and P. N. Szab


For an overdetermined inverse problem our goal is to determine the elements of the
model parameter vector by this equation

d = g(m),

(15)

where the functional dependence is generally nonlinear, and the number of the data
is greater then those of model parameters. Let us develop Eq. (15) in Taylor-series
in the neighbourhood of m 0 :

d = dl0) +GSm ,

where G - Jacobi matrix. This lead us to a linear set of equations

Sd = GSm ,

(16)

where the normalized terms owing to the different order of magnitude of the well
log data are

m.
jfobn)
a

8dk =dk

(obs")

m (0) 'dd<k>^
V

DM

, Y

Whereas Eq. (16) is inconsistent and the existence of measurement and model errors the prediction error vector of measured and calculated data is different from

122

The inversion of geophysical well logging data...


zero in any case. Therefore the problem can be solved by the minimalization by the
following norm

(17)

which makes the parameter correction vector in every iteration step during the (Iteratively Reweighted Least Squares) optimization. This is in the 17-th step:

=(^rW(9)GyGTW('l)Sd

m^

p-2

M
W

The goodness of inversion results are characterized in every depth-points separately with the relative model and data distances. They are in case o f p = 2 (Least
Squares Method)

1 L (W'"*1)

/ft"*)
100%

1(obx)

Dmod = J Y
'Mtr

M )

- mi'"" 0
m;(exact)

V
100%

where m(esl) - the estimated model vector, and mlaaa)

(18)

- the exact model in case of

inversion of synthetic data sets. The estimation error of the petrophysical parameters and the reliability of the solution can be characterized by the elements of the
covariance and correlation matrixes (MENKE, 1984).

123

M. Dobrka and P. N. Szab


3.1.1. Linearized interval inversion

Like in 3.1.1. the depth dependent layer parameters can be determined by means
of the minimalization the p-norm of the error vector of the overall observed and
calculated well log data:

DP

I
Sd*

(19)

/A

where DP - the number of depth-points in the whole observed interval. The goodness of result are

DP
DP lL f j
_/y('"/c) V
'hi
= J - J f t
100%
(oA.v)
VDPLtij
V
>"
/
1

Ddata

14

hi

hi

5
{exact)
rk

100%,

(20)

where S - the number of the layers in the interval.

3.2. Global optimization method - Simulated Annealing


The 3.1. methods are the most prevailing in the practice of inversion, because
in the case of having an initial model nearby the solution they are very quick and
effective algorithms. They are also capable of quality checking of the estimated
model parameters. After all they aren't absolute minimum searching methods, thus
they assign the solution to a local optimum of the (17) or (19) objective function

124

The inversion of geophysical well logging data...


with high probability. This problem is reformed by the Simulated Annealing (SA)
method, which performs the global optimization of the preference function with
random walking in the parameter space. SA was developed by METROPOLIS et al.
(1953) to model the thermal equilibrium state of solids.

In the metallurgy the removal of work-hardening of solids is realised by a


slow cooling manipulation from the temperature of liquid alloy state. This process
reduces progressively the kinetic energy of a large number of atoms with high
thermal mobility and then starts the crystallisation. Theoretically the perfect crystal
grating which has minimal overall atomic energy is produced by infinitely slow
cooling schedule. This is analogous with the stabilisation in the global optimum of
the objective function of a geophysical inverse problem. The quicker cooling
causes grating defects and the solid freezes in imperfect grid in a higher energy
state. It's similar to stagnating of the inversion process in a local extremum of the
objective function. However the atoms may escape from this higher energy state
owing to a special annealing process and then through a suitable slow cooling
schedule they can achieve the optimal crystal grating. The SA algorithmize this
progression to search the global optimum of the energy function (Metropolis algorithm).
The SA algorithm modifies the components of a model parameter vector in
every iteration step. The modification of they'-th model parameter can be write
m(;KW) = m(;'J) + b ,
where b - perturbation term. This small number varies between [b, femax], where
bmax decrease by
b(mw)
=b(M)
max
max

after a number of iteration steps (0< e < 1).

125

M. Dobrka and P. N. Szab


During the random walk the energy function of the relevant model is calculated and compared with the previous one in every iteration step. The acceptance
probability of the new model depends on the Metropolis criteria:

P(AE,T) =

It means that if the energy is lower in the new state than in the previous one, we
always accept the new model. If the energy of the new model has increased, there
is also a probability of acceptance depending on the value of the energy gain to escape from the local minima. If P(AE) > a fulfils - where a is generated with uniform probability from [0,1] - than the new model parameters are accepted but otherwise we reject them. During the process T temperature (only a control parameter)
must be decreased by an appropriate cooling schedule. The convergence of the inversion largely depends on the cooling schedule (g). We must balk too rapid cooling because the search can be frozen in a local minima, but too slow cooling is
neither preferred because of getting the correct solution much time later. The flowchart of the MSA algorithm can be seen in Fig. 1.

3.2.1.Global separated inversion

Let us see the definition the objective function of this inverse problem
which called energy function as we mentioned. If our data have Gaussian noise we
choose optimally

(21)

the L2 norm as energy function, which is accordant with the principle of least
squares. The (18) formulas are also eligible to quantify the goodness of results but

126

The inversion of geophysical well logging data...


the determination of the reliability of the estimated parameters nowadays is a
problem of long computational run time, therefore we did not deal with it.

3.2.2. Global interval inversion

From (12) the objective function is the following in case of Gaussian noise
1

DP

E22=yy
DP-Litf{

F J("*>)

J(CA!C)

r
d

()

START
Observed well logging data (d ( a h > )
Calculated data (d ( n k >)
Initial pctrophysica) modell ( n , )
Energy function ( E t )
Initial temperature ( T , )
Cooling schedule parameter ( g )
Maximal parameter perturbaticn ( b n u )
Perturhatioa reductive parameter (6 )
Maxin al number of iteration steps ( M A X S )
Maximal number of modification done (MC)

Selection of a parameter (m);


Generatioo of perturbation term (b);
m < " w) m (
+ b(oU)
:

AE = E '--E .

T>
bmax s bmax - B;

ESTIMATED
PETROPHYSICAL
MODELL

STOP

Figure 1.
The process flow diagram of the MSA algorithm

127

M. Dobrka and P. N. Szab


If we have also outliers in the data set it is more optimal choosing the following
function to optimise (Z,, norm)

(23)

This inverse problem is greatly overdetermined therefore we can determine the coordinates of formation boundaries in the inversion process. The Eq. (13) modifies
as follows
(24)

where the unknowns are also Xx,...,Xx

- layer boundary coordinates, besides B

coefficients and (13) contains these coordinates, too. The petrophysical and these
geometric parameters can be received by optimising (22) or (23).

4. Numerical results

In order to test the inversion algorithms in 3. we generated noisy synthetic well


logging data as quasi measured inputs. After processing them we calculated (18)
and (20) diagnostic values to quantify the goodness of the algorithms. We also
tested in-situ data collected in a Hungarian well in order to prove that these methods can also be applied on real geological structures.

4.1. Generation of synthetic data sets

To invert synthetic well logging data charged with some noise we defined a
series of strata that consists of four sedimentary layers. The petrophysical parameter of the model are shown in Table 1.
128

The inversion of geophysical well logging data...

H( m)

POR

SXO

SW

VCL

8.0
5.0
10.0
6.0

0.2
0
0.3
0.1

1.0
1.0
0.8
1.0

1.0
1.0
0.5
1.0

0.1
0.9
0
0.5

VSD
0.7
0.05
0.7
0.3

VLM
0
0.05
0
0.1

Table 1.
The four layered petrophysical model
The first column contains the layer thicknesses in meters and the petrophysical parameters of vector (1). As is seen, the third bed is the only HC reservoir with relatively large porosity. The synthetic data were calculated on this model by means of
eq. (4)-(10). We added 5 per cent Gaussian noise to these data to form the input
geophysical well logs to the inversion (see Fig. 2). This amount of noise is not less
then it is in case of real data. The sampling interval was 0.1 meter. Next we added
to the 25 per cent of these data additional 25 per cent random noise to simulate logs
charged with outliers. In Fig. 2a-b the curves with outliers are represented with
light line in the same diagram.

4.2. Inversion of synthetic data sets

The petrophysical parameters estimated by separated inversion with SA


method was determined by the optimalization of (21) for fixed layer boundaries.
This way we had only 7 data against 6 model parameters. In Fig. 3 it can seen that
the marginally overdetermined type of the inverse problem and the noise transformed from data to parameter space how different values of the parameters resulted in the same bed. Thus rather large uncertainty weighs on the quantities of
rock volumes and also the movable oil saturation (SXO-SW) in the third bed. The
data and parameter distances are the mean of (18) setting in Table 2 (where "o"
means outliers).

129

M. Dobrka and P. N. Szab


SP (mV)

GR (API)
20

40

60

SO

100

120

140

ct=01m

ck=Q1m

AC (MICS/M)
100

200

300

400

500

600

cb=aim
Figure 2a.
The synthetic data set charged with noises

130

The inversion of geophysical well logging data.

PORN(%)

DEN(G/CC)
0.5

1.0

L5

2.0

2.5

3.0

35

4.0

d=Q1m
RMLL(OHMM)

(t-Q1m
RLLD (OHMM)

10.0
10.00

100.00

30 - 1 -

ck=Q1m

ct=Q1m

Figure 2b.
The synthetic data set charged with noises

I3l

M. Dobrka and P.N. Szab


FaAn\\OLNCFR3CK
0.0

0.1

0.2

0.3

0.4

0.S

0.6

0.7

FHATKBaiA/ECFFOCK
0.8

0.9

1.0

0.0

0.1

0.2

0J

0.4

05

0.6

0.7

OJ

0.9

10

Figure 3.
Separated and interval inversion results

Well Logging
Inversion
Data (Noise)
algorithm
Synthetic (5%)
Separated
Synthetic (5%)
Interval
Interval
Synthetic (5% + o)
Synthetic (5% + o)
Interval
Synthetic (5%)
Interval
In-situ
Separated
In-situ
Interval
In-situ
Interval

LayerEnergy
thickness function
Fixed
e
Fixed
e j
Fixed
e
Fixed
e i
Unknowns
e
Fixed
e
Fixed
e ,
Unknowns
e i
2

2
2

Ddata

Dmodel

(%)

(%)

5.69
5.03
13.41
7.54
5.11
4.82
5.98
6.06

9.65
1.89
6.62
2.75
3.20

Table 2.
Accuracy of inversion results estimated by SA inversion methods

The SA procedure gave initial model independent and convergent solution. In


comparison with the LSQR method based on 3.1.1 (where Eq. (17) in case of p =

The inversion of geophysical well logging data...


2) where Ddata was 10.52 per cent and Dmodel was 5.65 per cent can be traced
that the global optimization improves the inversion.

The interval inversion beside unvarying layer thicknesses can also be found
in Table 2. In this case there were 2030 data against 24 unknowns, thus overdetermination highly enlarged in comparison to separated inversion. It can be seen in
Fig. 3 that interval inversion resulted more accurate parameter estimation compared local inversion. We mean under data and model distances (20) formulas. It
can be pointed out that circa 500 % improvement has been reached over separated
inversion, which implies a very accurate and reliable algorithm. On the other hand,
the LSQR method gave 2.27 per cent model distance in case of linearized interval
inversion which is less powerful than the global interval inversion procedure. In
practice it is not impossible that there are outliers in our well log data set. However, this fact can not be turned out to the detriment of accuracy of parameter estimation. Let us analyse the interval inversion of synthetic data charged with outliers. First we optimise (22) and then apply (23) which is famous for its resistance
against noises. The data distance is also defined in (23). In Table 2 it's detectable
that the results are adequate despite of the existence of outliers, therefore the SA
technique can be made resistant by the way of selecting an objective function based
on (17).

Let us take the layer-thicknesses into account as unknown model parameters in the interval inversion process. In Table 2 we can see that the goodness of
model parameter estimation is hardly fallen off for interval inversion beside fixed
layer-coordinates. These results are still three times accurate as in the case of separated inversion and therewith layer-thicknesses were estimated flawlessly and the
quickest (in the 3000. iteration step where the total number of iteration steps was
200000). As a consequence the interval inversion based on SA method estimates

133

M. Dobrka and P. N. Szab


more accurate oil saturation for the third bed which is very important task in petrophysical practice.

5. In-situ results

To invert real well logging data we chose some log measured in one high
part of a Hungarian exploratory bore-hole. The investigated stratigraphical complex was made of four unconsolidated sedimentary rocks, where the sand-bed was
an aquifer with relatively high porosity and a little amount of clay. Supposing simple lithology we treated POR, VSD, VCL parameters as unknowns for the fixed

Figure 4a.
The well logs as input data

values of SXO and SW The input data were the corrected values of SP, GR, DEN,
RMLL, CNC - compensated neutron - , RILD - deep penetration induction - logs.
The data set can be seen in Figure 4a-b and the results of separated and interval
inversion can be found side by side in Figure 5.

134

The inversion of geophysical well logging data...


mm

CHC (%)

' 125

RMLLO
(HMM)

RILD(OHMH)

ui 125
a

Figure 4b.
The well logs as input data

In case of separated (point by point) inversion there were 3 unknowns per depthpoint. Totally we had data in 195 points. In case of interval inversion we had to
determine 12 petrophysical and further 3 formation boundary-coordinates. From
135

M. Dobrka and P. N. Szab


Table 2 we can mention that the fitting in data space is satisfactory and the tendency is analogous with that in case of inversion of synthetic data. More accurate
parameters from the point of view of forward modelling can be get by the choosing
of more appropriate response functions and petrophysical constants. The interval
inversion can be got by choosing more suitable basis functions to (12) which describes better the vertical changes of the petrophysical parameter in the interval.
Lastly the interval inversion method has a large advantage over the separated inversion that it determines automatically the layer-thicknesses as it can be seen in
Figure 5.
RELATIVE VOLUME OF ROCK

RELATIVE VOLUME OF ROCK

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

Figure 5.
Separated and interval inversion results

6. Conclusions

We saw that the inversion of geophysical well logging data based on SA


results correct solutions. However, it is important to emphasise that the conver136

The inversion of geophysical well logging data...


gence of the process of this global optimalization method is largely influenced by
the setting of the control parameters like initial temperature, the cooling schedule,
etc. Another disadvantage is that SA require more computer run time than linearized methods. But stability, accuracy, independence of initial models are very
important advantages to form an effective algorithm. The interval inversion based
on SA with determination of layer-thicknesses has the capability to yield the ataunt
information about the investigated formation.

Acknowledgement

These investigations were carried out in the framework of the research program of
the MTA-ME Research Group for Geophysical Inversion and Tomography. The
authors thanks the support of the Hungarian Academy of Science.

References

W. MENKE, 1984: Geophysical Data Analysis: Discrete Inverse Theory. Academic


Press, INC.
M. DOBRKA: Egyttes inverzis algoritmusok bevezetse a mlyfrsi geofizikai
rtelmezsbe. ME, Geofizikai Tanszk, 1. Rszjelents (1991), II. Rszjelents (1993),
Zrjelents (1995).
P. N. SZAB, 1999: Mlyfrsi geofizikai mrsek kirtkelse globlis inverzis
mdszerrel. ME, Doktoranduszok fruma, Szekcikiadvny, pp 71-78.
P. N. SZAB, 2000: Mlyfrsi geofizikai adatok inverzija a Simulated Annealing
mdszer alkalmazsval. ME, Doktoranduszok fruma, szekci-kiadvny pp 5258.

137

You might also like