m~ê~ãÉíÉê=Éëíáã~íáçå=çÑ=íê~åëáÉåí=Ñäçï=éêçÄäÉãë

ïáíÜ=mbpq=áå=cbcilt

Jan Christian Kaiser

N
ITUPH Institut für Theoretische Umweltphysik, Munich, Germany

N m~ê~ãÉíÉê=Éëíáã~íáçå=çÑ=íê~åëáÉåí=Ñäçï=éêçÄäÉãë=ïáíÜ=mbpq=áå=cbcilt

for observable physical quantities, will profit from the

NKN fåíêçÇìÅíáçå application of concepts and methods of data assimila-
tion. In this analysis technique observed information
The integration of the PEST package for parameter (i.e. measured gauge curves of hydraulic heads) is used
estimation (Doherty et al.6) into the FEFLOW to improve the model state (i.e. FEFLOW's material
simulator4 was accomplished in FEFLOW's version 4.8 parameters) and to enhance the predictive power of the
for stationary flow processes. The coupling has been model. Two basic approaches to data assimilation are
realized via FEFLOW's interface manager (see known (Bouttier and Courtier1):
Gründler8) using the method of direct implementation.
The PEST module has access to FEFLOW's address • Sequential assimilation, which considers only
space and is called within its graphical user interface. observations from the past until the time of analy-
Data exchange is facilitated during run time with call- sis. This is the case for models of meteorological
back functions. A description of the PEST-FEFLOW weather forecast or flood forecast models in real-
coupling is given in Kaiser11. time applications.
• Retrospective assimilation, where observations
In FEFLOW's version 5.2 the range of applicability from any conceivable time range can be used for
of parameter estimation with PEST has been extended model calibration and validation. Groundwater
to transient flow problems. The extension required only models are typically treated with this approach.
minor changes in the PEST graphical user menu for
observations, since the time-dependent observation Both approaches can use the observed information
curves are treated as power functions, for which either intermittently or continuously in time. In the
FEFLOW provides a comprehensive data management. intermittent mode, observations are grouped in time
intervals, which are much smaller than the total simula-
From a theoretical point of view the analysis of tion time but much larger than the time scales of the
transient simulation models, which produce predictions underlying physical processes. Usually, this mode is

«aef=ö=ïïïKãáâÉéçïÉêÉÇÄóÇÜáKÅçã cbcilt=ö=N
m~ê~ãÉíÉê=Éëíáã~íáçå=çÑ=íê~åëáÉåí=Ñäçï=éêçÄäÉãë=ïáíÜ=mbpq=áå=cbcilt

technically convenient. In the continuous mode obser- n

  oi  t  – Mi  t p  
T –1
vations over the whole simulation time of the model J p = Ri  oi  t  – M i  t p   (1-1)
are considered to improve the model state, which is i=1
physically more realistic.
by summing over all n observation points. The vector
Most subsurface flow models are conveniently
treated with retrospective data assimilation in the con-
 o t  
tinuous mode. To make use of the PEST algorithm to  i 1 
minimize the objective function, the model parameters oi  t  =    (1-2)
 
must be constant in time. The general data assimilation  oi  tm  
concept includes time dependent model parameters as
well. However, in this case the Kalman filter technique
should be used for the parameter update. This tech- contains the recorded measurements of the hydraulic
nique is mathematically equivalent to the minimization head for m time steps of the gauge curve pertaining to
of the objective function†) (Bouttier and Courtier1). observation point i . For the vector M i  t p  an analo-
gous expression is valid. Measurement errors are mod-
In the present implementation complete use of the eled here in a simplified way. They are treated as
observed information is made by considering measure- uncorrelated and identical for a given observation
ments at each observation point i of the model and at point. After normalization they are represented as
each time step t j . Measured hydraulic heads o i  t j  can weights of the gauge curves. Note, that for each obser-
be compared with model predictions M i  t j p  by using vation point an individual sequence of time steps
the callback function OnTimeStepConstraint() exists, which may or may not differ from the sequences
of the interface manager. This functions ensures, that of other observation points. The set of FEFLOW mate-
the hydraulic head is calculated for the whole finite ele- rial parameters p min minimizes the objective function
ment grid at a prescribed time step irrespectively of the J  p  and can be considered as the optimal parameter
selected algorithm for time stepping. The vector p con- set of the parameter estimation problem. A full descrip-
tains the time-constant material parameters of tion of the mathematical framework of PEST and of the
FEFLOW, which are updated in the estimation process. algorithms to minimize the objective function is given
in Doherty et al.6.
Now the objective function J  p  can be constructed
with the residual vectors o i  t  – M i  t p  and the covari- In the following section four simple test examples
–1
ance matrix Ri of the measurement errors are presented to demonstrate the reliability of the
implementation of PEST in FEFLOW for transient
flow problems. The examples cover both confined and
unconfined aquifers with measured and synthetically
†) called 4D variational assimilation (4D VAR) in Bouttier generated observation data. In Theis' well problem for
and Courtier1.

O=ö=cbcilt
 NKO=qÉëí=bñ~ãéäÉë

an unconfined aquifer synthetically generated gauge explained in Tab. 1.1.

curves are used to retrieve the material parameters of a
FEFLOW benchmark problem. The pump tests of The time stepping algorithms of the FEFLOW
Breyell and Wichita have been conducted in a confined groundwater simulator have been successfully tested
and an unconfined aquifer. They were analyzed already against the Theis benchmark (Diersch5) on the finite
by Langguth and Voigt12 with graphical evaluation element grid of Fig. 1.1. Here the same grid with the
methods. Here they have been reanalyzed with PEST in same values for the simulation parameters (Tab. 1.1) is
FEFLOW to compare the results of both approaches. used to test the implementation of the PEST algorithm
Finally, a generic floodwave problem has been set up to for the estimation of parameters in transient flow prob-
assess the results of a parameter estimation, which is lems. The simulation was done with automatically con-
influenced by the consequences of non-uniqueness. trolled time steps based on the first order (FE/BE)
predictor-corrector techniques.

NKO qÉëí=bñ~ãéäÉë
NKOKN qÜÉáëD=ïÉää=éêçÄäÉã
Theis' problem of the lowering of a water table by a
pumping well is a famous benchmark problem for
numerical algorithms since it possesses an analytical
solution couched in the well-known well function
(Theis14). The complete solution of the governing par-
tial differential equation of second order

2
h- --------h 1--- h
--S- ----- – – ------ = 0 (1-3) Figure 1.1 FEFLOW grid with PEST observation points
T t r 2 r r for Theis' well problem.

for a confined aquifer in a radial symmetric geometry The coordinates of the observation points on the
with the boundary conditions of a constant flux q into FEFLOW grid are given in Tab. 1.2. At these points the
a pumping well of infinitely small radius and a constant drawdown of the Theis problem has been recorded. For
head h o at the outer fringe, which is very far away PEST 15 time steps at each observation point have
from the well, has been derived by Theis by exploiting been selected with an approximately constant distance
the analogy to a heat conduction problem (Carslaw and on a logarithmic scale (Fig. 1.3). The PEST estimation
Jaeger2). The initial condition h o  0 r  is a constant problem has been set up in one zone which comprised
head of zero at t = 0 in the whole plane, except at the all nodes of the FEFLOW grid. The transmissivity T
wellbore where h  0 r b  = -0.01 m . All symbols are and the storage coefficient S have been estimated

«aef=ö=ïïïKãáâÉéçïÉêÉÇÄóÇÜáKÅçã cbcilt=ö=P
m~ê~ãÉíÉê=Éëíáã~íáçå=çÑ=íê~åëáÉåí=Ñäçï=éêçÄäÉãë=ïáíÜ=mbpq=áå=cbcilt

simultaneously using the initial values of 10-4 m2s-1 The course of the PEST objective function during
and 10-3, respectively. The optimization control param- the optimization is shown in Fig. 1.2. With the starting
eters of PEST have been left at their default values. values for T and S the objective function reaches some
104 for in total 60 time steps from 4 observation points.
–6
Table 1.1 Simulation parameters for Theis' well After 26 model runs it has been minimized to 2.5  10
problem by passing through 10 orders of magnitude. The model
runs no. 27-39 are used by PEST to verify that a mini-
Name Symbol Unit Value mum of the objective function has been actually
reached.
wellbore radius rb m 0.3048

flow initial h  0 r  m 0

flow initial at wellbore h  0 r b  m -0.01

Q
well pumping rate q = -----------
2r b
m2d-1 638.75425

fixed head at outer h  t 304.8 m  m 0

boundary
–4
transmissivity T m2s-1 9.2903  10

storage coefficient S - 0.001

storage compressibility So - 0

Table 1.2 Coordinates and distance from the

pumping well of the observation points for Theis'
Figure 1.2 Minimization of the PEST objective func-
well problem tion for Theis' well problem.

Obser- Coordinates Distance

vation –4
point no. x m y m rm The estimated values from PEST are 9.2902  10
2 -1 –4
m s and 1.0002  10 for T and S , respectively. For
1 9.64 0 9.64 both parameters PEST produced highly accurate point
estimates which deviate only in the fifth digit from the
2 76.98 12.16 77.93
true values of Tab. 1.1. The correlation coefficient for
3 150.28 23.75 152.15 T and S is -0.57 and indicates a relatively weak corre-
lation favouring an effective parameter estimation. Fig-
4 223.57 35.33 226.34
ure 1.3 exhibits a perfect agreement of 'measured' and
Q=ö=cbcilt
 NKO=qÉëí=bñ~ãéäÉë

predicted hydraulic heads at the observation points NKOKO _êÉóÉää=éìãé=íÉëí

which should have been expected with the highly pre-
The pump test of the Breyell waterworks (Langguth
cise point estimates from PEST. This figure also dem-
and Voigt12) has been evaluated on the basis of the ana-
onstrates that after some 0.1 d the hydraulic head
becomes constant at all observation points owing to a lytical results of Theis14. The test has been conducted
simulation system of finite extent. On the other hand, in a confined aquifer in the Lower Rhine region near
with growing distance from the pumping well, the the city of Krefeld. The sequence of geological layers
begin of the drawdown is delayed. Therefore, the infor- has been subdivided in an upper unconfined aquifer
mation from observation point 4 alone is not sufficient which is separated by a layer of very impermeable clay
to estimate the material parameters, because of the from the lower unconfined aquifer. During the test the
–2
short drawdown phase. However, with the drawdown pumping rate was held constant at Q = 2.67  10
3 -1
curve of observation point 1 alone a very precise m s . The drawdown has been observed at three
parameter estimation is always possible. gauges: gauge 11b at a distance of 7.4 m, gauge 3b at a
distance of 23 m and gauge 6b at a distance of 139.6 m
from the well. The recorded drawdown curves, copied
from Langguth and Voigt12, are given in the Tables 1.3,
1.4 and 1.5. For gauge 11b also the rerise has been
recorded after the pump was stopped at t = 0.1690 d.
The measured values are shown in Tab. 1.6. All mea-
sured curves for the Breyell pump test are depicted in
Fig. 1.5.
Table 1.3 Measured drawdown at gauge
11b with r = 7.40 m for the Breyell
pump test

Time Drawdown
[10-4 d] [m]

1.968 0.30
Figure 1.3 Comparison of measured (symbols) and pre- 2.431 0.40
dicted (full lines) hydraulic heads for the four observa-
tion points (OP) of Theis' well problem. 3.125 0.50

4.051 0.70

6.134 0.93

«aef=ö=ïïïKãáâÉéçïÉêÉÇÄóÇÜáKÅçã cbcilt=ö=R
m~ê~ãÉíÉê=Éëíáã~íáçå=çÑ=íê~åëáÉåí=Ñäçï=éêçÄäÉãë=ïáíÜ=mbpq=áå=cbcilt

Table 1.3 Measured drawdown at gauge Table 1.4 Measured drawdown at gauge
11b with r = 7.40 m for the Breyell 3b with r = 23 m for the Breyell pump
pump test (continued) test

Time Drawdown Time Drawdown

[10-4 d] [m] [10-4 d] [m]

9.838 1.03 1.389 0.05

19.44 1.20 2.083 0.10

25.46 1.28 2.546 0.15

32.52 1.35 4.167 0.25

38.19 1.40 6.250 0.35

52.08 1.47 9.375 0.45

62.50 1.50 14.12 0.55

83.33 1.56 26.16 0.70

125.0 1.64 40.04 0.80

166.7 1.71 50.23 0.85

243.1 1.79 63.54 0.90

347.2 1.86 79.75 0.95

451.4 1.90 99.65 1.00

625.0 1.99 125.0 1.05

960.6 2.07 164.4 1.11

1690 2.18 208.3 1.15

250.0 1.20

319.4 1.25

402.8 1.30

506.9 1.35

S=ö=cbcilt
 NKO=qÉëí=bñ~ãéäÉë

Table 1.4 Measured drawdown at gauge Table 1.5 Measured drawdown at gauge
3b with r = 23 m for the Breyell pump 6b with r = 139.60 m for the Breyell
test (continued) pump test (continued)

Time Drawdown Time Drawdown

[10-4 d] [m] [10-4 d] [m]

611.1 1.39 555.6 0.59

937.5 1.48 746.5 0.65

1342 1.55 1001 0.71

1638 1.58 1325 0.76

1609 0.80

Table 1.5 Measured drawdown at gauge

Table 1.6 Measured rerise at gauge
6b with r = 139.60 m for the Breyell
11b after the pump stop at
pump test
t = 0.169 d
Time Drawdown
Time Drawdown
[10-4 d] [m]
[10-4 d] [m]
27.78 0.14
1690 2.18
41.67 0.20
1709 1.03
69.44 0.25
1716 0.95
90.28 0.29
1721 0.90
131.9 0.35
1727 0.85
173.6 0.40
1763 0.69
236.1 0.45
1775 0.65
319.4 0.50
1806 0.59
451.4 0.55
1830 0.55

«aef=ö=ïïïKãáâÉéçïÉêÉÇÄóÇÜáKÅçã cbcilt=ö=T
m~ê~ãÉíÉê=Éëíáã~íáçå=çÑ=íê~åëáÉåí=Ñäçï=éêçÄäÉãë=ïáíÜ=mbpq=áå=cbcilt

Table 1.6 Measured rerise at gauge Langguth and Voigt12 have given estimates for the
11b after the pump stop at transmissivity T and the storage compressibility S
t = 0.169 d (continued) using graphical evaluation methods of Theis14, and
Cooper and Jacob3, respectively. Theis' method is
Time Drawdown based on the graphical matching of a modified well
[10-4 d] [m] function with scaled curves of the recorded drawdown.
Cooper and Jacob3 use the exponential behavior of the
1873 0.47
drawdown which dominates the process asymptotically
1898 0.44 for sufficiently large times. A detailed description of
both methods is not the subject of this paper, it is given
2002 0.35
in Langguth and Voigt12. All estimates for T and S are
2072 0.30 summarized in Tab. 1.7. The averages over all esti-
–4 –4
mates are 97  10 m2s-1 for T and 3.7  10 for S .
2148 0.26
The corresponding maximal relative deviations are
2225 0.23 24 % for T and -43 % for S . Hence, the storage com-
pressibility S shows a larger variability and can only
2396 0.17
be estimated with less precision.
2697 0.09

Table 1.7 Summary of estimates for the transmissivity T and the storage compressibility S of the
Breyell pump test from Langguth and Voigt12

Gauges used for Transmissivity T Storage compressibility S

estimation [10-4 m2s-1] [10-4]

Theis’ superposition method

11b & 3b & 6b together 110 2.4

Time-drawdown-method (Cooper and Jacob3)

6b alone 120 2.1

3b alone 94 4.6

11b alone 89 3.7

Distance-drawdown-method (Cooper and Jacob3)

11b & 3b & 6b together 90 4.8

U=ö=cbcilt
 NKO=qÉëí=bñ~ãéäÉë

Table 1.7 Summary of estimates for the transmissivity T and the storage compressibility S of the
Breyell pump test from Langguth and Voigt12 (continued)

Gauges used for Transmissivity T Storage compressibility S

estimation [10-4 m2s-1] [10-4]

Distance-time-drawdown-method (Cooper and Jacob3)

11b & 3b & 6b together 89 4.6

Rerise method (Jacob9)

11b alone 84 n.a.

In FEFLOW the pump test is modeled in a 2D prob-

lem for confined aquifers on a mesh with 2138 3-noded
triangles. The outline of the mesh consists of a circle
segment with an opening angle of 30° and a radius of
1500 m. The mesh is based on two superelements. The
first superelement covers the zone near the well up to
the radius of 150 m with a higher spatial resolution of
the finite elements. This zone is shown in Fig. 1.4
together with the location of the pumping well and the
three gauges 6b, 3b, and 11b, respectively. The pump-
ing well is modeled with a boundary condition of the
4th kind at the tip of the circle segment. On the outer
boundary a constant head of 0 m is prescribed with a
boundary condition of the 1st kind. The initial head on
all nodes was set to zero.

«aef=ö=ïïïKãáâÉéçïÉêÉÇÄóÇÜáKÅçã cbcilt=ö=V
m~ê~ãÉíÉê=Éëíáã~íáçå=çÑ=íê~åëáÉåí=Ñäçï=éêçÄäÉãë=ïáíÜ=mbpq=áå=cbcilt

Figure 1.4 Finite element grid near the pumping well at the tip with gauges 11b at r = 7.40 m, 3b at r = 23.00 m and
6b at r = 139.60 m.

With PEST the transmissivity T and the storage of estimation runs has been conducted. Firstly, the
compressibility S have been estimated simultaneously drawdown curves were used separately to produce the
in one zone which covered the whole mesh. The initial parameter estimates. Then the drawdown curves of the
–4 –4
values were chosen to be 10 m2s-1 and 4  10 , Tables 1.3 to 1.5 have been used together in one esti-
respectively. The control parameters for the optimiza- mation run. Also the rerise curve of Tab. 1.6 was used
tion process have been left at their default values. Also in a separate run. Finally, a PEST problem has been set
the correlation structure of the two estimated parame- up using all recorded 75 observation points from the
ters has been calculated. To allow for a direct compari- Tables 1.3 to 1.6.
son with the results from Langguth and Voigt12 a series

NM=ö=cbcilt
 NKO=qÉëí=bñ~ãéäÉë

Table 1.8 contains the point estimates of the six the pumping well. Langguth and Voigt12 observe the
PEST estimation runs for the transmissivity T and the same trend. The PEST estimates for S differ by some
storage compressibility S . Owing to the narrow width factor of four whereas in Langguth and Voigt12 the
of the objective function near the minimum, the 95 % spread is markedly lower. For both T and S the aver-
confidence intervals (CI) are mostly around one digit. ages over all estimates agree well with PEST point esti-
Only for the PEST run using gauge 6b alone, which has mates of the last line of Tab. 1.8 which were obtained
the largest distance from the pumping well, the CI were by using all information of the 75 measured time steps
notably larger. In this run, which needed 73 model calls in one estimation run.
to converge, the correlation coefficient between T and In Fig. 1.5 the curves of the measured and predicted
S is almost one. In all other PEST runs the correlation drawdown are compared. To calculate the predicted
coefficient is negative and around 30 model calls were drawdown the point estimates of the last line of Tab.
necessary. The absolute value is minimal if the time 1.8 were used. The agreement is very good except for
steps for the rerise in gauge 11b is included in the esti- the late phase of the rerise in gauge 11b. During the
mation. Small correlation coefficients favour the stabil- pump test the pore structure of the aquifer may change
ity of the optimization process. so that drawdown and rerise cannot be described with
one pair of constant parameter values for T and S .
The point estimates for T exhibit a similar variabil- However, in general, with PEST in FEFLOW the
ity as those from Langguth and Voigt12 in Tab. 1.7. The results for the Breyell pump test of Langguth and
estimates of T decrease with decreasing distance from Voigt12 have been fully confirmed.
Table 1.8 Parameter estimates and statistical information of the PEST estimation runs for the
Breyell pump test

Storage Number of
Gauges used for Transmissivity T Objective Correlation
compressibility S observation J  N obs
estimation [10-4 m2s-1] function J coefficient
[10-4] points N obs
–3 –3
6b alone 117 2.35 1.7  10 14 3.0  10 0.99
–3
3b alone 90.7 5.12 1.9  10
–2
24 5.7  10 -0.78
–2
11b alone 79.8 7.91 2.0  10
–1
21 2.1  10 -0.82
–2
6b & 3b & 11b 88.2 5.03 3.7  10
–1
59 1.0  10 -0.77
–3
11b rerise only 96.8 1.81 2.5  10
–2
17 9.3  10 -0.65
–3
all observation 92.0 4.31 4.7  10
–1
75 9.2  10 -0.55
points

«aef=ö=ïïïKãáâÉéçïÉêÉÇÄóÇÜáKÅçã cbcilt=ö=NN
m~ê~ãÉíÉê=Éëíáã~íáçå=çÑ=íê~åëáÉåí=Ñäçï=éêçÄäÉãë=ïáíÜ=mbpq=áå=cbcilt

Figure 1.5 Measured (symbols) and predicted (lines) drawdown for the gauges 6b
(diamonds, dotted line), 3b (squares, broken line) and 11b (circles, full line) of the
Breyell pump test.

NKOKP táÅÜáí~=éìãé=íÉëí with a water-filled height of 8.20 m and a maximal

drawdown of 1.80 m at a gauge with a distance of
The Wichita pump test (Langguth and Voigt12) has 15.00 m from the pumping well. This value was
been conducted in an unconfined aquifer. If the recorded after 18 d when the well was operated at an
assumption of Dupuit (1863) is valid, the well formulas –2
average pumping rate of Q = 6.31  10 m3s-1. Values
of Theis14 can also be applied to aquifers with free
for in total six gauges, which have been arranged in
water tables with sufficient accuracy. In practical appli- two rows in the north and the south of the pumping
cations the maximal drawdown must be small (i.e. < 15 well, are given in Tab. 1.9. Langguth and Voigt12 have
% of the original water-filled height of the aquifer).
evaluated the pump test with a graphical method of
However, this is not the case for the Wichita pump test
NO=ö=cbcilt
 NKO=qÉëí=bñ~ãéäÉë

Cooper and Jacob3. To account for the deviation from instead, where B denotes the original water-filled
Dupuit's assumption they corrected the recorded draw- height of the aquifer (Jacob10).
2
down s and used a lower drawdown s' = s – s  2B

Figure 1.6 Finite element grid for the Wichita pump test with the location of the northern gauges 1-3 and the southern
gauges 4-6.

For the simulation of the pump test with PEST in pumping well at the tip. The pumping well was mod-
FEFLOW the same circle segment as for the Breyell eled with a well boundary condition of the 4th kind. At
pump test was used with a radius of 1500 m and an the outer radius a constant head boundary condition of
opening angle of 30°. A cut of the grid near the tip with 0 m has been prescribed. The problem has been set up
the location of the six gauges of Tab. 1.9 is shown in in 2d with an unconfined aquifer of -8.20 m bottom ele-
Fig. 1.6. Owing to the radial symmetry the results vation. The initial head at all nodes has been set to
depend only on the distance between the gauge and the zero. The final simulation time was 20 d.

«aef=ö=ïïïKãáâÉéçïÉêÉÇÄóÇÜáKÅçã cbcilt=ö=NP
m~ê~ãÉíÉê=Éëíáã~íáçå=çÑ=íê~åëáÉåí=Ñäçï=éêçÄäÉãë=ïáíÜ=mbpq=áå=cbcilt

mated simultaneously with the start values of 10-2 ms-1

and 0.2, respectively.
Table 1.9 Distance to pumping well and
observed drawdown at t = 18 d for the 6
The minimum of the objective function at 0.016
gauges of the Wichita pump test
was reached in 7 optimization iterations after 35 model
runs. In Tab. 1.10 the PEST point estimates are com-
Distance to
Drawdown pared with the results of the graphical evaluation
Gauge no. pumping well
[m] method of Langguth and Voigt12. For the hydraulic con-
[m]
ductivity the point estimate of PEST lies close to the
northern row estimate of the graphical evaluation method using the
corrected drawdown. The PEST estimate of the storage
1 15.00 1.80 compressibility overestimates value from the method
2 30.70 1.40 with the corrected drawdown by some ten percent. The
agreement of the estimates for both material parame-
3 57.70 1.04 ters from PEST is better with the estimates from the
southern row method using the corrected drawdown. This agreement
justifies the application of this method from hindsight.
4 14.95 1.67

5 30.60 1.31 The very large 95 % CI of the PEST point estimates

for the storage compressibility indicate a broad mini-
6 57.90 0.97 mum of the objective function. The estimated material
parameters show a large negative correlation of almost
-99 %. Both observations indicate that the Wichita
For PEST the six gauges of Tab. 1.9 were used as pump test poses a more difficult estimation problem
observation points. The definition of the time-depen- than the Breyell pump test.
dent power functions in FEFLOW requires at least two
time steps, but at the gauges measurements had been In Fig. 1.7 the predicted drawdown curves for the
taken only at one time step of 18 d. Therefore, to obtain three northern gauges are shown. Those for the south-
six valid power functions to each time step the initial ern gauges have been omitted because they are almost
hydraulic head 0 m at t = 0 was added. Note, that with identical within the drawing accuracy. The predicted
this measure no additional information has been intro- curves pass between measured points of the northern
duced into the estimation problem, because the initial and southern gauges which is an indication of the plau-
condition is already known without a need to run the sibility of the estimation result.
model. It does not depend on changes of the model
parameter. The two material parameters hydraulic con-
ductivity and storage compressibility have been esti-

NQ=ö=cbcilt
 NKO=qÉëí=bñ~ãéäÉë

Table 1.10 Estimates for the flow parameters of the Wichita pump test

Graphical evaluation (Langguth and

Parameter PEST in FEFLOW
Voigt12)
(95 % confidence
intervals in
with measured with corrected
Name Unit brackets)
drawdown drawdown

conductivity [10-4 ms-1] 23 28 27 (15;47)

storage compressibility [-] 0.44 0.30 0.34 (0.05;24)

Figure 1.7 Measured observation points at t = 18 d for both the northern row (open sym-
bols) and the southern row (full symbols) of gauges, predicted gauge curves only for the
northern row.

«aef=ö=ïïïKãáâÉéçïÉêÉÇÄóÇÜáKÅçã cbcilt=ö=NR
m~ê~ãÉíÉê=Éëíáã~íáçå=çÑ=íê~åëáÉåí=Ñäçï=éêçÄäÉãë=ïáíÜ=mbpq=áå=cbcilt

NKOKQ cäççÇï~îÉ=éêçÄäÉã 1.8. At the right edge a time-dependent boundary con-

dition of the 3rd kind has been applied. The floodwave
In this example a passing floodwave in a river is is modeled by a linear rise of the water table in the river
used to estimate the hydraulic properties of the river of 3 m during one day, followed by a subsequent drop
bed and the connected aquifer from the response of the to the initial value on the next day. It is shown in Fig.
hydraulic head. In FEFLOW a (quasi-) 1D problem has 1.9. At the left edge a no-flow boundary condition is
been set up for an unconfined aquifer of 10 m thickness applied.
on a finite element grid of 100 m length, using 100
4-noded quadrilateral mesh elements as shown in Fig.

Figure 1.8 Finite element grid for the floodwave problem with the location of the observation points 1 (at 1 m), 2 (at 5
m) and 3 (at 10 m); both shape and position of the time-dependent floodwave boundary condition of the 3rd kind are
indicated.

The observed gauge curves at the observation points in various combinations as indicated in Tab. 1.11. Note,
1 at 1 m, 2 at 5 m and 3 at 10 m distance from the right that the stationary estimation problem (for a confined
edge were calculated with the values of 10-5 ms-1, 5 d-1 aquifer) is ill-posed since only a head boundary is pre-
and 0.2 for the three material parameters conductivity, scribed. Now the conductivity is only defined up to an
transfer coefficient and storativity. They are shown in integration constant and cannot be estimated (Sun13,
Fig. 1.9 for the first ten days of the simulation. The Kaiser11). However, in a transient problem the conduc-
actual simulation time was 50 d where for each obser- tivity alone can be estimated with a relative error of
vation point 59 time steps have been recorded. Here to one percent (Tab. 1.11). In the estimation run for the
the transfer coefficients for inflow and outflow identi- transfer coefficient alone the point estimate exceeds the
cal values were always assigned. The initial head was true value by more than ten percent. Moreover, the true
set to zero at all nodes. With growing distance from the value is not included in the 95 % CI of the point esti-
river boundary, the maximum of the hydraulic head mate. If the conductivity and the transfer coefficient are
appears with a certain delay and the response to the estimated together the accuracy of the point estimate
floodwave becomes broader and weaker. for the conductivity remains unchanged. But for the
transfer coefficient the relative error is reduced to three
With PEST the conductivity, the transfer coefficient percent. When the conductivity and the storativity are
and the storativity have been estimated separately and estimated together, both point estimates fall signifi-

NS=ö=cbcilt
 NKO=qÉëí=bñ~ãéäÉë

cantly below the true values. When all three material

parameters are estimated the results become even
worse. Obviously, for this combination of parameters
there exists an issue with uniqueness.

Figure 1.10 Objective functions for the PEST optimization

runs to estimate the transfer coefficient alone (crosses), the
conductivity alone (triangles), the transfer coefficient and
conductivity together (diamonds), the conductivity and stor-
ativity together (squares), and all three parameters conduc-
tivity, storativity and transfer coefficient together (circles).
Figure 1.9 Hydraulic head curves at the right floodwave
boundary and at the observation points 1, 2 and 3 with dis-
tances of 1 m, 5 m and 10 m.

«aef=ö=ïïïKãáâÉéçïÉêÉÇÄóÇÜáKÅçã cbcilt=ö=NT
m~ê~ãÉíÉê=Éëíáã~íáçå=çÑ=íê~åëáÉåí=Ñäçï=éêçÄäÉãë=ïáíÜ=mbpq=áå=cbcilt

Table 1.11 Input parameter and parameter estimates of PEST in FEFLOW for the floodwave
problem, 95 % CI in brackets

Conductivity Transfer coefficient Storativity Objective function

[10-5 m s-1] [d-1] [-] [10-3 m2]

input values 1.0 5.0 0.2 n.a.

parameter estimates

conductivity 1.01 (1.00;1.02) 5.0 fixed 0.2 fixed 3.37

transfer coefficient 1.0 fixed 5.61 (5.55;5.66) 0.2 fixed 3.49

conductivity & transfer 1.01 (1.00;1.02) 5.16 (4.89;5.44) 0.2 fixed 3.48
coefficient

conductivity & storativity 0.754 (0.737;0.771) 5.0 fixed 0.149 (0.147;0.151) 3.51

all 3 parameters 0.481 (0.467;0.496) 2.49 (2.25;2.75) 0.095 (0.092;0.098) 3.63

NKP `çåÅäìÇáåÖ=oÉã~êâë various graphical evaluation methods based on Theis'

analytical solution (Langguth and Voigt12). Owing to
The ability of the PEST software package to esti- the simplification of the real hydrogeological setting,
mate material parameters of FEFLOW flow problems which produces unknown systematic errors, the pre-
has been tested with four simple examples. dicted parameter values show a higher variability. This
variability is reflected in the point estimates from both
For Theis' analytical problem the transmissivity and the PEST optimization process and the graphical evalu-
the storage coefficient have been estimated with excel- ation methods in a similar way. The agreement of the
lent precision. estimated parameter values is better for the transmis-
sivity than for the storage compressibility. Both estima-
In the Breyell pump test for a confined aquifer tion methods produced parameter estimates with
recorded drawdown data has been used to determine sufficient accuracy.
the transmissivity and the storage compressibility with

NU=ö=cbcilt
 NKP=`çåÅäìÇáåÖ=oÉã~êâë

The Wichita pump test has been conducted in an inant mass and heat transport processes. Release 5.2, Reference
unconfined aquifer and has been evaluated by Lang- Manual, WASY Ltd., Berlin, 2005.
6. Doherty, J., Brebber, L. and Whyte, P. PEST - model independent
guth and Voigt12 using the same approach as for an con-
parameter estimation, User's Manual, Watermark Computing,
fined aquifer with and without corrected drawdown. Corinda, Australia, 1994.
With PEST it could be shown that the evaluation with 7. Dupuit J. Etudes thèoriques et pratiques sur le mouvement des
the corrected drawdown produced a more accurate eaux dans les canaux dècouvertes et à travers les terrains per-
result. In this example the use of the graphical method mèables. 2nd edition., Dunod, Paris, 1863.
has been pushed to the limit of applicability, whereas 8. Gründler R. Interface manager - extensions and programming
interface for FEFLOW, WASY Ltd, Berlin, 2005.
with PEST the Wichita pump test has been evaluated
9. Jacob, C.E. The recovery method for determining the coefficient
straight forwardly. of transmissibility. In: Bentall, R., Methods of determining per-
meability, transmissibility and drawdown, Geol. Survey Water-
The numerical floodwave problem for the estima- Supply Paper 1536-I, 283-292, 1963.
tion of three material parameters showed that there was 10. Jacob, C.E. Determining the permeability of water-table aquifers.
no unique set of parameters to reproduce the measured In: Bentall, R., Methods of determining permeability, transmissi-
bility and drawdown, Geol. Survey Water-Supply Paper 1536-I,
curves of the hydraulic head. Although the deviations
245-271, 1963.
from the true parameters are not excessive this example 11. Kaiser, J.C. Kopplung von FEFLOW mit dem Programm PEST
demonstrated that the estimation of material parame- zur Parameterkalibrierung (Coupling FEFLOW to the program
ters with PEST does not automatically lead to reliable PEST for parameter calibration). In: 3. Fachtagung ’Grafik-
results. gestützte Grundwassermodellierung’, Conference proceedings,
WASY Ltd, Berlin, Germany, 51-65, 1998.
12. Langguth, H.-R. and Voigt, R. Hydrogeologische Methoden. in
German, Springer, Berlin, 1980.
oÉÑÉêÉåÅÉë 13. Sun, N.-Z. Inverse problems in groundwater modeling. Kluwer
Academics Publishers, Dordrecht, 1994.
1. Bouttier, F. and Courtier P. Data assimilation concepts and meth- 14. Theis C.V. The relation between lowering of the piezometric sur-
ods, Meteorological Training Course Lecture Series, Lecture face and the rate and duration of discharge of a well using
Notes of the European Centre for Meteorological Weather Fore- groundwater storage. Trans. Amer. Geophys. Union 16 (1935),
cast (ECMWF), 2002. 519-524.
2. Carslaw, H.S. and Jaeger, J.C. Conduction of heat in solids. 2nd
ed. Oxford: University Press, 1959.
3. Cooper, H.H. and Jacob, C.E. A generalized graphical method for
evaluating formation constants and summarizing well-field his-
tory. Trans. Am. Geoph. Union 27 (1946), 526-534.
4. Diersch, H.-J.G. Interactive, graphics-based finite-element simu-
lation system FEFLOW for modeling groundwater flow, contam-
inant mass and heat transport processes. Release 5.2, User’s
Manual, WASY Ltd., Berlin, 2005.
5. Diersch, H.-J.G. Interactive, graphics-based finite-element simu-
lation system FEFLOW for modeling groundwater flow, contam-

«aef=ö=ïïïKãáâÉéçïÉêÉÇÄóÇÜáKÅçã cbcilt=ö=NV