bêêçê= éêçé~Ö~íáçå= áå= íÜÉ= kÉïíçåJÄ~ëÉÇ= ëçäìíáçå

Åçåíêçä=çÑ=ìåë~íìê~íÉÇ=Ñäçï

H.-J. G. Dierscha & P. Perrochetb

N
a
WASY Institute for Water Resources Planning and Systems Research, Berlin, Germany
b
Centre d’Hydrogéologie, Université de Neuchâtel, Switzerland
simulations. As the major advantages the PVST is (1)
ABSTRACT unconditionally mass-conservative with respect to the
The Newton method represents the numerical core of the
primary variable switching technique (PVST) which has
time step size, (2) very effective and robust for dry ini-
shown superior to conventional approaches in both unsatu- tial conditions, (3) a Newton-based iteration method
rated flow and multi-phase flow modelling. In the context of with quadratic convergence, and (4) a general analysis
PVST empirically controlled strategies in time are rather method suitable for single- and multi-phase flow prob-
common, where the Newton convergence is attempted for a lems.
possibly large step size. This technique is known as the tar-
get-based full Newton (TBFN) time stepping strategy. In
comparison to adaptive techniques satisfying a predefined To control the overall iteration process Forsyth et al.
discretization error the TBFN results can be inaccurate in (1995) preferred an empirical target-based full Newton
spite of the achieved convergence in the Newton method. (TBFN) time stepping strategy. Recently, Diersch &
The present paper aims at analysing the cause of discrepan- Perrochet (1999) compared the TBFN with an adaptive
cies in simulating unsaturated flows. This is done in com-
parison of analytical solutions which are based on
temporally error-controlled predictor-corrector tech-
exponential constitutive laws. nique one-step Newton scheme (PCOSN). In their
N bêêçê=éêçé~Ö~íáçå=áå=íÜÉ=kÉïíçåJÄ~ëÉÇ=ëçäìíáçå=Åçåíêçä=çÑ=ìåë~íìê~íÉÇ=Ñäçï
extensive numerical benchmark analysis Diersch &
Perrochet (1999) found that, in spite of the achieved
NKN fåíêçÇìÅíáçå iteration convergence, TBFN results can be rather
depart from PCOSN findings, unless the target change
In contrast to Picard iteration schemes common in
parameters, and accordingly the step sizes, are kept
solving the Richards equation for unsaturated flow in
sufficiently small. In continuing the analysis the pres-
porous media, the Newton method in combination with
ent paper aims at a quantification of the resulting errors
appropriate strategies can reduce the solution effort by
along analytical solutions for the Richards equation
orders. This has been shown by Forsyth et al. (1995)
based on exponential saturation-pressure and conduc-
who introduced the idea of the primary variable switch-
tivity-pressure relationships.
ing technique (PVST) to saturated-unsaturated flow

«aef=ö=ïïïKãáâÉéçïÉêÉÇÄóÇÜáKÅçã cbcilt=ö=N
bêêçê=éêçé~Ö~íáçå=áå=íÜÉ=kÉïíçåJÄ~ëÉÇ=ëçäìíáçå=Åçåíêçä=çÑ=ìåë~íìê~íÉÇ=Ñäçï

NKO jçÇÉä=bèì~íáçåë s = f   = f s 

–1
 (1-2)
K r = g    = g  s  
The present finite-element model is based on the
Richards equation written in the following form
Here, van Genuchten or Brooks-Corey parametric
 s    models are common (cf. Diersch & Perrochet, 1999).
R  s   = S o  s    ------- +  -------------- (1-1)
t t Instead, if exponential constitutive laws are preferred
–    K r   K   +  1 +  e   – Q = 0 in the form

s–s  exp     +  a   for   a 

which has to be solved either for  or s . In (1-1), ------------r- = s e =  
1 – sr 1 for    a  (1-3)

 pressure head, (   0 saturated medium,   0 Kr = se 
unsaturated medium);
s    saturation, ( 0  s  1 , s = 1 if medium is satu-
analytical solutions of the nonlinear Richards equation
rated);
can be derived. In (1-3),  a is the air entry pressure
t time;
head, s r is the residual saturation and   0 is a con-
S o specific storage due to fluid and medium com-
stant.
pressibility;
 porosity;
K r    relative hydraulic conductivity ( 0  K r  1 ,
K r = 1 if saturated at s = 1 ); NKP kÉïíçå=jÉíÜçÇ=~åÇ=mspq
K tensor of hydraulic conductivity for the saturated
medium (anisotropy); The discretized form of the basic Richards equa-
 buoyancy coefficient including fluid density tions (1-1) yields
effects;
n+1
e gravitational unit vector; R X = 0 (1-4)
Q specific mass supply; to be solved for a primary variable
R residual;
X    s  (1-5)
Constitutive relationships are additionally required
(1) for the saturation s as a function of the pressure which can be either c or s at the new time level n+1.
(capillary) head  , as well as its inverse, the pressure Applying the Newton method to (1-4) we solve
head  as a function of the saturation s , and (2) for the
relative hydraulic conductivity Kr as a function of X
J  
n+1 n+1
 s X 
n+1
= –R
n+1
  s  (1-6)
either the pressure head  or the saturation s :

O=ö=cbcilt
 NKQ=qÜÉ=káííóJdêáííó

with the increment possibly large time step size. The step size is deter-
mined from a desired change in the variable per time
n+1 n+1 n+1 step given by user-specified targets.
X  = X + 1 – X (1-7)

X An important aspect of the iterative solution via the

and the Jacobian J expressed in indicial notation as PCOSN and TBFN schemes is the choice of an appro-
n+1 n+1 n+1
priate convergence criterion. Limiting the temporal dis-
X n+1 n+1 R I     s   cretization errors deviatory (change) error measures
J IJ     s  = ---------------------------------------------------
n+1
(1-8) n+1
X J d L p are the controlling criteria, which are func-
n+1 n+1
tions of the solution differences d   X  :
where  denotes the iteration number. The PVST n+1
d Lp  (1-9)
selects the primary variable in a dynamic manner
depending on inner nodal criteria of the solutions, c or
s . The derivatives of the Jacobian can be easily where  is a user-specified deviatory error tolerance.
switched between c and s in accordance with the Here, weighted RMS L 2 and maximum L  error
computational requirements. Their computations can norms can be chosen. Commonly, in the Newton
be done either analytically or numerically. method the deviatory error criterion (1-9) represents a
standard test to terminate the iteration within the time
step. In the PCOSN the temporal truncation and the
NKQ qÜÉ=káííóJdêáííó Newton termination error measures are equivalently
used. As a result, only one error criterion and one New-
Generally, the control of the solution of the resulting ton step per time step become necessary (  = 1 ).
n+1
highly nonlinear matrix system (1-6) is a tricky matter. Alternatively to the deviatory error estimate d  Lp ,
n+1
Both the choice of the time step size t n and the itera- the residual R  L p may be directly controlled, such
tion control of the Newton scheme significantly influ- as
ence the success and the efficiency of the simulation. In
n+1 n+1
the PCOSN scheme (Diersch & Perrochet, 1999) the R Lp  2 F Lp (1-10)
nonlinear matrix system is linearized by the predictor
solutions. Temporal truncation errors can be easily esti-
where an additional error tolerance  2 appears and an
mated by evaluating predictor and corrector solutions
appropriate normalisation of the residual (here with
which are the basis of an adaptive, error-controlled n+1
respect to the external supply F ) is required. In the
time stepping and iteration strategy. In contrast, the n+1
TBFN deviatory errors d  Lp and residual errors
TBFN (Forsyth et al., 1995) does not consider tempo- n+1
R Lp can be alternatively employed. Instead of a
ral truncation errors in the time and iteration control.
one-step Newton control so as done in the PCOSN the
The only criterion is the Newton convergence for a
predictor-corrector technique can also be extended to a

«aef=ö=ïïïKãáâÉéçïÉêÉÇÄóÇÜáKÅçã cbcilt=ö=P
bêêçê=éêçé~Ö~íáçå=áå=íÜÉ=kÉïíçåJÄ~ëÉÇ=ëçäìíáçå=Åçåíêçä=çÑ=ìåë~íìê~íÉÇ=Ñäçï

multiple step Newton (PCMSN) strategy satisfying   Kt - 

both criteria (1-9) and (1-10). To measure the global   z – --------------------
 so – si    1 – sr  
balance error we introduce s  z t  = s i + -------------------  erfc  --------------------------------- (1-14)
2   Kt 
 2 ------------------------ -
T     1 – s r 

 R  t  L2
dt
 Kt 
t------------------------------------
=0  z + -------------------- -
T = - (1-11)  1 – sr  
+ e erfc  --------------------------------- 
T z

 F t L 2 dt
Kt - 
 2 ------------------------
   1 – s r  
t=0 

for the ’accumulated loss’ of mass with respect to the It can be easily seen from (1-13) that with large  the
total external supply over the entire simulation period problem is dominated by advection. Otherwise, consid-
(0, T ). ering a fully implicit time discretization the temporal
numerical dispersion can be estimated to

NKR ^å~äóíáÅ~ä=pçäìíáçå n+1 t n K

D numdisp  -------- ---------------------
2
(1-15)
2   1 – sr 
In one dimension the Richards equation (1-1)

s     
 -------------- – ----- K r   K  ------- – 1 = 0 (1-12) NKS qÉëí=`~ëÉ
t z  z 

The problem is described in Fig. 1.1. For the lower

can be transformed into the linear advective-dispersive boundary a free drainage-type boundary condition is
equation of the form applied (Diersch, 1998). The 6 m column is discretized
2
by 120 linear elements, so the nodal spacing becomes
K - s
s- -------------------- K - s z = 5 cm.
---- + ----- – ------------------------ -------- = 0 (1-13)
t   1 – s r  z   1 – s r  z 2

for the exponential constitutive law (1-3) by using the NKSKN kÉïíçå= Åçåíêçä= Äó= íÜÉ= ÇÉîá~J
following assumptions: S o =  a = Q =  = 0 and z íçêó=Éêêçê=ÅêáíÉêáçå=ENJVF
is oriented downward in the direction of gravity.
The computed saturation profiles for two -parame-
With s  z 0  = s i and s  0 t  = s o the solution is ters in comparison with the analytical solution are

Q=ö=cbcilt
 NKS=qÉëí=`~ëÉ

shown in Fig. 1.2. Large conservation errors are NKSKO kÉïíçå=Åçåíêçä=Äó=íÜÉ=êÉëáÇì~ä

observed for the TBFN if the number of time and
accordingly Newton steps become small. For
–1
 = 5 m 20 and 61 time steps (49 and 116 Newton
steps) are needed for the constraints t max = 0.2 d and
t max = 0.05 d , respectively. It leads to the total inte-
gral balance errors (1-11) of   3 d   160%
and  40% , respectively. In contrast, the PCOSN took
240 variable time and Newton steps resulting an
acceptable balance error of only   3 d   0.06% with a
very good agreement with the analytical solution. Sim-
–1
ilar results appear for  = 200 m , where the TBF
gives   3 d   7% (74 time and 344 Newton steps) and
  3 d   2% (107 time and 301 Newton steps), respec-
tively, while the PCOSN obtains   3 d   0.09% after
360 time and Newton steps.
s o = 0.8
–1
K = 0.25 m d
 = 0.3
sr = 0
6m
Éêêçê=ÅêáíÉêáçå=ENJNMF
As outlined in Fig. 1.3 the conservative problems
disappears for the TBFN if the residual error criterion
–4
(1-10) is used with  2 = 10 . The adaptive PCMSN
and the TBFN gives comparable results which agree
quite well with the analytical solution. Here, the
PCMSN is now controlled by two the criteria (1-9) and
(1-10):  = 10
–4
for the time adaptation and
 2 = 10
–4
for the Newton termination, where more
than one Newton step per time increment can occur.
The TBFN needed 49 time (279 Newton) and 133 time
–1
(502 Newton) steps for  = 5 m and  = 200 m ,
–1
respectively, achieving total balance errors of
  3 d   0.0006% and  0.001% , respectively. Expect-
edly, the PCMSN took more time steps. We found 164
time (268 Newton) steps with   3 d   0.01% and 165
time (362 Newton) with   3 d   0.001% for
–1 –1
 = 5 m and  = 200 m , respectively.

–1
 = 5200 m
– 20
s i = 10


------- = 0
z
z
Figure 1.1 Sketch of the test problem.

«aef=ö=ïïïKãáâÉéçïÉêÉÇÄóÇÜáKÅçã cbcilt=ö=R
bêêçê=éêçé~Ö~íáçå=áå=íÜÉ=kÉïíçåJÄ~ëÉÇ=ëçäìíáçå=Åçåíêçä=çÑ=ìåë~íìê~íÉÇ=Ñäçï

a) 0.8 b)
0.8
0.7
TBFN, tmax = 0.2 d 0.7 TBFN, tmax = 0.2 d
TBFN, tmax = 0.05 d TBFN, tmax = 0.05 d
0.6 PCOSN PCOSN
0.6
analytical analytical
0.5

Saturation [1]
Saturation [1]
0.5

0.4
0.4

0.3
0.3

0.2 0.2

0.1 0.1

0.0 0.0
0 1 2 3 4 5 6 0 1 2 3 4 5 6
Depth [m] Depth [m]
–1 –1 –4
Figure 1.2 Saturation profiles at t = 3 d for a)  = 5 m and b)  = 200 m ,  = 10 with the deviatory
error criterion (1-9) for the Newton control, aggressive target change parameters are used for the TBFN with a
maximum time step constraint t max .

NKT `çåÅäìëáçåë
For the TBFN the residual error criterion should be
preferred rather than standard deviatory tests to avoid
conservation errors as long as the target change param-
eters allow large steps. On the other hand, the adaptive
PCOSN scheme sufficiently controls the solution pro-
cess by limiting time truncation errors and an addi-
tional residual test, so as done in the PCMSN scheme,
is not necessary in the most cases.
oÉÑÉêÉåÅÉë
1. Diersch, H.-J. G. (1998) Treatment of free surfaces in 2D and 3D
groundwater modeling. Math. Geologie 2, 17-43.
2. Diersch, H.-J. G. & Perrochet, P. (1999) On the primary variable
switching technique for simulating unsaturated-saturated flows.
Adv. Water Resour. 23 (1), 25-55.
3. Forsyth, P. A., Wu, Y. S. & Pruess, K. (1995) Robust numerical
methods for saturated-unsaturated flow with dry initial conditions
in heterogeneous media. Adv. Water Resour. 18(1), 25-38.

S=ö=cbcilt
 NKT=`çåÅäìëáçåë

a) 0.8 b)
0.8
0.7
TBFN TBFN
0.7
PCMSN PCMSN
0.6 analytical analytical
0.6

0.5

Saturation [1]
Saturation [1]

0.5

0.4
0.4

0.3
0.3

0.2 0.2

0.1 0.1

0.0 0.0
0 1 2 3 4 5 6 0 1 2 3 4 5 6
Depth [m] Depth [m]
–1 –1 –4
Figure 1.3 Saturation profiles at t = 3 d for a)  = 5 m and b)  = 200 m ,  =  2 = 10 with the
residual error criterion (1-10) for the Newton control.

«aef=ö=ïïïKãáâÉéçïÉêÉÇÄóÇÜáKÅçã cbcilt=ö=T