Download as pdf or txt
Download as pdf or txt
You are on page 1of 68

GEOS 5311 Lecture Notes: Introduction to

Modflow
Dr. T. Brikowski

Spring 2013

Version 1.22, April 3, 2013

About Modflow
I

Modflow was written to be the principal USGS program for


solving groundwater flow problems (Harbaugh, 2005;
Harbaugh et al., 2000; Harbaugh and McDonald, 1996;
McDonald and Harbaugh, 1988)
modularity is its main design feature
I
I

code can be easily modified and extended (e.g. Hill, 1990a)


tremendous improvement over earlier USGS codes (e.g.
Prickett and Lonnquist, 1971)
modularization is a good approach to take with any computer
program

Modflow is the de facto standard for modeling of hydrologic


flow problems. For example, a GeoRef search for Modflow
returns 2332 different articles.

Discretization in Modflow
I

Modflow uses a block-centered grid system (Fig. 1)

the modeled system is essentially discretized as a regular series


of cubic buckets (i.e. a block-centered grid, Fig. 2)

water mass balance is carried out by summing the water fluxes


across each side of the bucket (e.g. Fig. 3), plus internal
source/sinks (wells)

governing equation for a cell (p. 2-3, Harbaugh, 2005):


X
i

Qi = Ss

h
V
t
3

where Qi is a discharge into the cell ( LT ), Ss is the specific


storage of the cell, h is the head change over the time
interval t, V is the volume of the cell

Discretization in Modflow (cont.)

ultimately a matrix equation is derived in which the unknown


vector of cell heads is multiplied by a matrix of geometry and
rock-property dependent coefficients

the following notes detail the derivation of that matrix


equation, in order to clarify the significance of various
Modflow input parameters, matrix solver choices, etc.

Example Modflow Grid

Figure 1: Generic modflow grid. Arrays table lists typical variables


specified for each cell. After Anderson and Woessner (Fig. 3.9, 1992),
originally from McDonald and Harbaugh (1988)

Grid Types

Figure 2: Finite difference grid types. Modflow adopts the


block-centered approach for ease in describing cell-boundary fluxes. After
McDonald and Harbaugh (Fig. 2, 1988).

Cell Numbering

Figure 3: Cell numbering scheme in Modflow, where i is row number, j


is column, and k is layer. After McDonald and Harbaugh (Fig. 3, 1988).

Special Cells

not all cells in a finite-difference grid will be normal (i.e.


allow internal changes or flow across all sides)

in Modflow there are three basic cell types: variable-head


(active), constant-head (boundary condition) and inactive
(outside the problem domain). See Fig. 4

the status of all cells is stored in array IBOUND in Modflow

in GMS this array can be viewed from the 3D


Grid/Modflow/Global Options dialog

Cell Types

Figure 4: Basic cell types in Modflow. Variable-head cells are normally


called active, constant head cells can also be set using the General
Head Boundary (GHB) package. The grid boundary defaults to no-flow
(e.g. white cells along top row). After McDonald and Harbaugh (Fig. 8,
1988).

Layering and Grid Design

Figure 5: Two schemes for vertical discretization. (B) implemented by


computing material-averaged K for cubic cell, (C) implemented by
adjusting conductivity on a cell face to reflect true area (i.e. cubic grid,
with K on a face a function of its area). After Harbaugh (Fig. 2-8, 2005).

Modflow Packages in GMS


The modular components of Modflow are called packages, and
each Modflow run involves applying the appropriate package set.
Table 1: GMS-supported input/output control and basic problem
parameter packages for Modflow. BAS6 package is required.

Package Name
Basic Package

Abbrev
BAS6

Output
Control
Gage Package

OUT1
GAGE

Description
Used to specify the grid dimensions, the
computational time steps, and an array
identifying which packages are to be used.
Controls what and when information is to
be output from MODFLOW.

generates time-series output for selected


cell(s), as if a piezometer (gage) was installed there

Matrix Construction
GMS-supported matrix construction packages for Modflow. One
and only one of these must be selected.
Package Name
Block Centered Flow
Package

Abbrev
BCF6

Layer Propery
Flow
Package

LPF

Description
Performs the cell by cell flow calculations. The input to this package includes layer types and cell attributes
such as storage coefficients and transmissivity.
Parameters for sensitivity
analysis or parameter estimation are
NOT supported.
Performs the cell by cell flow calculations. The input to this package includes layer types and cell attributes
such as storage coefficients and transmissivity

Matrix Construction (cont.)

Hydrogeologic
Unit Flow
Package
Upstream
Weighting

HUF

UPW

Defines the model stratigraphy in a grid


independent fashion (in the vertical direction)
computes boundary values with emphasis on upstream value

Source/Sinks
GMS-supported source-sink packages for Modflow. Any number
of these can be selected, none are required.
Package Name
River Package

Abbrev
RIV1

Recharge
Package
Well Package
Multi-Node
Well

RCH

Description
Simulates river type boundary conditions (water source or sink, depending
on head difference between river and
aquifer)
Simulates areal recharge

WEL

Simulates injection/extraction wells

MNW

Well that spans multiple layers

Source/Sinks (cont.)

Drain Package

DRN

Evapotranspiration EVT
package
General
GHB
Head
Boundary
Package

Simulates drain-type conditions (water


extraction when head rises above specified elevation)
Simulates the effect of evapotranspiration from the vadose zone
Simulates a general purpose headdependent source/sink. Commonly used
to simulate lakes

Source/Sinks (cont.)
Stream/Aquifer
Interaction
Package

STR

Time Variant Specified Head


Package

CHD

Simulates the exchange of water between the aquifer and surficial streams.
Includes routing and automatic computation of stage (i.e. like River package but accounting for effect of aquifer
on river, and along-stream surface water
flow). Parameters for sensitivity analysis
or parameter estimation are NOT supported.
Simulates specified head boundary conditions where the head is allowed to vary
with time (e.g. a fluctuating lake level)

Source/Sinks (cont.)

Horizontal
Flow
Barrier
Package

HFB

Simulates the effect of horizontal flow


barriers such as faults/fractures (or engineered structures like sheet piles and
slurry trenches)

Matrix Solution
GMS-9.0-supported matrix solution packages for Modflow. One
and only one of these must be selected. Changing solvers may
allow a failed run to converge.
Package Name
Direct

Abbrev
DE4

Geometric Multigrid

GMG

Description
Modified Gaussian elimination
(Harbaugh, 1995). Quick, can
stall, but memory-intensive for
linear problems
Good for regular grids (generates
sub-grids based on input geometry) tricky with anisotropy

Matrix Solution (cont.)


Link
Algebraic
Multigrid Package

LMG

Strongly
Implicit
Package
Preconditioned
Conjugate Gradient
Package
Slice
Successive
Overrelaxation
Package

SIP1
PCG2

SSOR

Multi-grid independent of input


geometry (handles irregular grids
or anisotropy well), requires more
memory but converges 2-25 times
faster than previous MODFLOW
solvers (Detwiler et al., 2002)
Iterative solver, can develop oscillation or overshoot
Iterative solver based on the
preconditioned conjugate gradient
technique
Not available in Modflow-2005+,
iterative solver that solves sliceat-a-time, limiting communication between slices

Multi-Grid Approaches

initial head in a model can be approximated as a series of


periodic functions, e.g. a step function in head

matrix solution often fails because a given frequency of error


(high frequency for sound, low frequency for matrices) is
amplified, much like sound feedback (e.g. listen to start of
Beatles I Feel Fine)

use internally-computed coarser grids to dampen effect of


low-frequency errors (see slide 51)

see CFD-online summary

General Matrix Solution Guidance


In general try these methods in the following order
DE4 or SIP Direct, work well for simple problems
PCG2 More parameters to set, but more robust than DE4
LMG Handles most feedback (dry/rewet) errors well.
Best for large problems (e.g. Yucca Mountain)
NWT Newton-Raphson solution, available in
MODFLOW-NWT only (with UPW package)
I best for unsat. flow-dominated settings where
dry/rewet is frequent
I uses magnitude of change from last-computed
value at cell to accelerate convergence

Modflow Processes
Modflow2K has been structured in terms of four basic processes,
representing the general tasks involved in modeling (Fig. 6):
I

Global Process (GLO): general controls on the model (e.g.


grid definitions, etc.)

Ground-Water Flow (GWF): solution of the flow equation

Observation Process (OBS): calculates simulated values at


observation points (known head), and error statistics
(residuals, or difference between observed and simulated
value)

Sensitivity Process (SEN): uses OBS to determine which


parameters have the greatest impact on error

Parameter Estimation (PES): optimizes parameter values to


minimize error determined by OBS

Process Flowchart

Figure 6: Flowchart of Modflow process procedure. After Harbaugh et al.


(Fig. 1, 2000).

File Structure

Modflow utilizes a control file (name file), written by GMS


with the extention mfn (Fig. 8)

name file lists separate input or output files used by each


package

GMS, VisualModflow, GroundwaterVistas, etc. are all


packages to simplify creation of these input files, and reading
and interpretation of output files.

As such, these programs are just elaborate input/output


wrappers for the free (public-domain) Modflow software.

Modflow Example Grid

Figure 7: Example Modflow grid. After Harbaugh et al. (Fig. 10, 2000).

Modflow Namefile

LIST 6 twri.lst
BAS6 5 TWRI.ba6
BCF6 11 TWRI.bc6
WEL 12 twri.wel
DRN 13 twri.drn
RCH 18 twri.rch
SIP 19 twri.sip
OC 22 twri.oc
DIS 10 TWRI.dis
Figure 8: Example Modflow namefile. After Harbaugh et al. (pg. 90,
2000).

Modflow Discretization File

3 15 15 1 1 0 NLAY,NROW,NCOL,NPER,ITMUNI,LENUNI
1 1 0 CONSTANT 5.000000E+03 DELR
CONSTANT 5.000000E+03 DELC
CONSTANT 2.000000E+02 TOP of system
CONSTANT -1.500000E+02 Layer BOTM layer 1
CONSTANT -2.000000E+02 Confining bed BOTM layer 1
CONSTANT -3.000000E+02 Layer BOTM layer 2
CONSTANT -3.500000E+02 Confining bed BOTM layer 2
CONSTANT -4.500000E+02 Layer BOTM layer 3
8.640E+04 1 1.000E+00 SS PERLEN,NSTP,TSMULT,Ss/tr
Figure 9: Example Modflow discretization (grid-definition) file. After
Harbaugh et al. (pg. 90, 2000).

Modflow Input Instructions

Figure 10: Modflow discretization file input instructions. Each package file has such
instructions, making hand-editing possible. Note Modflow2K input is free-format
(spacing and length of entries is unimportant). After Harbaugh et al. (2000, pg. 45).

Governing Equation
The general governing equation solved by Modflow is:







h
h
Kxx
+
Kyy
+
Kzz
W = Ss
x
x
y
y
z
z
t
(1)
where
I

Kxx , Kyy , Kzz are hydraulic conductivity, with principal


direction oriented along a coordinate axis

W is volumetric source/sink, units of 1t (i.e. volumetric flux


per unit volume of aquifer). Areal source/sinks are converted
to these units

Ss is specific storage

1
L

Numerical Approximation
I

Recall that Modflow grids are block centered, and numbered


by the ith row, jth column and kth layer (Fig. 1).

Each block is ci wide, rj deep and Vk high

Flux across a cell face can be determined using Darcys Law


(i.e. flux into cell i, j, k across face at i, j 21 , k):
Qi,j 1 ,k
2

= Kri,j 1 ,k (ci vk )
2

hi,j1,k hi,j,k
rj 1
2

Cri,j 1 ,k (hi,j1,k hi,j,k )


2
| {z }

Conductance

where Kri,j 1 ,k is the hydraulic conductivity in the along-row


2

direction at the cell wall j 12 , over a cross-sectional area


ci vk

Numerical Approximation (cont.)

geometric and rock properties are combined into a


Conductance term that remains constant throughout the
solution process

Horizontal Discharge

Figure 11: Horizontal discharge calculation in Modflow. After (Fig. 4,


McDonald and Harbaugh, 1988).

Water Mass Balance


I

the total flux (continuity eqn) into cell i, j, k can then be


written using the equation on page 30:
Qtotal = Qi,j 1 ,k + Qi,j+ 1 ,k + Qi 1 ,j,k +
2

Qi+ 1 ,j,k + Qi,j,k 1 + Qi,j,k+ 1


2

Cri,j 1 ,k (hi,j1,k hi,j,k ) + Cri,j+ 1 ,k (hi,j+1,k hi,j,k )


2

Cri 1 ,j,k (hi1,j,k hi,j,k ) + Cri+ 1 ,j,k (hi+1,j,k hi,j,k )


2

Cri,j,k 1 (hi,j,k1 hi,j,k ) + Cri,j,k+ 1 (hi,j,k+1 hi,j,k )


2

adding storage and source/sinks to this equation, the final


mass balance is:
Qtotal + Qsource/sink = Ssi,j,k

hi,j,k
(rj ci vk )
t

where the fraction represents the time change of head, and


the quantity in the parentheses is the cell volume

(2)

Source-Sink Representation

source sinks are grouped into head-dependent and known-flux


plus known-head (where n is the source/sink number):
Qsource/sink =

Ns
X

(Pi,j,k,n hi,j,k ) +

qi,j,k,n

n=1

n=1

Ns
X

{z

HeadDependent

{z

KnownFlux/Head

head-dependent fluxes represent leakage, stream/river


interactions, etc.

known-flux source/sinks are wells, recharge, E.T., etc.

(3)

Finite Difference Equation


I

Modflow utilizes a backward-difference formulation, i.e. a


molecule in which all terms but one are evaluated at future
time m + 1

with this formulation, the final Modflow finite difference


equation (based on equation 2) is:




m+1
m+1
m+1
m+1
hi,j,k
+ Cri,j+ 1 ,k hi,j+1,k
hi,j,k
+
Cri,j 1 ,k hi,j1,k
2
2




m+1
m+1
m+1
m+1
+ Cri+ 1 ,j,k hi+1,j,k
+
Cri 1 ,j,k hi1,j,k
hi,j,k
hi,j,k
2
2




m+1
m+1
m+1
m+1
Cri,j,k 1 hi,j,k1
hi,j,k
+ Cri,j,k+ 1 hi,j,k+1
hi,j,k
+
2

Ns
X
n=1

m+1
Pi,j,k,n hi,j,k
+ Qi,j,k = Ssi,j,k (rj ci vk )

m+1
m
hi,j,k
hi,j,k

Finite Difference Equation (cont.)

the final step is to simplify by combining all the multipliers of


m+1
each hi,j,k
into a single coefficient HCOFi,j,k (which can be
adjusted for partially-saturated flow)
~,
these can be combined to make a matrix equation (A ~h = q
Fig. 12), which is solved repetitively in Modflow

Matrix Equation

Figure 12: Example matrix equation for 4x3x2 Modflow grid (see Fig.
13). After (Fig. 46, McDonald and Harbaugh, 1988).

Grid For Matrix Equation


9

13

10

14

11

15

12

16

Figure 13: 4x3x2 grid for matrix equation (Fig. 12)

Matrix Structure

Figure 14: Banded matrix structure of 4x3x2 Modflow grid example (see
Fig. 12), only non-zero diagonals are shown, D-E-F are adjacent. After
(Fig. 47, McDonald and Harbaugh, 1988).

Full vs. Quasi-3D


I

several choices are available for treatment of heterogeneity in


the vertical direction:
I

fully 3-D flow: when 3-D flow must be considered in the


lower-permeability units. Most useful when differences in
vertical conductivity between units is less than factor of 100
(see Fig. 15 and page 43)
quasi-3D: aquitards are treated as reduced conductances
between aquifers. Only vertical flow is considered in the
aquitard (see Fig. 16 and 43)

in Modflow, aquifer thickness is treated only indirectly as


part of the conductance (VCONT & HCOF) terms

it is important to understand the Modflow formulation of


VCONT (also known as leakance) in order to produce
successful models. This formulation is outlined below.

VCONT for Fully-3D

Figure 15: Geometry for computation of VCONT for fully-3D case. After
(Fig. 27, McDonald and Harbaugh, 1988).

VCONT for Quasi-3D

Figure 16: Geometry for computation of VCONT for quasi-3D case.


After (Fig. 28, McDonald and Harbaugh, 1988).

VCONT Formulas

in Modflow, conductance is calculated as


Cv =

Kz
(c r )
v

for convenience, let


VCONT =

Cv
Kz
=
c r
z

(4)

i.e. VCONT contains all the time-invariant parameters


affecting vertical flow in the cell (geometric and rock
properties, assuming unconfined saturated thickness constant
...)

VCONT Formulas (cont.)

For the fully 3-D case, vertical hydrauilc conductivity is given


by the harmonic mean of the two layers, and VCONT is:
Kz

VCONT

zk+ 1
2

vk+1 /2
Kk+1

vk /2
Kk

2
vk+1
Kk+1

vk
Kk

Note: VCONT incorporates variations in Kz and layer


thickness

VCONT Formulas (cont.)


I

for the quasi-3D case, Kz for the zone between points i, j, k


and i, j, k + 1 (see right side of Fig. 16) is calculated by
in-series formula for layered media, and VCONT (aka
Leakance) becomes (using the nomenclature in Fig. 16)
Kz

=
=

VCONT

ztotal
PN zi

i=1 Ki
Zu
2 + Zc
Zu
2
Kzu

Zc
Kzc

ZL
2

ZL
2
KzL

ZL
2
KzL

1
Zu
2
Kzu

Kzc
Zc

Zc
Kzc

for Kzc << Kzu , KzL

VCONT Summary
I

Modflow utilizes the leakance concept in many packages,


including BCF6, River, Stream, Drain, GHB
GMS and LPF
I

GMS interface to Modflow defaults to the LPF discretization


package
has no provision to generate the VCONT array (known as
VKCB, pg. 62 Harbaugh et al., 2000)
i.e. when using LPF in GMS aquitards are modeled as explicit
layers, or via the vertical anisotropy parameter
Gotcha: in MODFLOW2005 user can choose to recalculate at
each time step; gives different results than pre-MODFLOW2K
model runs

when BCF is selected for discretization, Leakance is calculated


from Map Module information in GMS, and can be edited
manually

Choosing a Matrix Package


I

as seen above, Modflow generates a matrix equation, which


can then be solved using a variety of techniques

different techniques will perform better under particular


circumstances
if a Modflow run fails to converge (see end of output file), do
the following:

adjust solver parameters via appropriate GMS dialog (pull


down Modflow/PCG2 ..., etc.
in particular, relaxing the maximum residual or head change
criterion can be effective. DONT increase this beyond about
1% of the value, since the criterion will be the minimum error
for the problem.
adjust other parameters, especially number of internal
iterations (e.g. PCG), etc.
change solver packages. In general LMG will be most robust
(but greatest memory demand), SSOR will most reliably
converge (but with higher error). See descriptions below

Choosing a Matrix Package (cont.)

see Barrett et al. (1994) for a good general summary of


methods

General Matrix Solution

as described above, in general the inverse of a matrix is sought

most methods involve some sort of factorization, transforming


the original matrix into one or more readily solvable forms

Gaussian Elimination, where rows of the matrix are combined


sequentially so as to generate an upper triangular matrix. In
greatly simplified form, the original matrix equation becomes:
A
|{z}

~h
|{z}

Conductances Unknown Heads

~
q
|{z}

Source/Sinks
Known Head

[U

~
L] ~h = q

General Matrix Solution (cont.)

many methods compute a residual (first estimate of change


between current solution and the last one), and use that to
accelerate convergence. This is done by multiplying the first
estimate of change by a number greater than 1 (the
acceleration parameter) and using that to make a revised
estimate

Comparative Performance
I

1-D case (Fig. 17)


I
I

I
I

for two-component error, one long and one short wavelength


SOR and SIP remove high-frequency error quickly, little effect
on low-frequency
PCG does better at reducing both
LMG removes both very quickly

2-D case (Fig. 18)


I
I

excessively sharp and high peaked initial condition


SIP, SOR and PCG converge slowly (initial peak is spread out
more slowly, i.e. only local nodes are affected during each
iteration)
PCG: Cholesky preconditioner spreads much more rapidly than
with polynomial, i.e. allows increased internode communication
again LMG gets near-perfect solution in one cycle (Fig. 19)

1-D Performance Comparison of Matrix Packages

Figure 17:

Comparison of Modflow matrix package performance, showing convergence rate and trends. The
problem is a 1-D grid with fixed head of 0 at both ends, and the indicated smooth (wavenumber k=1) and
oscillatory (k=15) components of error applied to the initial head condition (k=1+15). The exact solution is a
value of 0 everywhere in the grid. After Mehl and Hill (Fig. 1, 2001).

2-D Performance Comparison of Matrix Packages

Figure 18: Comparison of Modflow matrix package performance, showing


convergence rate and trends. The problem is a 1-D grid with fixed head of 0 at both
ends, and the indicated smooth (wavenumber k=1) and oscillatory (k=15)
components of error applied to the initial head condition. The exact answer is a value
of 0 everywhere in the grid. After Mehl and Hill (Fig. 2, 2001).

LMG 2D Performance

Figure 19: LMG Modflow matrix package performance. Excellent


solution is achieved in just one cycle. After Mehl and Hill (Fig. 2, 2001).

Strongly Implicit Method


I

this method modifies the matrix equation by adding C ~h to


both sides(where m is the iteration number:
m+1
m
~ + N~h
(M + N) ~h
= q

(5)

N is constructed so that (M + N) can be easily factored into


UL form (Chp. 12, McDonald and Harbaugh, 1988)
GMS user-settable parameters include (Fig. 20):
I

maximum iterations: increase this for runs that show steadily


decreasing residuals (see Modflow output file)
number of iteration parameters: sets number of values used for
weight parameter de-emphasizing influence of distant cells in
finite difference molecule. Larger values may aid convergence
in difficult cases. Essentially changes how N is built.
acceleration parameter: weight given to residual-correction
term

SIP Dialog

Figure 20: GMS interface for Modflow SIP (Strongly Implicit Package)
package showing most-commonly-set parameters.

SSOR Package
I

the slice-successive overrelaxation method solves


two-dimensional slices of the grid using Gaussian
Elimination, treating out-of-slice heads as known values.

Essentially de-emphasizes slice-to-slice links. Can work very


well if main flow is along-slice

slices are solved sequentially, iterating over the entire grid

the main goals of the method are to minimize memory usage


and numerical spikes from distant parts of the grid.

a good method for poorly-converging matrices, but normally


yields higher numerical errors

a residual scheme is used to accelerate convergence,


trial-and-error adjustment of the acceleration parameter can
optimize convergence

SSOR Dialog

Figure 21: GMS interface for Modflow SSOR


(Slice-Successive-Overrelaxation) package showing most-commonly-set
parameters.

PCG Package
I

like the SIP method, the pre-conditioned conjugate gradient


technique factorizes the A matrix into a sum of two matrices
M + N (matrix factorization)

M is as close to A as possible but is easily invertible (Hill,


1990a)

a pre-conditioner is used to obtain M so that it is


well-behaved during inversion. The conditioning function
may be set by the user in GMS (Fig. 22). Incomplete
Cholesky should give the quickest convergence, but
polynomial may be better for noisy problems (esp. with
parallel processing)

conjugate gradient refers to the inversion of M + N wherein


residuals in two directions in the matrix are minimized
simultaneously

PCG Package (cont.)

a residual acceleration scheme is used based on the previous


two estimates of head during iteration. The relaxation
parameter controls the rate of acceleration

Hill (1990b) indicates SIP may converge more quickly in most


cases

PCG Dialog

Figure 22: GMS interface for Modflow PCG (Preconditioned Congugate


Gradient) package showing most-commonly-set parameters.

LMG Method
I

main objective is to address different wavelength components


of the residual error using coarser (sub)grids than entered by
the user

errors that appear smooth on finer grids will appear oscillatory


on coarser grids, and can be more rapidly reduced

solution proceeds from the finest to coarsest grids, then


corrections are applied from the coarsest to the finest

note this is an algebraic multigrid method, and so technically


the matrix, rather than the physical grid, is refined

can be very efficient (converges in 10% of the time required


by PCG), but requires 3-8 times more memory (Mehl and Hill,
2001)

the solver of choice for large problems


the GMS interface (Fig. 23) allows:

LMG Method (cont.)

enable PCG iterations: do this if LMG fails to reduce residual


on repeated cycles (i.e. use PCG to reduce some error
components)
residual acceleration parameter: as above. Values < 1 retard
changes in h, values > 1 accelerate them

LMG Dialog

Figure 23: GMS interface for Modflow LMG (Linked Multi-Grid) package
showing most-commonly-set parameters.

Modflow Documentation

Here is a list of useful links to Modflow documentation/guides:


I

USGS online overview/guide for Modflow

USGS general Modflow page

References
Anderson, M.P., Woessner, W.W.: Applied Groundwater Modeling.
Academic Press, San Diego (1992)
Barrett, R., Berry, M., Chan, T.F., Demmel, J., Donato, J., Dongarra, J.,
Eijkhout, V., Pozo, R., Romine, C., der Vorst, H.V.: Templates for the
Solution of Linear Systems: Building Blocks for Iterative Methods.
SIAM, Philadelphia, PA (1994),
http://www.netlib.org/templates/templates.ps
Detwiler, R., Mehl, S., Rajaram, H., Cheung, W.: Comparison of an
algebraic multigrid algorithm to two iterative solvers used for modeling
ground water flow and transport. Ground Water 40(3), 267272
(MAY-JUN 2002)
Harbaugh, A.W.: Direct solution package based on alternating diagonal
ordering for the U.S. Geological Survey modular finite-difference
ground-water flow model. Open-File Report 95-288, U. S. Geol.
Survey, Denver, CO (1995), http:
//water.usgs.gov/software/MODFLOW/code/doc/ofr95288.pdf

References (cont.)
Harbaugh, A.W.: Modflow-2005, the u.s. geological survey modular
ground-water modelthe ground-water flow process. Techniques and
Methods Book 6-A16, U. S. Geol. Survey, Denver, CO (2005),
http://pubs.usgs.gov/tm/2005/tm6A16/PDF/TM6A16.pdf
Harbaugh, A.W., Banta, E.R., Hill, M.C., McDonald, M.G.:
MODFLOW-2000, the U.S. Geological Survey modular ground-water
model user guide to modularization concepts and the ground-water
flow process. Open File Rept. OFR00-92, U. S. Geol. Survey, Denver,
CO (2000), http://water.usgs.gov/nrp/gwsoftware/
modflow2000/ofr00-92.pdf, 121 p
Harbaugh, A.W., McDonald, M.G.: Users documentation for
MODFLOW-96, an update to the U.S. Geological Survey modular
finite-difference ground-water flow model. Open-File Report 96-485,
U.S. Geol. Survey (1996), http:
//water.usgs.gov/software/MODFLOW/code/doc/ofr96486.pdf
Hill, M.C.: Preconditioned conjugate-gradient 2 (PCG2), A computer
program for solving ground-water flow equations. Water-resour.
investig. rept. 90-4048, U.S. Geol. Survey, Denver, CO (1990a)

References (cont.)
Hill, M.C.: Solving groundwater flow problems by conjugate-gradient
methods and the strongly implicit procedure. Water Resour. Res. 26,
19611969 (1990b)
McDonald, M.G., Harbaugh, A.W.: A modular three-dimensional
finite-difference ground-water flow model. Techniques of Water
Resour. Investig. A1, Book 6, 200 (1988)
Mehl, S.W., Hill, M.C.: Modflow-2000, the u.s. geological survey modular
ground-water model user guide to the link-amg (lmg) package for
solving matrix equations using an algebraic multigrid solver. Open File
Report OFR 01-177, U. S. Geol. Survey, Denver, CO (2001), http://
water.usgs.gov/nrp/gwsoftware/modflow2000/ofr01-177.pdf
Prickett, T.A., Lonnquist, C.G.: Selected digital computer techniques for
groundwater resource evaluation. Bulletin, Illinois State Water Survey,
Urbana, IL (1971)

You might also like