Groundwater Flow and Solute Transport

Modeling

Ye Zhang

Dept. of Geology & Geophysics

University of Wyoming

⃝
c Draft date February 13, 2016
Contents

Contents i
0.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v
0.2 This Class . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi
0.2.1 Textbook . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
0.2.2 Tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii
0.2.3 Questions and Answers . . . . . . . . . . . . . . . . . . . viii
0.2.4 Homework, Projects, Exams and Grades . . . . . . . . . . viii
0.3 Basic Math Review . . . . . . . . . . . . . . . . . . . . . . . . . . ix
0.4 Homework 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix

1 Groundwater Flow Equations (Review) 1

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Darcy’s Law — Review . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.1 Darcy’s Law of This Course . . . . . . . . . . . . . . . . . 4
1.3 3D General Flow Equation (Confined Aquifer) . . . . . . . . . . 6
1.3.1 Less General Flow Equations . . . . . . . . . . . . . . . . 7
1.4 2D Planeview Flow Equation (Confined & Unconfined) . . . . . . 10

2 Modeling Overview for Groundwater Flow 15

2.1 General Methodology . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 Analytical versus Numerical Methods . . . . . . . . . . . . . . . 19
2.3 Validity of Numerical Solution . . . . . . . . . . . . . . . . . . . 23
2.4 Model Calibration and Parameter Estimation . . . . . . . . . . . 24
2.5 Interpreting Model Results . . . . . . . . . . . . . . . . . . . . . 27
2.6 Model Limitation and Uncertainty . . . . . . . . . . . . . . . . . 28
2.7 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

3 Mathematical Notations, Definitions and Theorems 31

3.1 Differential Equations . . . . . . . . . . . . . . . . . . . . . . . . 31
3.2 Scalar, Vector, Tensor, Field . . . . . . . . . . . . . . . . . . . . . 33
3.3 Taylor Series and Finite Difference Approximations . . . . . . . . 34
3.4 Numerical Error, Convergence, Stability . . . . . . . . . . . . . . 35
3.5 Matrix & Vector . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.6 Linear Algebra and Solution Techniques . . . . . . . . . . . . . . 39

i
ii CONTENTS

3.6.1 Direct Approaches . . . . . . . . . . . . . . . . . . . . . . 40

3.6.2 Iterative Approaches . . . . . . . . . . . . . . . . . . . . . 41
3.6.3 Comparison and Summary . . . . . . . . . . . . . . . . . 44
3.7 Homework 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.8 Gauss-Jordan Direct Solver with Pivoting (Optional) . . . . . . . 46

4 Lumped Parameter Models 53

4.1 Mass Balance Principle . . . . . . . . . . . . . . . . . . . . . . . 53
4.2 Groundwater Flow Model . . . . . . . . . . . . . . . . . . . . . . 55
4.2.1 Analytical Solution . . . . . . . . . . . . . . . . . . . . . . 56
4.3 Solute Transport Model . . . . . . . . . . . . . . . . . . . . . . . 57
4.3.1 Analytical Solution . . . . . . . . . . . . . . . . . . . . . . 58
4.4 Numerical Solution . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.4.1 Euler’s Method . . . . . . . . . . . . . . . . . . . . . . . . 59
4.4.2 Huen’s Method . . . . . . . . . . . . . . . . . . . . . . . . 60
4.4.3 Runga Kutta Methods . . . . . . . . . . . . . . . . . . . . 61
4.5 Homework 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.6 Project One — Flow & Transport Modeling . . . . . . . . . . . . 64

5 One-Dimensional Flow Modeling 67

5.1 1D Steady-State Flow . . . . . . . . . . . . . . . . . . . . . . . . 67
5.1.1 Homogeneous Porous Media . . . . . . . . . . . . . . . . . 67
5.1.2 Heterogeneous Porous Media . . . . . . . . . . . . . . . . 69
5.2 Homework 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.3 FD Solution by Computer Codes . . . . . . . . . . . . . . . . . . 71
5.3.1 Matrix-Based Approaches . . . . . . . . . . . . . . . . . . 71
5.3.2 Matrix-Free Approaches . . . . . . . . . . . . . . . . . . . 72
5.4 Irregular Grid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5.5 1D Transient Flow . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5.5.1 FD Explicit Formulation . . . . . . . . . . . . . . . . . . . 84
5.5.2 FD Implicit Formulation . . . . . . . . . . . . . . . . . . . 85
5.5.3 FD Weighted Formulation . . . . . . . . . . . . . . . . . . 87
5.6 Homework 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99

6 Two-Dimensional Steady-State Flow 103

6.1 Specified Head & No Flow boundaries . . . . . . . . . . . . . . . 103
6.1.1 Finite Difference Discretization . . . . . . . . . . . . . . . 107
6.2 Homework 6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
6.3 Specified Head & Specified Flux boundaries . . . . . . . . . . . . 116
6.3.1 Variable Flux Boundary Condition . . . . . . . . . . . . . 122
6.4 Project Two — 2D Steady State Flow Modeling . . . . . . . . . 123
6.5 General Boundary Conditions . . . . . . . . . . . . . . . . . . . . 129
6.5.1 Boundary Fluxes . . . . . . . . . . . . . . . . . . . . . . . 129
6.5.2 Internal Fluxes . . . . . . . . . . . . . . . . . . . . . . . . 131
6.6 Velocity Calculation . . . . . . . . . . . . . . . . . . . . . . . . . 131
6.6.1 Darcy Flux . . . . . . . . . . . . . . . . . . . . . . . . . . 131
CONTENTS iii

6.6.2 Mass Balance Calculation . . . . . . . . . . . . . . . . . . 135

6.6.3 Average Linear Velocity . . . . . . . . . . . . . . . . . . . 137
6.6.4 Groundwater Pathline Generation . . . . . . . . . . . . . 138
6.7 FD Formulations for Heterogeneous K Field . . . . . . . . . . . . 142
6.7.1 Project Two B (Optional) — 2D SS Modeling with Heterogeneous K142

7 Two-Dimensional Transient Flow 145

7.1 Mathematical Formulation . . . . . . . . . . . . . . . . . . . . . . 146
7.2 Finite Difference Formulations . . . . . . . . . . . . . . . . . . . 146
7.2.1 Explicit Formulation . . . . . . . . . . . . . . . . . . . . . 147
7.2.2 Implicit Formulation . . . . . . . . . . . . . . . . . . . . . 148
7.2.3 Weighted Formulation . . . . . . . . . . . . . . . . . . . . 148
7.3 Mass Balance Calculations . . . . . . . . . . . . . . . . . . . . . . 149
7.4 Project Three— 2D Transient Flow Modeling . . . . . . . . . . . 149

8 Solute Transport Modeling 159

8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
8.1.1 Advection & Dispersion . . . . . . . . . . . . . . . . . . . 159
8.1.2 Hydrodynamic Dispersion in Homogenous Medium . . . . 161
8.2 Derivation of the 1D Advection-Dispersion Equation (ADE) . . . 163
8.3 Finite Difference Discretization . . . . . . . . . . . . . . . . . . . 165
8.3.1 Explicit Formulation . . . . . . . . . . . . . . . . . . . . . 166
8.3.2 Implicit Formulation . . . . . . . . . . . . . . . . . . . . . 167
8.3.3 Weighted Formulation . . . . . . . . . . . . . . . . . . . . 167
8.4 Solute Transport Modeling . . . . . . . . . . . . . . . . . . . . . 168
8.5 Homework 7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
8.6 Extension of ADE to Higher Dimensions . . . . . . . . . . . . . . 173
8.7 Advanced Topics (Optional) . . . . . . . . . . . . . . . . . . . . . 174
8.7.1 Groundwater Pathline & Particle Tracking . . . . . . . . 174
8.7.2 Solute Macrodispersion in Heterogeneous Media . . . . . 174
8.7.3 Alternatives to ADE . . . . . . . . . . . . . . . . . . . . . 174

9 3D Models of Flow and Transport 177

9.1 Available Software . . . . . . . . . . . . . . . . . . . . . . . . . . 177
9.2 Project Four—3D Flow Modeling with Groundwater Vista . . . . 178
9.3 Final Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192

10 Parameter Estimation for Groundwater Models (Optional) 195

10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
10.2 Forward Estimation of Hydraulic Conductivity . . . . . . . . . . 197
10.2.1 Grid-Block Equivalent K* . . . . . . . . . . . . . . . . . . 199
10.2.2 Issues with Numerical Upscaling . . . . . . . . . . . . . . 201
10.3 Upscaling for Irregular Domains . . . . . . . . . . . . . . . . . . 203
10.3.1 Numerical Flow Experiments . . . . . . . . . . . . . . . . 206
10.3.2 Tensor Analysis . . . . . . . . . . . . . . . . . . . . . . . . 208
10.4 Inverse Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
iv CONTENTS

10.4.1 Notation Conventions . . . . . . . . . . . . . . . . . . . . 214

10.4.2 Least Square Objective Function . . . . . . . . . . . . . . 214
10.4.3 Weight Matrix . . . . . . . . . . . . . . . . . . . . . . . . 216
10.4.4 Gauss-Newton Nonlinear Regression . . . . . . . . . . . . 217
0.1. INTRODUCTION v

This is a set of course notes written by Ye Zhang for a Groundwater Flow

and Transport Modeling course.

Fall 2012
GEOL 5030
3 CREDITS
GRADING: S/U or A-F (It is up to each student.)
Lecture location: ESB1006
Lecture times: Tues + Thurs (9:35 am ∼ 10:50 am)
Office hours: M(4:00∼5:30 pm), F(3:00∼4:30 pm), GE 220
Email: [email protected]
Phone: 307-766-2981
—————————————————————————————————

Prerequisite:
Calculus I & II;
Geohydrology;
Matlab Programming language*
*This course emphasizes the fundamental development of mathematical mod-
els of groundwater flow and solute transport and their applications using com-
puter simulations. Students are expected to write small computer codes for
(up to 2D) problems, thus rudimentary skills in programming with Matlab are
necessary. If a student has not used Matlab before, please study an excellent
tutorial before attending this class (it will take around 3 hours):
http://faculty.gg.uwyo.edu/yzhang/teaching.html
—————————————————————————————————

NOTE: The lecture notes will be periodically posted on the Wyoweb

course website, usually 1 week before the relevant lectures. Please
make a habit of regularly checking for notes or other announcements
from the site. However, the lecture notes do NOT include: (1) so-
lutions to the exercises and homework; (2) proofs to theories and
equation derivations. These will be presented only during the lec-
tures. So, do not rely on the notes for everything — attendance and
in-class participation are key to doing well.

0.1 Introduction
Mathematical models of groundwater flow have been used since the late 1800s.
A mathematical model consists of differential equations developed from ana-
lyzing groundwater flow (or solute transport in groundwater) and are known
to govern the physics of flow (and transport). The reliability of model predic-
tions depends on how well the model approximates the actual situation in the
vi CONTENTS

Set of PDEs (Mathematical Set of Algebraic Equations

Model) Numerical Methods (Discrete Model)
(FDM, FEM, FVM)

Direct or
Calculus Iterative
Techniques Solution
Methods

Analytical Solution (Only Compare if possible Approximate Numerical

possible for a limited Solution (Only at discrete
number of cases) grid points)

Compare Compare

Field Observations

Figure 1: A flow chart showing the “big picture” of numerical modeling.

field. Inevitably, simplifying assumption must be made in order to construct

a model, because the field situation is usually too complicated to be simulated
exactly. In general, the assumptions necessary to solve a mathematical model
analytically are very restrictive. For example, many analytical solutions are de-
veloped for homogeneous, isotropic, and/or infinite geological formations where
flow is also steady-state (hydraulic head and groundwater velocity do not change
with time). To deal with the more realistic situations (e.g., heterogeneous and
anisotropic aquifer in which groundwater flow is transient), the mathematical
model is commonly solved approximately using numerical techniques.
Ever since the 1960s, when computers first become widely available, numer-
ical models have been the preferred type of model for analyzing groundwater
flow and transport. Figure 1 illustrates the relationship between mathematical
model, discrete algebraic model, analytical solution, approximate (numerical)
solution, and field observations.

0.2 This Class

In this course, the goal is the use of numerical models to simulate groundwa-
ter flow and solute transport. It offers an introduction to the popular Finite
0.2. THIS CLASS vii

Difference Method1 , a widely used numerical method to solve the governing

partial differential equations in hydrogeology. Fundamental understanding will
be emphasized, as a fairly rigorous mathematical treatment is presented. This
approach deviates from the typical introductory classes that may emphasize the
more applied aspects, i.e., software application. I feel that mastering of such
skills would come naturally after a student first develops a firm grasp of the fun-
damentals. This course is thus designed at the advanced upper undergraduate
and graduate level, appropriate for the level of mathematical rigor contained
herein. To comprehend the materials presented, a student should have suffi-
cient knowledge of college math, e.g., Calculus I & II (required), Linear Algebra
(optional), Differential Equations (optional). A prior class in Hydrogeology is
required, e.g., Geohydrology (GEOL4444/5444).
Throughout the course, many formulations and equations are developed us-
ing mathematics. The emphasis is on understanding how these equations are
obtained. However, you will rarely be tested on equation derivations in exams
(those few that you may be tested on — I’ll let you know), so it is unnecessary
to memorize a large number of formulas or solutions. Typically, the exams will
provide the necessary formulations so understanding what they mean and how
to use them is key.

0.2.1 Textbook
The set of lecture notes I developed include the essential contents for this course.
However, other books can serve as additional resources for self study:

• Introduction to groundwater modeling: finite difference and finite element methods,

H. F. Wang and M. P. Anderson, 1995, Academic Press, 237 p.

• Applied Contaminant Transport Modeling, C. Zheng, G. D. Bennett, 2002,

Wiley-Interscience, 656 p.

• Groundwater modeling by the finite element method, Jonathan Istok, 1989,

American Geophysical Union Publication, 495 p.

Some materials are obtained by assembling course notes prepared by others

(references will be given in the notes).
Due to time constraint, the current course cannot hope to cover every aspect
of the subject as presented in these books. For example, variable-density flow
is not presented, nor is flow in deforming porous media. Immiscible, multiphase
fluids (e.g., flow in the soil zone) are not covered. In terms of numerical ap-
proaches, only the Finite Difference Method is presented. Some of these topics
can be understood by independent study, others (e.g., Finite Element Method)
should be the subjects of specialized classes. Thus, most of this class is de-
voted to the study of single-phase (water), uniform-density flow moving
through non-deforming porous media (e.g., groundwater aquifers that are not
1 Due to time limitation, the Finite Element Method may be introduced in a separate future

class.
viii CONTENTS

going through compaction or subsidence). The solute transport equation and

solutions are developed for a nonreactive solute with small concentrations so as
not to affect the density of the groundwater in which it resides.

0.2.2 Tools
In this course, exercises, homework, labs, term project, exam problems can be
solved using a variety of approaches, e.g., by hand, using Excel spreadsheet,
writing Matlab codes, using software. The popular software packages (e.g.,
MODFLOW from USGS) will be introduced in the end, when we conduct 3D
modeling of steady-state groundwater flow and advective transport.

0.2.3 Questions and Answers

Students can ask questions: (1) during office hour; (2) during lectures. As
a rule, email is the last resort since I receive a lot of them every day and your
message stands a chance of being overlooked by mistake.

0.2.4 Homework, Projects, Exams and Grades

When working on homework/lab/exam problems, read each assignment care-
fully. Read it twice if you have to. Do not skip anything or you might find that
the later questions will not make sense. Of course, you should always point out
to me if anything is unclear in the problem descriptions. A study guide will be
provided before the exam.
Concerning the styles, 4 points must be emphasized: (1) Most problem
sets involve writing equations, so if appropriate, provide a complete analysis
rather than a single number. (2) Be professional in your presentations: write
down the unit for your numerical results and round off the final number to 1
or 2 decimal point.(3) You can discuss the problem sets with fellow students,
but complete your assignments by yourself. Copying other’s work is considered
cheating and no points will be given for that assignment. In particular, a student
caught cheating may receive a “F” for the class. (4) Hand in your work on time.
Unless otherwise stated, the general timeline is to hand in your homework in
the beginning of the class a week from the day it is assigned (see syllabus for
the detailed due dates). For larger lab assignments, you will be given 2 weeks
to complete it. If the assignment is not handed in on time, for every
day it is delayed, 10 points will be taken out of the 100 points of this
work until no points remain.
Grades: In this course, emphasis is placed on the homework problems, labs
and projects due to the time-consuming nature of most of these assignments.
The final grade will be given based on your homework, labs, project and/or
exams. The tentative percentage is shown:

• Homework 49% (7% × 7 homework)

0.3. BASIC MATH REVIEW ix

Table 1: Letter versus numerical grade

S U
>60 < 60

A B C D F
90-100 80-89 70-79 60-69 < 60

• Projects 36% (12% × 3 projects)

• Term Project (Or Final Exam) 15%

Note that each homework/lab/exam has a stand-alone grade of 100 points.

When determining the final grade, these will be normalized reflecting the per-
centage distribution above. The final letter grade is shown in Table(1).
Finally, I set high expectation in this class. Be prepared to come to class, pay
full attention, participate in exercises, work out the homework/lab by yourself,
hand in your assignments on time, write professionally (clear, precise, concise).
Please keep all course materials (notes, exercises, homework, exams, labs) to
yourself and do not pass them on to future students. They must, as you have,
work to earn the credit.

0.3 Basic Math Review

Please review the first handout: BasicMathReview.pdf, which is Chapter 1 Ba-
sic Math of the Geohydrology class (GEOL4444/5444) — for the mathematics
you should already know before taking this class. If you find the math in-
comprehensible, please consider taking some other courses first, e.g., Calculus,
Geohydrology. Please ignore 1.13 Test 1.

Gradient Tutorial

Please review the second handout: Gradient 123D.pdf, which is a tutorial of

how to compute hydraulic head gradient in 1D, 2D and 3D for any coordinate
system.

0.4 Homework 1
Exercise 1 and Exercise 2 of the first handout.
Chapter 1

Groundwater Flow
Equations (Review)

1.1 Introduction
The first step in developing a mathematical model of almost any system is to
formulate what are known as general equations. General equations are dif-
ferential equations that are derived from the physical principles governing the
process that is to be modeled. In the case of groundwater flow, the relevant
physical principles are Darcy’s law and mass balance. By combining the math-
ematical relation describing each principle, it is possible to come up with a gen-
eral groundwater flow equation, which is a partial differential equation (PDE).
Since in groundwater studies, the fluxes (Darcy flux, average linear velocity)
are macroscopic quantities which are related to the head gradient and hydraulic
properties of the aquifer (i.e., porosity, permeability, hydraulic conductivity) by
the Darcy’s law, the PDE developed for the general groundwater flow is thus
established for macroscopic flow in porous media.1
There are several different forms of the general flow equation depending on
whether the flow is two-dimensional or three-dimensional, isotropic or anisotropic,
and transient or steady state. In the following sections, we’ll first review the
most general flow equation (heterogenous, anisotropic and transient) and then
look at the simplified equations under different simplifying assumptions (ho-
mogenous, isotropic, steady-state). To make the derivations as understandable
as possible, a simple approach is adopted: first, the general equation is devel-
oped for the simple one-dimensional case; the results are then extended to the
three-dimensional case.
1 There are even more fundamental equations developed for flow at the microscopic scale

(i.e., pore scale), based on the theory of hydrodynamics. For example, with the Naiver-Stokes
equation, hydrologists have used various approaches to prove the macroscopic Darcy’s law.
These topics are however beyond this course. Interested students may find relevant discussions
in Section 5.10 of Bear (1988).

1
2 CHAPTER 1. GROUNDWATER FLOW EQUATIONS (REVIEW)

1.2 Darcy’s Law — Review

In 1856, engineer Henry Darcy was working on a project using sand to filter
water supply for a French town. He performed laboratory experiments to ex-
amine the factors that govern the rate of water through sand. The results of his
experiments defined the empirical principle of groundwater flow, in an equation
known as Darcy’s law. Darcy’s experiment consisted of a sand-filled column
with an inlet and an outlet for water (search Google image for Darcy’s law will
bring up many images of the Darcy’s experiment). Two manometers measure
the hydraulic head at two points within the column (h1 and h2 ). The sand is
fully saturated, and a steady flow of water is forced through it at a volumetric
rate of Q [L3 /T] (Q is sometimes called the flow rate or the discharge rate).
Darcy found that Q ∼ head difference ∆h between the manometers, inversely
proportional to the distance between manometers ∆s, and ∼ cross sectional
area of the sand column (A):

Q ∝ ∆h Q ∝ (1/∆s) Q∝A

Combining these observations and writing an equation in differential form

gives Darcys law for one-dimensional flow:

Q = −KA(dh/ds)

where Q is discharge rate in the s direction. The minus sign is necessary be-
cause head decreases in the direction of flow (i.e., water is always flowing from
higher hydraulic head to lower hydraulic head). If there is flow in the positive
s direction, Q is positive and dh/ds is negative2 . Conversely, when flow is in
the negative s direction, Q is negative and dh/ds is positive. The constant of
proportionality K is the hydraulic conductivity in the s direction, a property of
the porous medium and the fluid (water) filling the pores. The common units
for hydraulic conductivity are meters/year for regional studies, m/day for lo-
cal aquifer-scale studies, and cm/sec for laboratory studies. Therefore, in some
analysis, we often deviate from the rule of using the SI unit.
Another form of the Darcy’s law is written for the Darcy flux (or the
Darcy Velocity, or, the Specific Discharge) (q) which is the discharge rate per
unit cross-sectional area:

q = Q/A = −K(dh/ds)

The Darcy flux q has unit of velocity [L/T]. Darcy flux is not the actual fluid
velocity in the porous media, it is just discharge rate (Q) per unit (solid) cross-
sectional area. Since porous rock does not have a solid cross-sectional area,
Darcy flux is a not the fluid velocity.
In the real subsurface, groundwater flows in complex 3D patterns. Darcy’s
law in three dimensions is analogous to that of one dimension. In a Cartesian
2 In this case, head is decreasing with s (since flow is towards +s), so by definition: dh/ds ≃
h(s+∆s)−h(s)
∆h/∆s = ∆s
< 0, since h(s + ∆s) < h(s).
1.2. DARCY’S LAW — REVIEW 3

x, y, z coordinate system, it is commonly expressed as:

qx = −Kx ∂h/∂x
qy = −Ky ∂h/∂y
qz = −Kz ∂h/∂z

where Kx , Ky , and Kz are the hydraulic conductivity in each of the coordinate

direction, respectively. qx , qy , and qz are 3 components of the Darcy flux ⃗q.3
However, what exactly are the Kx , Ky , and Kz ? They are the directional
hydraulic conductivity evaluated along each of the coordinate axis. To estimate
these directional conductivities, Darcy test can be conducted along the x axis,
in which case a horizontal hydraulic conductivity along the x direction can be
determined: Kx . Same idea applies to estimating conductivities in the y and z
directions.

In general, for a given porous medium, Kx , Ky and Kz do not need to be

the same, in which case, the medium is called anisotropic. On the other hand,
if Kx = Ky = Kz , the medium is called isotropic.

In general, x, y, z can have any orientation, but it is common to set z vertical

and x and y horizontal. In three dimensions, all fluxes (Q ⃗ , ⃗q, ⃗v ) are vector quan-
tities (3 components), as is the hydraulic gradient I⃗ = [Ix , Iy , Iz ]T (Ix = ∂h/∂x,
Iy = ∂h/∂y, Iz = ∂h/∂z). The hydraulic head is still a scalar (1 component),
while the hydraulic conductivity is a tensor quantity (9 components):

 
Kxx Kxy Kxz
K =  Kyx Kyy Kyz 
Kzx Kzy Kzz

Notice that in the above equation, K contains more components than what
is shown in the previous 3D Darcy’s law with only Kx , Ky , Kz . Actually, the
previous Darcy’s law is established under a special requirement: the 3 coor-
dinate axes coincide with the principal axes of the hydraulic conduc-
tivity tensor (the direction of maximum, minimum, and intermediate
hydraulic conductivity). When this requirement is satisfied, K becomes a

3 Note that K , K , K , q , q , and q are all specific to the given coordinate. When
x y z x y z
we change or rotate the coordinate, the magnitude of these component quantities will change
with respect to the coordinate.
4 CHAPTER 1. GROUNDWATER FLOW EQUATIONS (REVIEW)

diagonal tensor4 and the Darcy’s law thus becomes simplified:

 
Kx 0 0
K =  0 Ky 0 
0 0 Kz

We can write: K = diag[Kx , Ky , Kz ] and Kx , Ky , and Kz are the three principal

components (or eigenvalues) of K.
When this coordinate alignment requirement is not satisfied, i.e., principal
conductivity axes do not coincide with the coordinate axes, the most general
form of 3D Darcy’s law is written as:

⃗q = −KI⃗

or, more explicitly:

    
 qx  Kxx Kxy Kxz  ∂h/∂x 
qy = −  Kyx Kyy Kyz  ∂h/∂y
   
qz Kzx Kzy Kzz ∂h/∂z

we can easily verify that previous 3D Darcy’s law (with 3 principal components:
Kx , Ky , Kz ) is just a reduced form of the above equation when K is a diago-
nal tensor (under the condition that the coordinate axes are aligned with the
principal axes of K).
If we have time, I’ll show you how to do tensor transformation with changing
coordinate, i.e., relating the principal components of K to its full tensor form.

1.2.1 Darcy’s Law of This Course

In this course, to facilitate calculations, we make the assumption that the coor-
dinate axes coincide with the principal axes of the hydraulic conductivity tensor,
so the hydraulic conductivity is diagonal. This will considerably simply the FD
formulations we’ll use. Thus, (in this course) we always use:
      
 qx  Kx 0 0  ∂h/∂x   Kx ∂h/∂x 
qy = −  0 Ky 0  ∂h/∂y =− Ky ∂h/∂y
     
qz 0 0 Kz ∂h/∂z Kz ∂h/∂z
4 If the time allows, I’ll derive and explain the principle behind tensor rotation and eigen

analysis. In turns out, a full tensor K of any coordinate system can be projected, by co-
ordinate transform, to a diagonal tensor K = diag[Kx , Ky , Kz ] which are defined along 3
orthogonal principal axes (the direction of each can be defined by the unit eigen vector of
K). Kx , Ky , and Kz are the corresponding eigenvalues. In hydrology, the above Kx , Ky ,
and Kz are also called principal components, the corresponding eigen vectors are sometimes
called principal axes. In Matlab, you can type “help eig” to find out how to do a eigen
analysis for a (n ×n) square symmetric matrix. Note that for tensors of n=2 or 3, we can
visualize the relations in 2D and 3D geometric coordinate (thus tensor rotation can be proved
by geometric means). But when n > 3, the eigen values and vectors cannot be visualized
geometrically. A formal introduction on eigen analysis can be found in this online textbook:
http://tutorial.math.lamar.edu/Classes/LinAlg/LinAlg.aspx (see the Chapter titled “Eigen-
values and Eigenvectors”.
1.2. DARCY’S LAW — REVIEW 5

You can see that the above relation reduces to the previous 3D Darcy’s law. It
is best if you memorize this equation, but keep in mind the underlying
assumptions.
Moreover, when analyzing 2D flow along the x-y plane (i.e., the velocity
component in the z direction does not change: qz = 0), the above relation
becomes:
{ } [ ]{ } { }
qx Kx 0 ∂h/∂x Kx ∂h/∂x
=− =−
qy 0 Ky ∂h/∂y Ky ∂h/∂y
This is because, according to Darcy’s law of this course, qz = −Kz (∂h/∂z) = 0,
since conductivity is always greater than 0, this leads to (∂h/∂z) = 0. Substi-
tuting both into the above full 3D equation leads to the current 2D form (since
the last equation becomes: qz = 0 = −Kz × 0 which no longer contains an
unknown variable).
A similar form can be written for 2D flow in the x-z plane (what do you
think it is?). Sometimes however the velocity is amenable to one-dimensional
analysis (e.g., flow in the Darcy Tube). In this case, only the one-dimensional
Darcy’s law is needed:
qs = −Ks (∂h/∂s)
Where Ks is hydraulic conductivity along the s direction (s can be aligned in
the x, y, z, or any direction of interest), ∂h/∂s is the hydraulic head gradient
along the s direction. For example, groundwater flow in consolidating clay tends
to be largely vertical, in this case, if we decide to conduct 1D analysis, we write:
qz = −Kz (∂h/∂z).
Finally, to facilitate analytical analysis by hand, many exercises in this class
make the assumption of an isotropic conductivity in the aquifer of interest,
which means: Kx = Ky = Kz = K, and the above general 3D equation can be
simplified to become:
   
 qx   ∂h/∂x 
qy = −K ∂h/∂y
   
qz ∂h/∂z
Similarly, depending on the dimension of the flow analysis, it can be reduced to
a 2D or 1D form. For example, 2D flow in the x-y plane of an isotropic aquifer
is written as: { } { }
qx ∂h/∂x
= −K
qy ∂h/∂y
Further, a 1D form is just the Darcy’s law for an isotropic medium:
qs = −K(∂h/∂s).
Compare this equation with the 1D Darcy’s law written above: in this form,
the conductivity without a subscript (K) implies isotropy; in the previous form
(Ks or Kz ), it implies anisotropy (though flow analysis may be one-dimensional,
spatially the conductivity can be anisotropic, Ks is just the conductivity along
the direction for which flow analysis is conducted).
6 CHAPTER 1. GROUNDWATER FLOW EQUATIONS (REVIEW)

z
qx(x) qx(x+ x)

x
x x+ x

Figure 1.1: A control volume element within the saturated zone with fixed
dimensions: ∆x, ∆y and ∆z.

1.3 3D General Flow Equation (Confined Aquifer)

In a typical mass balance analysis, the net flux of mass through the boundary
of a fixed control volume element is equated to the (time) rate of change of mass
within the element:
(1.1)
mass f lux in[M/T ] − mass f lux out [M/T ] = rate of change of mass[M/T ]

For example, if we pour water into a tub at 10 kg/min (mass flux in), but the
tub is leaking at 4 kg/min (mass flux out), then the water in the tub experiences
a rate of mass change at 10 −4 = 6 kg/min. This means at every minute, the
water in the tub increases by 6 kg (so the rate of change of mass is + 6kg/min).
This same mass balance principle can be applied to develop the groundwater
flow equation. Instead of the stationary tub, we’ll use the fixed and stationary
control volume again.
We will consider the mass balance for a small rectangular control volume
element within the saturated zone, as shown in Figure 1.1. The dimensions of
the element are fixed in space, regardless of compression or dilation of the aquifer
matrix or compression/expansion of pore water. For example, if the aquifer
compresses, more aquifer solids will be squeezed into the element and some
water will be displaced out of it. To make the derivation of the flow equations
as clear as possible, we will assume that the macroscopic flow in the vicinity of
this element is one-dimensional in the x direction: qx ̸= 0, qy = qz = 0. Using
mass balance and Darcy’s law, combined with the definition of specific storage
(Ss ) for a confined aquifer, we obtain for 1D flow (details given in class):
( )
∂ ∂h ∂h
(1.2) Kx = Ss
∂x ∂x ∂t

where Ss is the specific storage of a confined aquifer. Note that in developing

equation (1.2), the spatial gradient of the water density is considered to be
negligible (mostly fresh water).
1.3. 3D GENERAL FLOW EQUATION (CONFINED AQUIFER) 7

If the previous analysis is carried out without the restriction of one-dimensional

flow, there would be additional flux terms in the y and z directions (similar to the
flux term in the x direction). For three-dimensional flow, the general equation
is:
( ) ( ) ( )
∂ ∂h ∂ ∂h ∂ ∂h ∂h
(1.3) Kx + Ky + Kz = Ss
∂x ∂x ∂y ∂y ∂z ∂z ∂t

Thus the hydraulic head (h = h(x, y, z, t)) must obey this PDE to be con-
sistent with both the Darcy’s law and mass balance. Note that in equation
(1.3), the spatial gradient of the water density is considered to be negligible,
and, the principal conductivity axes must be aligned with the coordinate axes
(conductivity becomes a diagonal tensor: K=diag[Kx , Ky , Kz ]).
Equation (1.3) is the most universal form of the saturated flow equation for
a confined aquifer5 , allowing flow in three dimensions, transient flow (∂h/∂t ̸=
0), heterogenous conductivities (e.g., Kx , Ky , Kz are spatially variable), and
anisotropic porous medium (Kx ̸= Ky ̸= Kz ).
Equation (1.3) can be alternatively written in more condensed forms, e.g.,
Figure 1.2 shows two ways this equation is sometimes presented. On the other
hand, from the last of this equation:
∂h
−∇ · ⃗q = Ss
∂t
we can obtain the “most” general form of the 3D flow equation, see Figure 1.3.
Note that the most general 3D flow equation can also be written with the
Einstein summation (see Geohydrology, Advanced material, for explanation):

∂ ∂h ∂h
[Kij ] = Ss
∂xi ∂xj ∂t

Finally, equation (1.3) is just a reduced form of the most general equation under
the condition that conductivity principal directions are aligned with the axes of
the working coordinate.

1.3.1 Less General Flow Equations

Other, less general, forms of the flow equations can be derived from Equa-
tion(1.3) by making various simplifying assumptions. If the hydraulic conduc-
tivities are assumed to be homogenous (Kx , Ky , Kz are independent of x, y, z),
5 Actually, it is not. The most general form is when the coordinate axes are not aligned

with the conductivity principal axes, thus K has 9 components. This most general form of
groundwater flow equation will be presented in class, if time allows. However, many existing
groundwater programs and softwares solves equation (1.3), assuming that during the model
building stage, the coordinate system had been designed to align with the principal conduc-
tivity axes. This is a reasonable approach for many groundwater applications, however, for
complex geometry with multiple undulating layers or with intersecting structures, the most
general form may be necessary, allowing the more accurate calculations of the cross flows.
8 CHAPTER 1. GROUNDWATER FLOW EQUATIONS (REVIEW)

3D General Flow Equation in a Confined Aquifer

(6.10)

3D Law of this
course (assuming K prin-
cipal directions aligned
with coordinate axes)

Figure 1.2: Alternative forms of equation (1.3) by substituting the Darcy’s law
into the equation.

the general equation can be written as:

∂2h ∂2h ∂2h ∂h

(1.4) Kx 2
+ Ky 2
+ Kz 2
= Ss
∂x ∂y ∂z ∂t

This simplifies further when conductivity is also isotropic (Kx = Ky = Kz =

K):

∂2h ∂2h ∂2h Ss ∂h

(1.5) 2
+ 2 + 2 = ∇2 h =
∂x ∂y ∂z K ∂t

where ∇2 is called the Laplacian operator— a shorthand for the sum of the
2 2 2
second derivatives: ∇2 () = ∂∂x()2 + ∂∂y()2 + ∂∂z()2 .
In this section, all the previous equations have included the storage term
∂t ) which occurs only with transient flow ( ∂t ̸= 0). If however the flow
(Ss ∂h ∂h
∂h
is steady state ( ∂t = 0), the right-hand-side (RHS) of all the above equations
becomes zero. For example, under steady state, equation (1.5) becomes:

(1.6) ∇2 h = 0

This is a well-known PDE called the Laplace equation. Equation (1.6) has many
applications in fluid flow, heat conduction, electronics and elasticity.
The above equations in three-dimensional forms can be reduced to two or
one dimensions by simply dropping y and/or z terms from the equation. For
example, equation (1.3) can be reduced to solve a two-dimensional flow problem
in the horizontal plane (x,y axes only):
( ) ( )
∂ ∂h ∂ ∂h ∂h
Kx + Ky = Ss
∂x ∂x ∂y ∂y ∂t
1.3. 3D GENERAL FLOW EQUATION (CONFINED AQUIFER) 9

3D Most General Flow Equation in a Confined Aquifer

Most general Form of

3D Law (K
principal directions NOT
aligned with coordinate
axes)

When conductivity principle

directions are aligned with the
coordinate axes, K becomes a
diagonal tensor (Kx, Ky, Kz are
the principal components)

3D Most General Flow Equation Reduced to (6.10)

(6.10)

Figure 1.3: The most general form of the flow equation. Also shown is the
condition under which this equation is reduced to equation (1.3).
10 CHAPTER 1. GROUNDWATER FLOW EQUATIONS (REVIEW)
( ∂h )
This implies that ∂z ∂
Kz ∂z = 0 which becomes true if qz = −Kz ∂h ∂z = 0 or
∂h/∂z = 0 (envision a 3D flow field where velocity (⃗q) varies along the horizontal
plane, but its vertical component is zero).
Equation (1.3) can also be reduced to solve a two-dimensional flow problem
in the vertical plane (e.g., x,z axes only):
( ) ( )
∂ ∂h ∂ ∂h ∂h
Kx + Kz = Ss
∂x ∂x ∂z ∂z ∂t
( )
∂y = 0 which becomes true if qy = −Ky ∂y =
∂
Similarly, this implies that ∂y Ky ∂h ∂h

0 or ∂h/∂y = 0. A schematic diagram shows the conditions under which the

groundwater flow in aquifers can be modeled as two-dimensional or fully three-
dimensional (Figure 1.4).

1.4 2D Planeview Flow Equation (Confined &

Unconfined)
Groundwater flow in aquifers is often modeled as two-dimensional in the hori-
zontal plane. This is because most aquifers have large aspect ratio like a thin
pancake, with horizontal dimensions hundreds times greater than the vertical
thickness. In such a setting, groundwater more or less flows along the horizon-
tal plane, i.e., the z component of the velocity is comparatively small. There-
fore, a two-dimensional analysis is carried out in conjuncture with the use of
transmissivity, by assuming that h=h(x,y,t) only (h does not vary with z, thus
∂h/∂z = 0). This simplification of modeling 3D aquifer flow as horizonal two-
dimensional flow is called the Dupuit-Forchheimer approximation.
To picture what this simplification represents in a physical sense, imagine an
aquifer that is perforated by numerous tiny vertical tubes, each possessing an
infinitely large hydraulic conductivity (thus no resistance to flow in the vertical
direction). This leads to a constant hydraulic head long each vertical tube,
but laterally the head is still varying. Also, under this approximation, along
any vertical profile, the pressure is hydrostatic (we’ll prove this when working
on the homework). Figure 1.5 illustrates the difference between actual three-
dimensional flow and flow modeled with the Dupuit-Forchheimer approximation.
The general equation for the two-dimensional (horizontal) flow is developed
first for one-dimensional flow (qx ̸= 0, qy = 0), and then, it is extended to two-
dimensions. By assuming that the spatial gradient of water density is small
(i.e., ρw is mostly constant as in fresh water), a volume balance is done:
(1.7)
volume f lux in[L3 /T ]−volume f lux out [L3 /T ] = rate of change of volume[L3 /T ]

Consider a fixed control volume that is a vertical prism with cross-sectional

area ∆x × ∆y, extending the full saturated thickness of the aquifer (b) (Fig-
ure 1.6). By computing the discharge fluxes across the two sides of the prism as
well as computing another influx from recharge or leakage (N ([L/T]) is the net
1.4. 2D PLANEVIEW FLOW EQUATION (CONFINED & UNCONFINED)11

2D (x-z) Flow (qy=0)

y
z

river 2
river 1

3D Flow

y z
river

3D Flow To a Well
y
z

Figure 1.4: Planeview (x-y plane) of groundwater flow in aquifers with corre-
sponding cross-sectional view (x-z plane). The top panel illustrates the config-
uration where flow can be modeled as 2D (in this case, qy = 0); the bottom two
panels illustrates the configurations where 3D flow must be modeled (in these
cases, qx ̸= 0, qy ̸= 0, qz ̸= 0).
12 CHAPTER 1. GROUNDWATER FLOW EQUATIONS (REVIEW)

Figure 1.5: Cross-section of actual unconfined flow (left) and the same situa-
tion as modeled with the Dupuit-Forchheimer approximation (right). Hydraulic
head contours are dashed lines. In the Dupuit-Forchheimer model, there is no
resistance to vertical flow, resulting in constant head along the vertical lines
(∂h/∂z = 0).

flux of water entering the prism from the top and bottom), the net volume flux
across the prism boundaries is equated to the rate of change in volume stored
in the prism (which is related to the storativity (S) for a confined aquifer or the
specific yield (Sy ) for an unconfined aquifer). After some equation manipula-
tions and extending 1D flow to 2D (details given in class), we find:
( ) ( )
∂ ∂h ∂ ∂h ∂h
(1.8) Tx + Ty +N =S
∂x ∂x ∂y ∂y ∂t
This equation is the general equation for two-dimensional horizontal flow in
an aquifer (under the Dupuit-Forchheimer approximation), allowing for hetero-
geneity and anisotropy in transmissivity.
Similar to the 3D flow analysis of the last section, the above two-dimensional
general equation can be simplified under various assumptions. For example, if
transmissivity is homogeneous and isotropic (Tx = Ty = T = constant):
[ 2 ]
∂ h ∂2h ∂h
(1.9) T + +N =S
∂x2 ∂y 2 ∂t

This can be simplified further to ∇2 h + N/T = (S/T )∂h/∂t. If there is zero net
vertical recharge or leakage (N=0), this becomes:

∂2h ∂2h S ∂h
(1.10) ∇2 h = 2
+ 2 =
∂x ∂y T ∂t
If flow is steady-state but net vertical recharge/leakage is nonzero, this be-
comes:

(1.11) ∇2 h = −N/T

Equation (1.11) is also called the Poisson equation.

1.4. 2D PLANEVIEW FLOW EQUATION (CONFINED & UNCONFINED)13

top of aquifer (confined)

top of water table (unconfined)

b(x)

Qx(x) Qx(x+ x)

y
y base of aquifer
x x+ x

Figure 1.6: Prism control volume element of a two-dimensional aquifer.

When the recharge/leakage is zero, the Poisson equation further simplifies

to the Laplace equation:

(1.12) ∇2 h = 0

Unconfined Aquifer with Large Change in Saturated Thickness

For unconfined aquifer where its change in saturated thickness is small, the
previous equations derived for the confined aquifer is generally applicable (4.2
3D General Flow Equation). However, there are situations where the saturated
thickness of an unconfined aquifer experiences large variations. In these cases,
sometimes special formulations are obtained.
Flow in a homogeneous and isotropic unconfined aquifer underlain by a hori-
zonal impermeable base is a special case of the previous 2D analysis (4.3 2D
Planeview Flow Equation). If the base of the aquifer is considered the head
datum, then h = b, thus T = Kh, Equation (1.8) in this case becomes (details
given in class):
( ) ( )
∂ ∂h ∂ ∂h Sy ∂h
(1.13) h + h + N/K =
∂x ∂x ∂y ∂y K ∂t

If recharge and leakage can be ignored, the above relation becomes:

( ) ( )
∂ ∂h ∂ ∂h Sy ∂h
h + h =
∂x ∂x ∂y ∂y K ∂t
14 CHAPTER 1. GROUNDWATER FLOW EQUATIONS (REVIEW)

which is also called the Boussinesq equation.

The above two equations are nonlinear PDEs, since the terms involve h
multiplied by its derivative. Nonlinear equations are more difficult to solve than
linear ones. However, the nonlinearity disappears when writing these equations
for h2 . For example, equation (1.13) becomes (details given in class):
2N 2Sy ∂h
∇2 (h2 ) + =
K K ∂t
For steady flow this becomes the linear Poisson equation:
2N
(1.14) ∇2 (h2 ) = −
K
Further, if the recharge/leakge is zero, equation (1.14) becomes the linear
Laplace equation:

(1.15) ∇2 (h2 ) = 0

Note that in these two last equations, the unknown is the head squared h2 .
Chapter 2

Modeling Overview for

Groundwater Flow

Models of groundwater flow are widely used for a variety of purposes rang-
ing from water supply studies to designing contaminant cleanup. In general,
groundwater flow system can be divided into steady-state and transient. In
a steady-state flow system, the RHS of the flow equation is zero, while it is
nonzero in transient systems (as reviewed in the last chapter).1
There are dozens of free or commercial computer programs available to do
the calculations involved with these models. In this course, by working on
our own codes, we’ll understand the fundamental mathematical principles and
numerical methods behind these computer programs.

2.1 General Methodology

In general, mathematical modeling of groundwater flow and transport involves
the following 3 steps (Fitts, 2002):
• Review all the available data about the material properties (porosity, per-
meability, storativity), heads, and fluxes in the vicinity of the region to be
modeled.
• Develop a conceptual system that is simpler than the real flow system,
but captures the important overall features of the real system.
• Simulate the conceptual system developed in step (2) using a mathematical
model which is solved numerically with a computer code.
Usually developing a model requires several iterations through this process,
revisiting steps (1) and (2) in light of the results of the simulations developed
1 Solute movement in groundwater is generally always transient. A steady-state concentra-

tion field (the RHS of the transport equation is zero), though possible, is rarely solved. We
will defer the solution of the transient transport equations to Chapter 8.

15
16 CHAPTER 2. MODELING OVERVIEW FOR GROUNDWATER FLOW

in step (3). During this iteration, heads and fluxes predicted by the computer
model are compared to actual measurements of heads and fluxes observed in
the natural groundwater system. If significant discrepancies exist, the model
parameters (e.g., porosity, permeability, storativity) and boundary conditions
(specified head, specified flux) will be adjusted. The simulation is carried out
again until a suitable (and usually rather subjective) fit between predictions and
observations is achieved.
In the third step outlined above, the conceptual system should be examined
to determine which general flow equation applies. There are several different
general equations, each valid under certain conditions (as reviewed in the last
Chapter). Which one applies depends on circumstances such as:

1. Should the flow be represented as one-, two-, or three-dimensional?

2. Can the flow be approximated as steady state or should it be represented

as transient?

3. Is the flow in a confined or an unconfined aquifer?

For a given general equation, there is an infinite number of possible

solutions. For steady-state flow, the unique and appropriate solution
is one that matches the particular boundary conditions (BC) of the
conceptual model. (For transient flow system and for solute trans-
port, both initial condition and BC are required to obtain the unique
solutions of head and concentrations.)
Boundary conditions include things like heads at surface waters in contact
with the aquifer, the location and discharge rate of a pumping well or a leach-
ing irrigation field. For the no-flow boundary (which is a special type of flux
boundary with specified flux of zero), we need to investigate the involvement
of aquitard for the spatial/temporal scale of interest: no-flow boundary (small
scale engineering problems) versus low-K zone as part of model domain (larger
regional flow problems)?
The two most common types of BC are specified head and specified
flux conditions, as illustrated in Figure 2.1. At specified head boundaries, the
head is known. These are usually where the groundwater is in direct contact
with a surface water like a lake or a river. There are several types of specified
flux conditions, including impermeable boundaries that allow zero flux, recharge
boundaries at the top of the saturated zone, and wells that are pumped at a
known rate. The less common, so-called mixed boundary condition involves
some combination of head and flux specification2 .
Along a particular stretch of boundary, only one condition can be
specified. For example, it is not possible to specify both the head and the flux
along the same portion of a boundary (in class, we’ll show how one type of BC
2 For example, leakage through a silty river bed to an underlying aquifer is represented by

a flux that is proportional to the vertical conductivity of the silt layer and proportional to the
head difference from the river to the underlying aquifer (hint: write a vertical Darcy’s Law
for the silt layer)
2.1. GENERAL METHODOLOGY 17

Pumping
well
Recharge
Recharge
Silt layer below
river bottom

Sand aquifer

Low-K bedrock

Figure 2.1: Examples of flow domain boundary conditions. The pumping well
is a specified flux condition at the permeable section of the well. The recharge
is a specified flux applied at the water table. The low-K bedrock is considered a
special specified flux boundary (no-flow boundary). The leaky silt layer below
the river is a mixed condition where the flux through the layer is proportional to
the difference between the head in the river and the head in the aquifer beneath
the silt layer. Where the river is in direct contact with the aquifer, there is a
specified head condition.

can be translated to another type). You may specify either, but not both. Once
the general equation is known and the boundary conditions are assessed, review
the available modeling techniques and select one that can simulate the general
equation and boundary conditions of the problem at hand. Construct the model
and adjust its parameters and/or boundary conditions as necessary to fit the
observations of the real system. The parameters that are input (conductivity,
storativity, etc.) should be within the range of measured values, or if lacking
measurements, within the range of expected values for the geologic materials
present (Figure 2.2). Also the discharge fluxes that are modeled should be
reasonable. For example, the discharge to a river segment should be similar to
the measured or expected baseflow for that river segment.
Computer programs are now widely used to develop complex two- and three-
dimensional models. These programs implement numerical (approximate) or an-
alytic (exact) solutions to the general flow equations, and allow solutions with
diverse and irregular boundary conditions. This class focuses on the numerical
solutions, i.e., the Finite Difference Method (FDM), which solve the flow or
transport equations at discrete grid points, for the unknown hydraulic head,
groundwater fluxes or solute concentrations. (For transient problems, both ini-
tial and boundary conditions are needed, the flow and transport equations are
solved at both discrete grid points and discrete times.)
On the other hand, if exists, the analytical solutions predict these values
exactly at every continuous point in space (if steady-state flow) and continuous
point in time (if transient flow). For the same problem, the numerical solution
will generally converge to the analytical solution (if exists) when the numerical
model grid employed is dense (i.e., the spacing between model grid cells is small).
18 CHAPTER 2. MODELING OVERVIEW FOR GROUNDWATER FLOW

Hydraulic Conductivity (K) and Permeability (k) Range

for Unconsolidated Deposits and Rocks

Unconsolidated
Rocks Deposits
k (m2) K (m/s)
10-7 1

10-8 10-1

gravel
10-9 10-2

10-10 10-3
permeable basalt
karst limestone

clean sand

10-11 10-4
silty sand

10-12 10-5
metamorphic &
igneous rocks

10-13 10-6
fractured

limestone &

silt, loess

10-14 10-7
dolomite
sandstone

10-15 10-8
glacial till

10-16 10-9

10-17 10-10
marine clay

10-18 10-11
shale
metamorphic &
igneous rocks
unfractured

10-19 10-12

10-20 10-13

Figure 2.2: General range of hydraulic conductivity [L/T] and permeability

[L2 ] for different consolidated and unconsolidated geological materials (Freeze
& Cherry, 1979).
2.2. ANALYTICAL VERSUS NUMERICAL METHODS 19

As discussed by Fitts (2002), some groundwater models are very compli-

cated that require months of labor to create and adjust. At the other end of the
spectrum, simple models can be created in a matter of minutes. There still is
a role for hand calculations though (particularly those we’ve done in the Geo-
hydrology class), as it is often reasonable to neglect components of flow in one
or two directions, and use a 1D or 2D analysis of a simple conceptual model.
In these cases, a modeler can perform simple calculations, quickly. Hand cal-
culations can also teaches the modeler useful insights that can apply in general
ways to other, more complex situations. In Geohydrology, I introduced several
practical methods for analyzing 1D and 2D steady-state flow (Chapter 7) as well
as the commonly used methods for analyzing transient flow and pumping tests
(Chapter 8).

2.2 Analytical versus Numerical Methods

Mathematical models of groundwater flow and solute transport can be solved
generally with two broad approaches:

1. Analytical solution of the mathematical equation gives exact solution to

the problem, i.e., the unknown variable is solved continuously for every
point in space (steady-state flow) and time (transient flow).

2. Numerical solution of the mathematical equation gives approximate solu-

tion to the problem, i.e., the unknown variable is solved at discrete points
in space (steady-state flow) and time (transient flow).

For example, shown in Figure 2.3a is a regional confined aquifer: along the
top boundary of this aquifer, the hydraulic head is described by the potentiomet-
ric surface which is higher than the aquifer top. Within the aquifer, groundwater
flow forms several regional flow cells. In particular, the one near the middle is
the model domain of interest. To understand the hydraulic head and streamline
distribution in this region, mathematical model needs to be built and solved.
Toth (1962) derived an analytical solution for the hydraulic head in the
solution domain (Figure 2.3b):
∞
cL 4cL ∑ cos[(2m + 1)πx/L] cosh[(2m + 1)πz/L]
h(x, z) = h0 + − 2
2 π m=0 (2m + 1)2 cosh[(2m + 1)πh0 /L]

To obtain this solution, Toth (1962) made several simplifying assumptions. For
example, the aquifer is assumed homogeneous and isotropic for which the hy-
draulic conductivity is a scalar constant value. The flow is steady-state. Under
these assumptions, the flow equation becomes the Laplace equation (see Fig-
ure 2.3b). The potentiometric surface of the confined aquifer is assumed linear.
For this 2D problem, the hydraulic head of the top aquifer boundary is thus
linear: h(x, H) = cx + h0 . Finally, the aquifer domain is assumed to be a rect-
angle (in reality, real aquifer thickness usually varies). Note that the advantage
20 CHAPTER 2. MODELING OVERVIEW FOR GROUNDWATER FLOW

A Regional Confined Aquifer

hydraulic head (in meter) land s

urface
streamlines
potentiom
etric surf
piezometer (top aqu ace
ifer regio
perforation n only)
Confining Bed

1000 1200
1100

groundwater divide
groundwater divide

1150

Confined
Aquifer

Impermeable Base

sea level (head datum)

Model Domain & BC

z cL
h0
H
h=cx+h0

qx=0 2h 2h
qx=0
x2 + y2 =0

L
x
0 qz=0
Figure 2.3: A regional confined aquifer: (a) a geological transect of the aquifer
and its confining units; (b) the model domain and boundary conditions. Note
that the region to be modeled has a length of L and height of H, corresponding
to the “Solution Domain” shown in (a).
2.2. ANALYTICAL VERSUS NUMERICAL METHODS 21

of the analytical solution is that it is an exact solution for the above problem
under the various assumptions. A solution also exists for every point (x, z) in
space. The disadvantage is the various assumptions as required to obtain the
exact analytical solution. None of these assumptions is necessary to obtain a
numerical (approximate) solution.
For many problems, the assumptions that must be made to obtain an ana-
lytical solution will not be realistic. In these cases, to solve the mathematical
models, we must resort to approximate methods using numerical solution tech-
niques. In general, three main numerical methods are used in hydrogeology
(listed in the order of popularity): (1) Finite Difference Method (FDM); (2) Fi-
nite Element Method (FEM); (3) Boundary Element Method (AEM). The first
two methods are the most popular. They transforms the differential equation
that governs flow (or transport) into a set of algebraic equations.
Before computers became widely available, only hand calculations (develop-
ing analytical solutions for highly simplified and idealized problems) or analogue
studies were possible3 , and the numerical techniques (FDM, FEM, AEM) were
of limited value. Using computers, we can solve large number of algebraic equa-
tions by iterative methods or direct matrix-based methods (next chapter will
introduce these methods). We can test the numerical solution by comparing that
head (or solute concentration) distribution computed by the computer with that
determined from an analytical solution, if one is available. We can also check
the simulated head (or solute concentration) against the values observed in the
field. This procedure is summarized in Figure 1.
The mathematical formulations for the FDM and FEM are quite different:
the former is based on the Taylor’s Series Expansion; the later is based on the
Variational Principle which involves more advanced mathematics. In this class,
due to time limitation and the popular demand to use a software package (e.g.,
MODFLOW is based on FDM), only the FDM is introduced. The FDM has
certain advantage over FEM, e.g., it provides close approximation of mass conti-
nuity across grid cell boundaries (I’ll draw a plot on board to demonstrate what
this means). The traditional FEM’s approximation of mass continuity is weak
at the local cell boundaries. However, new approaches are continuously imple-
mented to address the FEM local mass balance problems. The hybrid Finite
Volume Method is becoming increasingly popular in the research community
since is combines the flexibility of FEM in grid cell shape (Figure 2.5) with the
local mass conservativeness of the FDM. However, FDM and FEM have simi-
larities: they are both grid-based. Compared to the analytical solution which
exists at any point in the solution domain, the numerical methods yield values
3 Before computers were commonplace and inexpensive, complex groundwater flow prob-

lems were modeled using physical models or analogs. These models would typically be a
miniature scaled model of the flow domain in a tank. The most common analog method uses
the flow of electricity through a network of wires and resistors, where voltage is analogous
to head, resistance is analogous to 1/K or 1/T, current is analogous to discharge, and ca-
pacitance is analogous to storage. However, computer simulation methods have essentially
replaced physical and analog modeling, so these methods are not discussed in this course.
Interested students can find more coverage of analog methods in Walton (1970) Groundwater
Resource Evaluation, McGraw-Hill, New York.
22 CHAPTER 2. MODELING OVERVIEW FOR GROUNDWATER FLOW

for only a fixed finite number of grid points (N ) in the solution domain. Us-
ing either FDM or FEM, we convert a partial differential equations to a set of
N algebraic equations involving N unknown values at these particular points.
However, it is important to note that there are advantages of using FEM to
solve the flow/transport problems. The most important one is its flexibility to
simulate complex internal geometry and irregular domain boundary since the
grid cells of the FEM can be of any shape. Depending on student demands, the
FEM can be taught in future, in an advanced class or seminar in hydrogeology.
For example, in Figure 2.5a, an aquifer is shown in map view, with the
locations of a well field, observation wells, and aquifer boundary. In Figure 2.5b,
c, d, three different grids are used to represent the model domain to solve
for either head (if only solving the flow equation) and concentration (if both
flow and solute equations are solved). We can immediately see that the FDM
can only give an approximate representation of the irregular boundary, while
the FEM can honor irregular boundary more realistically. Also, two different
grids are used in FDM: block-centered and mesh-centered. In this class, to
simplify the derivations, we use the block-centered grid to derive the
FD approximations of the flow and transport equations. Note that
in hydrogeology, MODFLOW, the most popular groundwater/transport code,
is based on block-centered FDM. Also, block-centered FDM is the preferred
method for most general purpose petroleum reservoir simulators (Lee et al.,
1998; Ertekin et al., 2001). The derivations for the mesh-centered grid are
very similar (only minor details vary): with the understanding for the former
method, the later one is easy to learn. Moreover, the grid spacing (∆x, ∆y) are
also of interest. The smaller we make ∆x and ∆y, the numerical approximate
solution converges to the analytical solution. The trade-off is that the number of
unknowns (N ) increases. For very large problems, e.g., radionuclide transport
at regional scale, oil reservoir simulations, the limitation on grid size is often set
by the computer code and memory. We generally select a N value that will not
be overwhelmingly large (exceeding the computation limit), but large enough
that most of the main hydrogeological features are represented.
Finally, a variety of solution techniques can be used to solve a set of linear
algebraic equations (separate classes on Numerical Linear Algebra exist, possi-
bly from the Math Department, which should cover these techniques in a more
comprehensive fashion). Based on my own past experiences, I’ve summarized
the ones often used in hydrogeology (Figure 2.4). In this class, due to time
limitation, we’ll use only some of the solution techniques. For example, though
we introduce the matrix-free iterative methods, due to their computation ineffi-
ciency, they’re no longer used very much for solving large systems of equations
which derive from large models. So, we’ll not reply on them too much. On the
other hand, the matrix-based iterative solver will only be mentioned in pass-
ing, as setting up these solvers will generally require access to a Unix cluster
on which the various libraries are usually installed (currently our PC labs do
not have these libraries). Working on Unix systems require knowledge on Unix
(e.g., compile and link to external libraries), which is beyond the purpose of this
course. Therefore, the most often used solution technique in this class is the
2.3. VALIDITY OF NUMERICAL SOLUTION 23

Full Matrix
Storage
(GaussElim.m)
Direct Solution
(Gaussian Banded Stor-
Matrix Elimination) age (Gentri.m)
Based
DGMRES from Half Band Stor-
IMSL (Matrix stored age (etc.)
as i,j,value)
Iterative Solu-
A Set of Linear
tion
Algebraic Many other solvers
Equations (e.g., LaPACK,
LinPACK, IMSL,
AZTEC, PETSc)
Matrix Free
(Iterative
Solution) Gauss-
Point Jacobi Seidel SOR

Figure 2.4: Commonly used linear solution techniques in hydrogeology and their
relations.

matrix-based direct solver (the various forms of Gaussian Elimination solvers,

tailed to the specific storage styles of the global coefficient matrix, more on this
later).

2.3 Validity of Numerical Solution

In solving simple examples, we can check the validity of the numerical solution
by comparing the numbers generated using the numerical solution with those
calculated from an analytical solution. However, analytical solutions are not
available for many problems of practical interest. Numerical methods allow us
to solve the governing equation in more than one dimension, for complex bound-
ary conditions and for heterogeneous and anisotropic aquifers, whereas most
analytical solutions are restricted to consideration of homogeneous, isotropic
aquifers.
An important point which is sometimes overlooked is the neces-
sity to verify the validity of every numerical solution. Therefore, sev-
eral checks on the correctness of the solution should be made routinely when
an analytical solution is not available. Specifically, it is a good practice to
check the sensitivity of the solution to (1) the choice of error tolerance when
matrix-free numerical techniques are used to solve the governing equations, e.g.,
Point-Jacobi, Gauss-Seidel and SOR (these are most suitable for simple domain
geometry and small to medium-sized grids); (2) nodal spacing. Note that both
matrix-free and matrix-based 4 numerical techniques are based on a grid. So
this sensitivity test applies to both groups. Moreover, it is desirable to perform
a mass balance calculation.
4 In the matrix-based numerical techniques, the coefficient matrix and right hand side vector

are assembled; their values sent to a direct Gaussian-elimination-based solver or an iterative

solver.
24 CHAPTER 2. MODELING OVERVIEW FOR GROUNDWATER FLOW

If a matrix-free technique is employed to solve the flow equation, as a general

rule, several computer runs should be made using successively smaller values
for the error tolerance until a small enough value is selected and the solution
does not change with the further decrease in error tolerance. In all numerical
techniques, the solution may also be improved through the use of smaller nodal
spacing. Several computer runs should be made with successively smaller nodal
spacing until the solution does not change with further decrease in nodal spacing.
However, it is generally a time-consuming process to modify the finite difference
grid and redefine the input. Instead, most modelers rely on the third type of
validity check—the mass balance calculation. A mass balance calculation is an
expression of the fact that at steady state, the amount of water entering
the system equals the amount leaving the system. If inflow does not
equal outflow for a steady-state computer simulation, there may be something
wrong with the numerical solution. However, a large error in the mass balance
may also indicate a lack of precision in the formula of mass balance calculation
itself.

2.4 Model Calibration and Parameter Estima-

tion
Model calibration is the process of adjusting the input parameters (hydraulic
conductivity, porosity, specific storage, well or natural recharge/discharge rate)
and boundary conditions (head or Darcy flux along sections of the boundary) of
a model to achieve a close fit to observed data (e.g., hydraulic head, flow rate)
in a real groundwater system. In flow model calibration, simulated heads and
Darcy fluxes are typically compared to their observed counterparts. If a model
is well calibrated, there will be some random deviations between simulated
and observed data, but there will not be systematic deviations. If there are
systematic deviations such as if most of the simulated heads exceed the observed
heads, the calibration is poor and adjustments should be made.
In thinking about how to calibrate a flow model, it helps to go back to the
basics — the Darcy’s law (here it is written in its 1D form along the x axis):
∂h
qx = −Kx
∂x
If the head gradients predicted by a model are too large, the relations in Darcy’s
law indicate that either the modeled fluxes are too large, or the modeled con-
ductivities are too low, or both.
Consider an example where the simulated heads in a regional flow model are
systematically higher than the observed heads, as shown in Figure 2.6 (Fitts,
2002). The aquifer is unconfined (so its heads correspond to the water table),
and water enters the aquifer as recharge and leaves as discharge to local streams.
Assuming that the constant heads assigned at the streams are correct, there are
two adjustments to the model that could be made to eliminate the systematic
error in the heads:
 2.4. MODEL CALIBRATION AND PARAMETER ESTIMATION
25

Figure 2.5: Finite difference and finite element representations of an aquifer region (from Wang & Anderson, 1995). (a) Map
view of aquifer showing a well field, observation wells, and aquifer boundary. (b) Finite difference grid with block-centered
nodes, where ∆x is the spacing in x direction, ∆y is the spacing in y direction, and b is the aquifer thickness at a particular
block (this thickness can vary in space). (c) Finite difference grid with mesh-centered nodes. (d) Finite element mesh with
triangular elements.
26 CHAPTER 2. MODELING OVERVIEW FOR GROUNDWATER FLOW

Fixed heads
Fixed heads

Stream -
Stream -
R

Simula
ted He Observed H
ads eads

h Unconfined
Aquifer

Impermeable Aquifer Base

Figure 2.6: Vertical cross section comparing modeled heads with observed heads.

1. Increase hydraulic conductivities in the aquifer, or

2. Decrease the rate of recharge applied.

If the fluxes to streams are known to be correct (the model predicted fluxes
to streams are equal to the observed fluxes within a small random error), then
the conductivities should be increased. Why this is so can be seen from Darcy’s
law; increasing K allows the same flux of water to be transmitted with smaller
head gradients. Thus, if we increase the value of K, to provide the same (correct)
Darcy flux to the streams, the head gradient will drop, thus the simulated head
will be lower than the current estimation and the simulated head gradient will
become smaller.
On the other hand, if the conductivities specified to the model aquifer are
known to be correct (for example, we have lots of detailed measurements of K
in the aquifer to back up this claim) and the predicted fluxes to streams are too
high compared to observed fluxes, the recharge rate should be lowered. Note
that smaller recharge rate specified to the model will allow lesser amount of
water to enter the aquifer for a fixed time period, so the predicted head mound
in between the steams will be lower.
The above calibration procedure involves a trial-and-error approach by the
modeler involving both the hard hydrogeological principles (relationship be-
tween head gradient, fluxes and conductivity) and soft intuition (which parame-
ter might be the most uncertain thus requiring extensive adjustment). However,
a recent development in flow modeling is the automated estimation of parame-
ters by special computer algorithms that will optimize the calibration of models
2.5. INTERPRETING MODEL RESULTS 27

(Hill, 1992; Doherty, 2000; Hill & Tiedeman, 2007). These techniques are based
on minimizing an objective function which is defined to be a measure of the
fit between model results and actual observations. The larger the computed
objective function is, the greater the discrepancy between the model results
and corresponding real observations. One simple and common definition of the
objective function (F ) is the sum of the squared residuals (differences between
modeled and observed heads):
∑
n
( obs )2
F = hi − hsim
i
i=1

where n is the number of observed heads, hobs i is the i th observed head and
hsim
i is the modeled head corresponding to the i th observed head. The above
equation can also be modified by multiplying weighting factors times each term
in the sum. The weighting factor depends on the importance of an observation to
the overall model validity and the confidence in that observation (“True Error”
defined in Chapter 9). Automated calibration techniques will find the optimal
set of parameter values that result in a minimal value of the objective function.
Such techniques can save a modeler some time in the calibration process, but
they are no substitute for careful thinking. The automated techniques can
yield unreasonable results if insufficient constraints are supplied. Chapter 9
will discuss the methods of parameter estimation, based on both the forward
upscaling theory and the inverse regression theory.

2.5 Interpreting Model Results

When interpreting flow model results, it is important to bear in mind the sim-
plifying assumptions that went into creating the conceptual model. Although
a well-calibrated mathematical model may simulate the conceptual model with
great accuracy and aesthetic appeal, it is likely to be only a crude representa-
tion of the real flow system. In some cases, there is no unique solution to the
calibration problem; several different combinations of input parameters will re-
sult in models that all equally fit the observations. Problems of non-uniqueness
are more prevalent when the observation data are limited. For example, in a
model where there are no known fluxes such as well rates or recharge rates, and
the only known observations are heads, there will be a range of recharge and
hydraulic conductivity combinations that can result in a similar calibration. As
long as the hydraulic conductivities (or transmissivities) and fluxes are increased
or decreased by the same proportion, the pattern of heads will remain the same.
The general equation for steady two-dimensional flow with recharge is shown
below to show why this is so: (please review GEOL4444/5444, Geohydrology)

∇2 h = −N/T

where N is the recharge rate, T is aquifer transmissivity. It is clear from

the above that the same pattern of h could be achieved with any number
28 CHAPTER 2. MODELING OVERVIEW FOR GROUNDWATER FLOW

of models that maintain a fixed ratio N/T . Therefore, for many modeling
projects, it is important to have some flux measurements, e.g., well rates, aquifer
recharge/discharge rates to lakes or streams, or precipitation-induced recharge
rate to the aquifer.
After a model is calibrated to observed existing conditions in the flow sys-
tem, it is then used to predict future conditions with a different set of boundary
conditions. For example, models are often calibrated to the known flow con-
ditions at a contaminated site prior to clean-up and then used to simulate the
response of the system to various proposed remediation designs including the
addition of pumping wells, barriers and/or drains. In the geological community,
there are also scientific studies to evaluate the vulnerability of regional aquifers
under future change in climate. For example, the Global Climate Models predict
that under Global Warming, the western US may become dryer (Levi, 2008).
This would mean less recharge water will reach the regional aquifers in this re-
gion which are sources of freshwater to many communities for both domestic,
industrial and agricultural uses. Then, if the west were to become dryer, there
will be less recharge water to the aquifer to balance the (possibly increasing)
water demands due to the increased population. We need a computer model to
predict the head decline under these hypothesized scenarios.

2.6 Model Limitation and Uncertainty

However, actual groundwater flow systems are much more complex than the
conceptual models can typically represent. The accuracy of predictive simula-
tions is thus difficult to assess, so it’s wise to assume a fair amount of uncertainty
when using models to make predictions. In particular, the subsurface has com-
plex distributions of materials with transient groundwater fluxes. No matter
how much effort is spent drilling, sampling, and testing the subsurface, only a
small fraction of it is ever sampled or tested hydraulically. The available data
will provide only an incomplete picture of the actual subsurface system. Be-
cause of the inherent difficulty of characterizing subsurface regions, substantial
uncertainty is always introduced in the conceptual model created.
Typically, the complex distribution of subsurface materials is represented
in the conceptual model as regions with locally homogeneous and isotropic hy-
draulic conductivity (anisotropic representation is also common, see Zhang et al
(2006), Zhang & Gable (2008)). The parameters assigned to these regions are
chosen to represent the large-scale average (or effective) hydraulic behaviors of
flow and transport. Complex transient fluxes like recharge or pumping rates are
represented in the conceptual model as either steady-state average values or as
transient rates that change in some simplified manner.
In a well-constructed mathematical model, most of the uncertainty in the
results stems from discrepancies between the real system and the conceptual
system. Most mathematical models provide a fairly accurate simulation of the
conceptual system. Therein lies the danger. Accurate simulation of the con-
ceptual system is often taken to mean accurate simulation of the real system.
2.7. REFERENCES 29

With sophisticated model-generated graphics, there is a tendency to forget the

unavoidable uncertainties in representing the real system with a simpler con-
ceptual model. Regardless of these limitations, models are usually the best way
to develop judgment when solving quantitative groundwater flow problems.

2.7 References
Toth J, 1962, A theory of groundwater motion in small drainage basins in Center
Alberta, Canada, Journal of Geophysical Research, Vol. 67, p. 4375-4387.
Lee S H, L J Durlofsky, M F Lough and W H Chen (1998) Finite differ-
ence simulation of geologically complex reservoirs with tensor permeabilities,
SPE Reservoir Evaluation and Engineering, December, p. 567-574.
Ertekin T, J H Abou-Kassem, G R King (2001) Basic Applied Reservoir Simulation,
SPE Textbook Series, 406 p.
Hill M C (1992) A computer program (MODFLOWP) for estimating param-
eters of a transient, three-dimensional ground-water flow model using nonlinear
regression, USGS Water Resources Investigation Report, 91-484.
Doherty J (2000) Manual for PEST2000, Watermark Numerical Computing,
Brisbane, Australia.
Levi B G (2008) Trends in the hydrology of the western US bear the imprint
of manmade climate change, Physics Today, p. 16-18, April 2008.
Ye Zhang, Carl W Gable, Mark Person, 2006, Equivalent Hydraulic Conduc-
tivity of an Experimental Stratigraphy - Implications for Basin-Scale Flow Simu-
lations, Vol. 42, W05404, Water Resources Research, doi:10.1029/2005WR004720.
Ye Zhang, Carl W Gable, 2008, Two-Scale Modeling of Solute Transport in
an Experimental Stratigraphy, Journal of Hydrology, Vol. 348, p. 395-411, doi:
10.1016/j.jhydrol.2007.10.017.
30 CHAPTER 2. MODELING OVERVIEW FOR GROUNDWATER FLOW
Chapter 3

Mathematical Notations,
Definitions and Theorems

This chapter introduces the basic mathematical notations, definitions and the-
orems. It supplements the handout of Basic Math Review given earlier but
focuses more specifically on the type of problems we’ll encounter in this course.

3.1 Differential Equations

Differential Equations The mathematical formulations of most flow and trans-
port problems involve the rates of change of a dependent variable (e.g., hydraulic
head, temperature, or concentration) with respect to one or more independent
variables (e.g., time, distance, angle). A mathematical equation that contains
a dependent variable and one or more of its derivatives is referred to as a Dif-
ferential Equation. The order of a differential equation is defined by its highest
derivative.
Ordinary Differential Equation (ODE): a type of differential equation that
contains only 1 independent variable (e.g., time or one spatial axis). For exam-
ple, a first-order ODE is:
dc
= −kc
dt
The solution is c = c(t), the one independent variable is time t. Note that for
ODE, the differential symbol used is d. By definition, the ordinary derivative:
dc/dt = lim∆t→0 [c(t + ∆t) − c(t)]/∆t.
Partial Differential Equation (PDE): a type of differential equation that con-
tains more than 1 independent variables (e.g., several spatial axes, time and
several spatial axes). For example, a second-order PDE is:

∂T ∂2T
(ρc) =λ 2
∂t ∂z

31
32CHAPTER 3. MATHEMATICAL NOTATIONS, DEFINITIONS AND THEOREMS

The solution is T = T (t, z), the two independent variables are time t and the
z axis. Note that for PDE, the differential symbol used is ∂ (partial derivative
symbol). By definition, the partial derivative: ∂T /∂t = lim∆t→0 [T (t + ∆t, z) −
T (t, z)]/∆t.

Initial and Boundary Conditions:

• Initial Condition: Constrains that are specified at the initial point, gen-
erally time point, are called initial conditions. Problems with specified
initial conditions are called initial value problems.

• Boundary Condition: Constrains that are specified at the boundary points,

generally space points, are called boundary conditions. Problems with
specified boundary conditions are called boundary value problems.

ODE: Separation of Variables:

Given a first-order ordinary differential equation

dy
(3.1) = f (x, y)
dx

If f (x, y) can be expressed using separation of variables as

f (x, y) = p(x)q(y)

then the equation can be expressed as

dy
= p(x)dx
q(y)

and equation (3.1) can be solved by integrating both sides to obtain:

∫ ∫
dy
= p(x)dx
q(y)

Any first-order ODE of the form

dy
(3.2) + p(x)y = q(x)
dx

has a solution (details given in class):

∫ ∫
p(x)dx
e q(x)dx +C
(3.3) y(x) = ∫
e p(x)dx
where C is a constant of integration.
3.2. SCALAR, VECTOR, TENSOR, FIELD 33

3.2 Scalar, Vector, Tensor, Field

Scalar: a numerical quantity which is defined by its magnitude alone. Examples
include concentration, temperature, pressure, hydraulic head.

Vector: a numerical quantity having both magnitude and direction. Exam-

ples include the Darcy flux vector, the average linear velocity vector. In 3D, the
Darcy flux has 3 components, e.g., qx is the normal projection of ⃗q along the x
axis.  
 qx 
⃗q = qy
 
qz

Magnitude of the vector above is:

√
|⃗q| = qx2 + qy 2 + qz 2

Tensor: a matrix quantity which relates one vector to another. One exam-
ple is the hydraulic conductivity tensor (K) which relates the hydraulic head
⃗ with Darcy flux (⃗q): ⃗q = −KI,
gradient (I) ⃗ or written out in full:
    
 qx  Kxx Kxy Kxz  ∂h/∂x 
qy = −  Kyx Kyy Kyz  ∂h/∂y
   
qz Kzx Kzy Kzz ∂h/∂z

This equation is also known as the general 3D Darcy’s law (where the coordinate
axes do not aligned with the principle components of the hydraulic conductivity
tensor).
The above equation can also be written out explicitly using the rule of
matrix-vector multiplication, one for each Darcy flux component,

∂h ∂h ∂h
qx = −Kxx − Kxy − Kxz
∂x ∂y ∂z

Field

A scalar field is created by simply assigning scalar quantities (numbers) to

each point in space. Think of temperature – each point in the room has a
different temperature.
A vector field is created by assigning vectors to each point in space. A flow
field is an example groundwater at each point in space has a certain speed and
direction as represented by the Darcy flux vector at that point.
A tensor field has a tensor corresponding to each point in space. An example
is the hydraulic conductivity of an aquifer: at each point in space there is a K
that relates the head gradient with Darcy flux at that point.
34CHAPTER 3. MATHEMATICAL NOTATIONS, DEFINITIONS AND THEOREMS

h(x)
h(x+∆x)
h(x)

∆x
x
x x+∆x

Figure 3.1: 1D Taylor series expansion along the x axis.

3.3 Taylor Series and Finite Difference Approx-

imations
Taylor Series Expansion: approximation of a function h at a point h(x + ∆x)
using derivatives of the function at h(x) (Figure 3.1):
• Forward Approximation:
∆x ∂h(x) ∆x2 ∂ 2 h(x) ∆x3 ∂ 3 h(x)
h(x + ∆x) = h(x) + + + + ...
1! ∂x 2! ∂x2 3! ∂x3
n
∆xn ∂ h(x)
If we truncate the higher order terms (i.e., n! ∂xn , n > 2), we get:
∂h(x)
h(x + ∆x) ≈ h(x) + ∆x ⇒
∂x

∂h(x) h(x + ∆x) − h(x)

(3.4) ≈
∂x ∆x
• Backward Approximation:
∆x ∂h(x) ∆x2 ∂ 2 h(x) ∆x3 ∂ 3 h(x)
h(x − ∆x) = h(x) − + − + ...
1! ∂x 2! ∂x2 3! ∂x3
If we truncate the higher order terms (n = 2), we get:
∂h(x)
h(x − ∆x) ≈ h(x) − ∆x ⇒
∂x

∂h(x) h(x) − h(x − ∆x)

(3.5) ≈
∂x ∆x
3.4. NUMERICAL ERROR, CONVERGENCE, STABILITY 35

If the function contains more than one independent variables, we will need
to fix one independent variable and evaluate the function in relation to the
other independent variable. For example, h = h(x, t), we can write the forward
approximation as:
• Forward Approximation:
∆x ∂h(x, t) ∆x2 ∂ 2 h(x, t) ∆x3 ∂ 3 h(x, t)
h(x+∆x, t) = h(x, t)+ + + +. . . |f ixt
1! ∂x 2! ∂x2 3! ∂x3
n
∆xn ∂ h(x,t)
If we truncate the higher order terms (i.e., n! ∂xn , n > 2), we get:
∂h(x, t)
h(x + ∆x, t) ≈ h(x, t) + ∆x ⇒
∂x

∂h(x, t) h(x + ∆x, t) − h(x, t)

(3.6) ≈ |f ixt
∂x ∆x
The Taylor Series is key to understanding the finite difference method. We’ll
see that by truncating the higher order terms in the Taylor Series, we can
approximate any ordinary or partial derivative in a ODE or PDE.
Integration by Parts (1D)1 :
∫ b ∫ b
udv = uv|ba − vdu
a a

∫b
Exercise 1 Use integration by parts, evaluate a
xex dx.

Exercise 2 Evaluate the Taylor Series expansion for h(x) = 3x + 2 + 9x2 at

x = 1 for truncation error of order ∆x2 and ∆x3 , with ∆x = 1. Hint: x0 = 0,
use Taylor Series expansion for h(x0 + ∆x) = h(0 + 1) = h(1). Repeat the
analysis for ∆x = 0.1: now h is evaluated at x = 0 + 0.1 = 0.1.

3.4 Numerical Error, Convergence, Stability

Numerical Error Two types of errors occur in computational science:
• Truncation Error— due to truncation of higher order terms in the Tay-
lor Series when approximating a derivative. In the previous forward or
backward approximation, the truncation error → 0 when ∆x → 0. For
n ∂ n h(x)
the above approximation where ∆x n! ∂xn (n > 2) terms are truncated,
we say that the truncation error is of order ∆x2 .
∫b ∫b ∫b ∫b ∫b
1 d(uv) = udv + vdu → a d(uv) = a udv + a vdu → uv|ba = a udv + a vdu
36CHAPTER 3. MATHEMATICAL NOTATIONS, DEFINITIONS AND THEOREMS

• Roundoff Error — due to the finite arithmetics of conducting math-

ematical operations using a computer. For example, when computing
x = 1/2 = 0.5, the memory of the computer where it is designated to
store a “real” number is 4 bytes (=4x8=32 bits). This size is sufficient
to hold the value 0.5 without loss of accuracy. However, when computing
y = 1/3 = 0.333 . . ., the memory of the computer where it is designated
to store a “real” number is also 4 bytes. This size is insufficient to hold
the value 0.33 . . .: instead 0.33333333 is stored, so all digits after the 8th
get rounded off. Thus, when we do an operation like z = y × 3, we get
z = 0.99999999 instead of the mathematically correct value of 1.0. This
type of error is called roundoff error and can either cancel out or accumu-
late during numerical operations done by a computer.

The above numerical errors influence whether a numerical approximation of

a differential equation converges and is stable:
• Convergence — a numerical approximation is said to converge if the
numerical solution approaches the analytical solution as ∆x approaches
zero.
• Stability (Only applicable to initial-value problems) — if the numerical
d h(x) is the exact analytical value of h at x, h(x)
error, e(x) = h(x)− h(x); d
is the numerical approximation of h(x) at x (x is a position vector), does
not grow through sequential calculations of the solution through time,
then the numerical approximation is said to be stable. If the error grows
without bound, then the numerical approximation is unstable.

3.5 Matrix & Vector

Consider the following matrix A consisting of n rows and m columns:
 
a11 a12 . . . a1m
 a21 a22 . . . a2m 
An×m =  ... ...

... ... 
an1 an2 . . . anm
Each entry in the matrix is called an element.
If n = m, the matrix of n × n is considered a square matrix:
 
a11 a12 . . . a1n
 a21 a22 . . . a2n 
An×n =   ... ... ...

... 
an1 an2 . . . ann
This course is mostly concerned with the solution of square matrices.
A square matrix is symmetric if:
aij = aji (i = 1, . . . , n; j = 1, . . . , n)
3.5. MATRIX & VECTOR 37

Note that the diagonal elements always satisfy aii = aii when j = i, so we only
look at the off-diagonal elements. For example, the matrix A1 is symmetric:
 
10.0 2.0 1.0
A1 =  2.0 23.0 5.0 
1.0 5.0 1.5

A matrix is diagonally dominant if:

aii ≥ |aij | (i = 1, . . . , n; j ̸= i)

This means, across any row in the matrix, the diagonal element is positive
and greater than the magnitude of the off-diagonal elements. For example, the
matrix A2 is diagonally dominant:
 
10.0 0.0 1.0
A2 =  −5.0 23.0 20.0 
−0.1 0.0 1.5

In A2 , for example, across the first row (i = 1), a11 > |a12 |, a11 > |a13 |. In the
second row (i = 2), a22 > |a21 |, a22 > |a23 |, etc.
A sparse matrix is one where most off-diagonal elements are equal to zero
(the diagonal elements are non-zeros), for example, the matrix A3 is considered
sparse:  
10.0 0.0 0.0 0.0
 3.0 23.0 0.0 0.0 
A3 =  0.0

2.0 1.5 0.0 
0.0 0.0 0.0 1.0
A special type of sparse matrix is the Banded Sparse matrix which arises in
finite difference (or finite element) discretization of PDEs describing groundwa-
ter flow and transport. Such banded matrices contain zeros except along the
diagonal where there are a band of elements, e.g., a 8 × 8 banded sparse matrix
can look like this:
 
a11 a12 0 0 0 0 0 0
 a21 a22 a23 0 0 0 0 0 
 
 0 a32 a33 a34 0 0 0 0 
 
 0 0 a43 a44 a45 0 0 0 
 
 0 0 0 a54 a55 a56 0 0 
 
 0 0 0 0 a65 a66 a67 0 
 
 0 0 0 0 0 a76 a77 a78 
0 0 0 0 0 0 a87 a88

The bandwidth m refers to the maximum number of non-zero columns occupied

across a single row of elements. In the above example, the bandwidth m = 3.
As we will show, the discretization of the 1D flow (or transport) equation
gives rise to a matrix with a bandwidth of 3. Such a matrix is also called a
38CHAPTER 3. MATHEMATICAL NOTATIONS, DEFINITIONS AND THEOREMS

tridiagonal matrix. It is the left-hand-side coefficient matrix of a linear set of

equations — results of turning the PDE into a set of discrete algebraic equations
(more on this later). In solving 2D to 3D problems, the coefficient matrices are
still Banded Sparse, but they have higher bandwidth.2
The determinant of a square matrix is defined as (only square matrices can
have determinants):
 
a11 a12 a13 . . . a1n  
 a21 a22 a23 . . . a2n
a22 a23 . . . a2n 
Det 
 ...
 = a11 Det  . . . . . . . . . . . . 
... ... ... ... 
an2 an3 . . . ann
an1 an2 an3 . . . ann
 
a21 a23 . . . a2n
−a12 Det  . . . . . . . . . . . .  + . . .
an1 an3 . . . ann
 
a21 a22 . . . a2,n−1
±a1n Det  . . . . . . . . . ... 
an1 an2 . . . an,n−1
[
]
a11 a12
For example, Det [a11 ] = a11 , Det = a11 a22 − a12 a21 . If the
a21 a22
determinant of a matrix is 0, the matrix is said to be singular.
A square matrix is positive definite if the determinants of itself and all its
submatrices are positive:

Det [a11 ] > 0

[ ]
a11 a12
Det >0
a21 a22
 
a11 a12 a13
Det  a21 a22 a23  > 0
a31 a32 a33
  ...
a11 a12 . . . a1n
 a21 a22 . . . a2n 
Det 
 ...
>0
... ... ... 
an1 an2 . . . ann
2 The characteristics of the matrix and bandwidth varies for groundwater flow and solute

transport problems. In solving the groundwater flow equation, a symmetric coefficient matrix
is produced, thus in a computer, commonly only half of the bandwidth is stored (this is a
practice common in the early days of numerical simulation where a computer has very limited
memory, e.g., 64KB). In using older codes, we’ll still see the definition of a reduced bandwidth:
mr = m/2, which defines the half band of a matrix. The same situation occurs in solving the
equation describing solute diffusion. However, whenever solving the typical solute transport
problem described by advection, dispersion, and diffusion, the coefficient matrix can be non-
symmetrical and sometimes off-diagonally dominated. Special solution technique are needed
to deal with such matrices, e.g., iterative solutions that are not restricted to any special matrix
characteristics, e.g., the IMSL DGMRES works for any general sparse non-symmetric matrix.
3.6. LINEAR ALGEBRA AND SOLUTION TECHNIQUES 39

Diagonally dominant positive definite square matrices have properties that

make them amenable to special solutions if they constitute the left-hand-side
coefficient matrix of a linear system of equations.
As shown in the handout in the beginning of the class, there is an impor-
tant relation between matrix and vector product, e.g., we can write the vector
outcome for a multiplication of a 3×3 matrix and 3× 1 vector:
    
a11 a12 a13  b1   a11 b1 + a12 b2 + a13 b3 
 a21 a22 a23  b2 = a21 b1 + a22 b2 + a23 b3
   
a31 a32 a33 b3 a31 b1 + a32 b2 + a33 b3

The resulting vector has 3 components or a 3× 1 vector. This can be extended

for the multiplication of n×n matrix and n×1 vector: the resulting vector is
n×1.

3.6 Linear Algebra and Solution Techniques

The previous properties of the matrix and matrix-vector product is the corner-
stone of linear algebra. The mathematics of linear algebra is broad and contin-
uously expanding. To study it systematically requires a sperate class which we
can ill afford. So, we select the important concepts and solution procedures from
linear algebra which are widely used in modeling hydrogeological problems.

What is a linear algebra problem?

A linear algebra problem typically consists of a set of n linear equations with

n unknowns, e.g.,


 a11 x1 + a12 x2 + . . . + a1n xn = b1


 a21 x1 + a22 x2 + . . . + a2n xn = b2
a31 x1 + a32 x2 + . . . + a3n xn = b3



 . ..

an1 x1 + an2 x2 + . . . + ann xn = bn

This equation can be written alternatively using matrix vector format as:
    
a11 a12 . . . a1n   x1 
  b1 
 a21 a22     

 . . . a2n 
 x2 b2
 ... ... =
... ...   ... 
  ... 
    

an1 an2 . . . ann xn bn

To write the above equation in a more condensed fashion, we use:

A⃗x = ⃗b

where the left-hand-side (LHS) matrix An×n is the coefficient matrix of the set of
linear equations, the solution vector ⃗x = {x1 , x2 , . . . , xn }T contains the unknown
40CHAPTER 3. MATHEMATICAL NOTATIONS, DEFINITIONS AND THEOREMS

values.3 The right-hand-side (RHS) vector ⃗b contains the known values.

The key in solving a set of linear equations is thus to find the solution vector
⃗x. In general, two broad solution techniques exist:4
• Direct Approach (Gaussian Elimination Schemes)
• Indirect Approach (Various Iterative Methods)

3.6.1 Direct Approaches

The direct approach have many variants. The classic Gaussian Elimination is
based on a two-step procedure:
1. Triangulation of Matrix: Eliminate all terms that occur below the diago-
nal elements (i.e., a11 , a22 , . . . , ann ). For example, after triangulation of a
3 × 3 matrix, we get:
    
a11 a12 a13  x1   b1 
 0 a22 a23  x2 = b2
   
0 0 a33 x3 b3

Note that compared to the original coefficient matrix (e.g., it can be a full
matrix), the new system will have different coefficient matrix and RHS
vector.
2. Back Substitution: After triangulation, the bottom equation (see above)
has only one unknown xn (i.e., x3 ). This can be solved for and the value
of xn substituted into the equation immediately above it to find xn−1 (i.e.,
x2 ). This process is repeated until we solve for the first equation to find
x1 .

Exercise 3 Consider the following system of linear equations — 3 equations

to solve for 3 unknowns. Use the two-step Gaussian Elimination to solve for
the unknowns. 
 x1 + 2x2 − x3 = 3 (1)
4x1 + x2 + 6x3 = 10 (2)

2x1 − 3x2 + 6x3 = 1 (3)

Exercise 4 Consider the following system of linear equations — 3 equations

to solve for 3 unknowns. Use the two-step Gaussian Elimination to solve for
3 In this class, a vector is considered a column vector: a vector of size n is thus of n × 1.

To allow easy writing, we write the column vector as the transpose of a row vector, as shown
above.
4 In hydrologically modeling, typically the Direct Approach is efficient in solving smaller

problems, e.g., 1D or 2D grid with less than 10,000 nodes. The Indirect Approach is commonly
used to solve much larger problems, e.g., 3D grid with more than 10,000 nodes.
3.6. LINEAR ALGEBRA AND SOLUTION TECHNIQUES 41

the unknowns. 
 x1 + x2 + x3 = 1 (1)
x1 + x2 + 2x3 = 2 (2)

x1 + 2x2 + 2x3 = 3 (3)

Note that in Exercise 4, the Gaussian Elimination is interrupted when we

found a22 = 0 after the first elimination. The procedure to deal with this is to
exchange this zero-diagonal equation to the one below, after which triangulation
continues to arrive at the final set which is then solved via back substitution.
This exchange (to allow triangulation) is also called pivoting.

3.6.2 Iterative Approaches

Iterative approaches also attempt to solve the system of linear equations, though
the methodology is quite different. Consider the following example of a 4 × 4
system of equations:


 a11 x1 +a12 x2 + a13 x3 + a14 x4 = b1
|{z}



 a21 x1 + a22 x2 +a23 x3 + a24 x4 = b2
|{z}

 a x + a32 2 + a33 x3 +a34 x4 = b3
x


31 1
|{z}

 a41 x1 + a42 x2 + a43 x3 + a44 x4 = b4
|{z}
We can make an initial guess for all of the unknowns, i.e., for the 4 × 4
system, they are xf1 , x
f2 , x
f2 , x
f4 , and re-arrange the equations: in sequence, we
solve for the diagonal terms with the underbraces xi ):
|{z}

 x1 = 1 [b1 − a12 x f2 − a13 xf3 − a14 x
f4 ]

 x = a111 [b − a x
2 a22 2 21 f1 − a 23 f
x3 − a24 f4 ]
x
(3.7)

 x 3 = 1
[b 3 − a 31 f
x1 − a 32 f
x2 − a34 f4 ]
x
 a33
x4 = a144 [b4 − a41 x f1 − a42 xf2 − a43 x
f3 ]

After a set of values are obtained using the above equations (x1 , x2 , x3 , x4 ), they
become the new “guesses” (f x1 , xf2 , x
f2 , x
f4 ), and are substituted back into the
same above equation to find the next iteration values. We repeat this recursive
process until there is almost no difference between the successive approximation
of the unknown vector ⃗x. That is, another round of iteration brings a negli-
gible change in the values of x f1 , x
f2 , x
f2 , x
f4 compared to the previous round of
values. The final converged values are considered the solution set for the linear
equations.
Note that before implementing the iterative method, the diagonal elements
must be all non-zero. This can be accomplished via pivoting. We have n equa-
tions for n unknowns (x1 , x2 , . . . , xn ), thus the coefficient for each unknown (aij )
must be non-zero at least once. Therefore we can always shift the order of the
equations to make the diagonal elements non-zero.
42CHAPTER 3. MATHEMATICAL NOTATIONS, DEFINITIONS AND THEOREMS

However, the recursive relation can be fine-tuned. There are several types
of Iterative Methods which differ in implementation details:

1. Point Jacobi

2. Gauss-Seidel

3. Successive Over-Relaxation

Point Jacobi is exactly what is described by equation (3.7) which can be

written concisely by the following recursive relationship:

1 ∑n
(3.8) xm+1
i = [bi − aij xm
j ] j ̸= i
aii j=1

where
xm — guess at beginning of an iteration step;
xm+1 — updated value at the end of an iteration step.
We continue the recursion (m → ∞) until some small tolerance (δ) has been
achieved:

hm+1 − hm
(3.9) max| i i
|≤δ
hmi

For example, equation (3.7) is just this condensed equation written out for
n = 4. For i = 1, we write: (j = 2, 3, 4)
xm+1
1 = 1
a11 [b1 2 − a13 x3 − a14 x4 ]
− a12 xm m m

For i = 2, we write:(j = 1, 3, 4)
xm+1
2 = 1
a22 [b2 1 − a23 x3 − a24 x4 ]
− a21 xm m m

For i = 3, we write:(j = 1, 2, 4)
xm+1
3 = 1
a33 [b3 − a31 xm
1 − a32 x2 − a34 x4 ]
m m

For i = 4, we write:(j = 1, 2, 3)
xm+1
4 = 1
a44 [b4 1 − a42 x2 − a43 x3 ]
− a41 xm m m

The above equations boil down to equation(3.7). Note that during a given
recursive round (m + 1), only the results from the last round (xm m m m
1 , x2 , x3 , x4 )
are used.
At the end of each update, we found xm+11 , xm+1
2 , xm+1
3 , xm+1
4 . These values
are compared to the previous results xm m m m
1 , x2 , x3 , x4 , for which we can find a rel-
ative difference for each unknown variable: (xm+1 1 − xm m m+1
1 )/x1 , (x2 − xm m
2 )/x2 ,
m+1
(x3 − xm3 )/xm
3 , (xm+1
4 − xm
4 )/x m
4 . From these 4 relative differences, we can
find a maximum absolute value (the highest magnitude of relative difference
3.6. LINEAR ALGEBRA AND SOLUTION TECHNIQUES 43

between iteration m and m + 1), if this maximum is still greater than δ, the
recursion continues until equation (3.9) is satisfied.

Gauss-Seidel Iteration Method is the same as Point Jacobi Method except

it uses the most recent updated value of xm+1 . It can be written concisely by
the following recursive relationship:

1 ∑
i−1 ∑n
xm+1
i = [bi − aij xm+1
j − aij xm
j ]
aii j=1 j=i+1

where
When j = 1, . . . , i − 1, xm+1
j are the updated value found during the current
(m + 1) iteration step;
When j = i + 1 . . . , n, xmj are the old values found during the last iteration
(m) step.
We continue the recursion (m → ∞) until the same small tolerance (δ) has
been achieved (note for each unknown variable, it is comparing its updated
value xm+1 with its previous old value xm , just like that for the Point Jacobi
Method):
hm+1 − hm
max| i i
|≤δ
hmi

For example, equation (3.7) can be rewritten for Gauss-Seidel Iteration. For
i = 1, we write: (j = 2, 3, 4)
xm+1
1 = 1
a11 [b1 2 − a13 x3 − a14 x4 ]
− a12 xm m m

For i = 2, we write (j = 1, 3, 4), now xm+1

1 is known:
xm+1
2 = 1
a22 [b2 − a21 xm+1
1 3 − a24 x4 ]
− a23 xm m

For i = 3, we write (j = 1, 2, 4), now xm+1

1 and xm+1
2 is known:
xm+1
3 = 1
a33 [b3 − a31 xm+1
1 − a32 xm+1
2 − a34 xm
4 ]

For i = 4, we write (j = 1, 2, 3), now xm+1

1 , xm+1
2 and xm+1
3 is known:
xm+1
4 = 1
a44 [b4 − a41 xm+1
1 − a42 xm+1
2 − a43 xm+1
3 ]
Note that in the Gauss-Seidel Method, during a given recursive round (m +
1), for x1 , only results from the last round (xm m m m
1 , x2 , x3 , x4 ) are used; however,
for x2 , x3 , . . ., increasingly, results from the current round are used.

Successive Over-Relaxation (SOR) — In this iterative method, we accelerate

the step change in xi from iteration m to m + 1. A two-step approach is used.
In step (1):
ci = xm+1
i (guess) − xm
i
44CHAPTER 3. MATHEMATICAL NOTATIONS, DEFINITIONS AND THEOREMS

xm+1
i (guess) is the initial guess for xm+1
i , estimated using the previous Gauss-
Seidel Iteration approach. In step (2), we use:

xm+1
i = xm
i + wci

where w is the Relaxation Factor (1 ≤ w ≤ 2). Note that when w = 1, SOR

collapses to Gauss-Seidel Iteration.
Combining step (1) and (2), we can also write a single SOR formulation:
   
 1 ∑
i−1 ∑n 
xm+1 = xm
i +w
bi − aij xm+1 − aij xm  − xm
i
 aii j j i

j=1 j=i+1

3.6.3 Comparison and Summary

1. Direct methods (Gaussian Elimination) provide the most reliable means of
solving a set of linear algebra equations. However, the chief shortcoming
of using direct methods is that it can be very flow in solving large matrix
problems, compared to the speed with which the same problems can be
solved by iterative methods. Typically, direct methods are used to solve for
a problem with 10,000 or less unknowns. Since numerical approximations
of equations of flow/transport are based on grids, this translates to a grid
with 10,000 or less grid cells (FD block centered) or mesh points (FD mesh
centered) or nodes (FEM).
2. Numerical round-off error occur in both direct and iterative methods since
the finite arithmetics of using limited number of bytes to store real num-
bers is unavoidable as long as computers are used to solve the problem.
3. In the matrix-free iterative methods (Point-Jacobi, Gauss-Seidel, SOR),
there is no matrix storage since the coefficient matrix does not need to be
explicitly formed (this will become apparent when you work on the Sec-
ond Program of Homework 2). In general, Gauss-Seidel converges twice as
fast as the Point Jacobi method. So, for the same set of linear equations,
using Gauss-Seidel iteration method to solve for the solution vector will
need only half of the computation time as the Point Jacobi method. The
convergence speed of the SOR is sensitive to the choice of w used. For a
given set of equations, methods for selecting an optimum value of w can be
found by trail-and-error or through theoretical means (see Remson, Horn-
berger and Molz (1971), Numerical Methods in Subsurface Hydrology).
4. Matrix-free iterative solver is nowadays rarely used, since they are much
slower compared to matrix-based iterative methods (IMSL, SPLIB, LA-
PACK, LINPACK, etc). In the matrix-based iterative methods, matrix
needs to be explicitly formed, and according to the requirement of each
solver, specific format is required in assembling the global coefficient ma-
trix. This can be time consuming at first but very useful in solving large
problems. Though matrix-based iterative solver is not covered in this
3.6. LINEAR ALGEBRA AND SOLUTION TECHNIQUES 45

class, you’re welcome to talk to me after class where I can show you some
examples of calling these solvers (mostly done on Unix clusters).

5. Compared to the Direct method, iterative methods can have trouble con-
verging or can develop instability (solution vector fluctuates greatly from
m to m + 1 without convergence as m → ∞), for highly deformed grids or
problems with large contrast in material properties. For example, in mod-
eling a multilayered aquifer-aquitard system, if the contrast in hydraulic
conductivity between aquifer and aquitard is large, iterative methods may
have trouble converging. In this case, investigate the condition number of
the matrix to see if it is approaching singular (e.g., calling a subroutine
in IMSL which computes the matrix condition number). If so, consider
improving the aspect ratio of the grid (so it is less deformed) and consider
adding buffer zones between very-high-K and very-low-K zones. More-
over, in solving transient problems (i.e., hydraulic head, temperature (T),
or velocity change with time), the convergence speed of iterative methods
can be sensitive to the choice of initial conditions.5 On the other hand,
direct methods seem generally less affected by these issues though there
is no guarantee.

6. For large problems where we’re not sure if iterative methods can provide
sufficiently accurate solutions, the same problem should be solved first by
a direct method (as long as the computer is sufficiently fast). By com-
paring the solutions using both approaches, the accuracy of the iterative
method will be ascertained. Since the iterative method can generally solve
bigger problems faster, it is still the best choice for large simulation mod-
els. However, if computer time is not excessive, verification with direct
methods is a prudent procedure to take.

7. The process of ensuring that an iterative method is giving the correct nu-
merical results (as compared to those predicted by the direct method for
the same problem) is called code verification. The process of ensuring that
an iterative method (or direct method) gives the results that are close to
the analytical (exact) solutions (when they exist) is called code validation.
However, for complicated problems with complex geometry, parameter
heterogeneity, and spatially-varying and/or time-dependent boundary con-
ditions, analytical solutions to such problems generally do not exist. In
such cases, though numerical simulations are used to model such prob-
lems, analytical solutions usually do not exist, thus the numerical codes
can be verified but not validated. In these cases, we conduct (1) grid res-
olution study (for transient problems both grid and temporal resolution
can be changed); (2) local and global mass balance analysis (for transient
problems, mass balance can be checked for every time step).
5 For example, to solve the transient responses of a geotherm field to groundwater infiltra-

tion and discharge, if the initial T field is chosen to reflect the final expected field more closely,
the iterative methods will converge to find the final solution faster than choosing an initial T
field that is far from reality (e.g., giving T a uniform value of 0 degree C).
46CHAPTER 3. MATHEMATICAL NOTATIONS, DEFINITIONS AND THEOREMS

3.7 Homework 2
1. Solve the below equations by hand to find the unknowns (x1 , x2 , x3 ) using
Gaussian Elimination. Hint: pivoting may be needed.

 x1 + x2 + x3 = 1.0 (1)
x1 + 1.0001x2 + 2x3 = 2.0 (2)

x1 + 2x2 + 2x3 = 1.0 (3)

2. Solve the below equations first by hand using Gaussian Elimination.

Then, write Matlab (alteratively you can write in Fortran or C) scripts to solve
for the unknowns by the following iterative methods: (a) Point-Jacobi; (b)
Gauss-Seidel; (c) SOR. Note that your earlier hand-calculated values (via Direct
Gaussian Elimination) help you determine if your codes are giving you correct
results. Due to numerical rounding errors, your numerical results (via each
Iterative Method coded by you) should be extremely close to the hand-calculated
values, but not necessarily the same.
For each iterative method, use a convergence tolerance δ=0.005 and initial
guesses of {0.0, 0.0, 0.0, 0.0}. Keep an account of how many iteration steps it
will take for each method to reach convergence. Which of the three methods
appear to converge the fastest? Which is the slowest? For SOR, try different
w values to see how it affects the speed of convergence, i.e., how many steps it
takes to converge. For example, you can try: w = 1.1, 1.3, 1.5, 1.7, 1.9. Which
value is optimal (giving the least number of steps towards convergence)? If the
above values do not determine an optimal value, you can further fine-tune w,
e.g., try w = 1.05. 

 4x1 − x2 − x3 = 1.0 (1)

−x1 + 4x2 − x4 = 2.0 (2)

 −x 1 + 4x3 − x4 = 0.0 (3)

−x2 − x3 + 4x4 = 1.0 (4)

For these exercises, you must hand in your codes as well. For hand calcula-
tions, present the full steps.
(Optional) You can also find out what happens to the speed of convergence
when you increase or decrease the tolerance.

3.8 Gauss-Jordan Direct Solver with Pivoting

(Optional)
Solving linear algebra problems for which the matrix is not well-behaved (e.g.,
pivoting is required as one of the class hand-exercises illustrates) is normally not
an issue with solving the groundwater flow and transport problems. Since after
3.8. GAUSS-JORDAN DIRECT SOLVER WITH PIVOTING (OPTIONAL)47

FD (or FE) discretization of the flow and transport equations, the matrix is gen-
erally well-behaved: for flow problem, the matrix is symmetric, positive-definite
(diagonally dominant), and sparse; for transport, the matrix is symmetric and
sparse, but can be off-diagonally dominant for advection-dominated problems
(in which case special solution technique is implemented, e.g., Modified Method
of Characteristics). In all these cases, we do not need to worry about switching
the orders of the unknowns as required by pivoting.
However, sometimes linear matrix problems do arise for which pivoting is re-
quired, or the matrix is not symmetric, positive-definite, or possesses any special
characteristics (for example, in solving the 1D flow equation, the coefficient ma-
trix is always tridiagonal for which a special direct Gaussian-Elimination solver
Gentri.m can be used which specifically works for this type of matrix). A case in
mind is solving the kriging or co-kriging equations (please take my Geostatistics
class to learn these methods) which have zero diagonal elements! Whenever
such cases happen (you’ll be surprised how often they do once you start doing
research), a more general solver can be used which has no special requirements
on matrix characteristics. A popular direct solver is the Gauss-Jordan method
which is capable of automatic pivoting (parallel to our hand-exercises when our
brain was doing the pivoting). For example, a full code solving a 5×5 problem
is shown below. If we write out the A⃗x = b and try to do this by hand, we soon
realize that we have to do pivoting. However, this code calls the Gauss-Jordan
method to do the pivoting and Gaussian-Elimination directly.

---------------------------------------------------------------------
% Gaussian-Jordan method of solving linear equations with full
% pivoting: this method is capable of solving non-symmetrical,
% non-positive-definite, non-diagonal-dominated matrix. IN general,
% any linear set of equations that are NOT singular (If the matrix is
% singular, there exist infinite sets of solution vectors for any one
% solve of Ax=b. That is, x cannot be uniquely determined).

PARAMETER(n=5, np=5, m=1, mp=1)

REAL*8 a(np,np), b(np,mp)

INTEGER i,j,k
% OPEN(UNIT=8, FILE=’debug_jordan.data’)
OPEN(UNIT=9, FILE=’output_jordan.data’)

a(1,1)=1
a(1,2)=1
a(1,3)=1
a(1,4)=1
a(1,5)=0

a(2,1)=0
a(2,2)=95.75