Journal of Hydrology 376 (2009) 567–578

Contents lists available at ScienceDirect

Journal of Hydrology
journal homepage: www.elsevier.com/locate/jhydrol

Improving parameter estimation for column experiments

by multi-model evaluation and comparison
Guoping Tang a,*, Melanie A. Mayes a, Jack C. Parker b, Xiangping L. Yin a,
David B. Watson a, Philip M. Jardine a
a
Environmental Sciences Division, Oak Ridge National Laboratory, P.O. Box 2008, MS-6038, Oak Ridge, TN 37831, United States
b
Department of Civil and Environmental Engineering, University of Tennessee, Knoxville, TN 37996, United States

a r t i c l e i n f o s u m m a r y

Article history: The equilibrium convection dispersion equation model is often unable to accurately simulate break-
Received 15 November 2008 through curves from column experiments. While the non-equilibrium convection dispersion equation
Received in revised form 5 June 2009 model may match the data well, uncertainty in parameter estimates is often large. In this work we inves-
Accepted 30 July 2009
tigate approaches to improve match for the equilibrium model and reduce parameter estimate uncer-
This manuscript was handled by P. Baveye,
tainty for the non-equilibrium model. Four column experiment data sets are selected from the
Editor-in-Chief literature for the illustration. For the equilibrium convection dispersion equation model, we show that
measurement error, presence of immobile water, and other mechanisms can cause mismatch between
Keywords:
model predictions and observations because the model is sensitive to water content. The mismatch
Sensitivity/uncertainty analysis may be overcome by calibrating the effective water content. For the non-equilibrium convection disper-
Nonlinear least squares sion equation model, simultaneous fitting of multiple tracers with reduced number of calibration param-
Equilibrium/non-equilibrium convective eters (e.g., assuming the dispersivity and mobile water fraction to be identical for different tracers, the
dispersion equation mass transfer coefficient to be proportional to tracer molecular diffusion coefficient) can reduce the
Monte Carlo analysis uncertainty in parameter estimate and better identify/quantify the non-equilibrium processes. By evalu-
ating and comparing the multiple estimates obtained with different choices of calibration parameters
(e.g., fixing or estimating water content), parameterizations and models (e.g., equilibrium or non-equilib-
rium), the reliability of the data interpretation can be improved by quantifying uncertainty in the exper-
iment, considering alternative transport processes, and following the principle of parsimony.
Published by Elsevier B.V.

Introduction water velocity (Ellsworth et al., 1996; Huang et al., 1995; Padilla
et al., 1999; Porro et al., 1993), the retardation factor (Anamosa
Column experiments are widely used to investigate subsurface et al., 1990; Comegna et al., 2001; Kamra et al., 2001; Langner
contaminant migration processes by fitting the breakthrough et al., 1999; Nkedi-Kizza et al., 1983; Porro et al., 1993; van
curves with transport models using nonlinear least squares meth- Genuchten and Wierenga, 1977; Wierenga and Van Genuchten,
ods. A simple model is often found to be inadequate to simulate the 1989), the water content (Schoen et al., 1999b), and/or the pulse
breakthrough curves, while uncertainty in parameter estimates duration or pore volume (Ellsworth et al., 1996; Kamra et al.,
can be large when a more complex model adequately matches 2001). The estimated retardation factor is often quite different
the observations. Reliable quantification of transport parameters from 1 for nonreactive tracers; the estimated dispersion coeffi-
from column experiments is essential for the identification of cient/dispersivity/Peclet number is sometimes not physically
multiple processes governing contaminant migration, and for the reasonable; and the estimated water content is often different
prediction of contaminant fate and transport. from the measured average water content. The failure of CDE to
The equilibrium convection dispersion equation model (CDE) is simulate breakthrough curves is often attributed to non-equilib-
probably the most widely used model to simulate breakthrough rium processes.
curves from column experiments, although poor agreement with The non-equilibrium convection dispersion equation model
data is often observed. To obtain a reasonable match with the (the mobile–immobile or the two-site model) (Toride et al.,
observations, published studies may estimate the average pore 1995), is often better able to match the breakthrough curves
because more parameters are estimated. The mobile zone disper-
* Corresponding author. Tel.: +865 574 7314; fax: +865 576 8646. sivity, mobile water fraction and mass transfer coefficient are
E-mail address: [email protected] (G. Tang). difficult to be independently measured. Simultaneous estimation

0022-1694/$ - see front matter Published by Elsevier B.V.

doi:10.1016/j.jhydrol.2009.07.063
568 G. Tang et al. / Journal of Hydrology 376 (2009) 567–578

of these parameters often results in ill-defined parameters (Come- highly correlated parameters. Even though the parameters may
gna et al., 2001) with excessively large confidence intervals or not be correlated in a physical sense, they can be dependent on
standard errors (Anamosa et al., 1990; Bajracharya and Barry, each other in terms of parameter estimation because they are de-
1997; Bond and Wierenga, 1990; Comegna et al., 2001; Maraqa rived from the same set of data. Correlation coefficients close to ±1
et al., 1997; Nkedi-Kizza et al., 1983) and high correlations (Koch are likely to signify over-parameterization. However, large uncer-
and Fluhler, 1993) among parameter estimates, or the estimates tainty and high correlation in the parameter estimate are often ig-
vary with initial inputs (Toride et al., 1995). In those cases, the esti- nored in the column experimental literature.
mated values may adequately describe the shape of the break- In addition to the goodness-of-fit statistics, Knopman and Voss
through curves but may not necessarily provide mechanistic (1988) propose to discriminate among competing transport mod-
information into the physical transport processes (Comegna els based on prediction error, systematic error, stability of param-
et al., 2001; Koch and Fluhler, 1993). The model may be overpa- eter estimates and measures of model fit between first and second
rameterized, and it is desirable to reduce overall complexity round experiments, where second round experiments were de-
according to the principle of parsimony. signed based on sensitivity analyses to differentiate between
Efforts are made to reduce the number of estimated parameters models. Ellsworth et al. (1996) use stability in parameter esti-
to avoid over-parameterization. The moment analysis method is mates and RMSE with a sequential parameter estimation method
used to ‘‘independently” determine some of the parameters, e.g., for model discrimination. For statistically reasonable estimates,
water content (Padilla et al., 1999), retardation factor (Brusseau, the F-test and Akaike’s Information Criteria (Burnham and Ander-
1992; Brusseau et al., 1989), and velocity (Huang et al., 1995; Porro son, 2004) are also used for model comparison. Even though sen-
et al., 1993). The value calculated from moment analysis is gener- sitivity analysis, uncertainty analysis, model evaluation and
ally similar to that from nonlinear least squares fitting of the comparison are well-established, and the principle of parsimony
breakthrough curves (Kamra et al., 2001) but its accuracy depends is widely respected in the literature of parameter estimation,
on sampling frequency, experimental error and truncation time inverse problem and model calibration (Bard, 1974; Burnham
(Das et al., 2005). Retardation factor is also determined by measur- and Anderson, 2004; Hill, 1998; Motulsky and Ransnas, 1987), it
ing the area above the breakthrough curve for a pulse input (Brus- appears not to be the common standard practice in the column
seau et al., 1989). The Peclet number is determined directly from experiment literature.
the slope of the breakthrough curve (Wellman et al., 2008). The
mobile water fraction is also estimated from fast-flow experiments
and then fixed in other experiments (Anamosa et al., 1990; Baj- (a) Observation
racharya and Barry, 1997). An average of estimated dispersion CDE1 (λ)
coefficients from experiments at various velocities is used by Baj- 0.5 CDE2 (λ & θ)
CDE3 (λ, θ, & t0)
Reduced concentration

racharya and Barry (1997). Because these parameters are derived

from the same set of breakthrough curve data, these methods do 0.4
not help overcome the nonuniqueness of parameter estimates or
make the inverse problem more well-posed. 0.3
Another common approach is to estimate some parameters, e.g.,
Peclet number/dispersion coefficient, from the breakthrough curve
0.2
for a nonreactive tracer, and then assume the same values apply for
other chemicals (Brusseau, 1992; Nkedi-Kizza et al., 1984; Toride
et al., 1995). It is known that the dispersion coefficient is not very 0.1
sensitive (Knopman and Voss, 1988) so uncertainty in the disper-
sion coefficient estimate is often large. Fixing the dispersivity 0
may result in underestimation of the uncertainty in other parame- 0 200 400 600 800 1000
ter estimates, or over-confidence in the model for reactive tracers. Time (h)
Multiple nonreactive tracers are used to investigate the non-equi-
librium processes (Mayes et al., 2003). The breakthrough curves for (b) Observation
different tracers are often fitted separately so that the identifiabil-
MIM1 (λ, β & ω)
ity problem remains, even though the tracer separation clearly 0.5 MIM2 (λ, β, ω, & t0)
shows the impact of non-equilibrium processes. Simultaneously
Reduced concentration

fitting multiple tracers is expected to reduce nonuniqueness of 0.4

the nonlinear least squares estimation.
To evaluate and compare physically reasonable nonlinear least
0.3
squares estimates, the first step is to plot the observations against
predictions to visually examine the match. This is important be-
cause some of the details in the information in the breakthrough 0.2
curves are treated as errors by the nonlinear least square method
(Beven et al., 1993). The sum of squared residuals (SSR), root mean 0.1
square of residuals (RMSE), R square (R2), and parameter estimation
statistics (standard errors, confidence intervals, and correlations)
0
are commonly calculated (Parker and van Genuchten, 1984; Toride 0 200 400 600 800 1000
et al., 1995). An R2 close to 1 is often used as an indicator of good Time (h)
fit. A very narrow confidence interval of the estimated parameters
may be suspect because it is based on the first-order approxima- Fig. 1. Observations (circles) and nonlinear least squares fits (lines) for Example 1
tion (Motulsky and Ransnas, 1987) and the assumption that the where (a) CDE = convection dispersion equation model and (b) MIM = mobile–
immobile water model with estimated parameters in (). The parameters are
measurement and model error are at the level of optimized RMSE. k = dispersivity, h = water content, and t0 = pulse duration, b = mobile water frac-
In contrast, wide confidence intervals may suggest over-parame- tion, x = mass transfer coefficient. The observation data are from Schoen et al.
terization or insensitivity of parameters to measurements, or (1999a,b).
 G. Tang et al. / Journal of Hydrology 376 (2009) 567–578 569

Table 1
Parameter estimates for Example 1.

k (cm) h t0 (h) b x RMSE R2 |q|max

a
CDE2 8.55 ± 0.38 0.217 ± 0.001 143.8 0.010 0.997 0.383
CDE3 8.92 ± 0.15 0.217 ± 0.001 147.2 ± 0.8 0.009 0.998 0.592
MIM1 8.55 ± 0.17 0.253 0.857 ± 0.003 0.000 ± 0.007 0.011 0.997 0.561
MIM2 8.37 ± 0.19 0.253 157.0 ± 1.2 0.833 ± 0.005 0.110 ± 0.018 0.007 0.999 0.934
a
Parameter estimate ± standard error. No ± means fixed parameter.

(a) 0.1 1

0.08
0. 5

0.06
RMSE

Scaled sensitivity
0
0.04

0.02 −0. 5

θ
0 t
0
0 5 10 15 20 −1
λ
λ (cm)
β
ω
(b) 0.1 −1. 5
0 200 400 600 800 1000
Time (h)
0.08
Fig. 3. Sensitivity for MIM2 for Example 1 where h = water content, t0 = pulse
duration, k = dispersivity, b = mobile water fraction and x = mass transfer
0.06 coefficient.
RMSE

0.04 estimations for column experiments. We will provide methods

for detecting and eliminating over-parameterization considering
uncertainty, sensitivity, and correlation between parameters, and
0.02
evaluating and comparing alternative models. Four typical column
experimental data sets from the literature are selected to illustrate
0 how parameter estimation can be improved by using these
0.7 0.75 0.8 0.85 0.9 0.95 1
techniques.
β

Theory
(c) 0.1
Transport models
0.08
The equilibrium convection dispersion equation with linear
adsorption is (Parker and van Genuchten, 1984)
0.06
RMSE

@c @2c @c
R ¼D 2m ð1Þ
0.04 @t @z @z
where R = 1 + qbkd/h is a retardation factor with qb as the bulk den-
sity [ML3], kd as the partition coefficient [M1L3] and h as the water
0.02
content []; c is the resident concentration [ML3] (i.e., averaged
over the solution volume within a representative pore volume), D
0
−4 −3 −2 −1
is a dispersion coefficient [L2T1], v is mean pore water velocity
10 10 10 10 [LT1], t is time, and z is distance from the inlet boundary. The initial
ω
condition is c(z, 0) = 0. A pulse input of concentration c0 is applied at
Fig. 2. Monte Carlo simulation results for MIM for Example 1. k (dispersivity), b the inlet boundary from t = 0–t0. Appropriate inlet and outlet
(mobile water fraction) and x (mass transfer coefficient) are sampled from boundary conditions are described by (Parker and van Genuchten,
uniformly distributed random numbers between 0–3, 0.8–1, and 10 to 2 for 1984; Toride et al., 1995). Eq. (1) in dimensionless variables is
ln(k), b and ln(x). The red dot represents the optimized value. (For interpretation of
the references to colour in this figure legend, the reader is referred to the web @C 1 @ 2 C @C
version of this article.) R ¼  ð2Þ
@T P @Z 2 @Z
The objective of this work is to provide examples and where T = vt/L, Z = z/L, P = mL/D = L/k, and C = c/c0with L as the length
recommendations to improve the standard practice in parameter of the column [L]; k as the dispersivity [L];.
570 G. Tang et al. / Journal of Hydrology 376 (2009) 567–578

(a) Expt. I @C 1 1 @ 2 C 1 @C 1
bR ¼   xðC 1  C 2 Þ ð3aÞ
1 @T P @Z 2 @Z
Obs @C 2
q = 111 cm/d CDE1 (λ) ð1  bÞR ¼ xðC 1  C 2 Þ: ð3bÞ
@T
0.8 CDE2 (λ & θ)
Reduced concentration

MIM (λ, β & ω) For the physical non-equilibrium (two-region, mobile–immo-

bile) model, b = (hm + fqbkd)/(h + qbkd) with hm as the mobile pore
0.6 water content, and f as the fraction of adsorption sites that equili-
brates with solute in mobile pore water; x = aL/(hv) is the dimen-
sionless mass transfer coefficient with mass transfer coefficient a
0.4 between mobile and immobile water [T1]; C1 and C2 are relative
concentrations in the mobile and immobile regions, respectively.
q = 2.71 cm/d
For the chemical non-equilibrium (two-site) model, b = (h + fqbkd)/
0.2 (h + qbkd), x = a(1b)RL/v with f as the fraction of equilibrium
adsorption sites and a as the first order kinetic rate coefficient [T1].
If f = 0, the two-site model reduces to one-site model. When b = 1,
0
0 1 2 3 4 5 6 7 Eq. (3) reduces to Eq. (2). For x = 0, Eq. (3) becomes
Effluent pore volume
@C 1 @ 2 C @C
bR ¼  : ð4Þ
@T P @Z 2 @Z
(b) Expt. II
1 In this case, MIM becomes an equilibrium model with bh repre-
Obs senting the effective water content. Therefore, the CDE can be con-
q = 36.7 cm/d CDE1 (λ) sidered as a special case of the MIM (Eq. (3)). Note in Eq. (4), b and
0.8 CDE2 (λ & θ) R are not differentiable from the nonlinear least square estimate.
MIM (λ, β & ω)
Reduced concentration

Analytical solutions for both CDE and MIM models were given in
(Toride et al., 1995) and are implemented in Excel using VBA in
0.6 CXTFIT/Excel to achieve the flexibility in using different choices
of calibration parameters, parameterizations, and models needed
for this work.
0.4

Nonlinear least squares method

q = 2.69 cm/d
0.2
The nonlinear least squares method adjusts M parameters (b1,
b2, . . ., bM) in the solution c(x, t, b1, b2, . . . bM) for Eq. (2) or Eq. (3)
0 to minimize the following objection function for matching N
0 1 2 3 4 5 6 7 observations (c1, c2, . . ., cN)
Effluent pore volume
X
N
2
Fig. 4. Observations (circles) and nonlinear least squares fits (lines) for Example 2. SSR ¼ ½ci  cðxi ; t i ; b1 ; . . . ; bM Þ : ð5Þ
(a) Expt. I-1 (q = 111 cm d1) and Expt. I-3 (q = 2.71 cm d1) where CDE = convec- i¼1
tion dispersion equation model and MIM = mobile–immobile water model with
estimated parameters in (), and (b) Expt. II-1 (q = 36.7 cm d1) and Expt. II-2 The optimization is performed using Microsoft Excel Solver. The
(q = 2.69 cm d1). The parameters are, k = dispersivity, h = water content, b = mobile sensitivities, Jji = oc(xi, ti, b1, b2, ...bM)/obj, are approximated by the
water fraction, and x = mass transfer coefficient. The data are from Anamosa et al. forward difference (Hill, 1998; Toride et al., 1995). To compare sen-
(1990). sitivities for parameters of different units and magnitudes, the
scaled sensitivity (ss) is defined as ssij = pjJij, and the composite
P 2
scaled sensitivity (css) is defined as css2j ¼ ssij =N (Hill, 1998).
The non-equilibrium convection dispersion model (MIM) is use- The covariance matrix for the estimated parameters is calculated
ful for describing transport in structured soils (Hu and Brusseau, by Cov = MSE[JTJ]1 with MSE = [SSR/(NM]1/2 as an approximation
1995). It can be cast in dimensionless terms as (Toride et al., 1995) of the observation variance (Bard, 1974). The variance of parameter

Table 2
Parameter estimates for Example 2.

k (cm) h b x RMSE R2 |q|max

Expt. I-1 CDE1 86.60 ± 6.62 0.044 0.979
q = 111 CDE2 36.39 ± 3.47 0.373 ± 0.013 0.026 0.993 0.874
MIM 25.00 ± 2.19 0.536 ± 0.017 0.193 ± 0.020 0.013 0.998 0.937
Expt. I-3 CDE1 15.63 ± 0.71 0.023 0.996
q = 2.71 CDE2 16.74 ± 0.42 0.560 ± 0.005 0.013 0.999 0.385
MIM 4.58 ± 3.26 0.373 ± 0.089 2.988 ± 0.339 0.012 0.999 0.942
Expt. II-1 CDE1 21.09 ± 0.64 0.016 0.998
q = 2.69 CDE2 22.00 ± 0.51 0.551 ± 0.005 0.012 0.999 0.396
MIM 9.89 ± 3.16 0.600 ± 0.089 1.195 ± 0.211 0.013 0.999 0.972
Expt. II-2 CDE1 100.03 ± 5.21 0.034 0.989
q = 36.7 CDE2 92.74 ± 4.67 0.438 ± 0.021 0.029 0.992 0.387
MIM 27.45 ± 2.75 0.467 ± 0.019 0.280 ± 0.020 0.013 0.998 0.945
 G. Tang et al. / Journal of Hydrology 376 (2009) 567–578 571

bi is s2(bi) = Cov(i, i). The coefficient of variation, a relative measure WA were collected following the methods of Pace et al. (2003)
of parameter uncertainty, is s(bi)/bi, and the correlation between bi and Mayes (2006). One core was taken parallel to the bedding
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi  
and bj is qij ¼ Covði; jÞ= Covði; iÞCovðj; jÞ ¼ Covði; jÞ= sðbi Þsðbj Þ . (HC pb) and the other across the bedding (HC xb). The cores con-
sisted of 93% sand, 4% silt and 3% clay. Nonreactive tracers bromide
When many parameters are estimated, the optimization can be
(Br) and piperazine-1, 4-bis (2-ethanesulfonic) acid (PIPES) were
trapped in a local minimum. The uncertainty and sensitivity
injected simultaneously under unsaturated flow conditions to
analysis is a first-order approximation based on assumed linearity
study the influence of layering on solute transport. Tracer separa-
at the optimized values, which can produce inaccuracies. To check
tion between Br and PIPES is found in the breakthrough curves
for local minima and inaccurate sensitivity and uncertainty
for both cores, which suggests non-equilibrium processes. While
approximation, Monte Carlo analyses are conducted.
either the CDE or MIM models can simulate the breakthrough
curves for the two tracers separately, our focus is on description
Materials and methods

The column experiment data from Anamosa et al. (1990), Mayes (a) 0.1
(2006), Schoen et al. (1999a) and Toride et al. (1995) are selected as
examples in this work. The first experiment involves transport of
2 0.08
H2O, bromide and chloride in a large undisturbed lysimeter
(Schoen et al., 1999a). The experimental site is located 40 km
northwest of Grenoble, France. The soil is an aggregated sandy 0.06

RMSE
and clayey loam with a high density of macropores. The lysimeter
has a surface area of 1.2 m2 and a depth of 1.5 m. Fluid was applied 0.04
using a rainfall simulator with grid of 30 dripping units. There was
no vegetation on the surface of the lysimeter during the experi-
ment, and the surface was covered to avoid evaporation. The goal 0.02
of this experiment was to quantify fluxes and mechanisms at a
scale close to field conditions to yield results that better depict 0
−1 0 1 2
reality of preferential flow and transport. 10 10 10 10
Experiment #1B involves a water flux of 0.105 cm h1, a mea- λ (cm)
sured water content of 0.253, a pulse duration of 143.8 h, and a
pulse 2H2O concentration of 4537 d%. The mass recovery is calcu- (b) 0.1
lated to be 105% for 2H2O by moment analysis. We digitize Fig. 3
in (Schoen et al., 1999a) for our analysis. Instead of choosing the
0.08
‘‘best” fit by eye as the authors did, we use the nonlinear least
squares method. For the CDE fit, we investigate the sensitivity of
water content and mass balance on the model prediction. For the 0.06
RMSE

MIM fit, we examine uncertainty in parameter estimates, by both

the nonlinear parameter estimate statistics, and Monte Carlo sim-
0.04
ulations. By model comparison and evaluation, we examine if the
experimental data support use of the MIM model or if CDE is ade-
quate to describe the breakthrough curve. 0.02
The goal of the second experiment was to determine if macro-
pore water flow and immobile-water regions should be considered
0
in describing solute movement while developing nutrient-man- 0 0.2 0.4 0.6 0.8 1
agement strategies (Anamosa et al., 1990). The soil is a clayey-skel- β
etal, oxidic isohyperthermic, typic Gibbsiorthox located 8 km north
of the University Center of Dschang in Western Province of Camer- (c) 0.1
oon. The two undisturbed soil columns were taken using two 80-
cm long, 9.6 i.d. PVC pipe from the top three horizons. The two
samples were separated in the field by about 2 meter, and have 0.08
similar characteristics. The lengths of columns are 71.6 and
68.9 cm. The water contents are 0.526 and 0.527. The break- 0.06
RMSE

through curves for Expt. I-1, I-3, II-1 and II-2 in Anamosa et al.
(1990) are digitized and reanalyzed. Fluxes for the four experi-
ments are 111, 2.71, 2.69, and 36.7 cm d1, and dimensionless 0.04
pulse durations are 1.43, 2.84, 2.88, and 2.59. To be consistent with
the original analysis, the retardation factor is fixed at 1.05. For the 0.02
CDE model, we consider the cases when water content is fixed or
estimated in addition to the dispersivity to be estimated. For
0
MIM, we focus on the uncertainties in the estimated parameters. 10
−1
10
0 1
10 10
2

The CDE and MIM estimates are compared for the two columns ω
at different flow rates to evaluate the significance of the non-equi-
librium processes. Fig. 5. Monte Carlo simulation results for 3H2O transport for MIM in Expt. I-3
In the third experiment, two large, intact, unconsolidated sedi- Example 2. k (dispersivity), b (mobile water fraction) and x (mass transfer
coefficient) are sampled from uniformly distributed random numbers between 2–
ment cores (25 cm in height and diameter) of the Hanford flood 5, 0–1, and 2–5 for ln(k), b and ln(x). The red dot represents the optimized value.
deposits from the Integrated Disposal Facility (IDF) on the U.S. (For interpretation of the references to colour in this figure legend, the reader is
Department of Energy (DOE) Hanford Reservation near Richland, referred to the web version of this article.)
572 G. Tang et al. / Journal of Hydrology 376 (2009) 567–578

of the tracer separation by simultaneously fitting of the break- Tortuosity is computed from water content following Millington
through curves for both tracers. To reduce the number of estimated and Quirk (1961) as
parameters for multiple tracers, the dispersion coefficient is
parameterized as (Fetter, 2001) s ¼ h7=3 =h2s ð7Þ

D ¼ sD0 þ mk ð6Þ with hs as the saturated water content. The mass transfer coefficient
is assumed to be proportional to molecular diffusion coefficient
where D0 is the free water molecular diffusion coefficient and s is similar to Hu and Brusseau (1995),
tortuosity. D0 for Br and PIPES are 18.7 and 7.8  1010 m2 s1, x ¼ gD0 ð8Þ
respectively (Mayes et al., 2003). Dispersivity k is assumed to be a
property of the porous medium and is identical for different tracers. where g is defined as a geometric factor.
The breakthrough curves in Fig. 7.9 of Toride et al. (1995) are
used as a typical example for non-equilibrium transport by van
(a) 0.1 Genuchten (1981) and Parker and van Genuchten (1984). It de-
scribes the transport of tritiated water (3H2O) and boron (B)
through a 30 cm long Glendale clay loam column. For 3H2O exper-
0.08 iment, h = 0.454, v = 37.5 cm d1, T0 = 3.102. D, b and x are esti-
mated as 15.5 cm2 d1, 0.822, and 0.85. For the boron (B)
0.06 breakthrough curve, v = 38.5 cm d1, R = 3.9. D and hm/h from
RMSE

3
H2O estimate were fixed, and x and f are estimated to be 0.70
and 0.49. It is recognized that estimate of D, b and x is dependent
0.04
on the initial input (Toride et al., 1995). The confidence interval
given by CXTFIT (Toride et al., 1995) is 7.1 to 24.0 cm2 d1 for D,
0.02 and 0.760 to 0.886 for hm/h for 3H2O. We asses the impact of this
uncertainty by fitting the breakthrough curves for 3H2O and B

0
0 1 2
10 10 10
λ (cm) (a) Expt. I−1
0.5
λ
(b) 0.1 β
ω
0.08
Scaled sensitivity

0.06
RMSE

0.04

0.02

0 −0.5
0.3 0.4 0.5 0.6 0.7 0.8 0.9 0 1 2 3 4 5
β Effluent pore volume

(c) 0.1
(b) Expt. I−3
0.5
0.08 λ
β
ω
0.06
RMSE

Scaled sensitivity

0.04
0
0.02

0
−2 −1 0
10 10 10
ω

Fig. 6. Monte Carlo simulation results for 3H2O transport for MIM in Expt. I-1 −0.5
Example 2. k (dispersivity), b (mobile water fraction) and x (mass transfer 0 1 2 3 4 5 6 7
coefficient) are sampled from uniformly distributed random numbers between 3– Effluent pore volume
9, 0.3–0.9, and 3–4 for ln(k), b and ln(x). The red dot represents the optimized
value. (For interpretation of the references to colour in this figure legend, the reader Fig. 7. Sensitivity for MIM for Expt. I-3 in Example 2 where k = dispersivity,
is referred to the web version of this article.) b = mobile water fraction and x = mass transfer coefficient.
 G. Tang et al. / Journal of Hydrology 376 (2009) 567–578 573

simultaneously with the same assumptions of identical k and hm/h CDE1). Water content was measured using a neutron probe with
for the two tracers. Eq. (8) is used to parameterize the mass trans- a coefficient of variation of about 10% (Schoen et al., 1999a), there-
fer coefficient with the value of D0 of 2.44  109 m2 s1 for 3H2O fore it is reasonable to try to calibrate water content. When both
(Wang et al., 1953) and 1.0  109 m2 s1 for B (Boudreau, 1997). dispersivity and water content are estimated, the CDE simulates
Also, we consider the 2-site and 1-site model for the description the breakthrough curve well (Fig. 1a CDE2). The effective water
of B transport because it is possible the dominant process is chem- content is estimated to be 0.217, which is 86% of the measured
ical non-equilibrium instead of physical non-equilibrium. water content (0.253). This can be due to measurement error
and/or presence of immobile water. The estimated standard errors
Results and discussions are reasonably narrow (Table 1), but the tail is not fully captured
(Fig. 1).
Example 1 transport of 2H2O in a large undisturbed lysimeter The MIM model fit (Fig. 1b MIM1) appears to be identical to
CDE2 fit. The mobile water fraction is estimated to be 0.86, which
As shown by Schoen et al. (1999b), the CDE fails to simulate the is the same as the effective water content of the CDE2 model (Table
breakthrough curve when only dispersivity is estimated (Fig. 1a 1). The estimated x is essentially zero, while the standard error for

(a) HC pb CDE (b) HC pb MI M

Br, obs Br, obs
1
PIPES, obs
1 PIPES, obs
Reduced concentration

Br, fit, sep. Br, fit, sep.

Reduced concentration
Br, fit, sim. Br, fit, sim.
0.8 PIPES, fit sep. 0.8 PIPES, fit sep.
PIPES fit sim. PIPES fit sim.

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 1 2 3 4 5 0 1 2 3 4 5
Effluent pore volume Effluent pore volume

(c) HC xb CDE (d) HC xb MIM

Br, obs Br, obs
1 PIPES, obs 1 PIPES, obs
Reduced concentration
Reduced concentration

Br, fit, sep. Br, fit, sep.

Br, fit, sim. Br, fit, sim.
0.8 PIPES, fit sep. 0.8 PIPES, fit sep.
PIPES fit sim. PIPES fit sim.

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 1 2 3 4 5 0 1 2 3 4 5
Effluent pore volume Effluent pore volume

Fig. 8. Observations (symbols) and fits (lines) using CDE (equilibrium convection dispersion equation model) and MIM (non-equilibrium mobile–immobile water model) for
Br and PIPES in Example 3, where ‘‘sep.” = separate fit of tracers and ‘‘sim.” = simultaneous fit of tracers.

Table 3
Parameter estimates for Example 3.

k (cm) b g RMSE R2 |q|max

pb CDE Br 4.599 ± 0.172 0.025 0.995
CDE PIPES 6.413 ± 0.319 0.035 0.989
CDE Br + PIPES 5.438 ± 0.194 0.035 0.989
MIM Br 3.172 ± 0.608 0.877 ± 0.064 0.127 ± 0.107 0.022 0.996 0.982
MIM PIPES 3.536 ± 0.667 0.818 ± 0.052 0.456 ± 0.240 0.028 0.993 0.971
MIM Br + PIPES 3.336 ± 0.228 0.819 ± 0.021 0.413 ± 0.086 0.027 0.994 0.864
xb CDE Br 0.634 ± 0.037 0.030 0.994
CDE PIPES 0.734 ± 0.061 0.049 0.993
CDE Br + PIPES 0.568 ± 0.032 0.042 0.994
MIM Br 0.449 ± 0.085 0.947 ± 0.030 0.143 ± 0.144 0.027 0.995 0.981
MIM PIPES 0.463 ± 0.046 0.936 ± 0.010 0.240 ± 0.094 0.032 0.997 0.911
MIM Br + PIPES 0.459 ± 0.029 0.937 ± 0.007 0.205 ± 0.054 0.029 0.997 0.894
574 G. Tang et al. / Journal of Hydrology 376 (2009) 567–578

Sensitivity analysis shows that the prediction is very sensitive

1 Obs to water content, but not to dispersivity or the mass transfer coef-
CDE
ficient (Fig. 3). This is quantified by the composite scaled sensitiv-
MIM
ity of 0.54 (h), 0.31 (t0), 0.10 (k), 0.57 (b) and 0.02 (x) for the MIM2
Reduced concentration

0.8
estimate. The sensitivities based on CDE2, CDE3 and MIM1 esti-
mates are similar because both the parameters and fits are close
0.6 to that of the MIM2 model. The insensitivity of the mass transfer
coefficient to the observations leads to a large uncertainty in the
x estimate. The high sensitivity of water content results in the fail-
0.4
ure of the CDE1 model because the water content is fixed at a level
significantly different from the effective water content. Therefore,
0.2 when the uncertainty of these experimentally controlled parame-
ters (h and t0) is suspect, estimating these parameters may be
appropriate.
0
0 1 2 3 4 5 6 7 8 Consideration of uncertainty in water content in transport
Effluent pore volume experiments has been reported in the literature. Effective water
content is calculated using moment analysis in Padilla et al.
Fig. 9. Observations (symbols) and fits (lines) using CDE (convection dispersion (1999) and fixed in breakthrough curve fitting, while mean water
equation model) versus MIM (mobile–immobile model) of 3H2O breakthrough velocity is estimated in Porro et al. (1993), Padilla et al. (1999)
curve for Example 4. The observation data are taken from Toride et al. (1995).
and Ellsworth et al. (1996). Estimating mean water velocity is
equivalent to estimating water content (Schoen et al., 1999a,b)
when dimensional time is used or dimensionless time is calculated
x is probably significant enough to suspect the reliability of this fit. using T = vt/L. When the dimensionless time is calculated from
The x estimate by Schoen et al. (1999b) is 0.086. Monte Carlo sim- accumulated effluent volume over total pore water volume
ulation of the model with ln(k), b and ln(x) uniformly randomly T = Veffluent/Vporewater (De Smedt and Wierenga, 1984; van Genuch-
distributed between 0 and 3, 0.8 and 1.0, and 10 and 2 suggests ten and Wierenga, 1977), estimating velocity is equivalent to esti-
a range of x values from 0 to 1 can yield a similar RMSE of 0.01 mating water content when the total pore water volume (Vporewater)
(Fig. 2). The mass transfer coefficient is therefore not well-deter- is calculated based on the effective water content instead of the
mined from the breakthrough curve data, even though the esti- measured mean water content. More often, the retardation factor
mated maximum correlation is less than 0.6 (Table 1). The is estimated (van Genuchten and Wierenga, 1977; Nkedi-Kizza
underestimation of the uncertainty by the nonlinear least squares et al., 1983; Wierenga and van Genuchten, 1989; Anamosa et al.,
is probably due to the inaccuracy of the first-order approximation 1990; Porro et al., 1993; Langner et al., 1999; Comegna et al.,
and underestimation of measurement error as RMSE. 2001; Kamra et al., 2001). Eq. (4) clearly shows that estimating
The mass recovery estimated by moment analysis is about 105% retardation factor and estimating water content are equivalent. A
(Schoen et al., 1999a). This can be due to a variation in the flux rate retardation factor less than 1 implies kd < 0, which is not physically
or concentration during the pulse duration. Since the experiments possible. It may indicate presence of stagnant water, anion exclu-
lasted for more than 50 days, it may not be trivial to maintain a sion or overestimation of the water content. An estimated effective
constant injection rate. Because mass balance error may have ad- water content greater than actual water content, on the other
verse impact on parameter estimation (Toride et al., 1995; van hand, may suggest retardation.
Genuchten, 1981), the pulse duration is estimated. The match re- Even though Schoen et al. (1999b) realize that 2-parameter CDE
mains about the same as for the CDE model (Fig. 1a CDE3) while (CDE2) would be satisfactory if neglecting the tail, the MIM is se-
the MIM model describes the tail better (Fig. 1b MIM2). The lected for further comparison with a convective lognormal transfer
parameter estimates are reasonable, except the correlation be- function model. They concluded that the MIM model performs bet-
tween b and x is high (0.934) for MIM. Monte Carlo simulation ter than CDE2. Our analysis suggests that the uncertainty in x is
confirms that the uncertainty in x is still very large for the too large to be physically meaningful.
MIM2 model, similar as for MIM1 (Fig. 2). The estimated pulse
duration is 102% and 109% of the actual pulse duration (Table 1)
for CDE3 and MIM2, respectively. In addition to the uncertainty Example 2 transport of 3H2O through an aggregated, gravelly Oxisol
in x estimate, the fact that MIM2 describes the tailing well may
not be very significant because the uncertainty in the tailing part Similar to Anamosa et al. (1990), both CDE and MIM models can
is relatively high due to low relative isotope concentrations simulate breakthrough curves reasonably well (Fig. 4). This is also
(Schoen et al., 1999a,b). Therefore, the RMSE of 0.007 may be far confirmed by an R2 close to 1 for all of the fits (Table 2). For the fast
less than the real experimental error. Further, there is no evidence flow cases, the MIM model matches the observations better than
to support the larger estimate of pulse duration. CDE, and estimating water content appears to improve the CDE

Table 4
Parameter estimates for Example 4 (physical non-equilibrium model).

k hm/h g f RMSE R2 |q|max

3
CDE1 ( H2O) 1.339 ± 0.088 0.030 0.995
MIM3 (3H2O) 0.415 ± 0.094 0.822 ± 0.028 0.414 ± 0.115 0.015 0.999 0.981
MIM2 (B) 0.415 0.822 0.806 ± 0.097 0.495 ± 0.019 0.055 0.970 0.756
MIM3 (B) 1.305 ± 0.438 0.822 0.534 ± 0.143 0.587 ± 0.050 0.048 0.978 0.925
MIM5 (B + 3H2O) 0.848 ± 0.173 0.913 ± 0.040 0.132 ± 0.118(3H2O) 0.512 ± 0.018 0.036 0.957
0.644 ± 0.085 (B) 0.991
MIM4 (B + 3H2O) 0.733 ± 0.171 0.824 ± 0.036 0.647 ± 0.091 0.537 ± 0.017 0.038 0.989 0.872
 G. Tang et al. / Journal of Hydrology 376 (2009) 567–578 575

4 (a) 0.1
λ
3 β
η
0.08
2
Scaled sensitivity

1 0.06

RMSE
0
0.04
−1

0.02
−2

−3
0
−2 −1 0
10 10 10
−4 λ (cm )
0 1 2 3 4 5 6 7 8
Effluent pore volume
(b) 0.1
Fig. 10. Sensitivity for MIM for 3H2O breakthrough curve for Example 4 where
k = dispersivity, b = mobile water fraction and x = mass transfer coefficient.
0.08

match. This is quantitatively shown in the reduction in RMSE as

more parameters are estimated (Table 2). For the slow flow cases, 0.06

RMSE
the improvement is not significant as more parameters are
estimated. 0.04
The largest absolute correlation coefficient is greater than 0.9
for all MIM estimates. For the slow flow cases, the uncertainty in
k (coefficient of variation of 71% and 32% for Expt. I-3 and Expt. 0.02
II-1) estimates is large (Table 2). Monte Carlo simulation shows
that the MIM model matches observations equally well for Expt. 0
I-3 within a range of k from 0.1 to 10 and b from 0.2 to 0.5 because 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1
the response surface of RMSE is flat (Fig. 5). This suggests that the β
MIM model is not well-defined by the observations and the CDE
model describes observations sufficiently well for the slow flow (c) 0.1
cases.
It is more difficult to discriminate between the CDE1 and the 0.08
CDE2 models based on the matches for the slow flow cases
(Fig. 4). When water content is estimated, the RMSE decreases
slightly (Table 2), while k estimate is statistically the same. The 0.06
RMSE

estimated water content is 107% and 105% of the saturated water

content for the Expt. I-3 and Expt. II-1 for CDE. Unless there is evi- 0.04
dence to support greater adsorption (retardation) of 3H2O in the
experiment, or the actual porosity is greater than 0.527, CDE1 de-
scribes the slow flow cases sufficiently well. 0.02
In contrast, the CDE1 model does not describe the fast flow case
very well in Expt. I-1 (Fig. 4). The estimated water contents (0.373 0
−2 −1 0 1
for Expt. I-1 and 0.438 for Expt. II-2) are less than the saturated 10 10 10 10
water content (0.527), suggesting only a fraction of pore water ac- ω
tively participates in the transport. Consequently the MIM model 3
Fig. 11. Monte Carlo simulation results for H2O transport for MIM for Example 4. k
may be necessary for the fast-flow experiments. The response sur-
(dispersivity), b (mobile water fraction) and x (mass transfer coefficient) are
face of k, b, and x is less flat for Expt. I-1 with the fastest flow case sampled from uniformly distributed random numbers between 6–2, 0.6–1, and
(Fig. 6) than that for Expt. I-3 with slow flow (Fig. 5). As long as the 3–4 for ln(k), b and ln(x). The red dot represents the optimized value. (For
experimental error is less than or at the level of 0.02 for the relative interpretation of the references to colour in this figure legend, the reader is referred
to the web version of this article.)
concentration measurements, the parameter estimates are reliable
even though the largest absolute correlation coefficient is as high
as 0.94 (Table 2).
Scaled sensitivity for the fast flow case shows that model pre- and x. Consequently the MIM model is not unequivocally defined
dictions are sensitive to all three parameters (k, b, and x, by the observations for the slow flow experiments.
Fig. 7a), and the sensitivity vectors (e.g., J1j, J2j, . . ., JNj, for parameter Anamosa et al. (1990) recognize that the equilibrium assump-
j, shown as curves in Fig. 7a) are linearly independent so that the tion may be satisfied for the slow flow cases. Similar to our results,
curves appear quite different from each other. Therefore, the the uncertainty in their estimate of x is large for Expt. I-3 even
MIM model is well defined by the observations, and parameter when b is fixed at 0.53, which is their estimate for the fastest flow
uncertainty is small. For the slow flow case, even though the pre- case. But the large uncertainty is ignored when the x estimate is
dictions are sensitive to the three parameters, the curves for b, used in the development of a linear relation between mass transfer
and x appear similar. The correlation between their sensitivity coefficient and flow rate. The large uncertainty in k and x for both
vectors is 0.95, resulting in very high parameter uncertainty for b slow flow cases shows that they can not be reliably quantified from
576 G. Tang et al. / Journal of Hydrology 376 (2009) 567–578

Obs
immobile water more slowly than Br during the infiltration
3
1 H2O Separate fit, 5 parameters process. Therefore, breakthrough is earlier and washout is later
Simultaneous fit, 5 parameters compared to Br.
Simultaneous fit, 4 parameters
The CDE model is able to simulate the breakthrough curves sep-
Reduced concentration

Simultaneous fit, 3 parameters

0.8
arately or simultaneously with Eq. (8), and the k estimate and the
match (Fig. 8) are similar (Table 3 CDE Br and CDE PIPES) whether
0.6 the two breakthrough curves are fitted separately or simulta-
neously. However, the CDE model does not simulate the tracer sep-
aration well even with fitting the two tracers separately. Separate
0.4
MIM fits of the two tracers capture the tracer separation and result
Boron in an improved match over CDE. However, the standard error for g
0.2 estimate is large and the correlation coefficient is close to 1 (Table
3 MIM Br and MIM PIPES). Simultaneous fits of the two tracers,
assuming the same dispersivity, mobile water fraction, and Eq.
0 (12) for mass transfer coefficient, result in a reduction in parameter
0 2 4 6 8 10 12 14 16 18 20
Effluent pore volume uncertainty (by half) and correlations (e.g., from 0.97 to 0.86 for pb
case, Table 3 MIM Br + PIPES), and the tracer separation is
Fig. 12. Observations (symbols) versus model fits (lines) for Example 4. described well.
In addition, the parameter estimate (k and g) can be used for the
model prediction of any other tracers as long as the molecular dif-
the data using nonlinear least squares method, which is further fusion coefficient is known. On the contrary, separately fitting the
confirmed by Monte Carlo analysis. two tracers will result in different k, b, and x estimate for each tra-
cer. It is therefore less certain to extrapolate these parameter esti-
mates for the prediction of other tracers, and the uncertainty is
Example 3 transport of Br and PIPES through intact Hanford sediment expected to increase because of the larger number of estimated
cores parameters.

In this example, the breakthrough curves for both nonreactive

tracers appear to be very close (Fig. 8). It can be seen that the PIPES Example 4 transport of 3H2O and B through Glendale clay loam
concentration is constantly higher than that of Br during the break-
through while the opposite is true during washout. This tracer sep- While the MIM model can simulate the 3H2O breakthrough and
aration is attributed to non-equilibrium processes by Mayes et al. tailing somewhat better than the CDE model (Fig. 9), the correla-
(2003) and Langner et al. (1999), i.e., the molecular size of PIPES tion is 0.94 between k and hm/h, and 0.98 between hm/h and g
is greater than that of Br, and therefore PIPES diffuses into the (which is used instead of x to facilitate comparisons in this

(a) 0.1 (b) 0.1

0.08 0.08

0.06 0.06
RMSE

RMSE

0.04 0.04

0.02 0.02

0 0
0 1 2 3 4 0.6 0.7 0.8 0.9 1
λ (cm) θm/θ

(c) 0.1 (d) 0.1

0.08 0.08

0.06 0.06
RMSE

RMSE

0.04 0.04

0.02 0.02

0 −2 0 2
0
10 10 10 0 0.2 0.4 0.6 0.8 1
η f

Fig. 13. Monte Carlo simulation results for simultaneous fit with MIM (4-parameter model) for Example 4. The red dot represents the optimized value. (For interpretation of
the references to colour in this figure legend, the reader is referred to the web version of this article.)
 G. Tang et al. / Journal of Hydrology 376 (2009) 567–578 577

Table 5
Parameter estimates for Example 4 (chemical non-equilibrium model).

k a f RMSE R2 |q|max
2-Site (B) 1.305 ± 0.438 0.430 ± 0.075 0.526 ± 0.050 0.048 0.978 0.889
2-Site (B + 3H2O) 1.336 ± 0.110 0.427 ± 0.047 0.528 ± 0.021 0.039 0.989 0.678

example), the coefficient of variation is close to 28% for g, and 23%

for k (Table 4, MIM3 3H2O). The large uncertainty in g is due to its
insensitivity (Fig. 10), with composite scaled sensitivity of 1.297 (a) 0.1
(k), 0.517 (hm/h), 0.049 (g), as well as the high correlations among
the three parameters. Monte Carlo simulation shows that the re- 0.08
sponse surface is flat for all of the three parameters (Fig. 11). Un-
less the experiment and model error is less than 2%, the
0.06

RMSE
uncertainty in the parameter estimate for MIM is quite high. Fixing
hm/h and k at the estimate from 3H2O in B breakthrough curve fit-
ting can be subject to question because the impact of the uncer- 0.04
tainty in hm/h and k on B is not considered.
With the assumptions of hm/h and k to be the same for 3H2O and
B following Toride et al. (1995), simultaneous fits of the two tracers 0.02
with five parameters (k, hm/h, g(3H2O), g(B), and f (B)) result in
similar matches (Fig. 12 and 5-parameter). The standard error of 0
g for 3H2O is large because the data for B do not contribute new 0 0.5 1 1.5 2 2.5 3 3.5 4
information. The correlation between hm/h and g remains high λ (cm)
(0.96, Table 4 MIM5)). With the assumption of Eq. (10), the uncer-
tainty of parameter estimates and correlations are significantly re- (b) 0.1
duced (Table 4 MIM4) while the match appears almost the same in
the 4-parameter fit (Fig. 12, k, hm/h, g, and f (B)). It should be noted 0.08
that the uncertainty in k and hm/h remains high, e.g., the coefficient
of variation for k is 23% (Table 4), and hm/h between 0.7 and 1 result
in RMSE of about 0.04 as shown in Fig. 13. 0.06
RMSE

An alternative explanation is that the adsorption of B is kinetic.

Kinetic sorption can be simulated using either one-site or two-site 0.04
non-equilibrium model. The one-site model does not work for this
example (not shown). Because the two-site non-equilibrium model
is equivalent to the MIM model with only different parameteriza- 0.02
tions (k, hm/h, x, and f for MIM and k, a, and f for 2-site model), the
match and k estimate are the same when only B data are consid- 0
−2 −1 0 1
ered (Table 4 MIM3). Simultaneously fitting of B with the two-site 10 10 10 10
model and 3H2O with the CDE model results in a reduction in the α
standard error and correlations for all estimated values (Table 5
and 2-Site (B + 3H2O), 3-parameter fit). Monte Carlo analysis con- (c) 0.1
firms that uncertainty for the three parameters is relatively small,
and the fitting error is sensitive to the three parameters (Fig. 14).
0.08
Comparing with the simultaneous fit of both tracers using MIM
(4-parameter), the number of estimated parameters reduces from
4 (k, hm/h, g, and f) to 3 (k, a, and f) while the RMSE only increases 0.06
RMSE

slightly from 0.038 to 0.039. The match appears to be very close to

others (Fig. 12and 3-parameter fit). With less assumption, smaller
0.04
parameter estimate uncertainty and correlation, and a match very
close to other models, this model description is acceptable and
preferable considering the principle of parsimony. 0.02
Toride et al. (1995) note that simultaneous estimates of k, b and
x result in different estimates with various initial guesses for B 0
breakthrough curve fitting, and use the dispersion coefficient and 0 0.2 0.4 0.6 0.8 1
b value estimated from 3H2O breakthrough curve in B break- f
through curve fitting. Our analysis clearly shows that the uncer-
Fig. 14. Monte Carlo simulation results for simultaneous fit with MIM (3-parameter
tainty in k estimate from 3H2O data is quite significant, and that
model) for Example 4. The red dot represents the optimized value. (For interpre-
separate fitting of the two tracers may significantly underestimate tation of the references to colour in this figure legend, the reader is referred to the
the parameter uncertainty, which may lead to over-confidence web version of this article.)
when these estimates are used for model predictions. Simulta-
neous fitting of the tracers helps maintain a comprehensive model relation while simulating the breakthrough curve data reasonably
(Hill, 1998), however, the parameter uncertainty remains high well. It appears that the data in this example do not have sufficient
even when the model is further simplified (number of parameters information content to unequivocally prove the significance of
reduced). A further simplified chemical non-equilibrium model physical non-equilibrium processes (mobile–immobile mass trans-
significantly reduces the parameter estimate uncertainty and cor- fer). A chemical non-equilibrium model may be inferred from the
578 G. Tang et al. / Journal of Hydrology 376 (2009) 567–578

data. Even though the predictions in all of these cases match the Brusseau, M., Jessup, R., Rao, P., 1989. Modeling the transport of solutes influenced
by multiprocess nonequilibrium. Water Resources Research 25 (9), 1971–1988.
observations reasonably well, the parameter estimates can be
Burnham, K.P., Anderson, D.R., 2004. Multimodel inference: understanding AIC and
quite different. Comparing these estimates improves our under- BIC in model selection. Sociological Methods and Research 33 (2), 261–304.
standing of the data and the influential processes. Comegna, V., Coppola, A., Sommella, A., 2001. Effectiveness of equilibrium and
physical non-equilibrium approaches for interpreting solute transport through
undisturbed soil columns. Journal of Contaminant Hydrology 50, 121–138.
Summary and conclusion Das, B.S. et al., 2005. Evaluation of mass recovery impacts on transport parameters
using least-squares optimization and moment analysis. Soil Science Society of
America Journal 69, 1209–1216.
Through analysis of four sets of column experiments in the liter- De Smedt, F., Wierenga, P.J., 1984. Solute transfer through columns of glass beads.
ature, we illustrate that the application of sensitivity/uncertainty Water Resources Research 20 (2), 225–232.
analysis, model evaluation and comparison can significantly im- Ellsworth, T.R., Shouse, P.J., Skaggs, T.H., Jobes, J.A., Fargerlund, J., 1996. Solute
transport in unsaturated soil: experimental design, parameter estimation, and
prove interpretation of column experiments, and potentially model
model discrimination. Soil Science Society of America Journal 60 (2), 397–407.
predictions. Because model predictions are very sensitive to the Fetter, C.W., 2001. Applied Hydrogeology. Prentice Hall, Upper Saddle River, NJ.
water content, measurement error in water content, presence of Hill, M.C., 1998. Methods and guidelines for effective model calibration, 98-4005,
US Geological Survey, Denver, Colorado.
stagnant water, and/or other mechanisms such as anion exclusion
Hu, Q., Brusseau, M.L., 1995. Effect of solute size on transport in structured porous
can cause failure of the equilibrium convection dispersion model media. Water Resources Research 31 (7), 1637–1646.
in simulating breakthrough curves from many column experiments. Huang, K., Toride, N., van Genuchten, M.T., 1995. Experimental investigation of
Estimating water content can help rectify poor matches between the solute transport in large, homogeneous and heterogeneous, saturated soil
columns. Transport in Porous Media 18, 283–302.
equilibrium model and effluent curves. This may also be applicable Kamra, S.K., Lennartz, B., Van Genuchten, M.T., Widmoser, P., 2001. Evaluating non-
to other sensitive parameters such as velocity, and pulse volume/ equilibrium solute transport in small soil columns. Journal of Contaminant
duration. Hydrology 48, 189–212.
Knopman, D.S., Voss, C.I., 1988. Discrimination among one-dimensional models of
Even though the non-equilibrium convection dispersion model solute transport in porous media: implications for sampling design. Water
is capable of capturing early breakthrough and long tail of break- Resources Research 24 (11), 1859–1876.
through curves, the parameter estimates are often nonunique Koch, S., Fluhler, H., 1993. Solute transport in aggregated porous media: comparing
model independent and dependent parameter estimation. Water, Air, and Soil
and/or the uncertainty and correlation in the parameter estimate Pollution 68, 275–289.
are excessively large. This can be due to the insensitivity of model Langner, H.W., Gaber, H.M., Wraith, J.M., Huwe, B., Inskeep, W.P., 1999. Preferential
prediction to the mass transfer coefficient and/or high correlations flow through intact soil cores: effects of matric head. Soil Science Society of
America Journal 63, 1591–1598.
among parameters. In these cases, the non-equilibrium model may
Maraqa, M.A., Wallace, R.B., Voice, T.C., 1997. Effects of degree of water saturation
be reduced to an equilibrium convection dispersion equation mod- on dispersivity and immobile water in sandy soil columns. Journal of
el to avoid over-parameterization. Simultaneous fit of multiple Contaminant Hydrology 25, 199–218.
Mayes, M.A., 2006. Solute transport and U(VI) reactivity in natural heterogeneous
tracers with fewer calibration parameters is able to reduce uncer-
sediments. PhD Thesis, University of Tennessee, Knoxville Tennessee.
tainty and correlation in the parameter estimates, and increase Mayes, M.A., Jardine, P.M., Mehlhorn, T.L., Bjornstad, B.N., Ladd, J.L., Zachara, J.M.,
confidence in data interpretation as well as model predictions. 2003. Transport of multiple tracers in variably saturated humid region
Our analysis highlights that the common practice of ignoring structured soils and semi-arid region laminated sediments. Journal of
Contaminant Hydrology 275, 141–161.
parameter uncertainty and high correlation in the column experi- Millington, R.J., Quirk, J.P., 1961. Permeability of porous solids. Transactions of the
ment literature may not be always appropriate, particularly when Faraday Society 57, 1200–1207.
more complicated models are used. Because multiple choices of Motulsky, H.J., Ransnas, L.A., 1987. Fitting curves to data using nonlinear regression,
a practical and non-mathematical review. FASEB 1, 365–374.
calibration parameters, different parameterizations and different Nkedi-Kizza, P. et al., 1983. Modeling tritium and chloride 36 transport through an
models may be equally able to simulate column experiment data, aggregated Oxisol. Water Resources Research 19 (3), 691–700.
it is necessary to evaluate and compare different models to reliably Nkedi-Kizza, P. et al., 1984. On the equivalence of two conceptual models for
describing ion exchange during transport through an aggregated Oxisol. Water
identify and quantify the processes, and thereby improve column Resources Research 20 (8), 1123–1130.
experiment data interpretation and model prediction. Pace, M.N. et al., 2003. Quantifying the effects of small-scale heterogeneities on flow
and transport in undisturbed cores from the Hanford formation. Vadose Zone
Journal 2, 664–676.
Acknowledgements
Padilla, I.Y., Yeh, T.-C.J., Conklin, M.H., 1999. The effect of water content on solute
transport in unsaturated porous media. Water Resources Research 35 (11),
We acknowledge the DOE Office of Biological and Environmen- 3303–3313.
Parker, J.C., van Genuchten, M.T., 1984. Determining transport parameters from
tal Research for support of this research through the Environmen-
laboratory or field tracer experiments. Bull. 84-3, Va. Agric. Exp. St., Blacksburg.
tal Remediation Sciences Program. We acknowledge Frederick M. Porro, I., Wierenga, P., Hills, R., 1993. Solute transport through large uniform and
Mann of the Tank Farm Vadose Zone Group and CH2 M Hill Han- layered soil columns. Water Resources Research 29 (4), 1321–1330.
ford Group, Inc. for providing support for the IDF work under con- Schoen, R., Gaudet, J.P., Bariac, T., 1999a. Preferential flow and solute transport in a
large lysimeter, under controlled boundary conditions. Journal of Hydrology
tract with the US DOE. Oak Ridge National Laboratory is managed 215, 70–81.
by the University of Tennessee-Battelle, LLC, under contract DE- Schoen, R., Gaudet, J.P., Elrick, D.E., 1999b. Modelling of solute transport in a large
AC05-00OR22725 with the US DOE. undisturbed lysimeter, during steady-state water flux. Journal of Hydrology
215, 82–93.
Toride, N., Leij, F.J., van Genuchten, M.T., 1995. The CXTFIT code for estimating
References transport parameters from laboratory or field tracer experiments. US Salinity
Laboratory.
Anamosa, P.R., Nkedi-kizza, P., Blue, W.G., Sartain, J.B., 1990. Water movement van Genuchten, M.T., 1981. Nonequilibrium transport parameters for miscible
through an aggregated, gravelly Oxisol from Cameroon. Geoderma 46, 263–281. displacement experiments. Research Report 119. US Salinity Laboratory.
Bajracharya, K., Barry, D.A., 1997. Nonequilibrium solute transport parameters and van Genuchten, M.T., Wierenga, P.J., 1977. Mass transfer studies in sorbing porous
their physical significance: numerical and experimental results. Journal of media: II. Experimental evaluation with tritium (3H2O). Soil Science Society
Contaminant Hydrology 24, 185–204. American Journal 41, 272–278.
Bard, Y., 1974. Nonlinear Parameter Estimation. Academic Press, New York. Wang, J.H., Robinson, C.V., Edelman, I.S., 1953. Self-diffusion and structure of liquid
Beven, K.J., Henderson, D.E., Reeves, A.D., 1993. Dispersion parameters for water. III. Measurement of the self-diffusion of liquid water with H2, H3, and O18
undisturbed partially saturated soil. Journal of Hydrology 143, 19–43. as tracers. Journal of the American Chemical Society 75, 466–470.
Bond, W.J., Wierenga, P.J., 1990. Immobile water during solute transport in Wellman, D.M., Gamerdinger, A.P., Kaplan, D.I., Serne, R.J., 2008. Effect of particle-
unsaturated sand columns. Water Resources Research 26 (10), 2475–2481. scale heterogeneity on Uranium(VI) transport in unsaturated porous media.
Boudreau, B., 1997. Diagenetic Models and their Implementation. Springer, Berlin. Vadose Zone Journal 7 (1), 67–78.
Brusseau, M., 1992. Nonequilibrium transport of organic chemicals: the impact of Wierenga, P.J., Van Genuchten, M.T., 1989. Solute transport through small and large
pore-water velocity. Journal of Contaminant Hydrology 9, 353–368. unsaturated columns. Ground Water 27 (1), 35–42.