You are on page 1of 34

JID: ADWR

ARTICLE IN PRESS [m5GeSdc;June 29, 2018;16:56]

Advances in Water Resources 000 (2018) 1–34

Contents lists available at ScienceDirect

Advances in Water Resources


journal homepage: www.elsevier.com/locate/advwatres

Determination of the diffusivity, dispersion, skewness and kurtosis in


heterogeneous porous flow. Part II: Lattice Boltzmann schemes with
implicit interface
Irina Ginzburg
Irstea, Antony Regional Centre, HBAN, 1 rue Pierre-Gilles de Gennes CS 10030, Antony Cedex 92761, France

a r t i c l e i n f o a b s t r a c t

Keywords: A simple local two-relaxation-time Lattice Boltzmann numerical formulation (TRT-EMM) of the extended method
Taylor dispersion of moments (EMM) is proposed for analysis of the spatial and temporal Taylor dispersion in d-dimensional
Skewness streamwise-periodic stationary mesoscopic velocity field resolved in a piecewise-continuous porous media. The
Kurtosis
method provides an effective diffusivity, dispersion, skewness and kurtosis of the mean concentration profile
Spatial and temporal moments
and residence time distribution. The TRT-EMM solves a chain of steady-state heterogeneous advection–diffusion
Residence-time distribution
Extended method of moments equations with the pre-computed space-variable mass-source and automatically undergoes diffusion-flux jump
Lattice Boltzmann ADE schemes on the abrupt-porosity streamwise-normal interface. The temporal and spatial systems of moments are computed
Interface jump within the same run; the symmetric dispersion tensor can be restored from independent computations performed
Numerical diffusion for each periodic mean-velocity axis; the numerical algorithm recursively extends for any order moment.
Stokes–Brinkman–Darcy porous flow We derive an exact form of the bulk equation and implicit closure relations, construct symbolic TRT-EMM
solutions and determine specific relation between the equilibrium and the collision degrees of freedom view-
ing an exact parameterization by the physical non-dimensional numbers in two alternate situations: “parallel”
fracture/matrix flow and “perpendicular” Darcy flow through porous blocks in “series”. Two-dimensional simu-
lations in linear Brinkman flow around solid and through porous obstacles validate the method in comparison
with the two heterogeneous direct LBM-ADE schemes with different anti-numerical-diffusion treatment which are
proposed and examined in parallel. On the coarse grid, accuracy of the three moments is essentially determined
by the free-tunable collision rate in all schemes, and especially TRT-EMM. However, operated within a single
periodic cell, the TRT-EMM is many orders of magnitude faster than the direct solvers, numerical-diffusion free,
more robust and much more capable for accuracy improving, high Péclet range and free-parameter influence
reduction with the mesh refinement. The method is an efficient predicting tool for the Taylor dispersion, asym-
metry and peakedness; moreover, it allows for an optimal analysis between the mutual effect of the flow regime,
Péclet number, porosity, permeability and obstruction geometry.

1. Introduction sification with respect to the dispersion and non-Gaussian properties of


the breakthrough curves.
Understanding of the structure and velocity influence on the mass Within the dispersion theory, established in fundamental works by
transport, prediction and optimization of the Taylor and environmen- Taylor (1953), Aris (1956) and Brenner (1980), the Gaussian description
tal dispersion, elongated tails of the averaged (upscaled) solute distri- applies to upscaled distributions with the longitudinal Taylor dispersion
butions and their time-rate, the residence-time distribution RTD, is a correction (Aris, 1956; Taylor, 1953) to molecular diffusion coefficient
task required in many engineering fields, such as chemical, polymer, [hereafter, DT and 0 , respectively] or, more generally, full dispersion
petrol, agricultural, ecological risk assessment and restoration, wastew- tensor (Brenner, 1980) due to the multi-dimensional gradients in
ater treatment. We propose a simple numerical method for prediction velocity field. The classical approach is focused on the spatial solute
of the first four moments of the solute distribution from the steady-state evolution after an instantaneous point release, when the distribution
velocity field established in the streamwise-periodic representative unit moments are computed via the spatial integration. The RTD introduced
cell. Especially, we keep in mind a direct application in X-ray micro- by Danckwerts (1953) is commonly monitored in the outlet of the chem-
tomography images of the double porosity media, like the carbonates, ical device (Cozewith and Squire, 2000), vegetation zone for pollutant
where the proposed method allows for the rock identification and clas-

E-mail address: irina.ginzburg@irstea.fr

https://doi.org/10.1016/j.advwatres.2018.05.006
Received 9 March 2018; Received in revised form 7 May 2018; Accepted 8 May 2018
Available online xxx
0309-1708/© 2018 Elsevier Ltd. All rights reserved.

Please cite this article as: I. Ginzburg, Advances in Water Resources (2018), https://doi.org/10.1016/j.advwatres.2018.05.006
JID: ADWR
ARTICLE IN PRESS [m5GeSdc;June 29, 2018;16:56]

I. Ginzburg Advances in Water Resources 000 (2018) 1–34

Fig. 1. Sketches for the heterogeneous obstacles or a composite material filled with the porous media, and a porous/solid/vegetation arrangement for dispersion
reduction within an unit periodic cell V.

remediation (Werner and Kadlec, 2000) or micro-channel (Cantu-Perez ideas (Vikhansky, 2008) and substantial developments (Vikhansky and
et al., 2008), and it can be characterized through the temporal moments Ginzburg, 2014), the EMM is elaborated (Ginzburg and Vikhansky,
integrated in time (Vikhansky, 2011). The Taylor regime is respectively 2018) in the form of the recursive algorithm for prediction of (i) ef-
characterized by a time and space linear growing of the variance, fective diffusion (structure) coefficient, (ii) Taylor dispersion dyadic,
like 𝜎 2 = 2(0 + 𝐷𝑇 )𝑡 at a dimensionless time 𝑡′ = 𝑡0 ∕ ≥≈ Pe, or and (iii) longitudinal coefficients of the high-order moments. Differ-
𝜎 2 = 20−3 (0 + 𝐷𝑇 )𝑥 at a dimensionless distance 𝑥′ = 𝑥∕ ≥≈ Pe ently from the Brenner’s averaging of the Brownian particles moments or
0  the volume-averaging approach (Valdés-Parada et al., 2016), the EMM
[hereafter, Pe = 0
is the characteristic Péclet number]. However,
searches for the solution of Eq. (1) in the form of a product of the low
many natural systems exhibit asymmetry, peakedness and heavy tails, a
frequency, slow monochromatic wave and a streamwise-periodic (say,
long time or distance after release. So far, the Taylor dispersion theory
along the x-axis) oscillating scalar field 𝐵(𝜔, 𝛾; 𝑟⃗):
fails to explain the non-Fickian behavior of the molecular propagators in
heterogeneous porous beds (Berkowitz et al., 2008; Bijeljic et al., 2011), 1
(𝑟⃗, 𝑡) = 𝐵(𝜔, 𝛾; 𝑟⃗) exp [𝑖(𝛾𝑥 − 𝜔𝑡)]. (2)
successfully reproduced by the numerical simulations in double porosity 2𝜋
carbonates (Bijeljic et al., 2013; Yang et al., 2013). Recent studies sug- A simultaneous perturbative expansion is performed either for 𝐵(𝜔(𝛾), 𝑟⃗)
gest that considering the vegetation as a porous zone offers a promising and temporal frequency 𝜔(𝛾), or for 𝐵(𝛾(𝜔), 𝑟⃗) and wavenumber 𝛾(𝜔); the
prediction of the pollutant retention RTD observed in the experimental mathematical algorithms based upon are referred to as “𝜔-form” and “𝛾-
vegetated channels and pond systems (see review Golzar, 2015). The form”, respectively:
open question lies in the optimal design (Su et al., 2009) of the remedia-
tion zones through their permeability (resistance), porosity and geomet- ∑
∞ ∑

“𝜔-form” ∶ 𝐵(𝜔, 𝛾; 𝑟⃗) = 𝐵𝑛 (𝑟⃗)(𝑖𝛾)𝑛 , 𝜔(𝛾) = −𝑖 𝜔𝑛 (𝑖𝛾)𝑛 , (3a)
ric obstruction. The high-order moments quantify the deviations from
𝑛=0 𝑛=1
the normal distribution: the skewness (Sk, third-order moment) and
kurtosis (Ku, fourth-order moment) are responsible for the asymmetry
and peakedness, respectively. Let us assume that a continuous velocity ∑
∞ ∑

field 𝑢⃗𝜙 (𝑟⃗) is resolved in a heterogeneous piece-wise-continuous poros- “𝛾-form” ∶ 𝐵(𝜔, 𝛾; 𝑟⃗) = 𝐵𝑛 (𝑟⃗)(𝑖𝜔)𝑛 , 𝛾(𝜔) = −𝑖 𝛾𝑛 (𝑖𝜔)𝑛 . (3b)
𝑛=0 𝑛=1
ity distribution 𝜙(𝑟⃗) [we keep in mind a sketch in Fig. 1]. A complete
evolution history needs to solve the d-dimensional advection–diffusion In both formulations, the (B-field) variable 𝐵𝑛 (𝑟⃗) solves a chain of
equation (ADE) for the continuous concentration (𝑟⃗, 𝑡): steady-state ADE with the recursively-built mass-sources. The two sets
𝜕𝑡 (𝜙) + ∇ ⋅ (𝑢⃗𝜙 ) = ∇ ⋅ (𝜙𝐃(0) ⋅ ∇), ∇ ⋅ 𝑢⃗𝜙 = 0, 𝑟⃗ ∈ 𝑉𝜙 . (1) {𝜔n } and {𝛾 n } are determined explicitly from the global mass conserva-
tion solvability condition; they sequentially determine the dispersion,
The Taylor dispersion coefficient DT is the same in the two sys- skewness and kurtosis (𝑛 = 2, 3, 4, respectively) in spatial and temporal,
tems of moments: spatial, ⟨𝜙⟩−1 ⟨𝑥𝑛 𝜙⟩ and temporal, ∫−∞ 𝑡𝑛 𝑃 (𝑥, 𝑡)𝑑𝑡,

respectively, system of moments; the solution procedure straight for-
𝑃 (𝑥, 𝑡) = 𝜕𝑡 ⟨(𝑥, 𝑡)⟩ [the brackets denote averaging over the fluid part wardly extends to any higher-order moment. Since the coefficients of
of a single periodic cell, {V𝜙 } ∈ V in Fig. 1]. However, the higher-order the two expansions are inter-related through simple algebraic formulae,
moments differ in the two systems; their computation requires specific the two sets of moments become determined within the same solution
initial/boundary set-up with the direct ADE solvers of Eq. (1) and path. In a streamwise-uniform duct flow, the EMM moments correspond
leads to a tedious numerical task combining the highly discontinuous (Ginzburg and Vikhansky, 2018) to the (upscaled) mean-concentration
diffusion coefficients in complex interface/boundary geometry with the solution obeying the high-order PDE (Mercer and Roberts, 1990; Ngo-
high Péclet numbers. In a periodic arrangement, the solute evolution Cong et al., 2015) without need to resorting for its solving.
is run through a long series of duplicated cells [V in Fig. 1]; since their The EMM allows for the symbolic moments prediction in continu-
number increases with Pe, the computational time to the Taylor regime ous parameter space. So far, the Taylor dispersion, skewness and kur-
grows as Pe2 , at least. tosis were exemplified (Ginzburg and Vikhansky, 2018; Vikhansky and
The Brenner’s B-method of moments (Brenner, 1980) circumvents Ginzburg, 2014) in (i) parabolic (Poiseuille) profile in a channel and
the problem: it restores the symmetric dispersion tensor 𝐃[𝑑 × 𝑑] = cylindrical capillary, (ii) non-Newtonian power-law flow in a capillary,
0 𝑡
𝑉
⟨∇(𝐁 − 𝑟⃗) ∇(𝐁 − 𝑟⃗)⟩ independently solving d steady-state advection– (iii) shallow profile through different cross-section shapes, (iv) Darcy–
𝜙
diffusion equations for space-periodic vector-variable B[d] inside a sin- Brinkman flow in stratified fracture/matrix layers and (v), a “perpen-
gle cell. The finite-difference scheme (Salles et al., 1993) validated dicular” Darcy flow through porous blocks. These solutions allow to
the B-method in microscopic three-dimensional Stokes flow through estimate the role of the porosity contrast and geometry aspect in the
the regular, fractal, random and reconstructed porous media. A sim- first four moments, also providing their asymptotic Pe-scaling. The ref-
ilar dispersion boundary-value problem was recently parameterized erence EMM solutions give the valuable benchmarks for direct solvers
(Valdés-Parada et al., 2016) with the Reynolds number in slow in- of Eq. (1) and numerical EMM formulations. Furthermore, applying
ertial flow through an uniform soil porosity. The extended method of the EMM decomposition to the effective, fourth-order-accurate mass-
moments (EMM) extends the dispersion procedure to heterogeneous conservation equation of a numerical scheme, the truncation interfer-
soil and any-order spatial or temporal moments. Originated from the ence with the physical moments may become quantified quasi-exactly

2
JID: ADWR
ARTICLE IN PRESS [m5GeSdc;June 29, 2018;16:56]

I. Ginzburg Advances in Water Resources 000 (2018) 1–34

in simple flow (Ginzburg, 2017b). The EMM copes with the station- and Ma, 2009) and the local BF/IBF approaches are inter-related with
ary (exact, numerical or approximate) incompressible velocity field of the help of the TRT collision operator (Ginzburg, 2016; Ginzburg et al.,
any nature. A periodic mesoscopic velocity field in porous beds and 2015b); the advanced boundary treatment is developed in work (Silva
random structures, e.g. computed with the two-relaxation-times (TRT) et al., 2017). The two-dimensional simulations will illustrate the EMM
Lattice Boltzmann (LBM) flow schemes in Newtonian flow (Ginzburg ability for device optimization through reduction of the Taylor disper-
et al., 2015b; Khirevich et al., 2014; Talon et al., 2012; Vogel et al., sion, porosity and Péclet number effect in moments. Once the moments
2015), non-Newtonian fluids (Talon and Bauer, 2013) or Darcy flow are predicted, the profiles can be restored from them, with the entropy-
in space-distributed permeability (Röding et al., 2016), is a natural in- maximization procedure (Ginzburg, 2017b; Vikhansky and Ginzburg,
put for EMM numerical computations. Another promising application 2014) based on the idea (Bandyopadhyay et al., 2004). In principle at
is the shallow flow within the rigid or flexible vegetation interactions least, the larger is the set of the predicted moments, the better is the
(Buxton, 2018; Gac, 2014; Yang et al., 2017). Very different numerical reconstruction, and the shorter is the time and length where it becomes
approaches apply for the effective diffusivity of the composite or porous valid.
structure, e.g., Monte-Carlo simulations (Trinh et al., 2000), numerical The predicted moments and reconstructed profiles will be compared
integration by random walks (Bolle et al., 2000), finite-element (Kim with their spatial and temporal counterparts by solving Eq. (1) directly.
et al., 1987) but also the direct solver (Ginzburg, 2005a; 2007) TRT- The optimal in stability TRT-ADE scheme (Ginzburg et al., 2010) has
ADE, e.g., in regular arrays (Chai et al., 2016a), unsaturated (Genty and been recently applied for uniform porosity dispersion coefficient mea-
Pot, 2014) and saturated (Genty et al., 2017) porous media or porous surement in Herschel–Bulkley non-Newtonian profile (Batôt et al., 2016)
polymer films (Gebäck et al., 2015). The EMM provides not only the ef- and duct flow (Ginzburg and Roux, 2015; Yan et al., 2017), reactive
fective diffusivity but also the associated high-order structure moments transport and bacterial chemotaxis (Yan et al., 2017) [see Ginzburg and
at zero velocity unit-cell simulations in the “𝜔-form”. Roux, 2015 for earlier dispersion LBM]. The pore-scale breakthrough
Our purpose is to propose a operationally simple numerical EMM curves simulations (Yan et al., 2017) in a randomly packed bed column
algorithm and to verify the EMM with its help in multi-dimensional het- confirm the TRT-ADE reliability in comparison with much more com-
erogeneous Stokes–Brinkman–Darcy velocity field. We propose TRT − plicated LBM and alternative solution methods (Yang et al., 2016). The
EMM formulation based on the TRT-ADE model for heterogeneous soil LBM-ADE unlaboredly extends for non-linear mass transport (Chai et al.,
(Vikhansky and Ginzburg, 2014). A subtle point is that the diffusive 2016b; Ginzburg, 2006; Hammou et al., 2011; Servan-Camas and Tsai,
flux −𝜙0 ∇𝐵𝑛 (𝑟⃗) should undergo the prescribed jump values on the 2010). There exist several LBM approaches for mass transfer in space-
streamline-normal interfaces when 𝜙(𝑟⃗) becomes discontinuous. The discontinuous porosity. Similarly as in mesoscopic flow, the interface-
TRT − EMM benefits from main LBM advantages: it locally prescribes implicit schemes with the space-variable collision rates run uniformly
the discontinuous diffusion values 𝜙0 (𝑟⃗) via one of its two relaxation through the whole domain (Wassén et al., 2014). The TRT-ADE linear
rates, and couples the diffusive-flux jump-vector with the advective ve- heterogeneous scheme for Eq. (1) was first employed for EMM validation
locity in local equilibrium distribution. The total advection–diffusive (Vikhansky and Ginzburg, 2014) in stratified flows where it confirmed
flux then automatically obeys the interfacial and solid-surface flux- the predicted deviations from the Gaussian solution in moments and
jump conditions, using the mass-conserving bounce-back reflection on reconstructed shapes. The TRT-ADE is supported by the fourth-order-
an impenetrable boundary. It should be recognized that such an im- accurate truncation analysis and stability bounds in stratified Darcy
plicit interface tracking shares the bounce-back deficiency concerning law (Vikhansky and Ginzburg, 2014); numerical dispersion, skewness
the inexact location of the closure continuity condition (Ginzburg, 2007; and kurtosis truncation estimates (Ginzburg, 2017b) in straight and
2016; Ginzburg et al., 2015b). An alternative second-order accurate radial Poiseuille flow. In alternative heterogeneous approach (Zhang
scheme (Guo et al., 2015) has been recently introduced for location et al., 2016), the single-relaxation-time SRT collision [which is a par-
of the concentration and flux jumps on the arbitrarily placed straight ticular TRT subclass where the two relaxation rates are the same] op-
and curved surface. We will restrict ourselves to stair-case midgrid wall erates with the uniform rate and incorporates heterogeneity with the
description and examine the difference between the fully-implicit node- semi-explicit partial bounce-back mechanism. The SRT approach (Perko
jump-tracking and semi-explicit link-jump-tracking only in pure diffusion and Patel, 2014) couples the advective and diffusive (non-equilibrium)
through porous blocks. Analytically, the TRT − EMM assessment is per- fluxes through an equilibrium. In uniform soil, the TRT overcomes SRT
formed by extending the exact Taylor-type analysis of the implicit clo- for stable velocity range due to the optimized dependency between its
sure relations and the symbolic solution construction (Ginzburg, 2007; free-tunable rate and diffusion coefficient (Ginzburg, 2012; Ginzburg
2016; Ginzburg et al., 2015b) from the streamline-invariant diffusion- et al., 2010; Kuzmin et al., 2011a). The respective stability and accu-
type interface problem to the Darcy advection through a streamline het- racy of the different heterogeneous approaches has been not evaluated
erogeneity in a “series” of porous blocks. A special attention is put on yet.
the consistent parameterization of the TRT − EMM solutions with the The TRT-ADE and TRT − EMM are very similar but while there is
dimensionless governing numbers in the presence of a space-variable no specific sources and flux-jumps in TRT-ADE, the TRT − EMM does
mass-source; a decisive role of the free-tunable equilibrium and colli- not need any anti-numerical diffusion corrections required in transi-
sion degrees of freedom will be demonstrated in dispersion, skewness tion and it can be accelerated for steady-state. The present works first
and kurtosis. This symbolic analysis is not restricted to TRT − EMM and compares the two approaches for the first four moments in a “perpen-
concerns any LBM-ADE steady problem. dicular” flow configuration. This study reveals that, unlike in a “paral-
The TRT − EMM is applied in both “𝜔-form” and “𝛾-form” with the lel” flow, the TRT-ADE modifies the constant advection velocity due
purpose to validate the equivalence of their moments solution and to to the heterogeneity in its anti-numerical-diffusion equilibrium cor-
examine their respective robustness. It was predicted (Ginzburg and rection. It has been already recognized that the discontinuity of the
Vikhansky, 2018; Vikhansky and Ginzburg, 2014) that the interface flux- concentration equilibrium weights results in discontinuous solute solu-
jumps and the associated mass-source terms become negligible when the tion, and different interface-explicit strategies have been attempted for
moments grow with Pe; this property would further simplify the method. a remedy (Ginzburg, 2006; Ginzburg and d’Humières, 2007; Yoshida
We will evaluate all these features in creeping Brinkman (1947) flow et al., 2014). In this work, an alternative anti-numerical-diffusion im-
through a “zigzag-type” solid and porous arrangements inspired from plicit MRT-r approach (Ginzburg, 2013), based on the idea of the non-
the micro-channel design (Vikhansky, 2011) and similar to one depicted diagonal multiple-relaxation-times matrix collision (Rasin et al., 2005;
in Fig. 1. The flow is solved with the local interface-implicit resistance- Yoshida and Nagaoka, 2010), will be extended from the homogeneous
force-based (improved) IBF scheme (Ginzburg et al., 2015b); the partial- to heterogeneous soil. In contrast with the TRT-ADE, the MRT-r copes
bounce-back “gray” link-based LBM-Brinkman models (see review Zhu with the continuous equilibrium and is able to remove the off-diagonal

3
JID: ADWR
ARTICLE IN PRESS [m5GeSdc;June 29, 2018;16:56]

I. Ginzburg Advances in Water Resources 000 (2018) 1–34

numerical-diffusion on the coordinate discrete-velocity stencil. On top 𝜇 1 (t) [hereafter, 𝜇 n (t) are normalized with the zero moment 𝜇0 (t)]. The
of the accuracy study, we support all schemes with the velocity bound Péclet number Pe is measured with the characteristic seepage veloc-
 
estimates and evaluate their efficiency for Pe-range. ity 0 and length , Pe ≈ 0 . At a long enough time after release,
0
Finally, although all presented schemes may apply the full (diago- 𝑡 ≈ Pe∕0 , the first four moments determine, respectively, the mean-
nal) velocity sets, like d2Q9, d3Q15 or d3Q19, we limit this work to seepage velocity component 𝑥 , longitudinal dispersion coefficient D,
the minimal schemes dDQ(2D+1). First, because they are sufficient for skewness Sk(t) and kurtosis Ku(t) of the averaged distribution 𝐶̄ (𝑥, 𝑡):
TRT − EMM in isotropic diffusion. Second, because the numerical dis-
persion, skewness and kurtosis are influenced by the diagonal weight 𝜇1 ( 𝑡 ) 𝜇 ⋆ (𝑡)
𝑥 = , 𝐷 = 2 , 𝐷 = 𝑒𝑓 𝑓 + 𝐷𝑇 , 𝐷𝑇 = 𝑘𝑇 0 ,
stencil in direct solvers (Ginzburg, 2017b). Third, because the bounce- 𝑡 2𝑡
back mass-conserving no-flux rule copes best with the coordinate sten- 1 𝜇 ⋆ (𝑡) 𝜇 ⋆ (𝑡)
cil on the mesh-aligned solid surface. In fact, the bounce-back creates Sk(𝑡) = Sk(𝑠) 𝑡− 2 = ( 3 ) , Ku(𝑡) = Ku(𝑠) 𝑡−1 = ( 4 ) − 3. (5)
3∕2 2
𝜇2⋆ (𝑡) 𝜇2⋆ (𝑡)
a spurious restriction of the tangential flux via the surface-inclined ve-
locity links: their accommodation layers modify the effective velocity, The effective diffusion coefficient 𝑒𝑓 𝑓 is measured at zero velocity; the
molecular diffusion and dispersion coefficients even in a straight chan- Taylor dispersion coefficient DT becomes apparent in motion because of
nel (Ginzburg, 2017a; Ginzburg et al., 2015a). the solute gradients due to the structure heterogeneity or the transverse
The rest of the paper is organized as follows. Section 2 discuses the velocity gradient; the variance (central second moment 𝜇2⋆ (𝑡)) grows lin-
computation of moments and introduces the direct solvers, LBM-ADE. early with time and the dispersion coefficient D sums 𝑒𝑓 𝑓 and DT . The
Section 3 summarizes the EMM and introduces its numerical formula- higher-order moments, skewness Sk(t) and kurtosis Ku(t) characterize
tion, TRT − EMM. Section 4 derives and analyses effective bulk equation the distribution asymmetry and peakedness, respectively, and tend to
and steady-state TRT − EMM interface conditions. Section 5 constructs zero as time increases as 𝑡−1∕2 and 𝑡−1 . An exact (Gaussian) solution of
apparent TRT − EMM moments solution for blocks in “series” and com- Eq. (4) gives Sk(𝑠) = Ku(𝑠) = 0. However, the high-order averaging of Eq.
pares the TRT − EMM and LBM-ADE. Section 6 examines the TRT − (1) gives them non-zero solution in the simplest configurations, like the
EMM symbolic solutions in channels: Poiseuille flow in Section 6.1, Poiseuille duct flow or a constant, parallel or perpendicular, Darcy flow
a fracture/matrix stratified bounded system in Section 6.2, and two- through heterogeneous blocks, (Ginzburg and Vikhansky, 2018; Vikhan-
layered double-periodic Brinkman flow in Section 6.3. Section 7 com- sky and Ginzburg, 2014).
pares the TRT − EMM and LBM-ADE moments in heterogeneous flow The “temporal dispersion” is quantified through the temporal mo-
around solid and through porous obstacles. Section 8 concludes the pa- ments of the residence-time distribution (RTD) due to Danckwerts
per. Appendix A provides the mass-source in “𝜔-form” and “𝛾-form”. (1953). The RTD𝑃 (𝑥, 𝑡) = 𝜕𝑡 ⟨(𝑥, 𝑡)⟩ is monitored at a (signed) distance
Appendix B resumes the reference diffusion solution in “series” and for- x from the inlet, e.g. Vikhansky (2011); the dimensionless distance
mulates its semi-explicit interface-jump tracking. Sections 2.1, 2.2, 3 and 𝑥∕ to the Taylor regime grows in proportion with Pe. The central
7.2 are self-consistent for the TRT − EMM implementation which is re- moments 𝜇𝑛⋆ (𝑥) = ∫0 (𝑡 − 𝑡̄(𝑥))𝑛 𝑃 (𝑥, 𝑡)𝑑𝑡 are accumulated in time; they

sumed in the algorithmic form in Appendix A.4. are restored from the raw moments with the centroid 𝑡̄(𝑥) = 𝜇1 (𝑥) =
1 ∞
∫ 𝑡𝑃 (𝑥, 𝑡)𝑑𝑡 [hereafter, the raw moments are normalized with the
𝜇 (𝑥) 0
2. A direct solving of the advection–diffusion equation (ADE) 0

zero moment 𝜇0 (𝑥) = ∫−∞ 𝑃 (𝑥, 𝑡)𝑑𝑡]. In the Taylor regime, the mean-
𝑥 𝜇2⋆ (𝑥)03
We introduce dimensionless transport coefficients in Section 2.1, seepage velocity 𝑥 = 𝜇1 ( 𝑥 )
and dispersion coefficient 𝐷 = 2𝑥
have
the TRT framework in Section 2.2 and the direct LBM-ADE solvers in the same solution as in Eq. (5); skewness Sk(x) and kurtosis Ku(x) of the
Section 2.3. RTD are determined as:

2.1. The spatial and temporal moments in Taylor regime 1 𝜇3⋆ (𝑥) 𝜇4⋆ (𝑥)
Sk(𝑥) = Sk(𝑡) |𝑥|− 2 = , Ku(𝑥) = Ku(𝑡) |𝑥|−1 = ( )2 − 3. (6)
(𝜇2⋆ (𝑥))3∕2 𝜇2⋆ (𝑥)
Consider a single cell volume 𝑉 = {𝑟⃗} in d-dimensional periodic
porous space. Let {V𝜙 } ∈ V be a set of penetrable volumes with smooth Exact solution to Eq. (4) with the Heaviside entry [(0, 𝑡) = 1 for √ t≥0
porosity distributions 𝜙(𝑟⃗) inside V𝜙 where concentration distribution
and (𝑥, 𝑡 = 0) = 0, x ≠ 0] gives dimensionless values Sk(⋆𝑡) = Sk(𝑥) |𝑥| =
obeys Eq. (1); 𝜙(𝑟⃗) is discontinuous on the penetrable interface A𝜙 be- √
tween V𝜙 and its neighbours, and on the solid surface As ∈ V𝜙 , see Fig. 1. 3 2(1 + 𝑘𝑇 )Pe−1 and Ku(⋆𝑡) = Ku(𝑥) |𝑥| = 30(1 + 𝑘𝑇 )Pe−1 . However, Sk(t)

Let ⟨𝜓⟩ = 𝑉𝜙 ∫𝑉 𝜓𝑑𝑉 denote the volume integration over a single cell. and Ku(t) generally differ from this solution. Indeed, the spatial and tem-
𝜙
In the frame of the classical (Aris, 1956; Taylor, 1953) and generalized poral moments, and hence their coefficients, are interconnected via sim-
(Brenner, 1980) dispersion theory extended to heterogeneous soil, the ple algebraic formulae established with the help of the extended method
cell-averaged mean concentration 𝐶̄ obeys the upscaled ADE in the de- of moments (EMM) [see Eqs. (31) and (32) in Vikhansky and Ginzburg,
veloped Taylor regime: 2014]. In this work, these relations will be verified in multi-dimensional
porous flow.
⟨𝜙⟩ ⃗ ⟨⃗𝑢𝜙 ⟩
𝜕𝑡 𝐶̄ + ∇ ⋅ ⃗ 𝐶̄ = ∇ ⋅ 𝐃∇𝐶̄ , 𝐶̄ = ,  = . (4)
⟨𝜙⟩ ⟨𝜙⟩
2.2. Lattice Boltzmann TRT framework
The averaged interstitial (seepage) velocity ⃗ is constant and it is ar-
gued (Brenner, 1980; Ginzburg and Vikhansky, 2018) that Eq. (4) copes The LBM is operated on the cuboid computational grid within the
with the symmetric constant dispersion tensor D. We assume that 𝜙(𝑟⃗) penetrable (fluid) d-dimensional domain 𝑟⃗ ∈ {𝑉𝜙 }; an equidistant mesh-
and 𝑢⃗𝜙 (𝑟⃗) are periodic over V, at least for one of the coordinate axes, size is set equal to one lattice unit; the Péclet number is kept the same
say the x-axis, hereafter referred to as a “streamwise direction”. Molec- in physical and numerical systems. The neighboring nodes are inter-
ular diffusion tensor in Eq. (1) is set isotropic: 𝐃(0) = 0 𝐈. We aim to connected by the d-dimensional discrete velocity set dDqQ; it consists
determine the longitudinal transport coefficients in spatial and temporal of a zero-amplitude vector 𝑐⃗0 and 𝑄𝑚 = 𝑄 − 1 vectors 𝑐⃗𝑞 , each vector
𝑄
systems of moments. 𝑐⃗𝑞 has the opposite one: 𝑐⃗𝑞̄ = −𝑐⃗𝑞 , 𝑞 = 1, … , 2𝑚 and the pair {𝑐⃗𝑞 , 𝑐⃗𝑞̄ }
The spatial dispersion characterizes the evolution of the point source is called a link. The local variable of the scheme is the set of Q real
(𝑟⃗, 𝑡 = 0) = 𝛿(𝑥 − 𝑥0 )𝛿(𝑡) through the spatial moments: nth-order raw numbers called “populations”: {𝑓𝑞 (𝑟⃗, 𝑡), 𝑞 = 0, … , 𝑄𝑚 }. The populations
moments 𝜇𝑛 (𝑡) = ∫−∞ (𝑥 − 𝑥0 )𝑛 𝐶̄ (𝑥, 𝑡)𝑑𝑥 are measured for n ≥ 0; their cen-

{𝑓𝑞 (𝑟⃗, 𝑡)} enter into the local collision, sums with their post-collision val-
tral counterparts 𝜇𝑛⋆ (𝑡) are built from them for n ≥ 2 with the centroid ues {𝑔𝑞 (𝑟⃗, 𝑡)} and then propagate to the neighbor sites according to their

4
JID: ADWR
ARTICLE IN PRESS [m5GeSdc;June 29, 2018;16:56]

I. Ginzburg Advances in Water Resources 000 (2018) 1–34

discrete velocities. The two populations with the opposite velocities 𝑑


∑ 𝑄𝑚
are decomposed into their symmetric and anti-symmetric components, 𝑒−
𝑞 (𝑟
⃗, 𝑡) = 𝑡𝑞 𝐾(𝑟⃗, 𝑡) 𝑢𝜙,𝛼 (𝑟⃗)𝑐𝑞𝛼 + 𝑒(𝑞𝑢,−) (𝑟⃗, 𝑡), 𝑞 = 1, … , , (10b)
2
{𝑓𝑞± = 12 (𝑓𝑞 ± 𝑓𝑞̄ )} and {𝑔𝑞± = 12 (𝑔𝑞 ± 𝑔𝑞̄ )}, 𝑞 = 1, … , 𝑄𝑚 ; in the present 𝛼=1
work, the symmetric component {𝑓𝑞+ } is updated with the same relax-
ation rate 𝑠+ 1 𝜌(𝑟⃗, 𝑡)
𝜙
for all links. The collision-stream algorithm then reads, 𝐾(𝑟⃗, 𝑡) = 𝐾 (𝑒𝑞) (𝑟⃗, 𝑡) + (𝑟⃗), 𝐾 (𝑒𝑞) = 𝛽𝑒 , (10c)
adding a time-independent mass source quantity (𝑟⃗) to the outgoing
2 𝜙(𝑟⃗)
immobile population 𝑓0 (𝑟⃗, 𝑡 + 1), as follows:
𝑄𝑚

𝑄𝑚 𝜌(𝑟⃗, 𝑡) = 𝑓𝑞 (𝑟⃗, 𝑡), (10d)
𝑓𝑞 (𝑟⃗ + 𝑐⃗𝑞 , 𝑡 + 1) = 𝑓𝑞 (𝑟⃗, 𝑡) + 𝑔𝑞+ (𝑟⃗, 𝑡) + 𝑔𝑞− (𝑟⃗, 𝑡), 𝑞 = 1, … , ,
2 𝑞=0
𝑄𝑚
𝑓𝑞̄ (𝑟⃗ + 𝑐⃗𝑞̄ , 𝑡 + 1) = 𝑓𝑞̄ (𝑟⃗, 𝑡) + 𝑔𝑞+ (𝑟⃗, 𝑡) − 𝑔𝑞− (𝑟⃗, 𝑡), 𝑐⃗𝑞̄ = −𝑐⃗𝑞 , 𝑞 = 1, … , , 𝑑
2 ∑ 𝑐𝑒 𝜙(𝑟⃗) 𝜙(𝑟⃗)
𝑄𝑚 ∕2 dDQ(2D+1) ∶ 𝑢2𝜙 (𝑟⃗) = 𝑢2𝜙,𝛼 ≤ , 𝑐𝑒 ≤ min . (10e)
∑ 𝛼=1
𝛽𝑒 𝑟⃗ 𝑑𝛽𝑒
𝑓0 (𝑟⃗, 𝑡 + 1) = 𝑓0 (𝑟⃗, 𝑡) + (𝑟⃗) − 2 𝑔𝑞+ (𝑟⃗, 𝑡) ; (7a)
𝑞=1 We note that Eq. (7a) does not need to compute 𝑒± 0
(𝑟⃗, 𝑡). When the
𝑄𝑚 + populations are initialized at the equilibrium, one prescribes 𝜌(𝑟⃗, 𝑡 = 0),
𝑔𝑞+ (𝑟⃗, 𝑡) = −𝑠+ (𝑓 + (𝑟⃗, 𝑡) − 𝑒+
𝜙 𝑞 𝑞 (𝑟
⃗, 𝑡)), 𝑞 = 1, … , , 𝑠𝜙 ∈]0, 2[. (7b) 𝑄𝑚
2 computes 𝑓𝑞 (𝑟⃗, 𝑡 = 0) = 𝑒+𝑞 + 𝑒𝑞 and 𝑓𝑞̄ (𝑟
− ⃗, 𝑡 = 0) = 𝑒+
𝑞 − 𝑒𝑞 , 𝑞 = 1, … , 2 ,

∑𝑄 𝑚 ∑𝑄𝑚
An exact mass-conservation equation means that 𝑞=0 𝑓𝑞 (𝑟⃗ + 𝑐⃗𝑞 , 𝑡 + e.g., with (𝑟⃗) = 0, then 𝑓0 (𝑟⃗, 𝑡 = 0) = 𝜌(𝑟⃗, 𝑡 = 0) − 𝑞=1 𝑓𝑞 (𝑟⃗, 𝑡 = 0). In
∑𝑄𝑚 what follows we address the minimal discrete velocity sets dDQ(2D+1):
1) = 𝑞=0 𝑓𝑞 (𝑟⃗, 𝑡) + (𝑟⃗) for any equilibrium distribution {𝑒𝑞 (𝑟⃗, 𝑡) = 𝑒+ 𝑞 + d1Q3, d2Q5 and d3Q7; they are built on the coordinate discrete ve-
𝑒−
𝑞 } . The anti-symmetric component { 𝑔 − (𝑟
𝑞 ⃗, 𝑡 )} will be computed either locity vectors {𝑐⃗𝑞 = ±1⃗ 𝛼 }, 𝛼 = 1, … , 𝑑. These schemes yield the same
with the (two-relaxation-times) TRT operator or (multiple-relaxation-
weight 𝑡𝑞 = 12 for a diffusion-flux variable 𝐾(𝑟⃗, 𝑡) in 𝑒+ 𝑞 (𝑟
⃗, 𝑡) and an
times) MRT-r operator. The TRT operator (Ginzburg, 2005a; 2007;
advective-flux 𝑢⃗𝜙 𝐾(𝑟⃗, 𝑡) in 𝑒−𝑞 (𝑟
⃗, 𝑡), because of the isotropic condition
Vikhansky and Ginzburg, 2014) applies the same relaxation rate 𝑠− 𝜙
for ∑𝑄 𝑚
all links and copes with Eq. (7c): 𝑡 𝑐 𝑐 = 𝛿𝛼𝛽 , ∀𝛼 , 𝛽. The macroscopic solution 𝐾(𝑟⃗, 𝑡) is defined
𝑞=1 𝑞 𝑞𝛼 𝑞𝛽
with the half mass-source value in Eq. (10c) for reduction of the trun-
𝑄𝑚 ± cation corrections (Kuzmin et al., 2011b) and correct parameteriza-
TRT ∶ 𝑔𝑞± (𝑟⃗, 𝑡) = −𝑠± (𝑓 ± (𝑟⃗, 𝑡) − 𝑒±
𝜙 𝑞 𝑞 (𝑟
⃗, 𝑡)), 𝑞 = 1, … , , 𝑠𝜙 ∈]0, 2[. (8)
2 tion of the modeled equation (d’Humières and Ginzburg, 2009). Be-
cause of this last objective, the weight distribution {tq ce } is also as-
The relaxation rates are space-dependent but time-independent in this
signed for the mass-term Λ+ 𝜙
(𝑟⃗) in Eq. (10a). The two scale param-
work. The two relaxation rates in Eq. (8) are defined from the locally
eters, ce and 𝛽 e , are set constant over the grid because of the leading-
prescribed positive eigenfunctions Λ±
𝜙
(𝑟⃗):
order continuity condition prescribed for 𝐾(𝑟⃗). When a local mass quan-
Λ𝜙 (𝑟⃗) tity 𝑡𝑞 𝑐𝑒  is directly added to 𝑔𝑞+ in Eq. (7c), an equivalent scheme
𝑠± (𝑟⃗) = (Λ± + 12 )−1 , Λ+ (𝑟⃗) = , Λ± > 0 , Λ𝜙 > 0 . (9)
𝜙 𝜙 𝜙 Λ−
𝜙
(𝑟⃗) 𝜙 reads with 𝑒+ ⃗, 𝑡) = 𝑡𝑞 𝑐𝑒 𝐾 (𝑒𝑞) + 𝑒(𝑞𝑢,+) . In the transient modeling of Eq.
𝑞 (𝑟
(1), (𝑟⃗, 𝑡) is set equal to 𝐾(𝑟⃗, 𝑡) and 𝛽 e ≡ 1. Since the transient term
It is convenient to parameterize Λ+𝜙
in Eq. (9) through a positive free- 𝛽𝑒 −1 𝜕𝑡 𝐾 vanishes at the steady state, the stationary solutions are 𝛽 e -
tunable parameter Λ𝜙 = Λ+ − , because Λ controls the truncation ac-
Λ
𝜙 𝜙 𝜙 independent but their convergence becomes accelerated when 𝛽 e > 1
curacy (Ginzburg, 2012; 2017b; Ginzburg and Roux, 2015), stability following (Ginzburg, 2006).
(Ginzburg, 2012; Ginzburg et al., 2010; Kuzmin et al., 2011a; Yan et al.,
2017) and interface/boundary accuracy (Cui et al., 2016; Demuth et al., 2.3. Direct transient LBM-ADE solvers
2016; Ginzburg, 2007; 2017a).
We extend the two classes of LBM-ADE isotropic schemes from
Remark 1. The same value 𝑠+ 𝜙
assigned for all the symmetric modes
the homogeneous to heterogeneous soil. A further extension to full
in Eq. (7c) is the most natural and quite stable choice according to the
anisotropic diffusion tensors may adopt either anisotropic-equilibrium
Von Neumann exact analysis (Ginzburg, 2013). The isotropic MRT mod-
(Chai et al., 2016b; Ginzburg, 2005a; 2012; Ginzburg and d’Humières,
els computed on the coordinate velocity sets dDQ(2D+1) then all re-
2007; Ginzburg et al., 2010; Servan-Camas and Tsai, 2010), link-wise
duce to the TRT operator (8). The “regularized” dDQ(2D+1) isotropic
collisions (Ginzburg, 2005a; Ginzburg and d’Humières, 2007) or non-
schemes (Wang et al., 2015) correspond to 𝑠+ 𝜙
≡ 1 and hence, Λ𝜙 = 12 Λ−
𝜙
.
diagonal matrix collisions (Dubois et al., 2016; Huang and Wu, 2014;
According to exact stability analysis (Ginzburg et al., 2010), Λ𝜙 = 14 , Rasin et al., 2005; Yoshida and Nagaoka, 2010); different techniques are
or the so-called optimal stability sub-class OTRT, may assure the suffi- combined, examined for numerical diffusion and supported with the cor-
ciency of the necessary stability conditions. The “regularized” choice is responding necessary stability bounds in homogeneous soil (Ginzburg,
not “optimal” but perhaps more robust than the SRT sub-class where 2013). In a space-variable velocity field, numerical diffusion is expected
𝑠+
𝜙
= 𝑠−
𝜙
, Λ𝜙 = (Λ− 𝜙
)2 and the non-negativity of the equilibrium function (Ginzburg and Roux, 2015; Huang and Wu, 2014) to be dominated by
{𝑒𝑞 = 𝑒+
𝑞 + 𝑒 − > 0} is proved (Kuzmin et al., 2011a) to become necessary
𝑞 the physical and numerical dispersion already in the moderate Péclet
1
in the limit Λ−
𝜙
→ 0. The OTRT with Λ𝜙 = 4
has larger velocity bounds range; the anti-numerical-diffusion treatment then becomes insignif-
in the isotropic and anisotropic schemes. icant. Nevertheless, let us start with two remarks on the numerical
diffusion because its role is important in diffusion-dominant processes
Remark 2. Another reason to keep Λ𝜙 independent from the trans-
through heterogeneous blocks in “series”.
port eigenfunction Λ−𝜙
is the consistent parameterization of the spatial
truncation correction and steady-state solutions by Pe; this property is Remark 3. The second-order Chapman–Enskog expansion, e.g.
guaranteed with the fixed value Λ𝜙 when the mass-source is constant Ginzburg (2005a), Servan-Camas and Tsai (2008), Chopard
(d’Humières and Ginzburg, 2009), and it will be examined for a space- et al. (2009), Ginzburg (2013) and Wang et al. (2015), pro-
variable source. vides the numerical-diffusion form Fnum . In heterogeneous
soil, Fnum sums with the right hand side (RHS) of Eq. (1):
In the present work, 𝑒± ∑𝑑 ∑𝑄 𝑚 ∑𝑑
𝑞 (𝑟
⃗, 𝑡) has the form: 𝐹 𝑛𝑢𝑚 = Λ−
𝜙 𝛼,𝛽=1 𝜕𝛽 𝜕𝑡 𝑢𝜙,𝛼  𝑡 𝑐 𝑐 = Λ−
𝑞=1 𝑞 𝑞𝛼 𝑞𝛽 𝜙 𝛼=1 𝜕𝛼 𝜕𝑡 𝑢𝜙,𝛼  [us-
𝑄𝑚 ing 𝐾 =  and 𝑒(𝑞𝑢,−) = 0 in Eq. (10b)]. Following Ginzburg
𝑒+
𝑞 (𝑟
⃗, 𝑡) = 𝑡𝑞 𝑐𝑒 (𝐾(𝑟⃗, 𝑡) + Λ+
𝜙
(𝑟⃗)) + 𝑒(𝑞𝑢,+) (𝑟⃗, 𝑡), 𝑞 = 1, … , , (10a)
2 (2005a), we replace 𝜕𝑡  by −𝜙−1 ∇ ⋅ 𝑢⃗𝜙  + 𝑂(𝜖 2 ) according to

5
JID: ADWR
ARTICLE IN PRESS [m5GeSdc;June 29, 2018;16:56]

I. Ginzburg Advances in Water Resources 000 (2018) 1–34

{ }
the first-order expansion, and obtain for a time-independent ve- 0 𝜙2 𝜙
∑ Λ− Λ− , Λ− = Λ− , 𝑢2𝜙 < 𝑐𝑒 𝜙.
locity field: 𝐹 𝑛𝑢𝑚 = −Λ− 𝜙
𝜙−1 𝑑𝛼,𝛽=1 𝜕𝛼 𝑢𝜙,𝛼 𝜕𝛽 𝑢𝜙,𝛽 . In a constant 𝜙 = 𝜙,1
=
(𝑐𝑒 𝜙 − 𝑢2𝜙 ) 𝜙,2 𝜙,3
=
𝑐𝑒
(12c)
or streamline-invariant duct flow an equivalent form reads:
𝐹 𝑛𝑢𝑚 = 𝐹 𝑢 = −Λ− ∇ ⋅ 𝐃(𝑢) ∇ with 𝐃(𝑢) = 𝜙−1 {𝑢𝜙,𝛼 𝑢𝜙,𝛽 }. We note
2
𝜙 The ADE-r operates with the continuous equilibrium distribution.
that the forward-time central-finite-difference Lax–Wendroff schemes
The matrix {𝑆𝛼𝛽
−1 } is the inverse of the collision matrix {S }; the d
𝛼𝛽
(Hindmarsch et al., 1984) complement the diffusion tensor with a sim-
1
Δ2 eigenfunctions {Λ− = − 12 } are built with the eigenvalues {𝑠− } of the
ilar anti-numerical-diffusion form, like Λ−
𝜙
𝐃(𝑢) Δ𝑡𝑥 |Λ− = 1 (see Ginzburg 𝜙 𝑠−
𝜙
𝜙
𝜙 2
collision matrix {S𝛼𝛽 } (cf. Ginzburg, 2013, Eq. (54)). This collision op-
et al., 2010). In LBM-ADE, by adopting anisotropic approaches, 𝐹 𝑢
2
erator is called “rotated” because its matrix becomes diagonal in the
can be cancelled either with the help of the “hydrodynamic” type axes aligned with the velocity vector 𝑢⃗𝜙 ; ADE-r reduces to ADE-1D in
u𝜙, 𝛼 u𝜙, 𝛽 -equilibrium correction on the full velocity sets (Chai et al., one-dimensional flow. An extension of Eq. (12) to full velocity sets and
2016b; Ginzburg, 2005a; 2012; Ginzburg et al., 2010), or with the anisotropic diffusion tensor may follow (Ginzburg, 2013). We note that
non-diagonal collision matrix (Ginzburg, 2013; Rasin et al., 2005), or the necessary stability line from Eq. (12c) coincides with the ADE-c sta-
else combining the two techniques (Ginzburg, 2013). The alternative bility bound from Eq. (10e), but Λ− unboundly increases on the sta-
𝜙,1
LBM techniques involve the pseudo-velocities (Servan-Camas and Tsai, bility line degrading the truncation accuracy. In this work, we fix an
2008) or “body force” strategy (Chai et al., 2016b; Chopard et al., uniform value Λ𝜙 and define Λ+ = Λ𝜙 ∕ max{Λ− } = Λ𝜙 ∕Λ− in Eq. (9).
𝜙 𝜙 𝜙,1
2009; Servan-Camas and Tsai, 2008; Wang et al., 2015). Following The MRT-r then intrinsically operates with two distinguished values, Λ𝜙
∑𝑑
this approach, one may set 𝑒(𝑞𝑢,−) = Λ− 𝑡
𝜙 𝑞 𝛼=1 𝜕𝑡 𝑢𝜙,𝛼 𝑐𝑞𝛼 in Eq. (10b), and (the smaller valued) Λ𝜙 Λ− ∕Λ− ; their disparity may degrade the
𝜙,2 𝜙,1
without any velocity field approximation, but with need to resorting accuracy and stability against the Λ𝜙 -optimized TRT schemes (Ginzburg,
for the backward finite-difference approximate to compute 𝜕𝑡  𝑢⃗𝜙 . 2013).
Remark 4. The recurrence equations (Ginzburg and Roux, 2015;
Vikhansky and Ginzburg, 2014) produce numerical-diffusion correction Remark 5. The two stability conditions in Eq. (10e) present a formal
to the RHS of Eq. (1) in the exact form: 𝐹 𝑛𝑢𝑚 = −Λ− ∇̄ 2 𝜙. The Chapman– extension to the necessary stability conditions established with the Von-
𝜙 𝑡
Enskog approximate is then matched applying twice ∇ ̄ 𝑡 𝜙 = −∇̄ ⋅ 𝑢⃗𝜙 , Neumann linear analysis (Ginzburg et al., 2010) in the absence of the
and replacing the central-differences by their differential counterparts. interface/boundary and mass-sources, but comprising the worst possible
This suggests that the removal of the entire numerical diffusion may scenario for the respective position of the wave and (constant) velocity
require the time-centered implicit anti-correction. vectors. Indeed, the velocity bound from Eq. (10e) assures the positive
We will consider two different anti-numerical-diffusion corrections semi-definiteness of the effective diffusion tensor Λ− 𝜙
(𝐈 − 𝐃(𝑢) ); the ADE-
𝑑
to the approximate form 𝐹 𝑢 (𝑟⃗) = −Λ− ∇ ⋅ 𝐃(𝑢) ∇, 𝐃(𝑢) = 𝜙−1 {𝑢𝜙,𝛼 𝑢𝜙,𝛽 } in
2
𝜙
j may exceed this bound by a factor 𝑑−1
in d2Q5 and d3Q7, especially
TRT-ADE schemes. They combine Eq. (8) with Eq. (10) on the coordinate 1
when Λ𝜙 = in Eq. (9). Numerical computations confirm the princi-
4
stencil dDQ(2D+1) in the three linear following forms, using (𝑟⃗, 𝑡) = pal stability bounds and trends in duct flow (Ginzburg and Roux, 2015;
𝐾(𝑟⃗, 𝑡), 𝛽 e ≡ 1 and 𝜙 = 𝜙0 : Yan et al., 2017) and multi-dimensional porous-scale velocity field (Yan
ADE-c ∶ 𝑒(𝑞𝑢,±) = 0, Λ−
𝜙 𝑐𝑒 = 𝜙 , 𝜙 = 𝜙0 , (11a) et al., 2017).
Finally, we recall that the effective mean-seepage velocity of the
numerical scheme is detected by the first raw moment, while the ef-
𝑑
∑ fective quantity of the numerical diffusion is detected by the variance.
ADE-j ∶ 𝑒(𝑞𝑢,+) = 𝑡𝑞 𝜙−1  𝑢𝜙,𝛼 𝑢𝜙,𝛽 𝑐𝑞𝛼 𝑐𝑞𝛽 , 𝑒(𝑞𝑢,−) = 0, Λ−
𝜙 𝑐𝑒 = 𝜙 , (11b) The direct solvers compute the solute evolution through a long array
𝛼=1
of duplicated periodic unit cells (cf. V in Fig. 1). A very useful check
( ) is that, in the spatial spread and Taylor regime, all minimal schemes
𝑢2𝜙,𝑥
ADE-1D ∶ 𝑒(𝑞𝑢,±) = 0, Λ−
𝜙 𝑐𝑒 − = 𝜙 , 𝑢2𝜙,𝑥 ≤ 𝑐𝑒 𝜙. (11c) should produce 𝜇1 (𝑡 + 𝛿𝑡 ) − 𝜇1 (𝑡) =  (𝑠𝑢𝑚) 𝛿𝑡 where an arithmetical-mean
𝜙 value  (𝑠𝑢𝑚) (over the cross-section) replaces 𝑥 in a straight strat-
The ADE-c operates with a continuous equilibrium, but it models the ified flow; the ADE-j, ADE-1D and ADE-r minimal schemes produce
diffusion tensor with the correction −Λ−
𝜙
𝐃(𝑢) . The ADE-j with 𝑒(𝑞𝑢,+) van- 𝜎 2 (𝑡 + 𝛿𝑡 ) − 𝜎 2 (𝑡) = 2𝜙 𝛿𝑡 exactly in 𝜙-uniform constant Darcy flow u𝜙, x ,
𝑢2𝜙,𝑥
ishes only the diagonal entries −Λ−
𝜙
𝐃(𝑢) on the coordinate velocity sets; while the ADE-c gives 𝜎 2 (𝑡 + 𝛿𝑡 ) − 𝜎 2 (𝑡) = 2𝜙 (1 − )𝛿𝑡 . We note that
𝜙𝑐𝑒
we note that 𝑒(𝑞𝑢,+) is subject to the jumps when the velocity vector is the Taylor regime is difficult to reach at high Pe because the number
perpendicular to the interface. This is because of the discontinuity in of cells grows linearly with Pe and the computational time increases as
, although 𝐹 𝑢 (𝑟⃗) is space-continuous together with the
2
coefficients Λ−𝜙 Pe2 , at least.
velocity field 𝑢⃗𝜙 (𝑟⃗). The ADE-1D in Eq. (11c) is only suitable for 1D The three schemes (11) will be examined for the quality of their pro-
where it removes 𝐹 𝑢 with the help of the velocity-dependent relax- files and the first four moments in a “perpendicular” one-dimensional
2

ation function Λ− ; this scheme has been compared with ADE-j in ho- Darcy flow through heterogeneous blocks, against the analytical EMM
𝜙
mogeneous soil (Ginzburg, 2013) and can be applied for Darcy flow in solution (Ginzburg and Vikhansky, 2018). The three schemes (11a),
“series”. In multi-dimensions, the MRT-r block-matrix collision allows (11b) and (12a) will be compared, between them and with the steady-
to remove 𝐹 𝑢 on the coordinate stencil (cf. Ginzburg, 2013, Eq. (44)). state TRT − EMM predictions, for dispersion, skewness and kurtosis in
2

We propose an extension to heterogeneous soil for 𝑔𝑞− (𝑟⃗, 𝑡) in Eq. (7a) as temporal and spatial moments in two-dimensional flow around solid and
follows: through porous obstacles.
𝑑
∑ 𝑄𝑚
MRT-r ∶ 𝑔𝑞− (𝑟⃗, 𝑡) = − 𝑆𝛼𝛽 (𝑟⃗)(𝑓𝛽− (𝑟⃗, 𝑡) − 𝑒−
𝛽 (𝑟
⃗, 𝑡)), 𝑞 = 𝛼 = 1, … , = 𝑑,
𝛽=1
2
3. Extended method of moments, EMM and TRT − EMM
(12a)
( ) The EMM provides an alternative way to compute the transport co-
0 𝜙 1 0 𝜙
𝑆𝛼𝛽
−1
= + 𝛿 + 𝑢 𝑢 , {𝛼, 𝛽} = 1, … , 𝑑. (12b) efficients in Eqs. (5) and (6). Section 3.1 formulates the corresponding
𝑐𝑒 2 𝛼𝛽 𝑐𝑒 (𝑐𝑒 𝜙 − 𝑢2 ) 𝜙,𝛼 𝜙,𝛽
𝜙 steady state boundary-value EMM problems; Section 3.2 proposes their
numerical formulation TRT − EMM based on the TRT operator (8) and
ADE-r ∶ 𝛽𝑒 = 1, 𝑒(𝑞𝑢,±) = 0, equilibrium distribution (10).

6
JID: ADWR
ARTICLE IN PRESS [m5GeSdc;June 29, 2018;16:56]

I. Ginzburg Advances in Water Resources 000 (2018) 1–34

3.1. The EMM solution for the spatial and temporal moments
𝑆𝑛 (𝑟⃗) = 𝜙 𝜕𝑥 𝑛−1 , 𝑛 ≥ 2, 𝐽⃗𝑛 (𝑟⃗) = 𝜙 𝑛−1 (𝑟⃗)1⃗ 𝑥 , 𝑛 ≥ 1, (16e)
The extended method of moments (Ginzburg and Vikhansky, 2018;
Vikhansky and Ginzburg, 2014) searches a solution to Eq. (1) in the de-
composed form (2), either expanded into perturbative series in “𝜔-form” ⟨𝑀𝑛 + 𝑆𝑛 ⟩
(3a) or in “𝛾-form” (3b). Assume first that the coefficients of the two 𝜔𝑛 = − , 𝑛 ≥ 1. (16f)
⟨𝜙0 ⟩
expansions, {𝜔n } and {𝛾 n }, are known and their dimensionless counter-
Let 𝑛⃗ be a normal vector on the boundary or interface and ||𝜓||𝐴𝜙 =
parts are defined giving the mean characteristic velocity 0 and char-
acteristic length , respectively: 𝜓𝐴+ − 𝜓𝐴− denote the jump of a scalar variable 𝜓(𝑟⃗) on the interface A𝜙
𝜙 𝜙
along 𝑛⃗. Solution 𝑛 (𝑟⃗) is constrained to the periodic, interface, bound-
𝜔𝑛 0−1 𝛾𝑛 0𝑛 ary and normalization conditions, respectively:
𝜔′𝑛 = and 𝛾𝑛′ = , 𝑛 ≥ 1. (13)
(𝑛−1) (𝑛−1)
𝑛 (𝑟⃗ + 𝐿𝑥 1⃗ 𝑥 ) = 𝑛 (𝑟⃗), (17a)
They determine the dimensionless transport coefficients in Eqs. (5) and
(6):
0 𝑛 ⋅ 𝐽⃗𝑛 ||𝐴𝜙 ,
||𝑛 ||𝐴𝜙 = 0, || − 𝜙 𝑛⃗ ⋅ ∇𝑛 ||𝐴𝜙 = ||⃗ (17b)
𝑥 = 0 𝜔′1 and , if 𝑥 ≠ 0, 𝑥 = , (14a)
𝛾1′

−𝜙 𝑛⃗ ⋅ ∇𝑛 |𝐴𝑠 = 𝑛⃗ ⋅ 𝐽⃗𝑛 |𝐴𝑠 , (17c)


𝐷 = −𝜔2 , 𝑒𝑓 𝑓 = −𝜔2 |𝑢⃗𝜙 =0 , (14b)

⟨𝑛 𝜙⟩
𝐷 3   = 0. (17d)
𝑘𝑇 (0 ) = − 1 = −𝜔′2 Pe − 1 = 𝑥 𝛾2′ Pe − 1, Pe = 0 , (14c) ⟨𝜙⟩
0 03 0
Eqs. (16a)–(16e) are linear with re-
√ spect to 𝑛 (𝑟⃗) because 𝑛 (𝑟⃗) and 𝐽⃗𝑛 (𝑟⃗) in
0 𝑡 3𝜔′3 0 𝑡 6𝜔′
Sk(⋆𝑠) = Sk(𝑡) (𝑠)
− 4, Eq. (16a) are built with the known solution 𝑘 (𝑟⃗), k < n, n ≥ 1;

= √ ( )3∕2 , Ku⋆ = Ku(𝑡)  = (14d)
2 |𝜔′2 | 𝜔′2 2 these terms are exemplified in Eq. (A.1). The set {𝜔n } is built in Eq.
(16f) from the solvability condition ⟨𝑛 ⟩ = 0 which implies global
√ √ mass conservation.
|𝑥| 3𝛾3′ 𝑥′ |𝑥′ | |𝑥| 6𝛾 ′ |𝑥′ |
Sk(⋆𝑡) = Sk(𝑥) = √ ( ) , Ku(⋆𝑡) = Ku(𝑥) = 4 ′ . (14e) Since 𝐵𝑛 (𝑟⃗), 𝑢⃗𝜙 (𝑟⃗) and hence, the advective flux 𝑢⃗𝜙 𝐵𝑛 are continuous,
 2 𝛾2′ 𝑥′
3∕2  𝛾2′ 2 𝑥 the normal component of the diffusion flux −𝜙 ∇𝐵𝑛 is subject to the
prescribed jump value: || − 𝜙∇ ̄ 𝑛 𝑛 ||𝐴 = ||𝜙𝑛−1 𝑛𝑥 ||𝐴 according to
In pure diffusion, 𝜔1 |𝑢⃗𝜙 =0 = 0 and the “𝜔-form” allows to compute 𝜙 𝜙

𝑒𝑓 𝑓 and the higher moments due to the structure; the “𝛾-form” becomes Eq. (17b) with Eq. (16e), when 𝜙(𝑟⃗) is discontinuous along the 𝑥−axis at
undefined. In motion, the two sets of the coefficients are inter-related: interface A𝜙 : ||𝜙(𝑟⃗)𝑛𝑥 ||𝐴𝜙 ≠ 0. Eq. (17b) comprises the solid side 𝑟⃗𝑠 ∈ 𝐴𝑠
with 𝜙(𝑟⃗𝑠 ) = 0, 𝐽⃗𝑛 (𝑟⃗𝑠 ) = 0, 𝑢𝜙,𝑥 (𝑟⃗𝑠 ) = 0 and prescribes the jump of the
1 𝜔′2 2𝜔′2 2 − 𝜔′1 𝜔′3
𝛾1′ = , 𝛾 ′
= − , 𝛾 ′
= , diffusion-flux normal component on the solid wall by Eq. (17c) with
𝜔′1 2 𝜔′ 3
3
𝜔′ 5 Eq. (16e): −𝑛⃗ ⋅ ∇𝑛 |𝐴𝑠 = 𝑛−1 𝑛𝑥 |𝐴𝑠 . The flux-jump term 𝐽⃗𝑛 (𝑟⃗) and the
1 1
mass-term 𝑆𝑛 (𝑟⃗) vanish in x-parallel stratified channels and ducts where
5𝜔′2 3 − 5𝜔′1 𝜔′2 𝜔′3 + 𝜔′1 2 𝜔′4
𝛾4′ = − . (15) the boundary-value problem reduces to diffusion equation with the
𝜔′1 7 transversely variable mass-source and continuous advection–diffusion
flux.
The back transform 𝜔′𝑛 (𝛾𝑛′ ) has exactly the same form, exchanging 𝜔′𝑛
The B-field is defined to a constant; it is fixed, without loss of general-
and 𝛾𝑛′ . In what follows, we have 0 = |𝑥 | with the x(component mean
ity, by the normalization condition (17d). Using Eq. (17d) for all orders
𝑥 of seepage velocity ⃗ from Eq. (4); then, when 𝑥 > 0, 𝜔′ = 𝛾 ′ = 1 ∑
1 𝑛 k < n in Eq. (16f), the terms ⟨𝜙 𝑛𝑘−1 =1 𝜔𝑘 𝑛−𝑘 ⟩ and ⟨𝜙 𝑛−2 ⟩ vanish for
in Eqs. (14a)–(15). In theory, the two formulations are equivalent and n > 2 and {𝜔n } can be computed as:
produce identical transport coefficients with the help of Eq. (15); it is
⟨𝑢𝜙,𝑥 1 ⟩ ⟨𝜙 𝜕𝑥 1 ⟩
shown (Ginzburg and Vikhansky, 2018) that the two problems are tack- 𝜔2 = −0 + − , (18a)
led through the linear combinations of the boundary-value problems. ⟨𝜙⟩ ⟨𝜙⟩
We focus on the “𝜔-form” because it is simpler and allows for pure dif-
fusion. ⟨𝑢𝜙,𝑥 𝑛−1 ⟩ ⟨𝜙 𝜕𝑥 𝑛−1 ⟩
𝜔𝑛 = − , 𝑛 ≥ 3. (18b)
In “𝜔-form”, the EMM looks for the scalar distribution {𝑛 (𝑟⃗)}, ⟨𝜙⟩ ⟨𝜙⟩
called B-field, which sequentially solves the steady-state ADE given by Eqs. (16f) and (18) allow to predict n longitudinal moments. The
Eqs. (16a)–(16e) inside a penetrable region 𝑟⃗ ∈ {𝑉𝜙 } ∈ 𝑉 of a single cell effective diffusivity 𝑒𝑓 𝑓 in Eq. (14b) and the associated higher-order
V, streamwise periodic along the x-axis: moments, due to the structure alone, are computed solving Eqs. (16)–
∇ ⋅ (𝑢⃗𝜙 (𝑟⃗)𝑛 (𝑟⃗) − 𝐽⃗𝑛 (𝑟⃗)) − 𝑛 (𝑟⃗) = ∇ ⋅ 𝜙 ∇𝑛 , 𝑛 ≥ 1, 𝜙 = 𝜙(𝑟⃗)0 , (17) with zero velocity 𝑢⃗𝜙 = 0. In order to obtain the three trans-
(16a) port coefficients in Eq. (14), a symbolic or numerical procedure solves
Eqs. (16)–(17) three times with respect to 𝑛 (𝑟⃗), 𝑛 = 1, 2, 3, …; the coef-
ficients 𝜔2 , 𝜔3 and 𝜔4 from Eqs. (16f) and (18) then allow to compute,
𝑛 (𝑟⃗) = 𝑀𝑛 (𝑟⃗) + 𝑆𝑛 (𝑟⃗) + 𝜙(𝑟⃗)0 𝜔𝑛 , 𝑛 ≥ 1, 0 = 1, (16b)
respectively, kT , Sk(⋆𝑠) and Ku(⋆𝑠) in Eq. (14) with the help of Eq. (13);
the temporal transport coefficients Sk(⋆𝑡) and Ku(⋆𝑡) from Eq. (14) are built
⟨𝑢𝜙,𝑥 ⟩ with the help of Eq. (15). The boundary value problem reads similar in
𝑀1 (𝑟⃗) = −𝑢𝜙,𝑥 0 , 𝑆1 = 0, 𝜔1 = = 𝑥 , (16c)
⟨𝜙⟩ the “𝛾-form” using Eqs. (A.2)–(A.4) with the non-zero mean-seepage ve-
locity 𝑥 . The boundary problem (16)–(17) can be solved independently
𝑛−1
∑ for each periodic direction of the unit cell V; the full-rank effective diffu-
𝑀𝑛 (𝑟⃗) = 𝜙 𝜔𝑘 𝑛−𝑘 (𝑟⃗) + 𝜙 𝑛−2 − 𝑢𝜙,𝑥 𝑛−1 , 𝑛 ≥ 2, (16d) sion and dispersion tensors can be restored from d runs with the help of
𝑘=1 the formulation (Ginzburg and Vikhansky, 2018); it applies either with

7
JID: ADWR
ARTICLE IN PRESS [m5GeSdc;June 29, 2018;16:56]

I. Ginzburg Advances in Water Resources 000 (2018) 1–34

an isotropic or anisotropic tensor D(0) ; at the first order 𝑛 = 1, this pro- steady-state solution of {𝑔𝑞± } exactly [see Ginzburg, 2017a, Eq. (20)];
cedure extends the B-equation (Brenner, 1980; Salles et al., 1993) from we will make the use of the two steady-state linear combinations:
homogeneous to heterogeneous soil. An extension to any-order longitu- {
̄ 𝑞 𝑒 − − Λ− Δ
𝑔𝑞+ (𝑟⃗) = [Δ ̄ 2 𝑒+ + (Λ𝜙 − 1 )Δ
̄ 2 𝑔 + ](𝑟⃗),
dinal coefficients is straightforward employing the recursive procedure 𝑞 𝜙 𝑞 𝑞 4 𝑞 𝑞
(Vikhansky and Ginzburg, 2014). − − − ̄ + 1 ̄2 −
Λ 𝑔 (𝑟⃗) = [Λ Δ𝑞 𝑒 − Δ 𝑒 − (Λ𝜙 − 1 )Δ ̄ 𝑞 𝑔 + ](𝑟⃗), 𝑟⃗ ∈ 𝑉𝜙 .
𝜙 𝑞 𝜙 𝑞 4 𝑞 𝑞 4 𝑞
(22)
3.2. Numerical formulation TRT − EMM
The central-differences apply link-wisely: Δ ̄ 𝑞 𝜓 (𝑟⃗, 𝑡) = 1 (𝜓 (𝑟⃗ +
2
The TRT − EMM solves Eqs. (16)–(17) inside a unit cell V with the ̄
𝑐⃗𝑞 , 𝑡) − 𝜓(𝑟⃗ − 𝑐⃗𝑞 , 𝑡)) and Δ𝑞 𝜓(𝑟⃗, 𝑡) = 𝜓(𝑟⃗ + 𝑐⃗𝑞 , 𝑡) − 2𝜓(𝑟⃗, 𝑡) + 𝜓(𝑟⃗ − 𝑐⃗𝑞 , 𝑡),
2
TRT collision operator from Eq. (8) and the equilibrium distribution ∀𝜓 = {𝑒± 𝑞 , 𝑔𝑞 }. Eq. (22) are derived with the space-uniform eigenfunc-
±
in Eq. (11), where it substitutes  = 𝑛 (𝑟⃗), 𝐾 = 𝐵𝑛 (𝑟⃗, 𝑡), 𝑒(𝑞𝑢,+) = 0 and tions Λ± ; in our context, they apply inside piece-wise continuous porous
∑ 𝜙
𝑒(𝑞𝑢,−) = −𝑡𝑞 𝑑𝛼=1 𝐽𝑛𝛼 𝑐𝑞𝛼 ; 𝑒(𝑞𝑢,+) is set equal to zero because numerical dif- block V𝜙 . Plugging first Eq. (22) with (19) into Eq. (21), an exact form
fusion vanishes at steady state; 𝑒(𝑞𝑢,−) is introduced in order to couple the of the modeled mass-conservation equation reads:
advective flux and diffusion-flux correction 𝐽⃗𝑛 and then to produce the ̄ ⋅ (𝑢⃗𝜙 𝐵𝑛 − 𝐽⃗𝑛 ) − 𝑛 + 𝑅𝑛 = ∇
∇ ̄ ⋅ 𝜙 ∇
̄ 𝐵𝑛 , 𝜙 (𝑟⃗) = Λ− (𝑟⃗)𝑐𝑒 , (23a)
𝜙
diffusive-flux jump in Eq. (17b) implicitly. Eq. (11) becomes:
𝑒+
𝑞 (𝑟
⃗, 𝑡) = 𝑡𝑞 𝑐𝑒 (𝐵𝑛 (𝑟⃗, 𝑡) + Λ+ (𝑟⃗)𝑛 (𝑟⃗)), (19a)
𝜙 ( )∑
𝑄𝑚
̄ 2 𝑛 (𝑟⃗) + Λ𝜙 − 1
𝑅𝑛 (𝑟⃗) = −Λ𝜙 𝑐𝑒 Δ ̄ 2 𝑔 + (𝑟⃗), 𝑛 ≥ 1.
Δ (23b)
4 𝑞=1 𝑞 𝑞
𝑑

𝑒−
𝑞 (𝑟
⃗, 𝑡) = 𝑡𝑞 (𝑢𝜙,𝛼 (𝑟⃗)𝐵𝑛 (𝑟⃗, 𝑡) − 𝐽𝑛𝛼 (𝑟⃗))𝑐𝑞𝛼 , (19b) Eq. (23a) presents the central-difference approximate of Eq. (16a) up
𝛼=1 to correction 𝑅𝑛 (𝑟⃗). The first component in 𝑅𝑛 (𝑟⃗) is non-zero when
𝑛 (𝑟⃗) is a non-linear function of the spatial coordinate; the term
1 𝜌(𝑟⃗, 𝑡) ∑𝑄𝑚 ̄ 2 +
𝐵𝑛 (𝑟⃗, 𝑡) = 𝐵𝑛(𝑒𝑞) (𝑟⃗, 𝑡) +  (𝑟⃗), 𝐵𝑛(𝑒𝑞) = 𝛽𝑒 , (19c) (Λ𝜙 − 14 ) 𝑞=1 Δ𝑞 𝑔𝑞 presents the generic TRT truncation correction (see
2 𝑛 𝜙(𝑟⃗)
Ginzburg, 2017b; Ginzburg and Roux, 2015). Eq. (23) is not specific
for TRT − EMM; it describes the steady-state equation modeled with the
𝑄𝑚
∑ TRT operator and linear equilibrium distribution (19); in that, 𝐽⃗𝑛 = 0,
𝜌(𝑟⃗, 𝑡) = 𝑓𝑞 (𝑟⃗, 𝑡), 𝑟⃗ ∈ {𝑉𝜙 }. (19d)
when the advective–diffusion flux is set continuous on the interface.
𝑞=0

The TRT − EMM can be applied in the form (19) with the full dis- 4. Bulk and interface analysis in a parallel and a perpendicular
crete velocity sets and isotropic weights {tq }; the minimal sets apply flow
𝑡𝑞 = 12 . An acceleration parameter 𝛽 e ≥ 1 from Eq. (11) should respect
the stability condition (10e); the steady state solutions are independent The analysis developed in this section applies exactly for any steady-
from 𝛽 e and initial distribution 𝜌(𝑟⃗, 𝑡 = 0). Like the direct ADE solvers, state solution of Eq. (23), provided that 𝐵𝑛 (𝑟⃗) varies along one lattice
the TRT − EMM runs across the heterogeneous domains {V𝜙 } without direction. In particular, the described situation takes place in a “series”
any explicit interface conditions; the diffusive-flux-jump condition on of porous blocks and stratified channels as sketched in Fig. 2.
the solid surface is modeled implicitly with the bounce-back boundary
rule [see Eq. (29b)]. The effective interface/boundary closure relations 4.1. Parameterization property
will be derived and compared to Eqs. (17b) and (17c). The obtained
steady-state solution 𝐵𝑛 (𝑟⃗) is normalized with Eq. (17d) yielding: When solution Bn (x) to Eq. (23) varies only along the x-axis, isotropic
solution 𝑔𝑞+ (𝑟⃗) to Eq. (21) reads:
⟨𝐵𝑛 𝜙⟩
𝑛 (𝑟⃗) = 𝐵𝑛 (𝑟⃗) − 𝐵𝑛 , 𝐵𝑛 = , 𝑛 ≥ 1. (20)
⟨𝜙⟩ 𝑔𝑞+ (𝑟⃗) = 𝑡𝑞 (𝑥)𝑐𝑞𝑥
2
, then (24a)
All the integral values < 𝜓 > in Eqs. (16), (18) and (20) are computed
with the summation over the grid points; hence, by construction, the 𝑄𝑚 𝑄𝑚
∑ ∑ ∑ ∑
mean-seepage velocity is set equal to  (𝑠𝑢𝑚) = 𝑟⃗ 𝑢𝜙,𝑥 (𝑟⃗)∕ 𝑟⃗ 𝜙(𝑟⃗). Ap- 𝑔𝑞+ (𝑟⃗) = (𝑥) and ̄ 2 𝑔 + (𝑟⃗) = Δ
Δ ̄ 2 (𝑥). (24b)
∑ 𝑞 𝑞 𝑥
plying the summation procedure, the global population mass 𝑟⃗ 𝜌(𝑟⃗), 𝑞=1 𝑞=1
∑𝑄 𝑚
𝜌(𝑟⃗) = 𝑞=0 𝑓𝑞 (𝑟⃗, 𝑡), is conserved exactly due to the discrete solvability Eq. (23) then becomes:

condition 𝑟⃗ 𝑛 (𝑟⃗) = 0; the sum of all populations over the grid should
remain the same during computations. The terms 𝑛 (𝑟⃗) and 𝐽⃗𝑛 (𝑟⃗) are ̄ ⋅ (𝑢⃗𝜙 𝐵𝑛 − 𝐽⃗𝑛 ) − 𝑛 − 𝛼𝑀 Δ
∇ ̄ 2 𝑛 = ∇ ̄ 𝐵𝑛 , 𝛼𝑀 = 1 − Λ𝜙 (1 − 𝑐𝑒 ).
̄ ⋅ 𝜙 ∇
𝑥 4
precomputed with Eqs. (16b)–(16e) exemplified in Eq. (A.1). In that,
(25)
𝑆𝑛 (𝑟⃗) is approximated with the help of the streamwise finite-differences
applied inside V𝜙 , e.g. using Eq. (A.6). We do not apply here the lo- We expect Eqs. (24a) and (25) to remain valid for any 𝑥−axis rotation
cal gradient solution, like 𝑐𝑒 ∇ ̄ 𝑥 𝐵𝑛 ≈ ∑𝑄𝑚 𝑔 − 𝑐𝑞𝑥 because of its intrinsic ∑𝑄𝑚 2
because of the isotropy of the term 𝑡𝑞 𝑞=1 𝑐𝑞𝑥 . Let us consider the fol-
𝑞=1 𝑞
correction due to the mass-source and velocity variations. The “𝛾-form” lowing dependency between Λ𝜙 and ce :
computes Eq. (19) with Eqs. (A.2) and (A.3) for 𝑛 (𝑟⃗) and Eq. (A.5) for
 1
𝐽⃗𝑛 (𝑟⃗); the set {𝛾 n } is constructed with Eq. (A.4). The numerical recipe is Λ𝜙 = ,  > 0, then 𝛼𝑀 = −  , (26a)
(1 − 𝑐𝑒 ) 4
formulated for TRT − EMM in Appendix A.4.
At steady state, 𝑓0 (𝑟⃗, 𝑡 + 1) = 𝑓0 (𝑟⃗, 𝑡) in Eq. (7a) and an exact mass-
1 1
conservation equation reads: 𝛼𝑀 = 0 if  = , then Λ𝜙 = . (26b)
4 4(1 − 𝑐𝑒 )
𝑄𝑚 ∕2
∑ Eq. (26a) allows to produce the same solution 𝐵𝑛 (𝑟⃗) for any individ-
(𝑟⃗) = 2 𝑔𝑞+ (𝑟⃗). (21)
ual values Λ−
𝜙
and ce when the diffusion coefficient 𝜙 = Λ− 𝑐 is fixed.
𝜙 𝑒
𝑞=1
In motion, Eq. (26a) allows for a consistent control of Eq. (25) by Péclet
The recurrence equations (d’Humières and Ginzburg, 2009; Ginzburg, number and porosity contrast. We say then that Eq. (26) allows for a
2012; Ginzburg and Roux, 2015) express both the transient and the proper parameterization of the modeled Eq. (25). However, 𝐵𝑛 (𝑟⃗) becomes

8
JID: ADWR
ARTICLE IN PRESS [m5GeSdc;June 29, 2018;16:56]

I. Ginzburg Advances in Water Resources 000 (2018) 1–34

Fig. 2. Sketches for two porous blocks in “series” and two-layered stratified flow; analysis applies along the x-axis with respect to midpoint interface position
𝑥0 = 𝑥±0 ∓ 12 𝑐𝑞𝑥 .

dependent upon a free-tunable value ; its specific solution  = 1


van- In the “series” configuration along the x-axis, the Darcy velocity 𝑢⃗𝜙 =
4
ishes the truncation correction. Since Λ𝜙 → 1
as ce → 0, this choice is 𝑢⟂ 1⃗ 𝑥 and the flux-jump term 𝐽⃗𝑛 = 𝐽𝑛 1⃗ 𝑥 are both perpendicular with the
interface 𝑥 = 𝑥0 ; the jump value ||𝐽⃗𝑛 ||𝑥0 in Eq. (17b) is non-zero when
4
expected to be robust.
the two blocks have distinguished porosity. In a stratified channel along
Remark 6. In principle, the mass-term in Eq. (19) can be applied with the y-axis, the velocity profile 𝑢⃗𝜙 = 𝑢𝜙,𝑥 (𝑥)1⃗ 𝑦 is parallel with the interface
an individual free-tunable scale parameter 𝑐𝑒(𝑀) ∈ [0, 1], as: [y replaces x in Eq. (16) for stratified flow in this section, see Fig. 2]; Eq.
(23) reduces to the diffusion equation and the interface analysis applies
1 𝑡𝑞 𝑐𝑒(𝑀)
𝑒+ in the (transverse) x-direction with 𝑢⟂ = 0 and 𝐽𝑛 = 0. Let us express Eq.
𝑞 (𝑟
⃗, 𝑡) = 𝑡𝑞 𝑐𝑒 (𝐵𝑛 (𝑟⃗, 𝑡) − 𝑛 (𝑟⃗)) + 𝑛 (𝑟⃗), then
2 𝑠+𝜙 (25) in terms of the (phase-dependent) grid Péclet number Pe⊥ :
1 (𝑀) (𝑀) ̄ 𝑥 𝐵𝑛 − ∇
̄ 𝑥 𝐽 ′ −  ′ − 𝛼𝑀 Δ
̄ 2 ′ = Δ
̄ 2 𝐵𝑛 , 𝑟⃗ ∈ 𝑉𝜙 ,
𝛼𝑀 = 𝜙 (𝑐𝑒 − 𝑐𝑒 )) − Λ𝜙 (1 − 𝑐𝑒 ).
(1 − 2Λ− (27) 𝑃 𝑒⟂ ∇ 𝑛 𝑛 𝑥 𝑛 𝑥 (31a)
4

Eq. (27) reduces to Eq. (19) when 𝑐𝑒(𝑀) = 𝑐𝑒 and 𝛼 M then reduces to 𝑢⟂
Eq. (26b). However, when 𝑐𝑒(𝑀) ≠ 𝑐𝑒 , Bn depends upon ce and Λ− 𝑃 𝑒⟂ = , 𝐽 ′ = −1
𝜙 𝐽 𝑛 ,  𝑛 =  𝜙  𝑛 ,  𝜙 = Λ𝜙 𝑐 𝑒 .
′ −1 −
sepa- (31b)
𝜙 𝜙 𝑛
rately even if their product is fixed.
We consider an interface-cut link 𝑐𝑞𝑥
2 = 1, 𝑐 3 = 𝑐 ; 𝑒± from Eq. (19),
𝑞𝑥 𝑞𝑥 𝑞
4.2. Interface conditions in a parallel and a perpendicular flow 𝑔𝑞− from Eq. (22) and 𝑔𝑞+ from Eq. (24a) then read:

𝑒+
𝑞 = 𝑡 𝑞 𝑐 𝑒 ( 𝐵 𝑛 + Λ𝜙 𝑐 𝑒  𝑛 ) ,

(32a)
Assume a flat interface 𝑥 = 𝑥0 midway the two grid nodes [𝑥− 0
, 𝑥+0
]
and, without loss of generality, fix the interface cut link 𝑐⃗𝑞 such that 𝑥±0
±
1
𝑐
2 𝑞𝑥
= 𝑥0 . The two steady state exact interface conditions of the TRT
𝑒−
𝑞 = 𝑡𝑞 𝑐𝑒 Λ𝜙 𝜖𝑛 𝑐𝑞𝑥 , 𝜖𝑛 = 𝑃 𝑒⟂ 𝐵𝑛 − 𝐽𝑛 ,
− ′ ′ ′
(32b)
operator (8) read ∀{𝑒±
𝑞 } (see Ginzburg, 2007; Ginzburg and d’Humières,
2007):
||𝑆𝑞 ||𝑥0 = 0, with ||𝑆𝑞 ||𝑥0 = 𝑆𝑞 |𝑥+ − 𝑆𝑞 |𝑥− , 𝑔𝑞+ (𝑥) = 𝑡𝑞 𝑐𝑒 Λ−
𝜙 𝑛 (𝑥)𝑐𝑞𝑥 ,
′ 2
(32c)
0 0
1 − 1 −
𝑆𝑞 |𝑥− = 𝑒+
𝑞 + 𝑔 − Λ+ 𝑔 + | − , 𝑆𝑞 |𝑥+ = 𝑒+
𝜙 𝑞 𝑥0 𝑞 − 2 𝑔𝑞 − Λ𝜙 𝑔𝑞 |𝑥+
+ +
, (28a)
0 2 𝑞 0 0
̄ 𝑥 𝐵𝑛 − 1 Δ
̄ 2 𝜖 ′ 𝑐 2 + 𝛼𝑀 ∇
̄ 𝑥 ′ ).
𝑔𝑞− (𝑥) = 𝑡𝑞 𝑐𝑒 𝑐𝑞𝑥 (∇ (32d)
4 𝑥 𝑛 𝑞𝑥 𝑛

||𝐺𝑞 ||𝑥0 = 0, with ||𝐺𝑞 ||𝑥0 = 𝐺𝑞 |𝑥+ − 𝐺𝑞 |𝑥− , The central-difference operators imply: Δ ̄ 𝑞𝜓 = ∇ ̄ 𝑥 𝜓 𝑐𝑞𝑥 = 1 (𝜓 (𝑥 +
0 0 2
̄ 2 ̄
𝑐𝑞𝑥 ) − 𝜓(𝑥 − 𝑐𝑞𝑥 )) and Δ𝑞 𝜓 = Δ𝑥 𝜓 𝑐𝑞𝑥 = 𝜓 (𝑥 + 𝑐𝑞𝑥 ) − 2𝜓(𝑥) + 𝜓(𝑥 − 𝑐𝑞𝑥 ),
2 2
1 + 1 +
𝐺𝑞 |𝑥− = 𝑒−
𝑞 + 𝑔 − Λ−
𝜙 𝑔𝑞 |𝑥0 , 𝐺𝑞 |𝑥0 = 𝑒𝑞 − 2 𝑔𝑞 − Λ𝜙 𝑔𝑞 |𝑥0 .

− +
− − −
+ (28b)
0 2 𝑞 ∀𝜓.
Applying Eq. (10), the effective continuity condition of the scalar
4.3. Effective continuity condition
variable 𝐾(𝑟⃗) is set by Eq. (28a); the effective advective–diffusion
flux-jump condition for 𝑢⃗𝜙 𝐾(𝑟⃗) − 𝜙 ∇𝐾(𝑟⃗) is set by Eq. (28b). The
Plugging Eqs. (32a) and (32c), Eq. (28a) reads with
anti-bounce-back (ABB) and bounce-back (BB) rules (Ginzburg, 2005b;
2017a) apply for the incoming population 𝑓𝑞̄ (𝑟⃗𝑏 , 𝑡 + 1) in the grid bound- 1 −
𝑆𝑞 |𝑥∓ = 𝑡𝑞 𝑐𝑒 (𝐵𝑛 − Λ𝜙 (1 − 𝑐𝑒 )′𝑛 𝑐𝑞𝑥
2
)± 𝑔 | ∓. (33)
ary node 𝑟⃗𝑏 : 0 2 𝑞 𝑥0
ABB ∶ 𝑓𝑞̄ (𝑟⃗𝑏 , 𝑡 + 1) = −[𝑓𝑞 + 𝑔𝑞+ + 𝑔𝑞− ](𝑟⃗𝑏 , 𝑡) + 2𝑒+
𝑞 (𝑟
⃗𝑠 ), (29a) We substitute Eq. (32d) for 𝑔𝑞− (𝑥), factorize tq ce and split 𝑆𝑞 (𝑡𝑞 𝑐𝑒 )−1 into
the two components: 𝑇𝑞(2) |𝑥∓ – the second-order central-difference Tay-
0

BB ∶ 𝑓𝑞̄ (𝑟⃗𝑏 , 𝑡 + 1) = [𝑓𝑞 + 𝑔𝑞+ + 𝑔𝑞− ](𝑟⃗𝑏 , 𝑡), 𝑟⃗𝑏 = 𝑟⃗𝑠 −


1
𝑐⃗ , 𝑟⃗ ∈ 𝐴𝑠 . (29b) lor expansion from 𝑥∓
0
to 𝑥 = 𝑥0 , and the remaining component 𝑅𝑞 |𝑥∓ .
2 𝑞 𝑠
0
Eq. (33) yields:
The BB applies for the no-flux condition in the direct ADE schemes
𝑆𝑞 |𝑥∓ = 𝑡𝑞 𝑐𝑒 (𝑇𝑞(2) |𝑥∓ + 𝑅𝑞 |𝑥∓ ), with (34a)
and TRT − EMM; the ABB applies for the Heaviside input in the direct 0 0 0
RTD computations. We note that ABB and BB steady state closure rela-
tions are comprised in Eqs. (28a) and (28b), respectively: 1̄ 1 ̄2
𝑇𝑞(2) |𝑥∓ = 𝐵𝑛 ± ∇ 𝐵 𝑐 + Δ 𝐵 𝑐2 | ∓ , (34b)
1 − 0 2 𝑥 𝑛 𝑞𝑥 8 𝑥 𝑛 𝑞𝑥 𝑥0
ABB ∶ 𝑒+
𝑞 + 2 𝑔𝑞 − Λ𝜙 𝑔𝑞 |𝑟⃗𝑏 = 𝑒𝑞 (𝑟
+ + +
⃗𝑠 ), (30a)
1 ̄2 1 ̄2 ′ 1 ̄ ′ 𝑐 | ∓ .
𝑅𝑞 |𝑥∓ = −Λ𝜙 (1 − 𝑐𝑒 )′𝑛 𝑐𝑞𝑥
2
− Δ 𝐵 𝑐2 ∓ Δ 𝜖 𝑐 ± 𝛼 ∇
1 0 8 𝑥 𝑛 𝑞𝑥 8 𝑥 𝑛 𝑞𝑥 2 𝑀 𝑥 𝑛 𝑞𝑥 𝑥0
BB ∶ 𝑒−
𝑞 + 𝑔𝑞+ − Λ−
𝜙 𝑔𝑞 |𝑟⃗𝑏 = 0.

(30b) (34c)
2

9
JID: ADWR
ARTICLE IN PRESS [m5GeSdc;June 29, 2018;16:56]

I. Ginzburg Advances in Water Resources 000 (2018) 1–34

When Λ𝜙 and ce are inter-related through Eq. (26a), Eq. (34c) be- key point is that 𝜕𝑥 𝐽𝑛 = 𝑆𝑛 = 𝑛 − 𝑀𝑛 − 𝜙𝜔𝑛 according to Eq. (16e).
comes: When 𝑃 𝑒⟂ = 0, 𝛼𝑀 = 0, Eqs. (31) and (34) with Eq. (35) then become,
( ) ( ) respectively:
1 ̄2 1 ̄2 ′ 1 1 ̄ 𝑥 ′ 𝑐𝑞𝑥 |𝑥∓ .
𝑅𝑞 |𝑥∓ = − ′𝑛 + Δ 𝐵 𝑐2 ∓ Δ 𝜖 𝑐 ± − ∇
0 8 𝑥 𝑛 𝑞𝑥 8 𝑥 𝑛 𝑞𝑥 2 4 𝑛 0 1 ̄2 𝑀 𝜔 𝜙 𝜔
−′𝑛 (𝑥) = (Δ 𝐵 − 𝑀𝑛′ − 𝜔̃ 𝑛 ), 𝑀𝑛′ = 𝑛 , 𝜔̃ 𝑛 = 𝑛 = 𝑛 , (37a)
(35) 2 𝑥 𝑛 𝜙 𝜙 0

Since ce is set constant over the grid, Eq. (28a), as well as Eqs. (34)–
1 1 ̄ 2 𝐽 ′ 𝑐𝑞𝑥 )||𝑥∓ , then
(35), become tq ce independent and parameterized by . A pure diffu- = ∶ ||𝐵𝑛(2) ||𝑥0 = ||(𝑀𝑛′ + 𝜔̃ 𝑛 ∓ Δ 𝑥 𝑛 (37b)
4 8 0
sion through a “series” of porous blocks is comprised with 𝑢⟂ = 0 and
||𝐽⃗𝑛 ||𝑥0 ≠ 0; the EMM then predicts the effective diffusivity and kurtosis
1
due to the porosity contrast (see Ginzburg and Vikhansky, 2018). In a = ∶ ||𝐵2(2) ||𝑥0 = 0, ||𝐵3(3) ||𝑥0 = 0. (37c)
semi-explicit diffusion scheme, formulated by Eqs. (B.4) and (B.5) for 4
comparison with the implicit jump-tracking via Eq. (19), 𝑒− 𝑞 = 0 and the Eq. (37b) is obtained by substituting  = 14 , Eq. (37a) and Δ ̄ 2 𝜖′ =
𝑥 𝑛
flux-jump value ||𝐽⃗𝑛 ||𝑥0 is interpolated to the interface and prescribed ̄ 2𝐽′
−Δ 𝑥 𝑛 into Eq. (35). When 𝑛 = 2, the RHS vanishes in Eq. (37b) be-
explicitly [then 𝜖 ′ ≡ 0 and ′ becomes replaced by ′ + ∇ ̄ 𝑥 𝐽 ′ in Eq. cause 𝜔̃ 𝑛 is constant ∀n, 𝑀2′ is constant and Δ ̄ 2 𝐽 ′ = 0. When 𝑛 = 3,
𝑛 𝑛 𝑛 𝑛 𝑥 2
(32)]. This scheme is expected to behave similar to the interface config- the RHS vanishes in Eq. (37b) because 𝐽3′ (𝑥) is the parabolic func-
uration which takes place in the stratified channels. tion, −2𝜕𝑥2 𝐽3′ − 𝜕𝑥 𝑀3′ = 𝜕𝑥3 𝐵3 when 𝛼𝑀 ( = 0) = 0 [using 𝜕𝑥 𝐽𝑛 = 𝑆𝑛 ],
Example 1 Stratified channels. In that case, 𝑃 𝑒⟂ = 0, Δ̄ 2 𝜖 ′ = 0 and ′ = and ∓ 1 Δ ̄ 2 𝐽 ′ 𝑐𝑞𝑥 = ∓ 1 𝜕 2 𝐽 ′ 𝑐𝑞𝑥 gives a continuation to the interface:
𝑥 𝑛 𝑛 𝑥 𝑛
̄ 2 𝐵 𝑛 − 𝛼𝑀 Δ
̄ 2 ′ is expressed from Eq. (31), the continuity condition 8 8 𝑥 3
−Δ 1
𝑥 𝑥 𝑛 ± 16 𝜕𝑥 𝑀3′ 𝑐𝑞𝑥 for the piece-wise linear function 18 𝑀3′ (𝑥) in the RHS of Eq.
(28a) gives with the help of Eqs. (34)–(35):
𝜕𝑥 𝐵3 (𝑥)𝑐𝑞𝑥 in the LHS of Eq. (37b). Then ||𝐵3(3) ||𝑥0 = 0 for
1 3
(37b), and ∓ 16
∀  ∶ ||𝐵𝑛(1) ||𝑥0 = 0 if 𝑛 (𝑥) = 0 ; (36a) the third-order polynomial solution B3 (x) [similarly as in Eq. (36c)].
We conclude that the semi-explicit and implicit interface-jump track-
ing impose the continuous polynomial solutions to pure diffusion equa-
1
= ∶ ||𝐵𝑛(2) ||𝑥0 = 0 if 𝑛 (𝑥) = 𝑐𝑜𝑛𝑠𝑡𝑛 ; (36b) tion (𝑛 = 2 and 𝑛 = 3) with two distinguished values,  = 18 and  = 14 ,
8
respectively. The choice  = 18 is the same as in stratified channels.
Example 3 Darcy velocity in a “series” when 𝑛 = 1. In that case, 𝑛 (𝑥)
1 ̄ 𝑥 ′ = 0, Δ
̄ 2 ′ = 0 and ∇
̄ 𝑥 𝐽 ′ = 0],
= ∶ ||𝐵𝑛(3) ||𝑥0 = 0 if 𝑛 (𝑥) = 𝑎𝑛 𝑥 + 𝑏𝑛 , with and Jn (x) are piece-wise constant [∇ 𝑛 𝑥 𝑛 𝑛
8
Bn (x) is the discrete exponential function and the term ∓ 81 Δ ̄ 2 𝜖 ′ 𝑐𝑞𝑥 =
1 1 𝑥 𝑛
𝑅𝑞 |𝑥∓ = ± 𝜕𝑥 ′𝑛 𝑐𝑞𝑥 |𝑥∓ = ∓ 𝜕𝑥3 𝐵𝑛 𝑐𝑞𝑥 |𝑥∓ ,
0 16 0 16 0 ̄ 2 𝐵1 𝑐𝑞𝑥 is non-zero in Eq. (35). When ′ (𝑥) is expressed from
∓ 81 𝑃 𝑒⟂ Δ
( ) 𝑥 1

∇ 𝐵 𝑐 =
1
𝜕 𝐵 +
1 3
𝜕 𝐵 𝑐 , then Eq. (31): ′ = 𝑃 𝑒⟂ ∇ ̄ 𝑥 𝐵1 − Δ
̄ 2 𝐵1 , the continuity condition given by
𝑥
2 𝑥 𝑛 𝑞𝑥 2 𝑥 𝑛 12 𝑥 𝑛 𝑞𝑥
1
Eq. (28a) with Eq. (33) takes the form of the asymmetric Taylor expan-
1 1 1 3 sion from the two interface sides 𝑥± = 𝑥0 ± 12 𝑐𝑞𝑥 :
𝑆𝑞 |𝑥∓ = 𝑡𝑞 𝑐𝑒 (𝐵𝑛 ± 𝜕𝑥 𝐵𝑛 𝑐𝑞𝑥 + 𝜕𝑥2 𝐵𝑛 𝑐𝑞𝑥
2
± 𝜕 𝐵 𝑐 )| ∓ ; (36c)
0 2 8 48 𝑥 𝑛 𝑞𝑥 𝑥0 0

̄ 𝑥 𝐵1 𝑐𝑞𝑥 + 𝛼 + Δ
𝑆𝑞 |𝑥∓ = 𝑡𝑞 𝑐𝑒 (𝐵1 ± 𝛼 − ∇ ̄ 2 𝐵1 𝑐 2 )|𝑥∓ , 𝑛 = 1, (38a)
0 𝑥 𝑞𝑥 0
1 1 1 ̄2
= ∶ ||𝐵𝑛(2) ||𝑥0 = ||′𝑛 ||𝑥0 = − ||Δ𝑥 𝐵𝑛 ||𝑥0 , ∀𝑛 (𝑥). (36d)
4 8 8
1 1
̄ 2 𝑛 = 0; 𝐵𝑛(𝑘) (𝑥) 𝛼 − (𝑥∓ )= ∓ Λ𝜙 (1 − 𝑐𝑒 )𝑃 𝑒⟂ 𝑐𝑞𝑥 , 𝛼 − → if Λ𝜙 → 0, (38b)
Eqs. (36a)–(36c) take into consideration that Δ 𝑥
0 2 2
means here the 𝑘th -order central-difference Taylor expansion from the
interface nodes to the interface. Eq. (36a) indicates that a piece-wise- 1
linear function Bn (x) is continuous on the interface. This addresses a 𝛼 + (𝑥∓
0
) = ∓ 𝑃 𝑒⟂ 𝑐𝑞𝑥 + Λ𝜙 (1 − 𝑐𝑒 ). (38c)
8
pure diffusion in “series” with 𝑛 = 1 in Eq. (25), either with the semi-
explicit or implicit interface [since J1 (x) is piece-wise constant, Δ̄ 2 𝜖′ = 0 When Pe⊥ → 0, then 𝛼 − = 12 ∀Λ𝜙 and the analysis reduces to the pure
𝑥 𝑛
in the two cases]. Eq. (36b) tells that the parabolic function Bn (x) is diffusion problem discussed above. However, the perpendicular flow in-
continuous on the interface when 𝑛 (𝑥) is a piece-wise constant only troduces the asymmetric (not equal) coefficients of Pe⊥ in 𝛼 ± for two
provided that  = 18 . This condition supports an exact parabolic solu- neighbor nodes and the leading-order coefficient 𝛼 − becomes accurate
tion Bn (x) to Eq. (25) either in a stratified Darcy flow when 𝑛 = 1 or only in the asymptotic limit Λ𝜙 → 0. This suggests that the midway conti-
with the semi-explicit jump-tracking diffusion scheme in “series” when nuity of the solution B1 (x) and the measured dispersion coefficients will
𝑛 = 2. Eq. (36c) extends the exact closure condition of  = 18 to the third- degrade with Pe⊥ , and that only very small values  in Eq. (26a) will
order polynomial solution Bn (x) governed by the piece-wise linear mass improve for this; however, such a choice is rather unstable (Ginzburg
source. This addresses 𝑛 = 3 (kurtosis) in Eq. (25) with the semi-explicit and Roux, 2015; Kuzmin et al., 2011a). This continuity analysis extends
jump-tracking diffusion scheme in “series” (Eqs. (B.4) and (B.5)). It fol- straightforwardly for the next orders.
lows that  = 18 is the most suitable choice for diffusion equilibrium
𝜖𝑛′ = 0 when the mass-source is either piece-wise constant or linear. Eq. 4.4. Effective flux-jump condition
(36d) indicates that when the correction of 𝛼 M vanishes due to  = 14 ,
the interface continuity condition is exact for parabolic distribution Bn Plugging Eq. (32) into Eq. (28b), an exact form of the interface conti-
nuity condition for the total flux component Φ𝑥 = (𝑢⟂ 𝐵𝑛 − 𝐽𝑛 − 𝜙 ∇ ̄ 𝑥 𝐵𝑛 )
only if ||′𝑛 ||𝑥0 = 0, that is, when 𝑛 scales with 𝜙 and ′𝑛 is continu-
ous. The described situation takes place in the next example. reads:
Example 2 Pure diffusion with the implicit jump-tracking in “series” of
||𝐺𝑞 ||𝑥=𝑥0 = 0, with
porous blocks. In pure diffusion through two heterogeneous blocks,
𝜔2𝑛−1 = 0, Jn (x) and 𝑛 (𝑥) are, respectively, 𝑛 − 1 and 𝑛 − 2 order poly- 1 ̄ 𝑥 𝑛 𝑐 2 ) + 1 Δ̄ 2 𝑒− 𝑐 2 | ∓
𝐺𝑞 |𝑥∓ = 𝑡𝑞 𝑐𝑞𝑥 (Φ𝑥 ±  𝑐 − 𝛼𝑀 ∇
nomials. Let us show that  = 14 in Eq. (36d) assures the continuous 0 2 𝑛 𝑞𝑥 𝑞𝑥 4 𝑥 𝑞 𝑞𝑥 𝑥0
1̄ ̄ 𝑥 𝑛 ± 1 Δ ̄ 2  𝑐 )| ∓
solution both for the parabolic function B2 (x) and the third-order poly- = 𝑡𝑞 𝑐𝑞𝑥 (Φ𝑥 ± ∇ Φ 𝑐 )| ∓ − 𝛼𝑀 𝑡𝑞 𝑐𝑞𝑥 (∇
nomial B3 (x). Eq. (19) matches the diffusion-flux jump implicitly via 2 𝑥 𝑥 𝑞𝑥 𝑥0 2 𝑥 𝑛 𝑞𝑥 𝑥0
̄2 ′ 1 ̄2 − 2 ̄ 𝑥 Φ 𝑥 =  𝑛 + 𝛼𝑀 Δ ̄ 2 𝑛 .
𝑒−
𝑞 = −𝑡𝑞 𝐽𝑛 𝑐𝑞𝑥 , hence Δ𝑥 𝜖𝑛 becomes non-zero for 𝑛 = 3 (kurtosis). The + Δ 𝑒 𝑐 | ∓ , where ∇ (39)
4 𝑥 𝑞 𝑞𝑥 𝑥0 𝑥

10
JID: ADWR
ARTICLE IN PRESS [m5GeSdc;June 29, 2018;16:56]

I. Ginzburg Advances in Water Resources 000 (2018) 1–34

̄ 2 𝑒− 𝑐 2 ||𝑥 = 0, Eq. (39) gives:


When ||Δ 𝑥 𝑞 𝑞𝑥 0
1 1
= ∶ Λ𝜙 = , (42b)
If ||∇̄ 𝑥 𝑛 ||𝑥 = 0, or ∀𝑛 if  = 1 , 𝛼𝑀 () = 0 ∶ 8 8(1 − 𝑐𝑒 )
0 4
̄
||(𝑢⟂ 𝐵𝑛 − 𝜙 ∇𝑥 𝐵𝑛 ) ||𝑥=𝑥0 = ||𝐽𝑛(1) ||𝑥=𝑥0 ,
(1)
5 5
= ∶ Λ𝜙 = . (42c)
1̄ 24 24(1 − 𝑐𝑒 )
with 𝐺𝑞 |𝑥∓ = 𝑡𝑞 𝑐𝑞𝑥 (Φ𝑥 ± ∇ Φ 𝑐 )| ∓ . (40)
0 2 𝑥 𝑥 𝑞𝑥 𝑥0 Firstly, Eq. (42a) vanishes 𝛼 M in Eq. (25) and provides continuous
Eq. (40) presents second-order accurate flux-jump condition of the parabolic and third-order polynomial pure-diffusion solution for B2 (x)
scheme; it is exact with respect to the interface location only when and B3 (x), respectively, in a “series” of porous blocks in the case of the
all flux components are linear functions in space. This is valid in the implicit interface tracking with Eq. (19). Secondly, Eq. (42b) assures
stratified Darcy flow when 𝑛 = 1 ∀ [𝑢⟂ = 𝐽𝑛 = 0, 1 is a piece-wise exact parabolic solution B1 (x), and hence the dispersion coefficient, in
constant, B2 is parabolic]; otherwise, Eq. (40) applies in a stratified a stratified Darcy flow. It also assures continuous, second and third-
flow with  = 14 . Eq. (40) applies in pure diffusion in a “series” when order polynomial functions B2 (x) and B3 (x) in diffusion process in a
𝑛 = {1, 2} ∀ [because 𝑛 is piece-wise constant when 𝑛 = {1, 2}]. How- “series” of the porous blocks when the semi-explicit jump-tracking from
̄ 2 𝑒− 𝑐 2 ||𝑥 is different from zero for all orders in the “perpen-
ever, ||Δ Eqs. (B.4) and (B.5) replaces Eq. (19). We note that Eq. (42b) is known
𝑥 𝑞 𝑞𝑥 0
dicular” Darcy flow and it is expected to modify the coefficient of the for the midway continuity of the parabolic solutions in the presence of
second-order correction in flux-jump effective location in Eq. (40). When the heterogeneous diffusion coefficients (Ginzburg, 2007; 2017b), but
𝑛 (𝑥) is a linear function and Bn (x) is the third-order polynomial, e.g., also for an exact location of the straight Dirichlet boundary with the anti-
in pure-diffusion problem, an exact midway location of the flux-jump bounce-back rule in parabolic distributions governed by the constant
condition is reached with the specific value  = 24 5
: mass-source (see particular solutions Chai and Zhao, 2014; Cui et al.,
2016 and general form Ginzburg, 2017b). When 𝑐𝑒 = 13 , Eqs. (42a) and

𝑛 = 𝑎𝑛 𝑥 + 𝑏𝑛 ,  =
5
, 𝛼 () =
1
∶ (42b) give, respectively, Λ𝜙 = 38 and Λ𝜙 = 16
3
; the first solution is known
24 𝑀 24 for its annihilation of the truncation coefficient due to the resistance
̃ (3) ||𝑥
||Φ = ||𝐽𝑛 ||𝑥0 , with
(2)
variation in the TRT Brinkman-flow scheme (Ginzburg et al., 2015b); the
𝑥 0
second solution assures the midway location of the interface-continuity
Φ̃ 𝑥 = −𝜙 𝜕𝑥 𝐵𝑛 = (Φ𝑥 + 1 𝜙 𝜕 3 𝐵𝑛 ), Φ𝑥 = −𝜙 ∇ ̄ 𝑥 𝐵𝑛 ,
6 𝑥 and the bounce-back no-slip condition in the parabolic velocity profile
̃ 𝑥 ± 1 𝜕𝑥 Φ
̃ 𝑥 𝑐𝑞𝑥 + 1 𝜕 2 Φ
̃ 𝑐 2 )| ∓ . (Ginzburg, 2007).
and 𝐺𝑞 |𝑥∓ = 𝑡𝑞 𝑐𝑞𝑥 (Φ (41)
0 2 8 𝑥 𝑥 𝑞𝑥 𝑥0 Thirdly, Eq. (42c) assures a midway location of the flux-continuity
and no-flux bounce-back boundary in the stratified flow governed by a
Eq. (41) is obtained using −𝛼𝑀 𝜕𝑥 𝑛 = 𝛼𝑀 𝜙 𝜕𝑥3 𝐵𝑛 in pure diffusion with piece-wise linear mass-source. It also assures the exact flux-jump condi-
linear source [∇̄ 𝑥 𝑛 = 𝜕𝑥 𝑛 , Δ
̄ 2 𝑛 = 0]: 𝐺𝑞 |𝑥∓ in Eq. (41) then gets an
𝑥 0 tion for the third-order polynomial diffusion solution in a “series” mod-
1
extra term (𝛼𝑀 + 8
− 16 )𝜙 𝜕𝑥3 𝐵𝑛 which vanishes with 𝛼𝑀 = 1
24
. The mod- eled with the semi-explicit jump-tracking. The implicit jump-tracking
𝑠𝑢𝑚)
eled third-order polynomial solution is exact with  = 24 5
in a stratified, from Eq. (19) needs however a parameter-dependent solution 𝐴(𝐾𝑢
either periodic or bounded by the bounce-back channel governed by from Eq. (B.3) for the same objective. These two (distinguished) solu-
the linear mass source. Eq. (41) also applies in a pure diffusion for 𝑛 = 3 tions provide an exact solution for the kurtosis [up to summation] in
(kurtosis) with the semi-explicit jump-tracking scheme where 𝑒− the pure-diffusion problem. Hence, on the one hand, the semi-explicit
𝑞 = 0,
jump-tracking makes the diffusion problem in “series” equivalent with
flux-jump value ||𝐽𝑛(2) ||𝑥0 is modeled exactly and the total linear mass-
the stratified channels. On the other hand, it does not vanishes the -
source includes 𝜕 x Jn (x) (see Eq. (B.5)).
dependency in parabolic and higher-order solutions Bn (x). To be con-
However, an implicit-jump diffusion scheme (19) incorporates
̄ 2 𝑒− 𝑐 2 ||𝑥 = − 1 𝑡𝑞 ||Δ
̄ 2 𝐽𝑛 𝑐𝑞𝑥 ||𝑥 = trasted with the local “node-centered” implicit tracking in Eq. (19), the
an additional non-zero term 41 ||Δ 𝑥 𝑞 𝑞𝑥 0 𝑥
4 0 semi-explicit jump-tracking is more involved numerically because it is
− 28 𝑡𝑞 ||𝜕𝑥2 𝐽𝑛 𝑐𝑞𝑥 ||𝑥0 with the parabolic distribution Jn (x) for 𝑛 = 3 “link-centered” and requires an interpolation of the prescribed jump-
(kurtosis); the difference with 𝐺𝑞 |𝑥∓ in Eq. (41) then reads: value to the interface.
0
𝑅𝑞 |𝑥∓ = 𝑡𝑞 𝑐𝑞𝑥 ((𝛼𝑀 − 18 )𝜕𝑥2 𝐽𝑛 + (𝛼𝑀 − 1
24
)𝜙 𝜕𝑥3 𝐵𝑛 )|𝑥∓ [using first Besides the space-variable mass sources, a new point in this inter-
0 0
face analysis is the presence of the non-zero interface-perpendicular ve-
−𝛼𝑀 𝜕𝑥 𝑛 = 𝛼𝑀 (𝜙 𝜕𝑥3 𝐵𝑛 + 𝜕𝑥2 𝐽𝑛 ), and then − 81 𝜕𝑥2 𝐽𝑛 𝑐𝑞𝑥 |𝑥∓ for con-
0 locity. It is predicted that the Darcy velocity u⊥ makes the continuity
tinuation of the parabolic solution Jn cqx in interface jump term
condition asymmetric from the two interface sides in “series”; this spe-
||𝐽𝑛(2) ||𝑥0 , Eq. (41)]. A particular (parameter-dependent) dependency cific deficiency only vanishes in the stability limit Λ𝜙 → 0; an impact of
(𝑟𝜙 , 𝑟ℎ ) from Eq. (B.3) then solves ||𝑅𝑞 ||𝑥0 = 0. this property on the moments will be examined in Section 5. Finally,
we note that the developed analysis procedure can be adopted for any
4.5. Summary linear equilibrium and collision operator in simple geometries, with a
possible extension for their rotation with respect to the lattice.
The bulk and interface analysis was developed in a common frame-
work for stratified channels and porous blocks in “series”. A specific 5. Series of porous blocks
dependency between  and ce from Eq. (26a) allows to parameterize
properly the modeled equation and the effective closure relations in We consider a periodic series of two porous blocks along the x-
these two situations. This analysis suggests that due to the linearity of axis with the interface at 𝑥0 = 0 (see Fig. 2). The porosity is 𝜙1 when
the truncation correction and the interface conditions with respect to the 𝑥 ∈] − ℎ1 , 0[ and 𝜙2 when x ∈ ]0, h2 [; the characteristic length is set
mass source 𝑛 and the flux-jump term 𝐽⃗𝑛 , the “𝜔-form” and “𝛾-form” equal to 𝐻 = ℎ1 + ℎ2 . The system can be parameterized through the
𝜙 ℎ
are numerically equivalent in these configurations when the control pa- porosity contrast 𝑟𝜙 = 𝜙1 and aspect ratio 𝑟ℎ = ℎ1 . Since the two blocks
2 2
rameter  from Eq. (26a) takes the same value in both formulations. are interchangeable, all results are invariant with respect to the simul-
Three physical-parameter independent solutions for  have been con- taneous transformation 𝑟𝜙 → 𝑟−1 and 𝑟ℎ → 𝑟−1 ; we consider 𝑟𝜙 ∈ [0, 1]
𝜙 ℎ
sidered: and rh > 0. Closed-form expressions are provided (Ginzburg and Vikhan-
1 1 sky, 2018): (i) the effective diffusivity 𝑒𝑓 𝑓 , skewness coefficients Sk(⋆𝑠)
= ∶ Λ𝜙 = , (42a)
4 4(1 − 𝑐𝑒 ) and Ku(⋆𝑠) (they are gathered in Eq. (B.1)); and (ii), for {𝜔′2 , 𝜔′3 }, and

11
JID: ADWR
ARTICLE IN PRESS [m5GeSdc;June 29, 2018;16:56]

I. Ginzburg Advances in Water Resources 000 (2018) 1–34

hence 𝑘𝑇 (𝑒𝑓 𝑓 ) and Sk(⋆𝑠) , in the presence of the constant Darcy flow (see Eq. (20). The continuity condition is elaborated in Eqs. (34)–(35), and
Eqs. (58), (C.1) and (C.2) in Ginzburg and Vikhansky, 2018). We will discussed in Example 2of Section 4.3; the effective flux-jump condition
measure the differences to the exact solution in the following forms: is developed in Eq. (39); the two conditions have the form of the central-
difference Taylor expansion from the neighbor interface nodes 𝑥0 ± 12 to
𝐷(𝑛𝑢𝑚) − 𝑒𝑓 𝑓 (1 + 𝑘𝑇 )
E(𝑘𝑇 ) = , when 𝑘𝑇 ≠ 0, (43a) x0 .
𝑒𝑓 𝑓 𝑘𝑇 When 𝑛 = 1, then 𝑀1 = 𝑆1 = 𝜔1 = 1 = 0, 𝐽1 = 𝜙 , 𝜕𝑥 𝐽1 (𝑥) = 0,
B1 (x) is piece-wise linear and Eq. (46) are exact [the superscript “(k)”
𝐷(𝑛𝑢𝑚) − 0 (1 + 𝑘𝑇 ) is dropped hereafter]. Hence, B1 (x) is exact and Eq. (16f), or Eq. (18),
E(D) = , (43b) ⟨𝜙 𝜕𝑥 1 ⟩
0 (1 + 𝑘𝑇 ) produces 𝑒𝑓 𝑓 = −𝜔2 |𝑢𝜙,𝑥 =0 = 0 + exactly. Recall, 𝑒𝑓 𝑓 is the
⟨𝜙⟩
length-weighted harmonic mean from Eq. (B.1a).
Sk(⋆𝑠,𝑛𝑢𝑚) Ku(⋆𝑠,𝑛𝑢𝑚) When 𝑛 = 2, then 𝑀2 = 𝜙 and 𝑆2 = 𝜙 𝜕𝑥 1 are piece-wise con-
E(Sk(⋆𝑠) ) = − 1, E(Ku(⋆𝑠) ) = − 1. (43c) stant, B2 (x) is parabolic, 𝐽2 (𝑥) = 𝜙 1 (𝑥) and the diffusive flux
Sk(⋆𝑠) Ku(⋆𝑠)
−𝜙 𝜕𝑥 2 (𝑥) are piece-wise linear. Eq. (46b) assures the flux-jump con-
Eq. (43a) is only applied for Darcy flow in “series”; we will show that dition midway the interface link with Eq. (40) and the solution of the
𝑒𝑓 𝑓 is the exact solution of the pure diffusion TRT − EMM and hence, system is exact ∀Λ𝜙 : 𝜔3 = 0, Sk(⋆𝑠) = 0. However, according to Eq. (37),
E(kT ) vanishes when kT → 0. This is to be contrasted with the direct ADE B2 (x) is discontinuous at 𝑥 = 𝑥0 except when Λ𝜙 obeys Eq. (42a).
solvers in the presence of the numerical diffusion, where E(kT ) diverges When 𝑛 = 3, 𝜔1 = 𝜔3 = 0, 𝑀3 (𝑥) = (𝜙𝜔2 + 𝜙 )1 (𝑥) and 𝑆3 (𝑥) =
when kT → 0. The blocks in “series” are challenging for both numerical 𝜙 𝜕𝑥 2 (𝑥) are piece-wise linear, B3 (x) is the third-order polynomial,
approaches, TRT − EMM and TRT-ADE, because the heterogeneity hap- 𝐽3 (𝑥) = 𝜙 2 (𝑥) and the diffusive flux −𝜙 𝜕𝑥 3 (𝑥) are parabolic. Ac-
pens in the streamwise direction and the associated non-Gaussian effects cording to Eq. (37), 𝐵3 (𝑥+
0
) = 𝐵3 (𝑥−
0
) is only valid provided that Λ𝜙 obeys
are relatively weak but non-zero. Eq. (42a); exact continuous solution B3 (x) of the scheme is exemplified
Section 5.1 constructs and examines the symbolic solutions of the in Fig. 3. In fact, when Λ𝜙 obeys Eq. (26a), the dimensionless B-field
TRT − EMM in pure diffusion. Section 5.2 extends this analysis for a interface jump-error is reduced with the second-order accuracy:
constant Darcy flow. Section 5.3 compares numerical solutions for the
′ (1 + 𝑟ℎ )(−1 + 𝑟𝜙 )(−1 + 4)
three transport coefficients in the Darcy flow between the TRT − EMM ||𝐵2 ||𝑥=𝑥0 = 𝐻 −2 ||𝐵2 ||𝑥=𝑥0 = , (47a)
and the three TRT-ADE schemes from Eq. (11). The TRT-ADE schemes 4(𝑟ℎ + 𝑟𝜙 )𝐻 2
are additionally examined on the continuity of their solutions.
′ 𝑟ℎ (−1 + 𝑟𝜙 )2 (−1 + 4)
||𝐵3 ||𝑥=𝑥0 = 𝐻 −3 ||𝐵3 ||𝑥=𝑥0 = . (47b)
5.1. Pure diffusion: symbolic TRT − EMM solutions 8(𝑟ℎ + 𝑟𝜙 )(1 + 𝑟ℎ 𝑟𝜙 )𝐻 2

In pure diffusion along the x-axis, Eq. (16a) reduces to the one- The two functions from Eq. (47) are displayed in Fig. 3; B2 (x) and
dimensional linear diffusion equation with respect to 𝐵𝑛(𝑘) (𝑥) inside each B3 (x) are continuous at the midway position 𝑥 = 𝑥0 only with  = 14 in
block k: agreement with the interface analysis in Eq. (37). However, according
to the third-order accurate Taylor analysis in Eq. (39), the flux-jump
−𝜕𝑥 𝐽𝑛(𝑘) − (𝑛𝑘) (𝑥) = 𝜕𝑥 𝜙𝑘 0 𝜕𝑥 𝐵𝑛(𝑘) (𝑥), 𝐽𝑛(𝑘) (𝑥) = 𝜙 𝜕𝑥 𝐵𝑛(𝑘−1
)
, 𝑛≥1. (44) condition || − 𝜙 𝜕𝑥 𝐵3 ||𝑥=𝑥0 = ||𝜙 2 (𝑥)||𝑥=𝑥0 is exact only with the
𝑠𝑢𝑚) 𝑠𝑢𝑚)
parameter-dependent value 𝐴(𝐾𝑢 (𝑟𝜙 , 𝑟ℎ ) from Eq. (B.3); 𝐴(𝐾𝑢 (𝑟𝜙 , 𝑟ℎ ) is
Mass-source (𝑛𝑘) obeys Eqs. (16b)–(16e) with 𝑢𝜙,𝑥 = 0. At the interface, 𝑠𝑢𝑚)
illustrated in Fig. 3. We note that 𝐴(𝐾𝑢 increases with the porosity con-
||𝐵𝑛 ||𝑥0 = 0 and || − 𝜙 𝜕𝑥 𝐵𝑛 ||𝑥0 = ||𝜙 𝑛−1 ||𝑥0 ; the periodic condition trast but remains confined to some suitable interval with realistic r𝜙 ,
is applied at the unit cell ends: 𝐵𝑛(1) |𝑥=−ℎ1 = 𝐵𝑛(2) |𝑥=ℎ2 . The TRT − EMM 𝑠𝑢𝑚)
e.g., 𝐴(𝐾𝑢 ∈ [0.29, 0.35] when 𝑟𝜙 ∈ [ 18 , 1[ for any aspect ratio.
solves Eq. (25) with 𝑢𝜙,𝑥 = 0:
The coefficient Ku(⋆𝑠) is estimated with Eq. (14d); its reference so-
̄ 𝑥 𝐽 (𝑘)
−∇ 𝑛 − (𝑛𝑘) (𝑥) − ̄ 2 (𝑘)
𝛼𝑀 Δ 𝑥 𝑛 = ̄ 2 𝐵 (𝑘) (𝑥),
𝜙𝑘 0 Δ 𝑥 𝑛 𝑛≥1. (45) lution is extended to the summation result Ku(⋆𝑠𝑢𝑚) in Eq. (B.2); Ku(⋆𝑠𝑢𝑚)
𝑠𝑢𝑚)
is the exact TRT − EMM solution only if using 𝐴(𝐾𝑢 from Eq. (B.3) in
B-field solution Bn (x) to Eqs. (44)–(45) is n-order polynomial; the 𝐽𝑛(𝑘) (𝑥) combination with the third-order accurate finite-differences from Eq.
and (𝑛𝑘) (𝑥) are, respectively, the polynomials of orders 𝑛 − 1 and 𝑛 − 2. (A.6) for 𝜔4 = −
⟨𝜙 𝜕𝑥 3 ⟩
in Eqs. (16f) and (18). Fig. 4 shows that Ku(⋆𝑠)
Hence, when n ∈ [1, 3], the central-difference operators ∇ ̄ 𝑥 𝐽𝑛(𝑘) and ⟨𝜙⟩
(𝑘) (𝑘) is twice more accurate with Eq. (42a) against Eq. (42b), the relative
̄ ̄
Δ𝑥 𝐵𝑛 (𝑥) are exact, Δ𝑥 𝑛 = 0 and 𝛼 M -correction vanishes. Interface
2 2
error with Eq. (42a) is rather small in very coarse resolution, but its
conditions from Eq. (28) read, e.g., with 𝑥0 = 0 and 𝑐𝑞𝑥 = 1 as (see
amplitude increases with the porosity contrast.
Fig. 2):
In resume, the symbolic solutions exemplify the TRT − EMM scheme
[ ](1) [ ](2) and confirm the more generic interface Taylor analysis. The effective
1 − 1 −
𝑞 + 2 𝑔 𝑞 − Λ𝜙 𝑔 𝑞
𝑒+ + +
|𝑥=− 1 = 𝑒+ 𝑞 − 2 𝑔 𝑞 − Λ𝜙 𝑔 𝑞
+ +
|𝑥= 1 (46a)
2 2 diffusivity and zero skewness are exact solutions of the scheme; the
flux-jump condition and kurtosis are only exact (up to summation)
[ ](1) [ ](2) with Eq. (B.3). In turn, the semi-explicit jump-tracking scheme from
1 + 1 +
𝑞 + 2 𝑔 𝑞 − Λ𝜙 𝑔 𝑞
𝑒− − −
|𝑥=− 1 = 𝑒−
𝑞 − 2 𝑔 𝑞 − Λ𝜙 𝑔 𝑞
− −
|𝑥= 1 (46b) Appendix B.2 matches the flux-continuity and Ku(⋆𝑠𝑢𝑚) with parameter-
2 2
independent solution from Eq. (42c). All these zero-velocity solutions
[ ](1) [ ](2) allow for an efficient verification of the numerical implementation of
1 − 1
𝑞 − 2 𝑔 𝑞 − Λ𝜙 𝑔 𝑞
𝑒+ + +
|𝑥=−ℎ + 1 = 𝑒+
𝑞 + 𝑔𝑞− − Λ+
𝜙
𝑔𝑞+ |𝑥=ℎ − 1 . (46c) Eqs. (16)–(17) at the first three orders, 𝑛 = 1, 2, 3.
1 2 2 2 2

Eqs. (46a) and (46c) are continuity conditions (28a); Eq. (46c) is
5.2. Darcy flow in “series”
the flux condition (28b); the flux condition on the periodic inter-
face is satisfied automatically. Eq. (46) read with Eq. (19) and (24a): ̄ 𝑃
The constant Darcy velocity 𝑢⃗𝜙 = 𝑢⟂ 1⃗ 𝑥 , 𝑢⟂ = −𝑘𝐷 𝜈𝜌𝑥 describes the

(𝑘)
𝑒+
𝑞
(𝑘)
= 𝑡𝑞 𝑐𝑒 (𝐵𝑛(𝑘) (𝑥) + Λ+
𝜙
(𝑛𝑘) (𝑥)), 𝑒−
𝑞
(𝑘) = −𝑡 𝐽 (𝑘) (𝑥)𝑐
𝑞 𝑛
+ (𝑘) =
𝑞𝑥 and 𝑔𝑞 0
flow in “series” with the length-weighted harmonic-mean permeabil-
𝑡𝑞 (𝑛𝑘) (𝑥)𝑐𝑞𝑥
2 ; 𝑔 − (𝑘) (𝑥) is expressed from them with Eq. (22) where
𝑞 ity value kD and a piece-wise linear, continuous pressure distribution
𝑘
̄ 𝑞 𝑒+ and Δ
( ) ̄ 𝑞 𝑔 + (𝑘) apply on the solution of the given block k. Solution
Δ 𝑞 𝑞 P(x). Exact EMM solution (Ginzburg and Vikhansky, 2018) shows that
{𝐵𝑛(1) (𝑥), 𝐵𝑛(2) (𝑥)} is determined by Eqs. (45)–(46) to a constant, fixed by 𝑘𝑇 (𝑒𝑓 𝑓 ) is asymptotically constant and all other transport coefficients

12
JID: ADWR
ARTICLE IN PRESS [m5GeSdc;June 29, 2018;16:56]

I. Ginzburg Advances in Water Resources 000 (2018) 1–34

3 1
Fig. 3. Pure diffusion with TRT − EMM in a “series” of two equal-length blocks (ℎ1 = ℎ2 = 4) with porosity contrast 𝑟𝜙 = 4
(solid, red), 𝑟𝜙 = 2
(dot–dashed, blue),
1 1
𝑟𝜙 = 4
(dotted, black), 𝑟𝜙 = 8
(dashed, magenta). First diagram: continuous solution 3 (𝑥) with Eq. (42a). Second and third diagrams: normalized interface jumps
𝑠𝑢𝑚)
||𝐵2 ||𝑥0 and ||𝐵3 ||𝑥0 from Eq. (47). Last diagram: 𝐴(𝐾𝑢 (𝑟ℎ , 𝑟𝜙 ) from Eq. (B.3) versus rh . (For interpretation of the references to color in this figure legend, the reader
is referred to the web version of this article.)

3 1 1
Fig. 4. Pure diffusion in a “series” of two equal-length blocks ℎ = 𝐻∕2 ∈ [4, 20] with the porosity ratio 𝑟𝜙 = 4
(solid, red), 𝑟𝜙 = 2
(dot–dashed, blue), 𝑟𝜙 = 4
(dotted,
black), 𝑟𝜙 = 1
8
(dashed, magenta). The relative TRT − EMM error E(Ku(⋆𝑠) ) to exact solution (B.1d) when  obeys Eq. (42a) (first diagram) and when  obeys Eq.
(42b) (second diagram). The two next diagrams are similar but with respect to Ku(⋆𝑠𝑢𝑚) from Eq. (B.2). (For interpretation of the references to color in this figure
legend, the reader is referred to the web version of this article.)

𝑥 𝐻 Eq. (38) has the form of the asymmetric Taylor expansion from the two
decay with Pe = 0
:
interface sides; this asymmetry of the leading-order coefficient scales
Pe → ∞ ∶ 𝜔′2 ∼ Pe−1 , 𝜔′3 ∼ Pe−2 , 𝜔′𝑛 ∼ Pe1−𝑛 and with 𝑃 𝑒(⟂𝑘) and vanishes only in the limit Λ𝜙 → 0. The dispersion coeffi-
𝛾𝑛′ ∼ Pe1−𝑛 , ∀ 𝑛 ≥ 1, cient 𝐷 = −𝜔(2𝑛𝑢𝑚) is computed with Eqs. (16f) and (18):

𝐷
𝑘𝑇 (𝑒𝑓 𝑓 ) = − 1 ∼ 𝑐𝑜𝑛𝑠𝑡, Sk⋆ ∼ Pe−1 , Ku⋆ ∼ Pe−1 . (48) 0 ∑ 𝑢 ∑ (𝑘)
𝑒𝑓 𝑓 𝐷(𝑛𝑢𝑚) = 0 + 𝜙𝑘 𝜕𝑥 (1𝑘) (𝑥𝑖 ) − ⟂  (𝑥 ). (52)
⟨𝜙⟩ 𝑘,𝑖 ⟨𝜙⟩ 𝑘,𝑖 1 𝑖
This behavior is principally different from the stratified flow (Ginzburg,
2017b) where the dimensionless coefficients grow with Pe [cf. Eq. (57)].
Numerical result D(num) is reproduced exactly using Eq. (A.6) for
5.3. Symbolic solutions with the TRT − EMM 𝜕𝑥 (𝑛𝑘) (𝑥𝑖 ).
Fig. 5 displays the analytical solution kT (r𝜙 , Pe) and the symbolic
The effective discrete equation (25) reads: prediction E(kT ) from Eq. (43a) versus space resolution per block ℎ =
𝐻
∈ [4, 12] when Pe ∈ [10, 500], 𝑟𝜙 ∈ [ 81 , 34 ]. We observe that E(kT ) in-
̄ 𝑥 𝐵 (𝑘) − ∇
𝑢⟂ ∇ ̄ 𝑥 𝐽 ( 𝑘 ) −  ( 𝑘 ) ( 𝑥 ) − 𝛼𝑀 Δ
̄ 2 (𝑘) = 𝜙𝑘 0 Δ
̄ 2 𝐵 (𝑘) (𝑥). (49)
2
creases with Pe and the porosity contrast; E(kT ) is relatively large, e.g.,
𝑛 𝑛 𝑛 𝑥 𝑛 𝑥 𝑛
E(kT ) ≈ 60% when 𝑟𝜙 = 18 with four nodes per block and the conver-
The discrete solution 𝐵𝑛(𝑘) (𝑥) to Eq. (49) reads:
gence is slower with the mesh refining when the porosity contrast is

𝑘)
1 + 𝑃 𝑒(⟂𝑘) ∕2 𝑢⟂
high. Since the modeled bulk equation is correct [up to discretization],
𝐵𝑛(𝑘) (𝑥) = 𝑃𝑛(𝑘) (𝑥) + 𝑃𝑛(−1 (𝑥)𝑟𝑥𝑘 , 𝑟𝑘 = , 𝑃 𝑒(⟂𝑘) = . (50) the deficiency might be attributed to the leading-order continuity con-
1− 𝑃 𝑒(⟂𝑘) ∕2 0 𝜙𝑘
dition in Eq. (38b) and it should then reduce with . Fig. 6 confirms
this prediction: E(kT ) reduces from ≈ 60% to ≈ 8% with  = 14 × 10−1
The coefficients of the n-order polynomials 𝑃𝑛(𝑘) (𝑥) are determined
by Eq. (49) by taking into account that ∇ ̄ 𝑥 𝑟𝑥 = 1 (𝑟2 − 1)𝑟𝑥−1 and and to ≈ 3 − 4% with  = 14 × 10−2 . Here 𝜙k diminishes in one of the
2
̄ 𝑟 = (𝑟 − 1) 𝑟 . In equal porosity blocks, all transport coefficients in
2 𝑥 2 𝑥 −1 two blocks with porosity contrast and the leading-order coefficient in
Δ 𝑥 continuity equation (38b) degrades when the grid Péclet number 𝑃 𝑒(⟂𝑘)
Eqs. (14c)–(14e) automatically reduce to zero with 𝜔2 = −0 and 𝜔𝑛 = 0
increases. Interestingly, the symbolic solutions reveal a very strong error
for n ≥ 3. In heterogeneous blocks, (1𝑘) (𝑥) = −𝑢⟂ + 𝜙𝑘 𝜔1 , 𝜔1 = 𝑥 =
variation between an even/odd node-number resolution per block, e.g.,
𝑢⟂ 𝜙̄ −1 , 𝑆 (𝑘) = 0, Eq. (49) and 𝐵 (𝑘) (𝑥) read, respectively:
1 1 the even node-number is much more accurate when  = 14 in Fig. 6 [a
𝑃 𝑒(⟂𝑘) ∇
̄ 𝑥 𝐵 (𝑘) + 𝑃 𝑒(⟂𝑘) (1 ̄ 2 𝐵 (𝑘) ,
− 𝜙𝑘 𝜙̄ −1 ) = Δ 𝑥 1 (51a) similar situation is confirmed numerically]. The symbolic solutions also
show that an additional scale parameter 𝑐𝑒(𝑀) >> 𝑐𝑒 from Eq. (27) may
1

improve kT .
⟨𝜙⟩ Hence, the discrete-exponential symbolic solutions provide, together
𝐵1(𝑘) (𝑥) = −𝑥(1 − 𝜙𝑘 𝜙̄ −1 ) + 𝑏𝑘 + 𝑎𝑘 𝑟𝑥𝑘 , 𝜙̄ = . (51b)
𝐻 with the summation and finite-difference approximate, TRT − EMM nu-
The coefficients {ak , bk } are determined with the three interface merical solution and allow to analyze it in the continuous parameter
conditions (46) and Eq. (20). The effective continuity condition from space. However, they become rather involved at the next orders and do

13
JID: ADWR
ARTICLE IN PRESS [m5GeSdc;June 29, 2018;16:56]

I. Ginzburg Advances in Water Resources 000 (2018) 1–34

Fig. 5. Darcy flow through a “series” of two equal-length blocks. Left diagram: Dispersion coefficient kT (r𝜙 ) versus 𝑟𝜙 ∈ [10−1 , 1] when Pe = 10 (solid, red), Pe = 20
(dot–dashed, blue), Pe = 102 (dotted, black), Pe = 5 × 102 (dashed, magenta)[kT is asymptotically constant and almost the same for Pe = 102 and 5 × 102 ]. Middle
diagram: the TRT − EMM error E(kT )[%] from Eq. (43a) when Pe = {10, 20, 102 , 5 × 102 } and 𝑟𝜙 = 14 . Right diagram: E(kT )[%] when Pe = 102 and 𝑟𝜙 = 34 (solid, red),
𝑟𝜙 = 12 (dot–dashed, blue), 𝑟𝜙 = 14 (dotted, black), 𝑟𝜙 = 1
8
(dashed, magenta). (For interpretation of the references to color in this figure legend, the reader is referred
to the web version of this article).

3 1 1
Fig. 6. Darcy flow through a “series” of two equal-length blocks with the TRT − EMM when Pe = 102 and 𝑟𝜙 = 4
(solid, red), 𝑟𝜙 = 2
(dot–dashed, blue), 𝑟𝜙 = 4
1
(dotted, black), 𝑟𝜙 = (dashed, magenta). From left to right: E(kT ) from Eq. (43a) with  =
8
{ 14 , 40
1 1
, 400
in Eq. (26a). These symbolic solutions are only valid when
}
h is integer. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article).

Table 1
The reference EMM values for kT , Sk(⋆𝑠) and Ku(⋆𝑠) from Eq. (14) in a series of two equal-length
porous blocks (𝑟ℎ = 1, ℎ1 = ℎ2 = 4, 𝐻 = 8) at the porosity contrast r𝜙 . The Darcy velocity u⊥ and
𝑢⟂ 𝐻 𝑥 𝐻
molecular diffusion coefficient 0 are the same in all experiments, 𝑥 = ⟨𝜙⟩
, Pe = 0
.
1 1 1 1 1 3
r𝜙 12 10 8 4 2 4

Pe 2.36 × 101 2.33 × 101 2.28 × 101 2.05 × 101 1.71 × 101 1.46 × 101

kT 8.38 6.71 5.05 1.81 3.83 × 10 −1


6.06 × 10−2
Sk(⋆𝑠) −6.46 × 10−2 −8.35 × 10−2 −1.09 × 10−1 −1.91 × 10−1 −1.52 × 10−1 −3.79 × 10−2
Ku(⋆𝑠) −8.67 × 10−3 −2.26 × 10−3 9.34 × 10−3 8.48 × 10−2 1.47 × 10−1 5. × 10−2

not concern stability. Next section complements the TRT − EMM sym- riodic cell and computes the set {𝜔2 , 𝜔3 , 𝜔4 } with Eq. (16f) or Eq. (18);
bolic analysis by numerical computations. their dimensionless counterparts in Eq. (13) are obtained with  = 𝐻
and 0 = 𝑥 ; they are then employed for the three transport coefficients
5.4. The TRT − EMM and TRT-ADE in Eq. (14). Their relative corrections to the reference solution are gath-
ered in Table 2 using  = 14 in Eq. (26a). Like in Fig. 6,  = 14 results
In this section, the TRT − EMM and the three TRT-ADE schemes (11) in very large errors at the high porosity contrast on the coarse grid, and
are examined on the base of their numerical solutions using the same the amplitude of the relative corrections grow with the order n, from
parameter range. The focus is put on the coarse resolution (ℎ1 = ℎ2 = 4, E(kT ) to E(Ku(⋆𝑠) ). Table 3 shows that the three moments improve with
1 1
𝐻 = 8) and high porosity contrast (𝜙1 ∈ [ 12 , 1], 𝜙2 = 1) [we set 𝑐𝑒 = 30 , Λ𝜙 = { 16 , 12
1
} when 𝑟𝜙 ≤ 14 (high contrast). This is especially true with
Λ− = 12 , Λ−
𝜙
= 𝜙Λ− , 𝑢⟂ = 0.02[6]]. The reference EMM solution is gath- |E(kT )|, which diminishes with Λ𝜙 = 12 1
up to the three orders of mag-
ered in Table 1 for kT , Sk(⋆𝑠) and Ku(⋆𝑠)and illustrated in Fig. 8 (bot- nitude in agreement with Fig. 6 and the prediction from the interface
tom row): whereas kT monotonously increases with the porosity contrast Taylor analysis in Eq. (38b).
(r𝜙 → 0), Sk(⋆𝑠) (𝑟𝜙 ) and Ku(⋆𝑠) (𝑟𝜙 ) are not monotonous and their values are The symbolic solutions allow for the search of the root
very small at the two limits, 𝑟𝜙 ≤ 10−1 and r𝜙 → 1: this makes it very dif- (𝑟𝜙 , 𝑟ℎ , 𝐻, Pe), which solves E(𝑘𝑇 ) = 0; as might be expected,
ficult to match them accurately. The kT is asymptotically constant and its value approaches to zero or becomes negative. As one exam-
its value slightly increases with Pe, the Sk(⋆𝑠) and Ku(⋆𝑠) decay with Pe. ple, Λ𝜙 = 2.62 × 10−3 assures E(𝑘𝑇 ) = 0 when 𝑟ℎ = 1, 𝑟𝜙 = 25 , 𝐻 = 8,
Pe = 18.29, in exact agreement with the numerical computations. Using
5.4.1. Numerical solutions by TRT − EMM this value, E(kT ) in Table 3 reduces by two-three orders of magnitude
Recall that TRT − EMM is resumed in the algorithmic form in A.4. when 𝑟𝜙 = { 14 , 12 }. Furthermore, this very small value Λ𝜙 improves
The TRT − EMM sequentially solves Eq. (49) for 𝑛 = 1, 2, 3 in an unit pe- drastically the kurtosis accuracy. To conclude, the TRT − EMM three

14
JID: ADWR
ARTICLE IN PRESS [m5GeSdc;June 29, 2018;16:56]

I. Ginzburg Advances in Water Resources 000 (2018) 1–34

Table 2
The TRT − EMM in small box ℎ1 = ℎ2 = 4, 𝐻 = 8: E(kT ), E(Sk(⋆𝑠) ) and E(Ku(⋆𝑠) ) from Eq. (43)
with respect to the reference values from Table 1. The TRT − EMM is applied with  = 14 in
Eq. (26a). Data are not in percents.
1 1 1 1 1 3
r𝜙 12 10 8 4 2 4

Pe 2.36 × 101 2.33 × 101 2.28 × 101 2.05 × 101 1.71 × 101 1.46 × 101

E(kT ) 1.3 × 101 1.01 × 101 7.39 2.37 4.58 × 10−1 7.03 × 10−2
E(Sk(⋆𝑠) ) 1.1 × 101 5.83 2.68 2.58 × 10−1 1.35 × 10−1 1.63 × 10−1
E(Ku(⋆𝑠) ) 3.29 × 102 8.75 × 102 −1.31 × 102 −2.75 −9.81 × 10−2 2.56 × 10−2

Table 3
1
The ratio of the errors in Eq. (43) with  = 4
from Table 2 to their values with  =
{ 16 , 12
1
, 2.62 × 10 } when 𝐻 = 8.
−3

1 1 1 1 1 3
r𝜙 12 10 8 4 2 4

𝜓 e𝑟𝑟(𝜓)|= 1 ∕e𝑟𝑟(𝜓)|= 1
4 6

kT 5.14 × 101 4.77 × 101 4.52 × 101 8.87 × 101 −9.92 −1.34
Sk(⋆𝑠) 9.01 6.73 4.59 1.36 1.19 −3.36
Ku(⋆𝑠) 1.44 1.5 1.59 1.77 1.58 × 10−1 −3.88 × 10−2
𝜓 e𝑟𝑟(𝜓)|= 1 ∕e𝑟𝑟(𝜓)|= 1
4 12
kT 1.19 × 103 −9.49 × 102 −2. × 102 −2.3 × 101 −3.67 −6.16 × 10−1
Sk(⋆𝑠) 3.28 4.07 8.4 −5.88 × 10−1 −2.65 × 10−1 −3.15 × 10−1
Ku(⋆𝑠) 4.32 4.74 5.37 3.02 1.34 × 10−1 −3.41 × 10−2
𝜓 e𝑟𝑟(𝜓)|= 1 ∕e𝑟𝑟(𝜓)|=2.62×10−3
4
kT 9.70 × 101 9.18 × 101 8.92 × 101 1.40 × 102 −5.89 × 102 1.85 × 101
Sk(⋆𝑠) −3.88 −3.05 −2.26 −8.32 × 10−1 −1.09 −2.10
Ku(⋆𝑠) −1.41 × 101 −1.27 × 101 −1.20 × 101 −3.00 × 102 6.81 × 10−1 −1.39 × 10−1

moments solution is very strongly impacted by the free-tunable param- using the TRT − EMM numerical parameters (given in the beginning of
1
eter  from Eq. (26a) in the perpendicular Darcy flow in “series”; this Section 5.3); the highest porosity contrast 𝑟𝜙 = 25 complements these
situation is similar to a very sharp Λ𝜙 dependency in bounce-back per- computations. Since these computations are conducted close to the sta-
1
meability measurements in dense solid arrays (Khirevich et al., 2014). bility line from Eq. (10e), the ADE-c loses stability with Λ𝜙 = 12 [be-
The “optimal”  value may differ in the three moments, but their cause of its negative numerical diffusion, in agreement with the predic-
accuracy improves with the space resolution where their dependency tions (Ginzburg et al., 2010; Kuzmin et al., 2011a)].
upon Λ𝜙 diminishes. Mean velocity 𝑥 . We observe that ADE-c and ADE-1D both repro-
𝑢⟂ 𝐻
duce 𝑥 = ⟨𝜙⟩ exactly, but this is not the case of ADE-j. Its relative er-
5.4.2. Numerical solutions by TRT-ADE  (𝑛𝑢𝑚)
ror E𝑢 = 𝑥
− 1 is displayed in Table 4. We find that asymptotically
We apply now the three TRT-ADE d2Q5 schemes (11) for the spatial 2Λ− 𝑢
dispersion in a series of porous blocks with the line release (𝑥, 𝑡 = 0) = E𝑢 ≈ − 𝐻 (𝑟𝜙 − 𝑟−1 𝜙

)2 in equal-length blocks and it vanishes in homo-
𝛿(𝑥 − 𝑥0 ). The system is set periodic in the y-direction; the ADE-1D from geneous soil 𝑟𝜙 = 1 in agreement with the previous results, e.g. Ginzburg
Eq. (11c) and ADE-r from Eq. (12) produce the same solutions because (2017a). These and other results show that: (i) Eu scales in proportion
the problem is one-dimensional. While TRT − EMM was operated in one to Λ− 𝑢⟂ ; (ii) Eu is independent of ce ; and (iii) Eu is independent of Λ𝜙 .
unit cell, the direct solvers need a long array of the identical cells [typi- Table 4 confirms that Eu decays linearly with H and that the linear con-
cally, about 102 − 2 × 102 cells]. We measure the central spatial moments vergence is reached faster with the smoother porosity contrast. Hence,
and derive {𝜔1 − 𝜔4 } from them (see Vikhansky and Ginzburg, 2014): E𝑢 ∝ Pe × 𝑐𝑒 × Λ− 2 [recall, Λ− 𝜙
= 𝜙Λ− ]; it reduces linearly when Pe in-
creases with 1∕Λ− , but remains the same when Pe increases with 1/ce .

𝜇 ( 𝑡 + 𝛿𝑡 ) − 𝜇 2 ( 𝑡 ) ⋆
𝜇 1 ( 𝑡 + 𝛿𝑡 ) − 𝜇 1 ( 𝑡 ) Further inspection reveals that an observed alteration of the mean-
𝜔1 = , 𝜔2 = − 2 , seepage velocity is caused by the discontinuity of the equilibrium term
𝛿𝑡 2 𝛿𝑡
𝑒(𝑞𝑢,+) in Eq. (11b). Indeed, the effective continuity condition (28a) pre-
𝜇3⋆ (𝑡 + 𝛿𝑡 ) − 𝜇3⋆ (𝑡)
𝜔3 = , dicts that the concentration will undergo an interface jump in the case
6 𝛿𝑡 of discontinuous weights, when ||𝑒+ 𝑞 ∕||𝑥0 ≠ 0. Concentration profiles in
𝜇4′ (𝑡 + 𝛿𝑡 ) − 𝜇4′ (𝑡) Fig. 7 confirm the presence of asymmetric interface jumps with ADE-j,
𝜔4 = − , 𝜇4′ (𝑡) = 𝜇4⋆ (𝑡) − 12(𝜇2⋆ )2 𝑡2 . (53)
24𝛿𝑡 while ADE-c and ADE-1D both produce very similar smooth profiles in
highly heterogeneous blocks.
The obtained solution (53) is non-dimensionalized with Eq. (13) and The dispersion coefficient kT . The numerical solution kT (r𝜙 ) and the
then employed in Eq. (14). Unlike the TRT − EMM, the direct solvers relative error E(kT ) from Eq. (43a) are displayed in Fig. 8 for three
do not reproduce exactly the Gaussian result Sk(⋆𝑠) = Ku(⋆𝑠) = 0 when schemes (11); their results are compared in Table 5. We observe that
𝑟𝜙 = 1, unless for very specific Λ𝜙 values which vanish the truncation 𝑘𝑇 (𝑒𝑓 𝑓 ) and E(kT ) increase together when the porosity contrast grows
corrections. In particular, they have been identified (Ginzburg, 2017b) 1
to 𝑟𝜙 = 25 . Nevertheless, all schemes match kT within 10–15% in a small
with ADE-j in the uniform porosity, constant velocity profile: E(Sk(⋆𝑠) )
1
cell ℎ1 = ℎ2 = 𝐻∕2 = 4. Despite the mean-velocity error, the ADE-j is
is predicted to vanish with Λ𝜙 = 12 , while the optimal kurtosis choice the most accurate among the three schemes using the stable choice
is Λ𝜙 = 16 . The present computations confirm these solutions with the Λ𝜙 = 14 . The accuracy of the direct solvers with Λ𝜙 = 14 is similar to
three schemes (11) in identical blocks (𝑟𝜙 = 1) where these two par- TRT − EMM results obtained with very small value  [see Tables 2 and
ticular choices reduce Sk(⋆𝑠) and Ku(⋆𝑠) by one order of the magnitude 3], and much better than those which TRT − EMM produces with  = 14 ,
against Λ𝜙 = 14 . In two heterogeneous blocks, the simulations are run Λ𝜙 |𝑐 = 1 ≈ 0.259. The ADE-c and ADE-1D have very similar results in
𝑒 30

15
JID: ADWR
ARTICLE IN PRESS [m5GeSdc;June 29, 2018;16:56]

I. Ginzburg Advances in Water Resources 000 (2018) 1–34

Table 4
The ADE-j relative mean-velocity error Eu (r𝜙 ) in two equal-length porous blocks in “series”;
Λ− 𝑢
E𝑢 (𝑟𝜙 ) ∝ 𝐻 ⟂ , here 𝑢𝜙,𝑥 = 0.02[6] and Λ− = 12 , 𝐻 = {8, 20, 40}. Mean-seepage velocity 𝑥
is exact with ADE-c and ADE-1D ∀r𝜙 , but only in homogeneous blocks 𝑟𝜙 = 1 with ADE-j.
1 1 1 1 3
r𝜙 12 8 4 2 4

E𝑢 |𝐻=8 −2.14 × 10−2 −1.55 × 10−2 −7. × 10−3 −1.66 × 10−3 −2.78 × 10−4
E𝑢 |𝐻=8 /E𝑢 |𝐻=40 3.22 3.82 4.68 4.98 5.03
E𝑢 |𝐻=20 /E𝑢 |𝐻=40 1.55 1.94 2.0 2.0 2.01

Fig. 7. Concentration distribution in two porous blocks in series at 𝑡 = 5 × 104 steps after point release. Two first diagrams: ADE-c (blue) and ADE-1D (red) together
1
when 𝐻 = 8 and 𝐻 = 40 (profiles coincide). Two last diagrams: ADE-j when 𝐻 = 8 and 𝐻 = 40. Data: 𝑟𝜙 = 1∕25, 𝑢⟂ = 0.01(3), 𝑐𝑒 = 60 , Λ− = 12 , Λ−𝜙 = 𝜙Λ− , Pe = 24.6.
(For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article).

1 3
Fig. 8. Three moments in two porous blocks of equal thickness 𝐻∕2 = 4 in series when 𝑟𝜙 ∈ [ 25 , 4 ]. Top row: the reference solution from Table 1 is plotted with the
solid line and numerical solution with symbols. Bottom row: E(kT ), E(Sk(⋆𝑠) ) and E(Ku(⋆𝑠) ) from Eq. (43); left diagram, dashed line: the prediction from Eq. (54) due to
the numerical diffusion in ADE-c. Left column: ADE-c (“lozenge”), ADE-j (“triangles”) and ADE-1D (“squares”) using Λ𝜙 = 14 . Middle column: ADE-j using Λ𝜙 = 12 1

1 1 1 1
(“lozenge”), Λ𝜙 = 4
(“triangles”), Λ𝜙 = 4
in bulk nodes and Λ𝜙 = 3 × 10−2 in interface nodes (“filled squares”). Right column: ADE-j using Λ𝜙 = 12
(“lozenge”); Λ𝜙 = 4
1
(“triangles”); Λ𝜙 = 6
(filled “circles”).

Table 5
Two equal-length porous blocks in “series” when 𝐻 = 8 and  = 14 in Eq. (26a). The table shows
the ratio of the relative errors: ADE-j to ADE-c (labeled “j/c”) and ADE-j to ADE-1D (labeled “j/r”),
1 3
versus 𝑟𝜙 ∈ [ 25 , 4 ].
1 1 1 1 1 3
r𝜙 25 12 8 4 2 4

E(kT ): “j/c” 8.42 × 10−1


8.56 × 10−1
9.99 × 10−1
− −1.13 × 10 −1
−1.47 × 10−2
E(kT ): “j/r” 8.02 × 10−1 7.05 × 10−1 6.93 × 10−1 7.21 × 10−1 9.27 × 10−1 1.31
E(Sk(⋆𝑠) ) ∶ “j/c” 1.00 1.06 1.08 1.04 9.8 × 10−1 9.7 × 10−1
E(Sk(⋆𝑠) ) ∶ “j/r” 9.77 × 10−1 1.01 1.02 1.02 1.03 1.00
E(Ku(⋆𝑠) ) ∶ “j/c” 1.2 1.84 1.94 8.19 × 10−1 9.1 × 10−1 9.62 × 10−1
E(Ku(⋆𝑠) ) ∶ “j/r” 1.05 1.38 1.59 1.1 9.92 × 10−1 1.

16
JID: ADWR
ARTICLE IN PRESS [m5GeSdc;June 29, 2018;16:56]

I. Ginzburg Advances in Water Resources 000 (2018) 1–34

highly heterogeneous blocks. However, while E(kT ) converges to zero is exact for 𝑥 and 0 but not for Sk(s) and Ku(⋆𝑠) in uniform porosity
when r𝜙 → 1 with ADE-j/ADE-1D, E(kT ) diverges with ADE-c because of Darcy flow. With heterogeneous blocks in “series”, ADE-j produces in-
the numerical diffusion D(u) (x). Since the averaged coefficient < D(u) > terface jumps in the concentration solution and they intrinsically mod-
sums with the numerical dispersion, then: ify the mean-seepage velocity. This effect is caused by the heterogene-
ity of the correction 𝑒(𝑞𝑢,+) in Eq. (11b); indeed, it plays the role of the
⟨𝐷(𝑢) ⟩
E(𝑘𝑇 )|ADE-c ≈ E(𝑘𝑇 )|ADE-j + E(𝑢) (𝑘𝑇 ), E(𝑢) (𝑘𝑇 ) = , discontinuous symmetric weights. The relative ADE-j velocity error re-
𝑒𝑓 𝑓 𝑘𝑇 duces linearly with the space resolution and diffusion eigenfunction
Λ− 𝑢2⟂ 𝐻 Λ− , but grows linearly with the velocity amplitude; ADE-c and ADE-
⟨𝐷(𝑢) ⟩ = − 𝜙 = 𝜙Λ .
, Λ− −
(54) 1D reproduce the mean-seepage velocity exactly. Nevertheless, ADE-j
⟨𝜙⟩
remains here most accurate for the dispersion coefficient. The ADE-
This estimate fits well E(kT )|ADE-c in Fig. 8. Hence, ADE-c is unsuitable c omits the discontinuous correction 𝑒(𝑞𝑢,+) . Consequently, it produces
for measurements of small valued kT but the three schemes produce much more continuous profiles, but becomes relatively inaccurate when
similar dispersion values at high porosity contrast, where the Taylor kT is small (small porosity contrast) and the numerical diffusion from Eq.
and numerical dispersion dominate numerical diffusion. (54) dominates the Taylor and numerical dispersion. The ADE-1D from
Skewness Sk(⋆𝑠) and kurtosis Ku(⋆𝑠) . The ADE-j results for skewness and Eq. (11c) manifests itself as a good compromise in the examined param-
kurtosis are displayed in Fig. 8 with different Λ𝜙 values; the ratio of eter range. The three schemes are expected to have similar accuracy at
ADE-j relative errors to ADE-c and ADE-1D is compiled in Table 5 for high Pe but differ for stability. In the present computations, the ADE-j
Λ𝜙 = 14 . In that case, the ADE-c is the most accurate for the two high is the most stable among the three schemes when Λ𝜙 takes small values
moments at the high porosity contrast, where it is followed by the ADE- and also improves the skewness coefficient.
1D, but the ADE-j gains at relatively small contrast. On the whole, the To conclude, the direct solvers behave more accurately with the sta-
distinctions between the three schemes are very small for skewness, and ble choice Λ𝜙 ≈ 14 than the TRT − EMM with  = 14 , Λ𝜙 ≈ 0.259. How-
their results are comparable with the TRT − EMM in Table 2. Although ever, they are restricted to the coarse resolution where their solutions
the difference between three schemes increases up to a factor of two for (especially, skewness and kurtosis) strongly depend on Λ𝜙 , in compli-
kurtosis, they remain much more accurate than TRT − EMM. ance with the predicted (Ginzburg, 2017b) moments truncation impact
Let us compare now the results in “series” with the truncation predic- in constant and parabolic profiles. Although the TRT − EMM needs a
1
tion (Ginzburg, 2017b) in homogeneous soil. Like in that case, Λ𝜙 = 12 finer resolution for the strong heterogeneity in “series”, it is much more
reduces E(Sk(⋆𝑠) ) by the one order of magnitude in small porosity contrast, capable than the direct solvers for high Pe and weak Λ𝜙 -dependency
but only twice in highly heterogeneous blocks where E(Sk(⋆𝑠) ) is domi- with the help of the grid refining.
nated by the implicit interface rather than bulk truncation. In turn, the
computations with very small value Λ𝜙 = 3 × 10−2 , uniquely placed in 6. Stratified flow
the interface nodes, show a drastic improvement of E(Sk(⋆𝑠) ). The ADE-
j and ADE-1D show very similar results reducing Λ𝜙 but ADE-c be- It has been demonstrated (Ginzburg and Vikhansky, 2018) that in
the case of a constant Darcy flow, the non-Gaussian effects are much
comes unstable in this limit. Fig. 8 displays E(Ku(⋆𝑠) ) with Λ𝜙 = { 12
1 1 1
, 6 , 4 }:
1
stronger in the heterogeneous stratified channels than in porous blocks
Λ𝜙 = 6 remains most accurate in highly heterogeneous blocks; the three in “series”. The two systems are sketched in Fig. 2; in this section, the
schemes show here similar results. stratified system is “rotated” and the velocity field is parallel with the
x-axis in compliance with Eq. (16). The EMM and TRT − EMM symbolic
5.5. Summary solutions will be compared in: (i) Poiseuille flow; (ii) bounded system of
open and diffusive layers; and (iii) stratified double-periodic Brinkman
Construction of the exact steady-state symbolic solutions is extended flow, respectively, sketched in Figs. 9, 10 and 14.
from the streamline-invariant (channel) configurations (Ginzburg, 2007; In a stratified channel or duct flow 𝑢⃗𝜙 = 𝑢𝜙,𝑥 (𝑦, 𝑧)1⃗ 𝑥 , the B-field solu-
Ginzburg et al., 2015b) to the streamwise heterogeneity in the presence tion Bn (y, z) is streamwise-invariant, ∇ ⋅ 𝑢⃗𝜙 𝐵𝑛 = 0 and Eqs. (16b)–(16e)
of a space-variable mass-source and an interface flux-jump. The sym-
read with 𝐽⃗𝑛 = 0, 𝑆𝑛 = 0. Thereby, Eq. (16a) reduces to the diffusion
bolic TRT − EMM solutions validate the interface Taylor analysis devel-
equation:
oped in a more general framework. The numerical results confirm that
the steady-state solutions can be parameterized by the dimensionless −𝑛 (𝑦, 𝑧) = 𝜙 (𝜕𝑦2 + 𝜕𝑧2 )𝐵𝑛 (𝑦, 𝑧), 𝑛 ≥ 1. (55)
group {r𝜙 , rh , Pe} provided that the two free parameters, Λ𝜙 and ce , are
inter-related through Eq. (26a); the transport coefficients then all be- Interface conditions (17b) prescribe the continuity for Bn (y, z) and for
come dependent on one free parameter , except the effective diffusiv- the normal-flux on the flat surface A𝜙 (y, z) separating any two porous
ity 𝑒𝑓 𝑓 and Sk(𝑠) = 0 which are exact solutions of the TRT − EMM pure- layers: ||𝐵𝑛 ||𝐴𝜙 (𝑦,𝑧) = 0 and || − 𝜙 ∇𝑛 𝐵𝑛 ||𝐴𝜙 (𝑦,𝑧) = 0; Eq. (17c) prescribes
diffusion scheme at the two first orders 𝑛 = 1 and 𝑛 = 2, respectively. At zero-diffusion flux on the solid surface. Eq. (18) becomes:
order 𝑛 = 3, Ku(𝑠) = 0 is exact in the homogeneous porosity; otherwise,
⟨𝑢𝜙,𝑥 1 ⟩ ⟨𝑢𝜙,𝑥 𝑛−1 ⟩
the two distinguished solutions  produce the reference value Ku(⋆𝑠𝑢𝑚) 𝜔1 = 𝑥 , 𝜔2 = −0 + , 𝜔𝑛 = , 𝑛 ≥ 3. (56)
from Eq. (B.2) with the implicit and the semi-explicit interface-jump ⟨𝜙⟩ ⟨𝜙⟩
trackings; they are useful for implementation check. It is confirmed nu-
The sets {𝜔′𝑛 , 𝛾𝑛′ } from Eq. (13) and the transport coefficients from
merically that the “𝜔-form” and “𝛾-form” produce identical moment so-
Eq. (14) obey (see Ginzburg and Vikhansky, 2018; Vikhansky and
lutions in the Darcy flow in “series” using transform (15), provided that
Ginzburg, 2014):
 takes the same value in both formulations. In the Darcy flow, the
inexactness of the closure relations penalizes the strong porosity con- 𝑘𝑇 ∼ 𝑃 𝑒2 ;
trast and the discrepancy typically increases with the order n, from kT
Pe → ∞ ∶ 𝜔′𝑛 ∼ Pe𝑛−1 , 𝛾𝑛′ ∼ Pe𝑛−1 ,
to Ku(⋆𝑠) ; we find that only very small values  improve the transport √ √
coefficients on the coarse grid. Since this choice may cause instability, Pe → ∞ ∶ Sk(⋆𝑠) ∼ Pe, Ku(⋆𝑠) ∼ Pe, Sk(⋆𝑡) ∼ Pe, Ku(⋆𝑡) ∼ Pe. (57)
we should conclude that the TRT − EMM needs relatively fine grids in
“series”. The TRT − EMM operates in the stratified geometry with 𝐽⃗𝑛 = 0 in Eq.
At the same time, the three direct TRT-ADE schemes (11) are ex- (19) and 𝑆𝑛 = 0 in Eq. (16e). The system is closed with the continuity
amined in heterogeneous series of porous blocks. We note that ADE-j condition (28a) and either a symmetric, periodic or bounce-back no-flux

17
JID: ADWR
ARTICLE IN PRESS [m5GeSdc;June 29, 2018;16:56]

I. Ginzburg Advances in Water Resources 000 (2018) 1–34

Fig. 9. The TRT − EMM with Eq. (42a) in Poiseuille flow sketched in the first diagram. From left to right: the relative errors (in log-scale) E(D) , E(Sk(⋆𝑠) ) and E(Ku(⋆𝑠) )
from Eq. (43) versus H.

Fig. 10. Reference solution in fracture/diffusive-matrix two-layered bounded system sketched in the first diagram. From left to right: 𝑘𝑇 Pe−2 and asymptotic
ℎ1
solution Sk(⋆𝑠) Pe− 2 and Ku(⋆𝑠) Pe−1 from Eq. (14) versus 𝑟ℎ =
1
1 1
ℎ2
at 𝜙 = 10−2 (solid, red), 3
× 10−1 (dot–dashed, blue), 10−1 (dotted, black) and 2
(dashed, magenta);
𝑟ℎ = 0 corresponds to Poiseuille flow. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article).

condition (29b) on the horizontal ends. The effective interface continu- 𝛼𝑀 Pe2 Pe2
The second-order correction − 5𝐻 2
to the Taylor value 𝑘𝑇 = 210
ity condition ||𝐵𝑛 ||𝑥=𝑥0 = 0 is elaborated in Eqs. (34)–(36); it is exact
has place because of the parabolic distribution 1 (𝑦); when  = 14
with the midway interface position when Bn (y) is linear ∀Λ𝜙 , and when
and 𝛼 M vanishes with Eqs. (26a) and (42a), the remaining correction
Bn (y) is parabolic with  = 18 , Eq. (42b). The B-field B1 (y) is parabolic
in Eq. (58b) decays with fourth-order rate. In fact, Eq. (58b) is the
in a stratified Darcy flow because the mass-source 1 (𝑦) is a piece-
discrete counterpart of the Taylor solution, when the Taylor Ansatz
wise constant. A recursive construction of the symbolic solution of the
(Taylor, 1953) is solved with the discrete Laplacian and integration
TRT − EMM follows the EMM exact procedure (Ginzburg and Vikhan-
is performed via summation (see Ginzburg and Roux, 2015; Ginzburg
sky, 2018) but the averaging integration in Eqs. (16) and (20) becomes
⟨𝑢𝜙,𝑥 (𝑦)⟩
and Vikhansky, 2018). Using  = 14 , this property remains valid for all
performed via summation of the grid values, with 𝜔1 = ⟨𝜙(𝑦)⟩
=  (𝑠𝑢𝑚) . higher orders: the difference with the exact solution is only due to the
discrete Laplacian and the summation effect using Eq. (42a). Fig. 9 dis-
plays the corresponding relative errors versus channel resolution when
6.1. Straight Poiseuille flow Pe ∈ [1, 103 ] for E(D), E(Sk(⋆𝑠) ) and E(Ku(⋆𝑠) ) from Eq. (43). It is shown
that |E(D)| is very small and it becomes Pe-independent as Pe ≥ ≈ 102 .
In Poiseuille flow 𝑢𝜙,𝑥 (𝑦) = 32 𝑥 (1 − 4𝑦2 ∕𝐻 2 ), 𝜙(y) ≡ 1, the EMM and At the same time, |E(Sk(⋆𝑠) )| and |E(Ku(⋆𝑠) )| increase progressively with the
TRT − EMM solutions Bn (y) to Eq. (25) are order 4n polynomials. The order n, but they decrease very rapidly with the grid refining and agree
reference EMM solutions for {𝜔′𝑛 }, {𝛾𝑛′ } and the transport coefficients within 5% with the EMM when H ≈ 10. The relative corrections are al-
from Eq. (14) are provided (Ginzburg and Vikhansky, 2018, see Ta- most Pe-independent because kT , Sk(⋆𝑠) and Ku(⋆𝑠) obey the Pe-scale from
bles 2 and 3). The coefficients of Bn (y) are set by Eq. (25) except the Eq. (57) on the numerical solution. The effective transport coefficients of
first and zeroth orders. The coefficient of the linear-term is set by the the ADE-j scheme (11) in a straight/radial Poiseuille flow have been de-
bounce-back closure, but its value does not affect the solution for {𝜔n } rived (Ginzburg, 2017b; Ginzburg and Roux, 2015). Using a coordinate
in Eq. (56) [because < 𝑢𝜙,𝑥 (𝑦)𝑦 >= 0 ]; hence, the transport coefficients stencil in a straight channel, the first moment produces mean-velocity
are not affected by an exact location of the no-flux boundary inside one  (𝑠𝑢𝑚) exactly; the second moment reproduces Eq. (58b) only provided
mesh spacing. When 𝑛 = 1, 1 (𝑦) =  (𝑠𝑢𝑚) − 𝑢𝑥 (𝑦), Δ ̄ 2 1 = −𝜕 2 𝑢𝜙,𝑥 (𝑦),
𝑦 𝑦 that the numerical dispersion vanishes thanks to Λ𝜙 = 14 (1 − 𝑐𝑒 (Λ− )2 )
2 ̄
B1 (y) solves: −1 + 𝛼𝑀 𝜕𝑦 𝑢𝜙,𝑥 = 0 Δ𝑦 𝐵1 (𝑦) and Eq. (56) produces the
2
(see Eqs. (72)–(76) in Ginzburg and Roux, 2015). With this choice,
dispersion value 𝐷(𝑛𝑢𝑚) = −𝜔(2𝑛𝑢𝑚) . The dispersion coefficient kT from Λ𝜙 → 14 when ce → 0 (Pe → ∞) and the TRT − EMM with Eq. (42a) shares
Eq. (14c) obtains solution: the same property; an advanced stability (Ginzburg et al., 2010) of
Λ𝜙 = 14 was confirmed (Ginzburg and Roux, 2015) in Poiseuille pro-
Pe2 (20 − 21𝐻 2 )Pe2 (4 − 5𝐻 2 + 𝐻 4 )Pe2  𝐻
𝑘(𝑇𝑛𝑢𝑚) = + − 𝛼𝑀 , Pe = 𝑥 , file at high Pe. We have also compared the highly-accurate truncation
210 210𝐻 6 5𝐻 6 0 estimate (Ginzburg, 2017b) for Sk(⋆𝑠) and Ku(⋆𝑠) against the TRT − EMM
(58a) solutions with  = 14 in the limit ce → 0: the ADE-j with Λ𝜙 = 14 shows
very slightly more accurate results than the TRT − EMM. To conclude,
and to be contrasted with the Darcy flow in “series”, it is underlined
Pe2 (20 − 21𝐻 2 )Pe2 1
𝑘(𝑇𝑛𝑢𝑚) = + if Λ𝜙 = , 𝛼 = 0. (58b) that in Poiseuille flow the direct solvers and TRT − EMM produce very
210 210𝐻 6 4(1 − 𝑐𝑒 ) 𝑀
similar results around stable choice Λ𝜙 = 14 .

18
JID: ADWR
ARTICLE IN PRESS [m5GeSdc;June 29, 2018;16:56]

I. Ginzburg Advances in Water Resources 000 (2018) 1–34

Fig. 11. Symbolic TRT − EMM solutions in fracture/diffusive-matrix with two equal-thickness layers at Pe = 102 . From left to right: E(D), E(Sk(⋆𝑠) ) and E(Ku(⋆𝑠) ) from Eq.
(43) (in log scale) versus channel width H when 𝜙 = 10−2 (solid, red), 13 × 10−1 (dot–dashed, blue), 10−1 (dotted, black) and 12 (dashed, magenta). (For interpretation
of the references to color in this figure legend, the reader is referred to the web version of this article).

6.2. A bounded stratified open/diffusive system plies numerically that H is even. Figs. 11–13 display |E(D)|, |E(Sk(⋆𝑠) )|
and |E(Ku(⋆𝑠) )| at Pe = 102 towards H. In Fig. 11, the two layers have the
We consider a system of two parallel layers of the thickness h1
same thickness (𝑟ℎ = 1), but 𝜙 varies, 𝜙 ∈ [10−2 , 12 ]. In Figs. 12 and 13,
and h2 bounded at 𝑦 = 0 and 𝑦 = 𝐻 = ℎ1 + ℎ2 ; the system is sketched
𝜙 = 10−1 , but the porous layer is thinner (rh ≤ 1) and thicker (rh ≥ 1),
in Fig. 10: the bottom layer has porosity 𝜙 and zero velocity 𝑢⃗𝜙 = 0;
respectively, than the open one. Fig. 11 shows that kT is more accurate
the top layer mimics a fracture (𝜙 = 1) and velocity profile u𝜙, x (y) is
with the smaller 𝜙; however, Sk⋆ and Ku⋆ become less accurate when
parabolic there. The transport coefficients in Eq. (14) are set by Pe, 𝜙
𝜙 diminishes. The results displayed in Fig. 12 show that E(D) improves
and an aspect ratio 𝑟ℎ = ℎ1 ∕ℎ2 ; they obey Pe-scaling from Eq. (57) and
when rh reduces to zero and the system approaches the Poiseuille flow;
reduce to the Poiseuille solution when either 𝑟ℎ = 0 or 𝜙 → 0. The set
E(Sk(⋆𝑠) ) and E(Ku(⋆𝑠) ) are not monotonous with rh : 𝑟ℎ = 0 and 𝑟ℎ = 1 are
{𝛾′(n) } from Eq. (13) is constructed (Vikhansky and Ginzburg, 2014,
more accurate than 𝑟ℎ ∈ [0, 1]. Fig. 13 shows that when an open layer is
Eqs. (124)–(132)) and extended for a radial capillary; the reference so-
not resolved sufficiently, e.g., 𝑟ℎ = 4, ℎ2 = 1 when 𝐻 = 5 or ℎ2 = 2 when
lution (Ginzburg and Vikhansky, 2018) for {𝜔′2 − 𝜔′4 } is employed here
𝐻 = 10, kT becomes very inaccurate. At the same time, Sk(⋆𝑠) and Ku(⋆𝑠)
(Ginzburg and Vikhansky, 2018, Eqs. (C.10)–(C.12)) with 𝑟𝑏 = 𝑟−1 ℎ
). The
are much less affected by rh .
described here open/diffusive two-layered bounded system is similar,
In resume, because of an inexact interface continuity condition, the
but not identical with a fracture embedded into diffusive matrix, also
TRT − EMM is less accurate than in open flow, but the errors decay
extended Ginzburg and Vikhansky (2018) to the porous Brinkman flow
rather rapidly with the space resolution When 𝜙 = 10−1 , the three trans-
in a fractured layer.
port coefficients agree with the reference result within ≈ 5% with 5
Since the solute may diffuse into the impermeable layer, the
mesh spacings per layer using  = 14 . This is comparable with the mea-
fracture/diffusive-matrix manifests strong deviations from the Gaussian
surements of the kurtosis in pure diffusion in “series” with porosity con-
prediction. Fig. 10 illustrates the porosity and aspect ratio effects in Pe-
1 trast 𝑟𝜙 = 10−1 , and it is much more accurate than in a perpendicular
independent solutions: kT /Pe2 and the two asymptotes, Sk(⋆𝑠) Pe− 2 and Darcy flow (cf. Table 2).
Ku(⋆𝑠) Pe−1 . It is shown that at fixed Pe, kT monotonously increases with
rh and 𝜙. In contrast, when the diffusive layer is not too much thick, Sk(⋆𝑠) 6.3. Porous Brinkman flow
and Ku(⋆𝑠) manifest very sharp peaks where their amplitudes grow when
𝜙 diminishes. Unlike in Poiseuille profile, where Sk(⋆𝑠) < 0 (the profiles We construct the reference EMM solutions in stratified Brinkman
flow in Section 6.3.1 and compare them with the symbolic TRT − EMM
are shifted right), Sk(⋆𝑠) may change its sign in the presence of the diffu-
solutions using the exact and discrete velocity profiles in Section 6.3.2.
sive matrix. The direct computations (Vikhansky and Ginzburg, 2014)
with the ADE-j scheme (11b) of the spatial release and RTD in temporal
6.3.1. The EMM is stratified porous flow
process demonstrate very sharp distribution profiles and agree with the
Consider a stratified system of two porous layers of porosity 𝜙i ,
EMM prediction, e.g., within 5% for all three moments at small porosity
permeability ki and width hi , 𝑖 = 1, 2, separated by a flat interface
𝜙 = 10−1 and Pe ≈ 14 in a channel of 𝐻 = 12. The TRT − EMM solves
at 𝑦 = 𝑦0 = − 𝐻2 + ℎ1 , 𝐻 = ℎ1 + ℎ2 . The velocity profile 𝑢⃗𝜙 = 𝑢𝜙,𝑥 (𝑦)1⃗ 𝑥
Eq. (25) for n ≥ 1; 𝐵𝑛(𝑘) (𝑦) is the 2n-polynomial in a porous layer (𝑘 = 1),
solves Brinkman equation (Brinkman, 1947) in the form (Ginzburg et al.,
and 4𝑛−polynomial in an open layer (𝑘 = 2). The symbolic system is
2015b):
closed by Eq. (28a) at the interface and by the bounce-back closure re-
lation (28b) in two boundary nodes, 𝑦 = 12 and 𝑦 = 𝐻 − 12 ; a constant 𝜈𝜌0 (𝑖) 𝜈𝜌0 2 (𝑖)
𝑢 +∇̄ 𝑥𝑃 = 𝜕 𝑢 , (59)
is fixed with Eq. (20). The closure relations are all prescribed with the 𝑘𝑖 𝜙 𝑓𝑖 (𝜙) 𝑦 𝜙
use of Eqs. (19), (22) and (24a), similar as in Eqs. (32)–(35); the trans- 𝜈
verse central differences are all computed on the one-layer solution. Eq. fi (𝜙) ∈ ]0, 1] determines Brinkman viscosity 𝑓𝑖 (𝜙)
, lim𝜙𝑖 →1 𝑓𝑖 (𝜙) = 1 and
(36d) tells us that when Λ𝜙 obeys Eq. (42a), then the continuity con- typically, 𝑓𝑖 (𝜙) = 1 or 𝑓𝑖 (𝜙) = 𝜙𝑖 . Velocity profile reads in the ith layer:
dition ||𝐵𝑛 ||𝑦=𝑦0 = 0 is not located midway at 𝑦0 = ℎ1 . As a result, the √
dispersion coefficient kT obtains second-order parameter-dependent cor- (𝑖)
̄ 𝑥𝑃
∇ 𝑓𝑖 (𝜙)
𝜉𝑖 𝑦 − 𝜉𝑖 𝑦
𝑢𝜙 (𝑦) = − (𝑘 − (𝑎𝑖 𝑒 + 𝑐𝑖 𝑒 )), 𝜉𝑖 = . (60)
rection which vanishes for Poiseuille flow in Eq. (58b). The constructed 𝜈𝜌0 𝑖 𝑘𝑖
symbolic solutions for the three transport coefficients agree exactly with
the TRT − EMM numerical solution and they are employed to explore The velocity and shear stress are set continuous on the interface:
the accuracy of the scheme within a continuous parameter space. The ||𝑢𝜙,𝑥 ||𝑦=𝑦0 = 0 and || 𝑓 𝜈(𝜙) 𝜕𝑦 𝑢(𝜙𝑖) ||𝑦=𝑦0 = 0, respectively. We consider a
𝑖
symbolic solutions are formally defined for any rh and H ≥ 2; the nu- double periodic system along the x and y axes which is sketched in
merical system has an integer number of cells per layer, e.g., 𝑟ℎ = 1 im- Fig. 14 together with the profile 𝑢(𝜙𝑖) (𝑦). A bounded system can be ad-
dressed similarly, with the limiting case of the open/diffusive-matrix

19
JID: ADWR
ARTICLE IN PRESS [m5GeSdc;June 29, 2018;16:56]

I. Ginzburg Advances in Water Resources 000 (2018) 1–34

Fig. 12. Fracture/diffusive-matrix. Similar as in Fig. 11 but when an open layer layer is thicker than the diffusive one, 𝑟ℎ = ℎ1 ∕ℎ2 = { 14 , 12 , 1} and 𝜙 = 10−1 ; 𝑟ℎ = 0
corresponds to Poiseuille flow.

Fig. 13. Fracture/diffusive system. Similar as in Fig. 11 but when a diffusive layer is thicker than the open one, 𝑟ℎ = ℎ1 ∕ℎ2 = {1, 2, 4} and 𝜙 = 10−1 ; 𝑟ℎ = 0 corresponds
to Poiseuille flow.

1 1
1
Fig. 14. Stratified periodic Brinkman flow is sketched on the first diagram; the three next diagrams show results at D𝑎− 2 ≈ 56.5 (dotted, black) and D𝑎− 2 ≈ 20 × 56.5
(dashed, red), when 𝑘1 ∕𝑘2 = 10−3 and 𝑓 (𝜙) = 1, 𝜙1 = 1 and 𝜙2 = 1 . Second diagram: velocity profile (60). Third diagram: 𝜎 𝐶̄ (𝑥̄ , 𝑡) is reconstructed from Sk(𝑠) and
30 3 ⋆

Ku(⋆𝑠) when t/T ≈ 0.49Pe, Pe ≈ 740. Fourth diagram: RTD profile 𝜎P(x, 𝜏) is reconstructed from Sk(⋆𝑡) and Ku(⋆𝑡) when x/H ∼ 1.14Pe, Pe ≈ 740. Data at D𝑎− 2 ≈ 56.5
1

(dotted, black): Sk(⋆𝑠) Pe− 2 ≈ −0.88, Ku(⋆𝑠) Pe−1 ≈ 0.99, Sk(⋆𝑡) Pe− 2 ≈ 1.09, Ku(⋆𝑡) Pe−1 ≈ 1.7. Data at D𝑎− 2 ≈ × 56.5 (dashed, red): Sk(⋆𝑠) Pe− 2 ≈ −0.51, Ku(⋆𝑠) Pe−1 ≈ 0.41,
1 1 1 1
1
20
Sk(⋆𝑡) Pe− ≈ 0.74, Ku(⋆𝑡) Pe−1 ≈ 0.9. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article).
1
2

from Section 6.2, when permeability tends to zero in one layer and be- and Vikhansky, 2018; Vikhansky and Ginzburg, 2014) through strati-
comes infinite in the other. In the double-periodic system, the four con- fied porosity and in a single bounded porous channel (Ginzburg and
stants in Eq. (60) are derived from two interface conditions and two Vikhansky, 2018). The Darcy system presents an interesting example
symmetry relations in the middle of each layer. Prescribing {fi (𝜙)}, the where the complex dispersion behavior is solely due to the porosity
velocity profile can be formulated in terms of the dimensionless Darcy contrast; this system manifests much stronger non-Gaussian effects than
𝑘
number D𝑎 = 𝐻𝐷2 (with 𝑘𝐷 = (𝑘1 ℎ1 + 𝑘2 ℎ2 )∕𝐻), the permeability con- the “perpendicular” Darcy flow in series at the same porosity contrast
trast and the aspect ratio. When the permeability has the same value in and Pe. In a single porous channel, the transport coefficients are set by
the two layers, 𝑘𝑖 = 𝑘𝐷 , the double-periodic system reduces to a con- two dimensionless parameters, f(𝜙)/Da and Pe; the relative dispersivity
̄ 𝑥𝑃
stant (“parallel”) Darcy flow 𝑢(𝜙𝑖) (𝑦) = − 𝑘𝑇 Pe−2 is predicted to decrease from the Stokes to the Darcy regime and

𝜈𝜌0
𝑘𝐷 . The Brinkman profile
(60) falls to the Darcy regime at small Da, and to the Stokes regime to increase when f(𝜙) decreases with 𝜙 at fixed Da. Exact solution for
at very high Da; the interface effects become important at the interme- the transport coefficients in a fracture embedded into porous Brinkman
diate (Brinkman) regime, e.g., see Silva and Ginzburg (2016). flow has been constructed and confirmed numerically (Ginzburg and
The EMM exact solution for the transport coefficients in Eq. (14) has Vikhansky, 2018); this system can be regarded as the limit system of
been constructed and analyzed in a “parallel” Darcy flow (Ginzburg

20
JID: ADWR
ARTICLE IN PRESS [m5GeSdc;June 29, 2018;16:56]

I. Ginzburg Advances in Water Resources 000 (2018) 1–34

Fig. 15. Effect of fk (𝜙) in stratified periodic Brinkman flow, Eq. (59). From left to right: the ratio of the Brinkman permeability to its Darcy value, kB /kD [ 𝑓𝑘 (𝜙) = 1
(solid red), 𝑓𝑘 (𝜙) = 𝜙𝑘 (dot–dashed, blue)] and the three reference values: 𝑘𝑇 Pe−2 , Sk(⋆𝑠) Pe− 2 and Ku(⋆𝑠) Pe−1 from Eq. (14) versus D𝑎− 2 . Legend: Pe = 102 (red, solid)
1 1

1
and Pe = 103 (black, dotted) when 𝑓𝑘 (𝜙) = 1; Pe = 102 (dot–dashed, blue) and Pe = 103 (dashed, magenta) when 𝑓𝑘 (𝜙) = 𝜙𝑘 . Data: 𝑘1 ∕𝑘2 = 10−3 , 𝜙1 = 30 , 𝜙2 = 13 ,
ℎ1 = ℎ2 = 𝐻∕2. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article).

√ √
the double-periodic Brinkman flow system. All these stratified systems versus 𝑥̄ = (𝑥 − 𝑥0 − 𝑥 𝑡)∕𝜎, 𝜎 = 𝜇2⋆ (𝑡) = 2𝐷𝑡 and (ii), the RTD𝜎P(x,
obey the Pe-scaling from Eq. (57). √ √
In Brinkman flow, 𝐵𝑛(𝑖) (𝑦) solves Eq. (55) where the source term 𝜏) versus 𝜏 = (𝑡 − 𝑡̄)∕𝜎, 𝑡̄ = 𝑥∕𝑥 , 𝜎 = 𝜇2⋆ (𝑥) = 2 𝐷3 𝑥. As has been
𝑥
𝑛 (𝑦) is computed with Eq. (16b) on the profile (60). Solution 𝐵𝑛(𝑖) (𝑦)
expected from the predicted moments, the profiles are more asymmet-
reads in terms of the 2k-order polynomials 𝑃𝑖(2𝑘) (𝑦) and 𝑋𝑖(2𝑘) (𝑦) as fol-
ric and sharp in the Darcy zone. Mean-concentration profiles are skewed
lows (see also dimensionless form Ginzburg and Vikhansky, 2018):
right since Sk(⋆𝑠) < 0. The RTD are typically skewed left, Sk(⋆𝑠) > 0 because
𝑛
∑ 𝑛
∑ of the Heaviside entry. These profiles can be compared to their coun-
𝐵𝑛(𝑖) (𝑦) = 𝑃𝑖(2𝑘) (𝑦)𝑒(𝑛−𝑘)𝜉𝑖 𝑦 + 𝑋𝑖(2𝑘) (𝑦)𝑒(𝑘−𝑛)𝜉𝑖 𝑦 , 𝑛 ≥ 1. (61) terparts in Poiseuille flow and in the limit case of a fracture embedded
𝑘=0 𝑘=0
into diffusive matrix (see results for Pe ≈ 740 in Figs. 10, 22 and 23 in
When 𝑘1 = 𝑘2 , 𝐵𝑛(𝑖) (𝑦) reduces to the 2n-order polynomial in the Darcy Ginzburg and Vikhansky, 2018). The profiles are clearly more asymmet-
flow. The coefficients in Eq. (61), except the zeroth and first-order terms, ric in the stratified Brinkman system than in an open flow; in the limit
are set by Eq. (55). The three remaining coefficients are fixed through system, their form depends on the porosity in the diffusive layer. The
the interface continuity condition ||𝐵𝑛(𝑖) ||𝑦=𝑦0 = 0 and the two symmetry constructed EMM moments are employed now to verify the TRT − EMM.
conditions; the remaining constant is fixed by Eq. (20). The final expres-
sions are lengthy but they allow to explore the transport coefficients in 6.3.2. The TRT − EMM in stratified porous flow
the full continuous parameter space. Once the set {𝜔n } is constructed, We construct symbolic TRT − EMM solutions in Brinkman velocity
the set {𝛾 n } is derived from it with Eqs. (13) and (15); the two sets de- profile extending the known classes of the exact solutions, the polyno-
termine the transport coefficients in Eq. (14); they obey the Pe-scale in mials (Cui et al., 2016; Ginzburg, 2007; 2017a) and the discrete expo-
Eq. (57). nents (Ginzburg et al., 2015b). The symbolic solutions allow to examine
Let us exemplify the role of the Brinkman viscosity in the two sys- the response of the transport coefficients on the numerical formulation,
tems, 𝑓𝑖 (𝜙) = 1 and 𝑓𝑖 (𝜙) = 𝜙𝑖 . Fig. 15 displays the relative effective per- but also on the quality of the velocity field, e.g., when Brinkman flow is
⟨𝑢𝜙,𝑥 ⟩𝜈𝜌0 computed with the discrete numerical scheme. Let us assume that the ve-
meability kB /kD , 𝑘𝐵 = − ̄ 𝑥𝑃 and the three transport coefficients ver-

locity profile 𝑢(𝜙𝑖) (𝑦) solves the discretized Brinkman equation (Ginzburg
− 12 𝜙1
sus D𝑎 = √𝐻 in two equal-thickness layers with 𝑟𝜙 = 𝜙2
= 10−1 and et al., 2015b) in the layer i:
𝑘𝐷
𝑘
𝑟𝑘 = 𝑘1 = 10−3 . It is shown that kB reduces with fi (𝜙) in the transition 𝜈𝜌0 (𝑖) 𝜈(1 + 𝛿𝑖 )
2 𝑢 +∇̄ 𝑥𝑃 = ̄ 2 𝑢(𝑖) ,
𝜌 Δ (62)
− 12
(Brinkman) zone D𝑎 ∈ [10, 50]. The dispersion coefficient kT decays 𝑘𝑖 𝜙 𝑓𝑖 (𝜙) 0 𝑦 𝜙
towards the Darcy regime and it only weakly depends upon Brinkman and reads
viscosity. The difference between 𝑓𝑖 (𝜙) = 1 and 𝑓𝑖 (𝜙) = 𝜙𝑖 monotonously
increases in Sk(⋆𝑠) and Ku(⋆𝑠) towards the Darcy regime, together with their ̄ 𝑥𝑃

𝑢(𝜙𝑖) (𝑦) = − (𝑘 − (𝑎𝑖 𝑟𝑦𝑖 + 𝑐𝑖 𝑟− 𝑦
𝑖 )),
− 12 𝜈𝜌0 𝑖
absolute values. In the Stokes zone, Sk(⋆𝑠) Pe and Ku(⋆𝑠) Pe−1 almost co- √

incide for Pe = 102 and Pe = 103 , meaning that the asymptotic regime 1 + 𝑏𝑖 𝜉𝑖2 𝑓𝑖 (𝜙)
(57) has been reached, like in an open flow. However, the asymptotes 𝑟𝑖 = √ , 𝑏𝑖 = 2 , 𝜉𝑖 = . (63)
1 − 𝑏𝑖 𝜉𝑖 + 4(1 + 𝛿𝑖 ) 𝑘𝑖
need much higher Pe in the Darcy zone. In resume, like in the single
porous channel (Ginzburg and Vikhansky, 2018), 𝑘𝑇 Pe−2 decreases from 𝜈
As seen, the Brinkman viscosity 𝑓𝑖 (𝜙)
gets the relative correction {𝛿 i }
the Stokes to the Brinkman–Darcy regime, but the non-Gaussian effects with the different numerical approaches, like the finite-elements, finite-
are predicted to grow towards the Darcy zone. differences or LBM schemes because the truncation term Δ ̄ 2 [ 𝜈𝜌0 𝑢(𝑖) (𝑦)]
𝑥 𝑘 𝜙 𝑖
Fig. 14 illustrates the reconstructed mean-concentration and the
due to the Laplacian of the resistance sums with the viscous term (see
RTD profiles. In spatial and temporal systems, the two profiles are dis-
1 Ginzburg et al., 2015b and Ginzburg, 2016). This correction is similar
played together, one in Darcy’s range D𝑎− 2 ≈ 56.5 and another one with the 𝛼 M -term in Eq. (25). The coefficients {ai , ci } in Eq. (63) are set
1 − 12
in highly permeable zone D𝑎 ≈ 20 × 56.5; porosity and permeabil- by the effective interface conditions of a given scheme. In principle, the
ity contrasts correspond to Fig. 15. The reconstruction of the pro- symbolic solutions allow to examine an impact of the correction {𝛿 i },
files is based on the entropy maximization (Bandyopadhyay et al., or the implicit interface flow conditions, on the transport coefficients.
2004), respectively, constrained to the predicted spatial or temporal In this work we only verify for the discrete effect between an analytical
system of the first four moments. Following Vikhansky and Ginzburg velocity profile (60) and a model finite-difference profile, referred to as
(2014), Ginzburg and Vikhansky (2018), the normalized (dispersion- (“ideal”) IFD: it solves Eq. (62) with 𝛿𝑖 = 0 and second-order accurate
independent) distributions are displayed in a moving frame: (i) 𝜎 𝐶̄ (𝑥̄ , 𝑡) Taylor expansion in interface/boundary conditions (see Ginzburg et al.,

21
JID: ADWR
ARTICLE IN PRESS [m5GeSdc;June 29, 2018;16:56]

I. Ginzburg Advances in Water Resources 000 (2018) 1–34

Fig. 16. Impact of the velocity profile in stratified Brinkman flow. The two cases from Table 6 are displayed together: (a) D𝑎 = 10−2 and (b) D𝑎 = 10−4 , using analytical
velocity profile (60) (“Th”) and numerical profile (63) (“Num”). Legend: (a) “Th” (solid, red), (b) “Th” (dot–dashed, blue) and (a) “Num” (dotted, black), (b) “Num”
(dashed, magenta).

Table 6 small Pe, the numerical corrections may become relatively large. Fig. 17
Two-layered porous Brinkman flow: parameters in compares TRT − EMM with Λ𝜙 from Eqs. (42a) and (42b) on the ana-
four symbolic experiments with 𝑘1 ∕𝑘2 = 10−2 , 𝜙1 = lytical velocity profile (60) with the parameter setting from Table 6. On
10−1 , 𝜙2 = 45 , 𝑓1 (𝜙) = 𝑓2 (𝜙) = 1.
the one hand, 𝛼𝑀 = 0 in Eq. (25) using Eq. (42a), the transport coeffi-
Case Pe Da kT /Pe2 Sk(⋆𝑠) Ku(⋆𝑠) cients corrections are then solely due to the implicit interface tracking
and the summation effect. On the other hand, kT is exact in the Darcy
(a) 10 2
10 −2
2.98 × 10 −3
−4.59 29.99
(b) 102 10−4 2.39 × 10−3 −7.72 73.86 flow using Eq. (42b) [see Eq. (36b)].
(c) 10 10−2 2.98 × 10−3 −0.17 0.17
Fig. 17 shows that, effectively, Eq. (42b) is more accurate for the
(d) 10 10−4 2.39 × 10−3 −0.22 −0.22 three coefficients on the coarse grid in the Darcy regime D𝑎 = 10−4
(bottom row); this is especially true with kT . In the intermediate zone
D𝑎 = 10−2 (top row), the situation is reversed for all the three coeffi-
2015b and Silva et al., 2017). Note that Eq. (63) coincides with Eq. cients and Eq. (42a) is one-order more accurate on the coarse grid, es-
(60) in grid nodes provided that 𝑟𝑖 = 𝑒𝜉𝑖 . This condition can be formally pecially for Sk(⋆𝑠) and Ku(⋆𝑠) . The results for Pe = 10 are similar in the two
fulfilled prescribing Eq. (63) with a model value 𝑓𝑖(𝑛𝑢𝑚) : cases (c) and (d) from Table 6.
To conclude, second-order accurate Brinkman numerical velocity
𝑓𝑖(𝑛𝑢𝑚) = 2(1 + 𝛿𝑖 )𝑘𝑖 (cosh[𝜉𝑖 ] − 1). (64) profile causes additional corrections in the three transport coefficients
of the same order magnitude as the TRT − EMM with Eq. (42a) has in
Thereby, once the TRT − EMM symbolic solution is built for the discrete the exact profile (cf. bottom row in Figs. 16 and 17). Using the ana-
profile (63), it describes the numerical solutions computed with the ex- lytical velocity profile in the Brinkman regime, the TRT − EMM retains
act profile using Eq. (64) and exact coefficients {ai , ci } from Eq. (60). its solution within ≈ 10% when having H/2 ≈ 5 mesh spacings per layer.
Solution Bn (y) to Eq. (25) is searched in the form: Such a precision is quite comparable with the limit case of an open layer
𝑛
∑ 𝑛
∑ adjacent to the diffusive-matrix from Section 6.2, and it is much better
𝐵𝑛(𝑖) (𝑦) = 𝑃𝑖(2𝑘) (𝑦)𝑟(𝑖𝑛−𝑘)𝑦 + 𝑋𝑖(2𝑘) (𝑦)𝑟(𝑖𝑘−𝑛)𝑦 . (65) than what we have observed for the Darcy flow in “series”. The Darcy
𝑘=0 𝑘=0 regime may require to reduce Λ𝜙 towards Eq. (42b), because when the
Similarly as in EMM solution (61), the coefficients are determined by velocity profile is relatively flat, the interface effects dominate the trun-
Eq. (25), except the zero- and the first-order term. The three remaining cation due to the velocity-variation in the mass-source. We note that the
coefficients are fixed through the continuity condition, Eq. (28a) with non-uniform and flow-regime-dependent Λ𝜙 distribution has been also
Eqs. (22), (24a), similar as in Eqs. (32)–(35), and the symmetry con- suggested in TRT Brinkman flow modeling (Ginzburg et al., 2015b).
ditions in the middle of each layer; the remaining constant is set with
Eq. (20). The constructed solutions for 𝜔2 − 𝜔4 and 𝛾2 − 𝛾4 coincide with 6.4. Summary
the TRT − EMM numerical solutions for the different-type numerical
profiles (63) and analytical profile (60). As one example, the four ex- The TRT − EMM numerical formulation (19) was examined on
periments in two equal-thickness layers combine D𝑎 = {10−2 , 10−4 } and the base of its symbolic solutions in several stratified systems: open,
Pe = {10, 102 }; the reference EMM values are gathered in Table 6. Fig. 16 fracture/diffusive-matrix and heterogeneous porous flow. For the sake
displays the relative differences TRT − EMM to EMM together for Eqs. of their validation, the EMM exact solutions (Ginzburg and Vikhansky,
(60) and (63) in Brinkman zone D𝑎 = 10−2 and Darcy zone D𝑎 = 10−4 . 2018) were extended to the stratified Brinkman flow. The TRT − EMM
The first diagram shows that the relative velocity error  (𝑠𝑢𝑚) to 𝑥 is behaves similarly to the direct solvers and much more accurately in a
most significant on the numerical profile in the Brinkman zone, and it “parallel” flow than in a “perpendicular” flow. The TRT − EMM trans-
reduces to about 5% only when we have H/2 ≥ 10 nodes per layer; the port coefficients are identical with the “𝜔-form” and “𝛾-form” [using
numerical profile then produces the most large corrections in the three Eq. (15)] in stratified channels provided that the free collision product
transport coefficients. This example suggests that an accurate descrip- Λ𝜙 and the truncation coefficient 𝛼 M in Eq. (25) take the same values in
tion of the velocity gradients is essential for the accuracy in the first the two formulations. Their solutions are entirely parameterized by Pe
four moments. These symbolic experiments were also run with Pe = 10, when Λ𝜙 and ce are inter-related through Eq. (26a). The transport coef-
cases (c) and (d) from Table 6. In these and other experiments, the ficients then obey the Pe-scale from Eq. (57) exactly for any individual
TRT − EMM with Pe = 10 is more accurate for the dispersion coefficient value of the characteristic velocity, ce , Λ− and Λ𝜙 , but they depend upon
against Pe = 102 , but the situation is reversed for E(Sk(⋆𝑠) ) and E(Ku(⋆𝑠) ) on the free-tunable parameter  from Eq. (26a). In the Stokes–Brinkman
the coarse grid, e.g., |E(Sk(⋆𝑠) )| is larger by a factor of three, and E(Ku(⋆𝑠) ) regime,  = 14 is efficient because it vanishes the principal truncation
is twice larger, with Pe = 10 against Pe = 102 . The described behavior is correction due to the mass-source variation with velocity. However, in
similar to the truncation estimate (Ginzburg, 2017b) of the moments in the Darcy zone,  = 18 is preferable for its advanced interface accuracy
ADE-j scheme (11b), because when Sk(⋆𝑠) and Ku(⋆𝑠) are very small with in kT measurements. It was shown that the bad quality of the velocity

22
JID: ADWR
ARTICLE IN PRESS [m5GeSdc;June 29, 2018;16:56]

I. Ginzburg Advances in Water Resources 000 (2018) 1–34

Fig. 17. Stratified Brinkman flow from Table 6. Top row: case (a), D𝑎 = 10−2 . Bottom row: case (b), D𝑎 = 10−4 . The TRT − EMM relative errors in three moments
are compared for Λ𝜙 = 1∕(4(1 − 𝑐𝑒 )) (solid, red), Eq. (42a), and Λ𝜙 = 1∕(8(1 − 𝑐𝑒 )) (dot–dashed, blue), Eq. (42b). (For interpretation of the references to color in this
figure legend, the reader is referred to the web version of this article.)

field may predominate over the truncation accuracy in the presence of The velocity field is solved with the d2Q9 IBF − LBM interface-implicit
the high velocity gradients and impair the prediction of moments. scheme (Ginzburg et al., 2015b) using the bounce-back no-slip rule on
As a reminder, the direct solver (11b) produces the first moment the solid surface. The resistance is prescribed with the permeability
with the effective velocity  (𝑠𝑢𝑚) in a straight bounded channel only value kf in a surrounding flow and km inside a porous obstacle; the ratio
if using the coordinate discrete-velocity stencil; otherwise the effective 𝑘𝑓
𝑘𝑚
= 108 is the same in the two flow experiments, hereafter referred to
mean velocity becomes modified in proportion to the diagonal velocity 1
weight tq in Eq. (10b) (see Ginzburg, 2017a). In addition to the reported as Exp. I and Exp. II, with {𝑘𝑓 , 𝑘𝑚 } = 4 × {107 , 10−1 }[l.u], D𝑎− 2 ≈ 63.25 ×
1
results, we observe that the direct ADE schemes extend this deficiency {10−4 , 1} in Exp. I, and {𝑘𝑓 , 𝑘𝑚 } = 4 × {105 , 10−3 }[l.u], D𝑎− 2 ≈ 63.25 ×
to stratified heterogeneous interfaces. This property is to be contrasted {10−3 , 10} in Exp. II. The two Darcy numbers {kf /(H × l)2 , km /(H × l)2 }
with the TRT − EMM: although its symbolic solutions were built with are kept fixed in l-times refined cell. Without obstacles, the channel flow
the d2Q5 discrete velocities, the TRT − EMM stratified flow solutions are approaches the Stokes (Poiseuille) limit. The streamlines in Fig. 18 al-
identical for d2Q5 and d2Q9 because of the streamwise-invariance of the low to observe the flow effect due to the low permeable obstacles in
steady-state solutions Bn (y) in straight channels. Thereby, the weight- Exp. II. Table 7 compiles the ratio of the effective mean-permeability 𝑘̄
independence of the transport coefficients in the basic configurations is to its open flow value kP ; the case of the solid obstacle is very similar
another advantage of the steady-state TRT − EMM approach against the in the two experiments with 𝑘̄ ≈ 7.57 × 10−3 𝑘𝑃 ; otherwise, the mean-
direct simulation of moments. permeability contrast between the two experiments is about 2.53. The
porosity 𝜙 impacts the velocity field only through the Brinkman viscos-
7. Heterogeneous porous flow ity value 𝜈𝐵 = 𝜈∕𝑓 (𝜙); the effective permeability depends very weakly
on 𝜙 and the transport coefficients are practically the same for the flow
We examine the TRT − EMM and three direct ADE schemes for tem- computed with 𝜈𝐵 = 𝜈∕𝜙 and 𝜈𝐵 = 𝜈.
poral and spatial dispersions in Brinkman flow around solid and porous
obstacles. Section 7.1 describes geometry and flow. Sections 7.2 and
7.3 address the numerical aspects of TRT − EMM and LBM-ADE, respec- 7.2. The TRT − EMM: numerical set-up
tively. Section 7.4 examines the physical insight on the obtained trans-
port coefficients. The TRT − EMM applies algorithm from Appendix A.4 and it solves
Eqs. (16)–(17) using the TRT collision (8) with the equilibrium (19); the
7.1. Geometry and flow mass-source 𝑛 (𝑟⃗) and the flux-jump 𝐽⃗𝑛 (𝑟⃗) in Eqs. (16b)–(16e) are com-
puted in “𝜔-form” with Eq. (A.1) and in “𝛾-form” with Eqs. (A.2)–(A.5).
We consider two-dimensional steady-state flow through a periodic We aim to verify the equivalence of the two formulations in this two-
porous system consisting of two equivalent rectangular obstacles de- dimensional field. The velocity field is precomputed in the unit cell; due
⟨𝑢𝜙,𝑥 ⟩
picted in Fig. 18. The two obstacles are either solid (𝜙 = 0) or porous, to the linearity of the flow, its arithmetical mean value  (𝑠𝑢𝑚) = ⟨𝜙⟩
𝜙 ∈ { 18 , 14 , 12 , 34 , 19
20
}. The “basic cell” has the size [𝑋 × 𝐻] = [56 × 40] is rescaled to the same value in all configurations. Prescribing 𝜙(𝑟⃗),
(l.u); it is periodic along the x-axis and bounded by two horizontal solid the velocity field and Pe =  (𝑠𝑢𝑚) 𝐻∕0 , the computations of Eqs. (16)–
walls; the two obstacles [𝑋 × 𝑌 ] = [8 × 32] are distanced by 8 mesh spac- (17) are performed three times sequentially for 𝑛 = 1, 2, 3; their steady-
ings; their centers are located at (𝑥, 𝑦) = (20, 16) and (𝑥, 𝑦) = (36, 24). The state solution 𝐵𝑛 (𝑟⃗) is then normalized with Eq. (20) and 𝑛 (𝑟⃗) is used
linear Brinkman velocity field 𝑢⃗𝜙 (𝑥, 𝑦) is governed by a constant body to precompute 𝑛 (𝑟⃗) and 𝐽⃗𝑛 (𝑟⃗) for the next orders, namely, 𝑛 = 2, 3, 4.
force 𝐹⃗ = 𝐹𝑥 1⃗ 𝑥 : One does not need to solve the system at 𝑛 = 4, because {𝜔n } is de-
termined with Eq. (16f) or Eq. (18), and {𝛾 n } is determined with Eq.
𝜈𝜌0 𝜈𝜌0 2
𝑢⃗ + ∇𝑃 − 𝐹⃗ = Δ 𝑢⃗𝜙 . (66) (A.4) from the previous-level solution. No-flux condition is enforced
𝑘 𝜙 𝑓 (𝜙)
23
JID: ADWR
ARTICLE IN PRESS [m5GeSdc;June 29, 2018;16:56]

I. Ginzburg Advances in Water Resources 000 (2018) 1–34

Fig. 18. Porous flow. From left to right: basic periodic cell [𝑋 × 𝑌 ] = [56 × 40]; streamlines around two solid obstacles and permeable obstacles, in the Exp. I and
Exp. II, when the “basic cell” is refined by the factor of five.

Table 7
The ratio of the mean permeability 𝑘̄ to its open flow value 𝑘𝑃 = 𝐻 2 ∕12 in the two flow
systems, Exp. I and Exp. II, versus porosity 𝜙; the solid obstacle is comprised for 𝜙 = 0. The
𝑘
permeability contrast is 𝑘𝑓 = 108 is both systems but kf and km are 102 times smaller in
𝑚
Exp. II.
1 1 1 3 19
𝜙 0 8 4 2 4 20

Exp. I 7.57 × 10−3 1.98 × 10−2 2.03 × 10−2 2.08 × 10−2 2.12 × 10−2 2.14 × 10−2
Exp. II 7.57 × 10−3 7.83 × 10−3 7.93 × 10−3 8.11 × 10−3 8.29 × 10−3 8.43 × 10−3

Fig. 19. The TRT − EMM in porous flow, Exp. I, Pe ≈ 74. From left to right: numerical solution {𝜔(𝑛𝑛𝑢𝑚) }, {𝛾𝑛𝑛𝑢𝑚 } and the relative difference of {𝜔𝑛 (𝛾𝑛(𝑛𝑢𝑚) )} computed
with Eq. (15) to numerical solution {𝜔(𝑛𝑛𝑢𝑚) } with “𝜔-form”. Data: 𝑛 = 2 (solid, “triangles”), 𝑛 = 3 (dot–dashed, “lozenge”), 𝑛 = 4 (dotted, “squares”).

with the bounce-back rule; the total mass in the system is conserved this condition was not satisfied when Pe ≥ ≈ 74, e.g., in “basic-cell” flow
at each time step because of the solvability condition ⟨𝑛 ⟩ = 0. While around the solids and in Exp. II [using 𝑥(𝑟𝑒𝑓 ) =  (𝑠𝑢𝑚) ≈ 0.0355, 𝑐𝑒 = 30
1
],
the exact global mass conservation provides a very valuable numerical stable results have been most of the time achieved using Λ𝜙 from Eq.
test, we find it more difficult to validate exactly with the “𝛾-form”, be- (42a) [Λ(𝜙𝑟𝑒𝑓 ) = Λ𝜙 |𝑐 1 ≈ 0.259]. As it might be expected, Eq. (10e) is
𝑒 = 30
cause of the round-off summation effect due to very large increase of the
not a necessary stability condition in a space-variable velocity field in
values {𝛾 n } with n. As one example, the two sets {𝜔′𝑛 (𝑛𝑢𝑚) } and {𝛾𝑛′ (𝑛𝑢𝑚) }
the rigorous sense, but it provides a very reasonable estimate, e.g., the
are illustrated in Fig. 19. In a given configuration, the dimensionless co-
𝜔𝑛 𝛾𝑛 𝑥𝑛 smallest porosity value 𝜙 = 18 becomes most unstable when Λ− 𝜙
reduces
efficients in Eq. (13), {𝜔′𝑛 = 𝑥 𝐻 (𝑛−1)
} and {𝛾𝑛′ = 𝐻 (𝑛−1)
} should remain for higher Pe.
fixed by Pe; it has been confirmed that {𝜔′𝑛 (𝑛𝑢𝑚) } and {𝛾𝑛 (𝑛𝑢𝑚) } follow this
′ The d2Q5 scheme operates with four free parameters: ce , Λ− , Λ𝜙
dimensionless scaling when 𝑥 and H vary. Further, we confirm that the and 𝛽 e . Recall that, when ce and Λ𝜙 are linked through Eq. (26a), the
set {𝜔′𝑛 (𝛾𝑛′ (𝑛𝑢𝑚) )}, obtained with Eq. (15), practically coincides with the solution becomes independent of the two separated values ce and Λ− ,
numerical solution {𝜔′𝑛 (𝑛𝑢𝑚) } obtained with the “𝜔-form”; the relative provided that their product 0 = 𝑐𝑒 Λ− is fixed, either in a “perpendic-
difference between the two sets is exemplified in the last diagram in ular” or a “parallel” one-dimensional flow. This property remains valid
Fig. 19. We note that Eq. (15) provides another very useful implemen- in one-dimensional flow, when the obstacles have the same porosity
tation check. Therefore, the two numerical formulations, “𝜔-form” and and permeability as the surrounding √ flow. For example, √ if we employ
“𝛾-form”, produce the same solution for the dispersion coefficient kT
1
the two sets {𝑐𝑒 , Λ− , Λ𝜙 }: { 30 , 1∕3, Λ(𝜙𝑟𝑒𝑓 ) } and { 13 , 1∕3 × 10, 38 } with
and any one of them can be run for the prediction of the four transport the same value 0 = 𝑐𝑒 Λ− , the difference {𝛿𝐷, 𝛿Sk(⋆𝑠) , 𝛿Ku(⋆𝑠) } with the
coefficients in Eq. (14). open flow will be identical in the two cases and very small [because
Hereafter, we will display the relative differences of the trans- the flow approaches the open regime], as {0.07, −2.21 × 10−7 , 0.13}[%]
port coefficients to their open-flow counterparts at the same Pe; they in the “basic cell” at Pe ≈ 74, Exp. II. However, this exact parame-
are labeled as 𝛿D for 𝐷 = (1 + 𝑘𝑇 )0 , {𝛿Sk(⋆𝑠) , 𝛿Ku(⋆𝑠) } for {Sk(⋆𝑠) , Ku(⋆𝑠) } terization property of Eq. (26a) is not valid in a two-dimensional
and {𝛿Sk(⋆𝑡) , 𝛿Ku(⋆𝑡) } for {Sk(⋆𝑡) , Ku(⋆𝑡) }, respectively. The reference open flow, e.g., in the presence of the solids, the relative difference be-
(Poiseuille) flow solutions is given in Table 3 (Ginzburg and Vikhansky, tween the two sets is {−8.7, 0.58, −0.15}[%] in the “basic cell”, but only
2018). The simulations are run at Pe ≈ {7.4, 74, 740} in the basic cell {−0.29, 0.01, −0.01}[%] for a five-times finer cell. Similarly, in the ab-
and Pe ≈ 7.4{ × 10, × 102 , × 103 } in five times finer cell; stability bound sence of solids when 𝜙 = 34 , the relative difference between the two pa-
from Eq. (10e) is adopted as an approximate. Although in several points

24
JID: ADWR
ARTICLE IN PRESS [m5GeSdc;June 29, 2018;16:56]

I. Ginzburg Advances in Water Resources 000 (2018) 1–34

Fig. 20. The TRT − EMM in porous flow. From left to right: number of time steps to steady state for dispersion (𝑛 = 1), skewness (𝑛 = 2) and kurtosis (𝑛 = 3) is
displayed together for four configurations in Eq. (10): 𝛽𝑒 = 1, 𝑥 = 𝑥(𝑟𝑒𝑓 ) ∕10 (“filled triangle”); 𝛽𝑒 = 5, 𝑥 = 𝑥(𝑟𝑒𝑓 ) ∕10 (“filled lozenge”); 𝛽𝑒 = 10: 𝑥 = 𝑥(𝑟𝑒𝑓 ) ∕10
(“empty lozenge”) and 𝛽𝑒 = 1, 𝑥 = 𝑥(𝑟𝑒𝑓 ) (“empty triangle”) [some points are absent when ce does not satisfy Eq. (10e)].

rameter sets is {1.48, −4.15, −0.46}[%] in the “basic cell” and it reduces the spatial spread, an evolution of the plume distribution (𝑥 = 𝑥0 , 𝑦, 𝑡 =
to {0.07, −0.05, −0.02}[%] in the refined cell. 0) = 1 is run through 102 − 2 × 102 basic cells; the raw moments 𝜇𝑛 (𝑡) =
∑ ∑ 𝑛
This loss of the exact parameterization is because Eq. (24a) is not 𝑥 𝑦 𝜙(𝑥, 𝑦)(𝑥, 𝑦, 𝑡)(𝑥 − 𝑥0 ) 𝑑 𝑥𝑑 𝑦 are normalized with 𝜇 0 (t) and em-
exact any more and Eq. (25) obtains a different truncation component. ployed to compute the central moments 𝜇𝑛⋆ (𝑡) for 𝑛 = 2, 3, 4. The set
The effect of Λ𝜙 decays here approximately as 𝐻 −2 for all moments. {𝜔(𝑛𝑛𝑢𝑚) } is derived from the moments with Eq. (53) and normalized
Since the TRT − EMM applies in a single cell, the insensitivity to the with Eq. (13); it determines  (𝑛𝑢𝑚) , 𝑘𝑇 = 𝐷 − 1 and Sk(⋆𝑠) , Ku(⋆𝑠) with Eq.
0
free-parameters can be achieved over fine grids. It is important to stress (14). As the reference example for ADE-j in open Poiseuille flow through
that the transport coefficients are all exactly parameterized by Pe and ce , a “basic cell” [𝐻 = 40], we obtain  (𝑛𝑢𝑚) =  (𝑠𝑢𝑚) and the relative er-
meaning that they remain the same when the velocity field and diffusion rors {0.027, 0.53, 0.1}[%] for D, Sk(⋆𝑠) and Ku(⋆𝑠) , respectively, using the
eigenfunction Λ− scale accordingly.
porous flow parameter set [Pe ≈ 74, 𝑐𝑒 = 30 1
, 𝑥 = 𝑥(𝑟𝑒𝑓 ) , Λ𝜙 = 14 ]. The
The steady-state solution 𝑛 (𝑟⃗) is independent of the acceleration pa-
TRT − EMM convergence criteria is adopted, but the stop-value is much
rameter 𝛽 e in Eq. (10). When 𝛽 e > 1, ce , and velocity amplitude should
higher, typically 𝜖 ≈ 10−6 . We note that the ADE schemes then only ap-
reduce approximately linearly with 𝛽 e , according to stability estimate
proximately reach the Taylor regime with the same order of steps as
in Eq. (10e). The TRT − EMM computations are stopped when the rel-
the TRT − EMM needs to converge to the steady state. It follows that
ative difference in 𝛿D, 𝛿Sk⋆ and 𝛿Ku⋆ is less than 𝜖 = 10−11 over suc-
operating in a single cell, the TRT − EMM is hundreds time faster al-
cessive 5 × 102 steps. Fig. 20 shows that the number of steps to steady
ready in the small Pe range. Further, since the time to the Taylor regime
state is nearly the same in the three equations (𝑛 = 1, 2, 3); the number
and, accordingly, the number of the unit cells, increases linearly with
of steps decreases linearly with the 𝑥 or 𝛽 e increase; that is, nearly
Pe, the ADE computational time increases as Pe2 . We also note that
the same computational time is required with {𝑥 , 𝛽𝑒 } = {10𝑥(𝑟𝑒𝑓 ) , 1}
TRT − EMM with Λ𝜙 |𝑐 = 1 ≈ 0.259 from Eq. (42a) behaved much more
or {𝑥 , 𝛽𝑒 } = {𝑥(𝑟𝑒𝑓 ) , 10}. Since the small velocity amplitudes are more 𝑒 30
accurate because of the truncation corrections, 𝛽 e is expected to be help- robustly than the three direct solvers in spatial dispersion. As one exam-
ful. The results in Fig. 20 correspond to Exp. II in “𝜔-form”; the “𝛾-form” ple, the TRT − EMM reaches Pe ≈ 740 in the “basic-cell” [except 𝜙 = 18
typically requires a slightly larger number of steps. We note that when in Exp. I] while ADE-r becomes undefined in Exp. II with all examined
Pe increases linearly due to the mesh refining, the number of steps to porosity values when Pe ≥ 74 because Λ− 𝜙,1
becomes negative in several
steady state increases slightly. points (cf. Eq. (12c)).
Finally, it has been suggested (Ginzburg and Vikhansky, 2018; The temporal dispersion is modeled with the zero initial distribu-
Vikhansky and Ginzburg, 2014) that the mass-source component Sn tion (𝑥, 𝑦, 𝑡 = 0) = 0; the Dirichlet boundary condition (𝑥 = 0, 𝑦, 𝑡) =
and flux-jump 𝐽⃗𝑛 from Eq. (16e) can be dropped at sufficiently high 𝑏 is prescribed with the anti-bounce-back rule (29a) in boundary
Pe provided that the sets {𝜔′𝑛 } and {𝛾𝑛′ } grow with Pe for n ≥ 2, e.g., nodes 𝑟⃗𝑏 = (𝑥 = 12 , 𝑦) using  = 𝑏 = 1 in 𝑒+ 𝑞 (𝑟
⃗𝑠 ). The RTD distribution
like in Eq. (57) with stratified flow. In fact, these two terms originate 𝑃 (𝑥𝑖 , 𝑡) ≈ 1 (𝐶̄ (𝑥𝑖 , 𝑡) − 𝐶̄ (𝑡 − 1, 𝑥𝑖 )) is monitored at each time step in a pre-
2
from the diffusion term in Eq. (1) and they cannot be omitted in a selected set of the grid points xi ; the RTD central moments 𝜇𝑛⋆ (𝑥) ≈

diffusion-dominated process, like in “series” of heterogeneous porous ̃ ̄ 𝑛 ̃ 1
𝑡 (𝑡 − 𝑡(𝑥𝑖 )) 𝑃 (𝑥𝑖 , 𝑡)𝛿𝑡 , 𝑡 = 𝑡 − 2 , 𝛿𝑡 = 1, are accumulated in time using the
blocks where {𝜔′2 } is asymptotically constant and decays with Pe for effective centroid value 𝑡̄(𝑥𝑖 ) = 𝜇1 (𝑥𝑖 ); the raw moments are all normal-
n ≥ 3, see Eq. (48). Fig. 21 displays the relative differences in 𝛿D, 𝛿Sk(⋆𝑡) ized with the actual value 𝜇 0 (xi ). Since the Dirichlet anti-bounce-back
and 𝛿Ku(⋆𝑡) between two EMM solutions, with and without {𝑆𝑛 , 𝐽⃗𝑛 }, in rule is not exact, 𝜇 0 (xi ) may slightly deviate from the steady-state exact
the two flow experiments. In both of them, the three relative differ- solution 𝜇 0 (xi ) ≡ 1. The set {𝛾 n } is extracted from the central moments
ences diminish from the interval [−20%, 40%] at Pe ≈ 7.4 to about 1% at with the formulae (Ginzburg and Vikhansky, 2018; Vikhansky and
𝑥
Pe ≈ 740, but they are still noticeable for 𝛿Sk(⋆𝑡) and 𝛿Ku(⋆𝑡) at Pe ≈ 74 in Ginzburg, 2014); it determines  (𝑛𝑢𝑚) (𝑥𝑖 ) = 𝛾1 = 𝜇 (𝑥𝑖 ) , 𝑘𝑇 (𝑥𝑖 ) = 𝐷 − 1,
1 1 𝑖 0
Exp. I. Since the spatial and temporal moments are inter-related through Sk(⋆𝑡) (𝑥𝑖 ) and Ku(⋆𝑡) (𝑥𝑖 ) in Eq. (14). When the Taylor regime is reached,
Eq. (15), a similar situation will take place with 𝛿Sk(⋆𝑠) and 𝛿Ku(⋆𝑠) . These typically at xi /H ≈ Pe, these values downstream remain approximately
simulations confirm that in the presence of the fractured velocity path, the same. Within the “basic cell”, the RTD simulations are more robust
Sn and 𝐽⃗𝑛 can be omitted at sufficiently high Pe making numerical im- than the spatial spread and the three schemes reach Pe ≈ 740 in Exp. I;
plementation much more simple. ADE-c/ADE-j most of the time achieved for this as well in Exp. II. Since
the downstream distance should increase linearly with Pe, the computa-
7.3. The direct ADE schemes: numerical set-up tional time grows as Pe2 , like in the spatial spread. Let us recall that the
mean seepage velocity  (𝑠𝑢𝑚) is set with the TRT − EMM in any geome-
We model the spatial and temporal dispersion with the three d2Q5 try by construction; the direct ADE schemes produce  (𝑛𝑢𝑚) =  (𝑠𝑢𝑚) on
schemes: ADE-c and ADE-j from Eq. (11), and ADE-r from Eq. (12); the the coordinate stencil in stratified channels, but  (𝑛𝑢𝑚) differs from the
bounce-back is applied on the horizontal walls and solid obstacles. For

25
JID: ADWR
ARTICLE IN PRESS [m5GeSdc;June 29, 2018;16:56]

I. Ginzburg Advances in Water Resources 000 (2018) 1–34

Fig. 21. Temporal dispersion, Exp. I (top row) and Exp. II (bottom row). From left to right: Pe ≈ {7.4, 74, 740}. The relative differences of the results in
𝛿(0 )(red, circle), 𝛿(Sk(⋆𝑡) )(blue, square) a𝑛𝑑 𝛿(Ku(⋆𝑡) )(black, lozenge) when TRT − EMM applies without Sn and 𝐽⃗𝑛 terms in Eq. (16) to “standard” solution with them.
The computations are run in the “basic cell” cell 56 × 40. (For interpretation of the references to color in this figure legend, the reader is referred to the web version
of this article.)

Table 8
Temporal dispersion in Exp. I. Mean-velocity error Eu [in percents] at Pe ≈ 740 with the ADE-
c and ADE-j from Eq. (11) and ADE-r from Eq. (12a). The computations are run in a series of
“basic cells” [56 × 40], 𝑥 = 𝑥(𝑟𝑒𝑓 ) ≈ 0.0355 and 𝑐𝑒 = 301
. The ADE-r is undefined for 𝜙 = 0
because of the stability condition Λ−𝜙,1 > 0 in Eq. (12c); the ADE-c and ADE-j are unstable for
𝑟𝜙 = 18 .
1 1 1 3 19
𝜙 0 8 4 2 4 20

ADE-c[%] 4.25 − 1.74 × 10−2 5.18 × 10−2 3.29 × 10−2 1.44 × 10−2
ADE-r[%] −2.28 × 10−1 − −7.56 × 10−2 −4.53 × 10−2 −6.39 × 10−2 −8.09 × 10−2
ADE-j [%] 3.02 − −2.28 × 10−1 −2.13 × 10−1 −2.63 × 10−1 −3.13 × 10−1

prescribed constant Darcy value with ADE-j in porous blocks in “series” nant role of the Taylor and numerical dispersions. When Pe ≈ 74, the
[see Section 5.3.2]. Fig. 22 shows that the three ADE schemes mod- three schemes produce similar results only provided that ADE-j oper-
ify the prescribed mean-velocity value in two-dimensional flow, both ates with small velocity, as  (𝑠𝑢𝑚) = 𝑥(𝑟𝑒𝑓 ) ∕10 according to Fig. 22.
in the spatial and temporal systems. When Pe ≈ 7.4 [velocity is small, Finally, as it might be expected from the mean-velocity results in
 (𝑠𝑢𝑚) = 𝑥(𝑟𝑒𝑓 ) ∕10], ADE-c/ADE-r manifest quasi-identical relative ve- Table 8, the difference with the TRT − EMM is very similar in the three
locity error Eu ; it slightly differs with ADE-j but it has the same de- schemes at Pe ≈ 740. We note, however, that the difference between the
pendency upon 𝜙. However, when Pe ≈ 74 [ (𝑠𝑢𝑚) = 𝑥(𝑟𝑒𝑓 ) ], ADE-j pro- TRT − EMM and ADE increases with Pe; this can be partly related to
duces relatively large velocity errors and it reproduces the ADE-c only the fact that the steady-state Taylor regime is more difficult to reach
with much smaller velocity, as  (𝑠𝑢𝑚) = 𝑥(𝑟𝑒𝑓 ) ∕10. Thus, although ADE- with ADE when Pe grows. On the whole, in this and all other spatial
j is often more stable than ADE-c close to the stability line, thanks to and temporal dispersion experiments, a similar difference within 5%
the reduction of the (negative) numerical diffusion, it needs to operate between the TRT − EMM and the direct schemes was mostly observed
on a smaller velocity range in heterogeneous soil for the correct advec- in the “basic cell”. We conclude that the three ADE schemes produce
tion in the intermediate Pe range. Table 8 shows that when Pe ≈ 740, similar solutions when their advection process is accurate enough, and
the velocity error becomes small with all schemes using the same ve- then confirm the TRT − EMM moments predictions.
locity [ (𝑠𝑢𝑚) = 𝑥(𝑟𝑒𝑓 ) ]. As a reminder, a reduction of the velocity error

with Λ− = 𝑐 0 was first observed with ADE-j in “series” of porous blocks.
𝑒 7.4. The transport coefficients
Remarkably, this makes the advection process more accurate when Pe
grows decreasing Λ− .
Now, using the TRT − EMM solutions, we examine how the pres-
Fig. 23 displays the three relative differences 𝛿D, 𝛿Sk(⋆𝑡) and 𝛿Ku(⋆𝑡)
ence of the solid and porous obstacles impacts the transport coeffi-
between the direct schemes and TRT − EMM. The three direct schemes
cients. Fig. 24 displays the relative differences with respect to open
produce nearly identical results at Pe ≈ 7.4, where their difference with
flow in Exp. I, both in the spatial and temporal dispersions. The dis-
TRT − EMM is within 0.5% for all three moments, and the largest dis-
persion coefficient 𝐷 = 0 (1 + 𝑘𝑇 ) is the same in the two systems and
crepancy happens for kT at the smallest porosity value 𝜙 = 18 . The anti-
𝛿D is displayed together for Pe ≈ 7.4{ × 10, × 102 , × 103 } in the first
numerical-diffusion mechanism does not play a role, either because
diagram. The four relative differences, 𝛿Sk(⋆𝑠) , 𝛿Ku(⋆𝑠) and 𝛿Sk(⋆𝑡) , 𝛿Ku(⋆𝑡) ,
of the relatively smaller velocity  (𝑠𝑢𝑚) = 𝑥(𝑟𝑒𝑓 ) ∕10 or the predomi-
are displayed together in the three next diagrams at fixed Pe. We ob-

26
JID: ADWR
ARTICLE IN PRESS [m5GeSdc;June 29, 2018;16:56]

I. Ginzburg Advances in Water Resources 000 (2018) 1–34

Fig. 22. Relative mean-velocity error E(U) in LBM-ADE. Exp. I: spatial (red, circle), temporal (blue, square); Exp. II: spatial (black, lozenge), temporal (magenta,
triangle). Top row: Pe = 7.4, ADE-c, ADE-r, ADE-j with 𝑥 = 𝑥(𝑟𝑒𝑓 ) ∕10. Bottom row: Pe = 74, ADE-c [𝑥(𝑟𝑒𝑓 ) ], ADE-j [𝑥(𝑟𝑒𝑓 ) ∕10], ADE-j [𝑥(𝑟𝑒𝑓 ) ]. (For interpretation
of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 23. Temporal dispersion, Exp. I, Pe ≈ {7.4, 74, 740} [ from left to right]. The relative differences of the LBM-ADE to TRT − EMM in D (red, circle), Sk(⋆𝑡) (blue,
square) and Ku(⋆𝑡) (black, lozenge). Top row: ADE-c from Eq. (11a). Middle row: ADE-r from Eq. (12a). Bottom row: ADE-j from Eq. (11b) with 𝑥 = 𝑥(𝑟𝑒𝑓 ) ∕10 when
Pe = 74. The results correspond to Fig. 22 and Table 8. (For interpretation of the references to color in this figure legend, the reader is referred to the web version
of this article.)

27
JID: ADWR
ARTICLE IN PRESS [m5GeSdc;June 29, 2018;16:56]

I. Ginzburg Advances in Water Resources 000 (2018) 1–34

Fig. 24. Porous flow, Exp. I, TRT − EMM prediction for the relative difference with the open flow. First diagram: 𝛿D together for Pe ≈ 74 (red, “circle”),
740 (blue, “square”), 7400 (black, “lozenge”). The next three diagrams: Pe ≈ 7.4{ × 10, × 102 , × 103 }. The results are displayed together for 𝛿Sk(⋆𝑠) (red, empty circles),
𝛿Ku(⋆𝑠) (blue, empty square), 𝛿Sk(⋆𝑡) (black, filled circles) and 𝛿Ku(⋆𝑡) (magenta, filled squares). The computations were run in the single periodic cell 280 × 200. (For
interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

serve that all moments are quite different for solid and highly-permeable Fig. 27 displays the RTD and the mean-concentration profiles com-
porous blocks. Remarkably, all transport coefficients become porosity- puted with the ADE-c at Pe ≈ 74. According to the predictions in
independent at Pe ≈ 740. The relative difference to open flow also be- Fig. 26 for Pe ≈ 74, 𝛿D ≈ 0 when 𝜙 = 18 and the RTD profiles almost
comes Pe-independent; that means that all transport coefficients obey coincide with their open flow counterparts. However, 𝛿D ≈ 0.6 when
the same (stratified flow) Pe-scaling from Eq. (57) in highly-permeable 𝜙 = 1920
and both the RTD and the mean-concentration profiles show
porous flow. A very interesting result in Fig. 24 is that the highly- that the porous system becomes more dispersive than the open flow. In
permeable porous system from Exp. I is very efficient in the suppres- self-similar coordinates, the reconstructed and numerical profiles agree
sion of the Taylor dispersion, independently of the porosity value. In- to each other and look very similar to the open-flow profiles because
deed, while 𝛿𝐷 ≈ −0.1 when Pe ≈ 7.4 (not shown), 𝛿D becomes ≈ −0.8 𝛿Sk(⋆𝑡) ≈ 0 and Ku(⋆𝑡) keeps its sign. We note that the concentration pro-
for Pe ≈ 74 and this difference further decreases with Pe towards 𝛿𝐷 ≈ files look more irregular in the less permeable system (Exp. II); they are
𝑘𝑇
(𝑝) − 1 ≈ −1 or kT ≈ 0. These results predict the dispersion coefficient more oscillating with the ADE-j scheme from Eq. (11b) because of the
𝑘𝑇
interface jumps [see “perpendicular” flow in series Fig. (7)]. At the same
to diminish by ≈ 80% at Pe ≈ 74, and almost to vanish at Pe ≈ 740 and
time, the ADE-c and ADE-r produce very similar moments and profiles
Pe ≈ 7400 towards 𝑘𝑇 = 0 in porous flow. The TRT − EMM also pre-
in both spatial and temporal systems.
dicts the relative degree of the asymmetry and peakedness in the mean-
Finally, Fig. 28 displays the normalized numerical RTD profiles at
concentration and RTD distributions against the open flow. Recall that
Pe ≈ 740; they are monitored closer to the inlet, when the Taylor regime
Sk(⋆𝑠) and Ku(⋆𝑠) are both negative in the spatial spread in open (Poiseuille)
has not been yet reached. We note that ADE-c and ADE-j produce very
flow where the mean-concentration profiles are hence skewed right
similar profiles at Pe ≈ 740 in agreement with the moments prediction.
and relatively flat (see Ginzburg and Vikhansky, 2018; Vikhansky and
The asymmetry of the profiles reduces with the porosity, following the
Ginzburg, 2014). Fig. 24 predicts that Sk(⋆𝑠) remains negative in porous
Taylor regime predictions in Fig. 26 [Sk(⋆𝑡,𝑝) > 0 and 𝛿Sk(⋆𝑡) (𝜙) ∈ [−1, −0.7]
flow but its absolute value increases [𝛿Sk(⋆𝑠) ≈ 2.5]; the kurtosis is pre-
monotonously increases with 𝜙]. The reconstruction from the four mo-
dicted to change the sign meaning a larger peakedness. The RTD profiles
ments is not accurate enough close to the inlet, but it improves down-
are typically skewed left because of the Dirichlet inlet condition [Sk(⋆𝑡)
stream and predicts RTD correctly already at 𝑥∕𝑋 ≈ 5 × 10−2 Pe. Accord-
and Ku(⋆𝑡) are both positive in open flow]. Fig. 24 predicts them to hold ing to results (Vikhansky and Ginzburg, 2014), an accurate reconstruc-
approximately the same form in the presence of the highly-permeable tion of very sharp distributions needs more moments. The positive as-
obstacles, because 𝛿Sk(⋆𝑡) ≈ 0 and 𝛿Ku(⋆𝑡) ≈ 0.5. pect is that any order moments can be predicted with the TRT − EMM,
Fig. 25 displays the mean-concentration distribution 𝐶̄ (𝑥) and RTD applying Eqs. (16)–(17) sequentially.
profiles P(t) in Exp. I at Pe ≈ 74; they are produced by the direct solver
ADE-c when 𝜙 = 18 and compared with the reconstructed profiles (see
7.5. Summary
Section 6.3.1). Here, the reconstruction is performed from the exact
Poiseuille moments and from the TRT − EMM moments computed in
The TRT − EMM and three direct ADE d2Q5 schemes have been ex-
porous flow. The direct computations confirm the drastic reduction of
amined in two-dimensional Brinkman flow around solid and porous ob-
the Taylor dispersion with respect to an open flow. They also con-
stacles: the porosity contrast between the fracture and matrix was varied
firm that the self-similar (dispersion-independent) distribution 𝜎 𝐶̄ (𝑥̄ ) re-
in the interval 𝜙 ∈ [ 81 , 19
20
]; their permeability contrast was fixed to 108 ,
mains skewed right, but it is much sharper than in open flow. The recon-
two flow scenarios with Darcy numbers varying by a factor of 102 have
struction from the TRT − EMM moments practically coincides with the
been considered.
RTD numerical profiles; in self-similar coordinates, they look very alike
The Péclet number was varied in the interval Pe ∈ [7.4, 740] in
to open flow, in agreement with the moments predictions. The direct
the “basic cell” [8 mesh spacings per obstacle and the gap], and up
computations also confirm that the profiles are almost 𝜙-independent.
to Pe ≈ 7400 with the TRT − EMM in five-times finer cell. Whereas the
Fig. 26 displays the TRT − EMM predictions for the spatial and the
TRT − EMM was operated within a single periodic cell; the direct sim-
temporal dispersion in Exp. II. As has been expected from the mean-
ulations were run through one to two hundred of the duplicated cells.
permeability values in Table 7, the TRT − EMM predictions for flow
The TRT − EMM predictions were validated by the direct numerical sim-
around the solids are nearly the same in the Exp. I and Exp. II, but this
ulations; an agreement within ≈ 5% between the two approaches was
case is accommodated smoothly in Exp. II, because the porous obstacles
found in “basic cell” computing the relative, temporal (RTD) and spatial
are low permeable. A striking difference with the Exp. I in Fig. 24 is
(mean-concentration), moment differences with the open flow. The rela-
that kT increases almost linearly with 𝜙, hence the less-porous low-
tive accuracy of the two approaches agrees well with our error estimate
permeable blocks reduce the Taylor dispersion more efficiently. The
in stratified channels, but it is much better than it might be expected
Taylor dispersion reduction increases with Pe, but it remains weaker
from the analysis in “series”, for the same porosity contrast, Pe and dis-
than in Exp. I.
cretization level. Despite the space-variable mass-source and interface

28
JID: ADWR
ARTICLE IN PRESS [m5GeSdc;June 29, 2018;16:56]

I. Ginzburg Advances in Water Resources 000 (2018) 1–34

Fig. 25. Comparison of the profiles in porous and open flow, Exp. I, Pe ≈ 74, 𝜙 = 18 . All diagrams: numerical profile in porous flow (blue, solid), reconstruction
from the TRT − EMM moments in porous flow (black, dashed) and from the√EMM moments in open flow (red, dotted). Two first diagrams: 𝐶̄ (𝑥) at t/T ≈ 0.41Pe
[𝑇 = 𝐻∕𝑥 , 𝐻 = 40] versus x/H and 𝜎 𝐶̄ (𝑥) versus 𝑥̄ = (𝑥 − 𝑥0 − 𝑥 𝑡)∕𝜎, 𝜎 = 2𝐷𝑡. Two last diagrams: RTD profiles 𝑃 ′ (𝑡′ ) = 𝑃 (𝑡)∕𝑇 versus 𝑡′ = 𝑡∕𝑇 at the distance

𝑥∕𝑋 = {0.54Pe, 0.68Pe, 0.81Pe} downstream, and 𝜎P(𝜏) versus 𝜏 = (𝑡 − 𝑡̄)∕𝜎, 𝜎 = 2𝐷𝑥−3 𝑥 when 𝑥∕𝑋 = 0.54Pe [or 40 periods of length 𝑋 = 56 downstream]. (For
interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 26. Similar as in Fig. 24 but in Exp. II. First diagram: 𝛿D together for Pe ≈ 74 (red, “circle”), 740 (blue, “square”), 7400 (black, “lozenge”). The next three diagrams:
Pe ≈ 7.4{ × 10, × 102 , × 103 }, respectively. The results are displayed together for 𝛿Sk(⋆𝑠) (red, “empty circles”), 𝛿Ku(⋆𝑠) (blue, “empty square”), 𝛿Sk(⋆𝑡) (black, “filled
circle”) and 𝛿Ku(⋆𝑡) (magenta, “filled square”). The computations were run in a single periodic cell 280 × 200. (For interpretation of the references to color in this
figure legend, the reader is referred to the web version of this article.)

Fig. 27. Comparison of the profiles in porous and open flow, Exp. II, Pe ≈ 74. All diagrams: numerical profile in porous flow (blue, solid), reconstruction from
the TRT − EMM moments in porous flow (black, dashed) and from the EMM moments in open flow (red, dotted). Two first diagrams: dimensionless RTD profiles
𝑃 ′ (𝑡′ ) = 𝑃 (𝑡)∕𝑇 versus 𝑡′ = 𝑡∕𝑇 at the distance 𝑥∕𝑋 = {0.54Pe, 0.68Pe, 0.81Pe} downstream when 𝜙 = 18 and 𝜙 = 19
20
, respectively. Third diagram: RTD in self-similar
coordinates when 𝜙 = 19 , 𝑥∕𝑋 = 0.54Pe [or 40 periods of the length 𝑋 = 56 downstream]. Last diagram: 𝜙 = 19 , 𝐶̄ (𝑥) at 𝑡′ = 0.4Pe versus x/H. (For interpretation of
20 20
the references to color in this figure legend, the reader is referred to the web version of this article.)

flux-jump, the TRT − EMM behaved more stably than the direct solvers and the flux-jump term 𝐽⃗𝑛 can be omitted already at Pe ≈ 740, in agree-
in the spatial spread. ment with the dimensional analysis (Vikhansky and Ginzburg, 2014) for
The TRT − EMM solutions follow the dimensional analysis from fractured-type flow. This property greatly simplifies the TRT − EMM be-
Eq. (13); the “𝜔-form” and “𝛾-form” produced almost identical solu- cause, on the one hand, the diffusion flux becomes continuous on the
tions for dispersion coefficient kT and, with the help of Eq. (14), for the porous and solid obstacles and, on the other hand, the finite-differences
skewness and kurtosis in temporal and spatial dispersions. In a given can be avoided in precomputing Sn . The EMM can be then handled by
grid, Pe, ce and Λ𝜙 fix the steady state solutions for any reference ve- any standard linear ADE solver based on the continuous flux condition
locity 𝑥(𝑟𝑒𝑓 ) and diffusion eigenfunction Λ− ; however, an exact (ce - and local mass-source.
independent) parameterization with Eq. (26a) is only valid asymptot- Among the three ADE schemes, the ADE-c from Eq. (11a) and ADE-r
ically in multi-dimensions. The scale parameter 𝛽 e from Eq. (10) accel- from Eq. (12) have shown very similar results within the ADE-r stability
erates the convergence to steady state and allows to operate efficiently at range: 𝑢2𝜙 (𝑟⃗) < 𝑐𝑒 𝜙(𝑟⃗). Using Λ𝜙 = 14 , the ADE-c and ADE-j could remain
small velocity amplitudes. It was shown that the non-local mass term Sn stable outside this interval, but their effective stability is more difficult

29
JID: ADWR
ARTICLE IN PRESS [m5GeSdc;June 29, 2018;16:56]

I. Ginzburg Advances in Water Resources 000 (2018) 1–34

Fig. 28. Temporal dispersion in Exp. II, Pe ≈ 740. All diagrams: numerical profile in porous flow (blue, solid) and reconstruction from the TRT − EMM moments (black,
dashed). Three first diagrams: the RTD numerical profile in open flow (solid, red), porous flow at 𝜙 = 19 20
, and porous flow at 𝜙 = 18 at the distance 𝑥∕𝑋 = 0.034Pe
1
[or 25 periods of the length 𝑋 = 56] downstream the inlet. Last diagram: 𝜙 = 8
at 𝑥∕𝑋 = 0.054Pe [or 40 periods]. (For interpretation of the references to color in
this figure legend, the reader is referred to the web version of this article.)

to estimate. The ADE-j needs a smaller velocity amplitude to reproduce more robustly than the spatial spread with the proposed LBM-ADE di-
correctly the averaged seepage velocity in heterogeneous soil at the in- rect schemes.
termediate Pe-range. However, the three ADE schemes produce very It has been recognized that, because of the truncation correction due
similar results at higher Pe where ADE-j reduces its velocity correction to the non-linear mass-source space variation, the fixed value of the
reducing Λ− . At fixed Pe and grid, the transport coefficients in direct eigenfunction product Λ𝜙 is not sufficient to parameterize the steady-
simulations depend upon all numerical parameters: 𝑥(𝑟𝑒𝑓 ) , ce , Λ− and state TRT solutions by Pe exactly; a specific dependency on the diffusion-
Λ𝜙 separately, because of the truncation corrections (Ginzburg, 2017b). coefficient-scale parameter ce : Λ𝜙 (𝑐𝑒 ) = 1−𝑐 , guarantees this property in
𝑒
A highly permeable “micro-channel”-type porous arrangement was a “parallel” and a “perpendicular” flow for any mass-source; future anal-
found most efficient for dispersion reduction in creeping Brinkman flow, ysis will attempt to extend this functionality to the multi-dimensional
independently of the matrix porosity 𝜙 at the same mean-seepage veloc- flow. A robust choice  = 14 eliminates the mass-source truncation cor-
ity value 𝑥(𝑟𝑒𝑓 ) . When the Pe-behavior of moments becomes established, rection and assures continuous Bn solution for the three first orders with
e.g. fractured type Pe-scale from Eq. (57), the asymptotic solutions can the pure-diffusion implicit-jump tracking in “series”. However, the “per-
be estimated from the TRT − EMM computations performed at the in- pendicular” Darcy flow through porous blocks presents an increasing dif-
termediate Pe-range, saving drastically computational resources. The ficulty benchmark for dispersion and higher-order moments where the
TRT − EMM is especially suitable for mesh refining and clearly advan- leading order midway location of the continuity condition requires a
tageous for the prediction of the dispersion and the higher moments in rather “fragile” (unstable) limit  → 0 [or Λ𝜙 → 0]. The choice of Λ𝜙 is
the Taylor regime. In transition to the Taylor regime, the EMM approach therefore decisive for the high porosity contrast modeled on coarse grids
allows to predict qualitatively an effect of different scenarios, e.g., due where very small values Λ𝜙 are expected to improve implicit interface
to the flow regime and geometry aspects. location in X-ray tomography “series-block” type geometry. In strati-
fied flow and fractured Brinkman flow through permeable obstacles,
8. Concluding remarks the TRT − EMM and direct solvers behave much more accurately than
in “series” and have shown much smaller dependency on Λ𝜙 ;  = 14 is
The two-relaxation-times Lattice Boltzmann linear advection–
recommended for TRT − EMM in fractured Brinkman flow, while  = 18
diffusion scheme TRT − EMM has been proposed for the numerical for-
is more suitable in heterogeneous Darcy flow. Operated in a single cell,
mulation of the extended method of moments (EMM) in d-dimensional
the TRT − EMM copes very well with the grid refining for a solution in-
heterogeneous soil. The TRT − EMM incorporates locally the recursively
dependence upon free numerical parameters and high Péclet range. In
precomputed space-variable mass sources and automatically accounts
our parallel work we attempt to reduce Λ𝜙 -impact on the unresolved
for the diffusion-flux interface jumps in the streamwise-discontinuous
continuity conditions with the help of the conjugate interface correc-
porosity; the scheme allows for an optional acceleration to steady-state
tions.
with relatively small velocity amplitudes. A sequentially solved chain
In the direct schemes, numerical diffusion becomes insignificant
of the boundary-value problems, independent for each mean-velocity
against the Taylor dispersion at sufficiently high Pe and heterogene-
direction, provides three longitudinal Taylor coefficients: dispersion,
ity contrast. The two examined direct solvers, with the equilibrium
skewness and kurtosis; the full dispersion tensor can be restored from d-
and collision anti-numerical-diffusion treatment, and the scheme with-
independent runs and a recursive extension to the next order longitudi-
out any treatment, then show similar results but different stability in
nal moments is straightforward. Although the whole evolution picture is
compliance with the predictions. In diffusion-dominant heterogeneous
not described by the Taylor moments, they might indicate qualitatively
problems, the collision anti-numerical-diffusion collision treatment be-
the relative role of the different factors for transition regimes. Provided
comes preferable to the equilibrium approach because the discontinuous
that the moments amplitudes grow with Pe, the diffusion-flux jump 𝐽⃗𝑛
equilibrium weights cause concentration jumps and modify the mean-
and the associated non-local mass-source Sn become negligible and the
seepage velocity. This preference is expected to hold for the equilibrium
method can be run by any heterogeneous ADE solver based on the con-
and collision models with heterogeneous anisotropy (Ginzburg, 2006;
tinuous diffusion flux. It has been confirmed that the (spatial) “𝜔-form”
Ginzburg and d’Humières, 2007). In the presence of solid and porous ob-
and the (temporal) “𝛾-form” are numerically equivalent in providing the
stacles, the prescribed mean-seepage velocity is not kept exact with all
two systems of moments; the “𝜔-form” is preferable because of its abil-
direct schemes. To be contrasted, the TRT − EMM guarantees advection
ity to the fully diffusive problems, its simplicity and higher numerical
velocity by construction and its steady solutions are numerical-diffusion
reliability. The algebraic transforms between the temporal and spatial
free. The proposed scheme applies with the stair-wise geometry descrip-
systems of moments are also applicable to direct ADE computations:
for example, the RTD simulations under the Heaviside entry behaved

30
JID: ADWR
ARTICLE IN PRESS [m5GeSdc;June 29, 2018;16:56]

I. Ginzburg Advances in Water Resources 000 (2018) 1–34

𝑛

tion; it will be coupled with the more accurate interface and boundary
tracking for a fine description of solid and porous objects in future work. 𝑆𝑛 (𝑟⃗) = 𝜙 𝛾𝑘 𝜕𝑥 𝑛−𝑘 , 𝑛 ≥ 2,
𝑘=1
To conclude, the proposed simple TRT − EMM is readily applicable
to tomographic rock images and reconstructed media for a systematic 𝑆1(𝛾) = 0, 𝑆2(𝛾) = 𝜙 𝛾1 𝜕𝑥 (1𝛾) , 𝑆3(𝛾) = 𝜙 (𝛾2 𝜕𝑥 (1𝛾) + 𝛾1 𝜕𝑥 (2𝛾) ),
classification of their dispersion properties, and for optimization design 𝑆4(𝛾) = 𝜙 (𝛾3 𝜕𝑥 (1𝛾) + 𝛾2 𝜕𝑥 (2𝛾) + 𝛾1 𝜕𝑥 (3𝛾) ), … . (A.3)
under desired moments constraints in composite material, wetland or
chemical device. Our artificial porous arrangement exemplified such The solution for {𝛾 n } reads:
kind of application: it has been found that the porous, highly permeable
⟨𝑀𝑛(𝛾) + 𝑆𝑛(𝛾) ⟩
obstacles reduce very efficiently the Taylor dispersion independently of 𝛾𝑛 = , 𝑛 ≥ 1, (0𝛾) = 1, then (A.4a)
their porosity, and that the moments then asymptotically obey the same ⟨𝑢𝜙,𝑥 (0𝛾) ⟩
Pe−scale as in open flow. However, the Péclet number and porosity de-
pendencies both become more complicated with the weakly penetra- ⟨𝜙⟩ 1
𝛾1 = = , and , for 𝑛 ≥ 2 ∶ (A.4b)
ble obstacles and they may even enhance the Taylor dispersion against ⟨𝑢𝜙,𝑥 ⟩ 𝑥
the open flow. Brinkman creeping flow in the two Darcy regimes was
∑𝑛−1 ∑𝑛
𝑘=1 𝛾𝑘 < 𝑢𝜙,𝑥 𝑛−𝑘 > + < 𝑆𝑛 > +⟨𝜙 ⟩( 𝑚=1 𝛾𝑚 𝛾𝑛−𝑚 )
simulated, but further, the numerical EMM simulations in inertial flow −
may clarify the influence of the viscous and inertial resistance, turbu- 𝛾𝑛 = . (A.4c)
⟨𝑢𝜙,𝑥 ⟩
lence, shear layers or wake effects, as well as the role of heuristic flow-
model parameters, such as the Brinkman viscosity in the present study. The jump-term 𝐽⃗𝑛 (𝑟⃗) reads:
Nonetheless, the flow should be described quite accurately, because it 𝑛

was demonstrated that the prediction of moments is quite sensitive to 𝐽⃗𝑛 (𝑟⃗) = 𝜙 ( 𝛾𝑘 𝑛−𝑘 )1⃗ 𝑥 , 𝑛 ≥ 1,
the quality of the heterogeneous velocity field. 𝑘=1

𝐽1(𝛾) = 𝛾1 𝜙 1⃗ 𝑥 , 𝐽2(𝛾) = 𝜙 (𝛾2 + 𝛾1 (1𝛾) )1⃗ 𝑥 ,


𝐽3(𝛾) = 𝜙 (𝛾3 + 𝛾2 (1𝛾) + 𝛾1 (2𝛾) )1⃗ 𝑥 , … . (A.5)
Acknowledgments
A.3. Computation of the 𝑆𝑛 (𝑟⃗)
The author is grateful to Alexander Vikhansky and Gonçalo Silva for
The mass term 𝑆𝑛 (𝑟⃗) in Eqs. (A.1) and (A.3) requires to precompute
their kind help.
𝜙 (𝑟⃗)𝜕𝑥 𝑘 (𝑟⃗) for k < n. When the three to four grid neighbors are avail-
able at point x along the longitudinal x-axis inside the continuous poros-
ity block V𝜙 , the third-order-accurate finite-differences apply:

Appendix A. The EMM: implementation details 2 1


𝜕𝑥 𝑓 (𝑥) ≈ (𝑓 (𝑥 + 1) − 𝑓 (𝑥 − 1)) − (𝑓 (𝑥 + 2) − 𝑓 (𝑥 − 2)) + 𝑂(𝜖 4 ),
3 12
(A.6a)
A.1. The “𝜔-form”
1 1 1
The three first orders 𝑛 = 1, 2, 3 in Eqs. (16b)–(16e) read locally with 𝜕𝑥 𝑓 (𝑥) ≈ ∓ 𝑓 (𝑥) ± 𝑓 (𝑥 ± 1) ∓ 𝑓 (𝑥 ∓ 1) ∓ 𝑓 (𝑥 ± 2) + 𝑂(𝜖 4 ), (A.6b)
2 3 6
(superscript “(𝜔)′′ is employed in “𝜔-form”):
11 3 1
𝜕𝑥 𝑓 (𝑥) ≈ ∓ 𝑓 (𝑥) ± 3𝑓 (𝑥 ± 1) ∓ 𝑓 (𝑥 ± 2) ± 𝑓 (𝑥 ± 3) + 𝑂(𝜖 4 ). (A.6c)
(𝑛𝜔) = 𝑀𝑛(𝜔) + 𝑆𝑛(𝜔) + 𝜙𝜔𝑛 , 𝑛 ≥ 1, 𝜔1 = 𝑥 , 6 2 3
These operators are exact on the third-order polynomial solution.
𝑀1(𝜔) = −𝑢𝜙,𝑥 (0𝜔) , (0𝜔) = 1, 𝑆1(𝜔) = 𝜙 𝜕𝑥 (0𝜔) = 0, 𝐽1(𝜔) = 𝜙 1⃗ 𝑥 ,
In particular, in a series of porous blocks of the minimal thickness of
𝑀2(𝜔) = 𝜙𝜔1 (1𝜔) + 𝜙 (0𝜔) − 𝑢𝜙,𝑥 (1𝜔) , 𝑆2(𝜔) = 𝜙 𝜕𝑥 (1𝜔) , 𝐽2(𝜔) = 𝜙 (1𝜔) 1⃗ 𝑥 , four nodes per block, Eq. (A.6b) applies with all points except the cor-
ners where Eq. (A.6c) applies; this allows to match pure diffusion solu-
𝑀3(𝜔) = 𝜙(𝜔1 (2𝜔) + 𝜔2 (1𝜔) ) + 𝜙 (1𝜔) − 𝑢𝜙,𝑥 (2𝜔) ,
tions for kurtosis from Eq. (B.2). When only one or two neighbors are
𝑆3(𝜔) = 𝜙 𝜕𝑥 (2𝜔) , 𝐽3(𝜔) = 𝜙 (2𝜔) 1⃗ 𝑥 , 𝑆4(𝜔) = 𝜙 𝜕𝑥 (3𝜔) . (A.1) available inside the same porosity block, the standard two-three points
finite-differences apply for 𝜕 x f(x), like

Eqs. (16)–(17) with Eq. (A.1) are sequentially solved for (1𝜔) , (2𝜔) and 1
𝜕𝑥 𝑓 (𝑥) ≈ ± (3𝑓 (𝑥) − 4𝑓 (𝑥 ∓ 1) + 𝑓 (𝑥 ∓ 2)) + 𝑂(𝜖 3 ),
(3𝜔) ; {𝜔n } is computed recursively with Eq. (16f) or Eq. (18) for 𝑛 =
2
1
2, 3, 4. 𝜕𝑥 𝑓 (𝑥) ≈ ± (𝑓 (𝑥 ± 1) − 𝑓 (𝑥 ∓ 1)) + 𝑂(𝜖 3 ),
2
𝜕𝑥 𝑓 (𝑥) ≈ ±(𝑓 (𝑥) − 𝑓 (𝑥 ∓ 1)) + 𝑂(𝜖 2 ). (A.7)

The two first operators (A.7) are exact for parabolic distributions, the
A.2. The “𝛾-form” last one is only exact for a linear function f(x).

The three first orders 𝑛 = 1, 2, 3 in Eqs. (16b)–(16e) read with (super- A.4. The TRT − EMM algorithm
script “(𝛾)′′ applies in “𝛾-form”):
We resume the principal TRT − EMM steps following Section 3.2. The
TRT − EMM computes sequentially the steady state numerical solution
(𝑛𝛾) = 𝑀𝑛(𝛾) + 𝑆𝑛(𝛾) − 𝛾𝑛 𝑢𝜙,𝑥 (0𝛾) , 𝑛 ≥ 1, (0𝛾) = 1, 𝑀1(𝛾) = 𝜙,
𝑛 (𝑟⃗) to Eqs. (16)–(17) for 𝑛 = {1, 2, 3}. This solution is used to compute
𝑀2(𝛾) = (𝜙 − 𝛾1 𝑢𝜙,𝑥 )(1𝛾) + 𝜙 𝛾12 (0𝛾) , the next-order mass-source terms 𝑀𝑛+1 (𝑟⃗) and 𝑆𝑛+1 (𝑟⃗), and then 𝜔𝑛+1
with Eq. (16f) in “𝜔-form” [or 𝛾𝑛+1 with Eq. (A.4a) in “𝛾-form”]; the
𝑀3(𝛾) = (𝜙 − 𝛾1 𝑢𝜙,𝑥 )(2𝛾) − 𝛾2 𝑢𝜙,𝑥 (1𝛾) + 𝜙 (2𝛾1 𝛾2 + 𝛾12 (1𝛾) ), (A.2)
two sets {𝜔n } and {𝛾 n }, 𝑛 = {2, 3, 4}, define the transport coefficients in
Eqs. (14b)–(14e) with the help of Eqs. (13) and (15). One computational
and run for 𝑛 = {1, 2, 3} consists from the following steps:

31
JID: ADWR
ARTICLE IN PRESS [m5GeSdc;June 29, 2018;16:56]

I. Ginzburg Advances in Water Resources 000 (2018) 1–34

1. Pre-initialization step: (f) In “𝜔-form”, 𝛾𝑛′ +1 (𝜔′𝑛+1 ) is computed with Eq. (15);
This step is the same for 𝑛 = {1, 2, 3}. In “𝛾-form”, 𝜔′𝑛+1 (𝛾𝑛′ +1 ) is computed with the inverse to Eq. (15);
(a) Prescribe the formulation: either “𝜔-form” or “𝛾-form” [use “𝜔- (g) If 𝑛 = 1 ∶ compute kT with Eq. (14c);
form” for pure diffusion in Eq. (1), note that “𝜔-form” is simpler]; If 𝑛 = 2 ∶ compute Sk(⋆𝑠) with Eq. (14d), and/or Sk(⋆𝑡) with
(b) Prescribe a single cell V and its longitudinal periodic direction, Eq. (14e);
say 1⃗ 𝑥 ; If 𝑛 = 3 ∶ compute Ku(⋆𝑠) with Eq. (14d), and/or Ku(⋆𝑡) with
(c) Prescribe porosity field 𝜙(𝑟⃗) ∈ [0, 1] and steady-state velocity Eq. (14e);
field 𝑢⃗𝜙 (𝑟⃗)[𝑙.𝑢], periodic along 1⃗ 𝑥 and resolved on the regular com- (h) Check steady-state criteria with the prescribed stop-value 𝜖,
putational grid inside the penetrable region 𝑟⃗ ∈ {𝑉𝜙 } ∈ 𝑉 , see 𝑘𝑇 (𝑡)−𝑘𝑇 (𝑡−𝑡(𝑟𝑢𝑛) )
e.g. | 𝑘𝑇 (𝑡)
| < 𝜖.
Eq. (1). (i) Repeat Computation step if neither the steady state nor the maxi-
2. Initialization step: mum number of iterations has been reached;
(a) Prescribe constant equilibrium parameters in Eq. (19): ce > 0 and Go to Output step otherwise.
𝛽 e ≥ 1 such that 𝑐𝑒 ≤ min𝑟⃗ 𝜙( 𝑑𝛽
𝑟⃗)
, 𝑑 = {1, 2, 3} with d1Q3, d2Q5 and 5. Output step.
𝑒
d3Q7, respectively; see Eq. (10e); Save 𝑛 (𝑟⃗), 𝜔𝑛+1 [and/or 𝛾𝑛+1 ] for the next pass if 𝑛 = {1, 2}.
(b) Optionally: rescale 𝑢⃗𝜙 [𝑙.𝑢] in linear (Stokes–Brinkman–Darcy) 6. Final step.
flow, and/or modify ce and 𝛽 e , to satisfy an approximate stability Go to Initialization step if 𝑛 = {1, 2};
∑ 𝑐 𝜙(𝑟⃗)
condition: 𝑢2𝜙 (𝑟⃗) = 𝑑𝛼=1 𝑢2𝜙,𝛼 ≤ 𝑒𝛽 , see Eq. (10e). When ce and Stop computations if 𝑛 = 3.
𝑒
𝛽 e are fixed, this condition is satisfied if 𝑢⃗𝜙 is multiplied by the Remark A.1. Steady-state fields 𝑛 (𝑟⃗), 𝐽⃗𝑛 (𝑟⃗) and Λ± (𝑟⃗) are precom-
√ 𝜙
𝑐𝑒 𝜙(𝑟⃗) puted during initialization in the present algorithm. However, they can
scale value 𝑠 ∈]0, min𝑟⃗ 2 ]: 𝑢
⃗𝜙 → 𝑢⃗𝜙 × 𝑠.
𝛽𝑒 𝑢𝜙
∑ be recomputed at each time step for memory saving.
𝑟⃗∈{𝑉𝜙 } 𝑢𝜙,𝑥 (𝑟⃗)
(c) Precompute  (𝑠𝑢𝑚) = ∑ ;
𝑟⃗∈{𝑉𝜙 } 𝜙(𝑟⃗) Remark A.2. One pass 𝑛 = 1 is sufficient for the dispersion coefficient
(d) Prescribe Pe in Eq. (1) and compute Λ− from condition: kT : since 𝑆1 = 0 in “𝜔-form” and “𝛾-form”, the finite-difference operators
 (𝑠𝑢𝑚) [𝑙.𝑢] from Eqs. (A.6) and (A.7) are not needed and 1 (𝑟⃗) is computed locally.
Pe = 0 [𝑙.𝑢]
, 0 [𝑙.𝑢] = 𝑐𝑒 Λ− where [𝑙.𝑢] is the characteristic
length; Two runs 𝑛 = {1, 2} are required for Sk(⋆𝑠) and/or Sk(⋆𝑡) ; the three passes
(e) Prescribe 𝐵0 (𝑟⃗) = 0 (𝑟⃗) ≡ 1 and 𝜔1 with Eq. (16c) [or 𝛾 1 with 𝑛 = {1, 2, 3} are needed for Ku(⋆𝑠) and/or Ku(⋆𝑡) .
Eq. (A.4b) in “𝛾-form”], replacing 𝑥 by  (𝑠𝑢𝑚) ;
Remark A.3. In principle, the model parameters: ce , 𝛽 e , Λ− and Λ𝜙 (𝑟⃗)
(f) If 𝑛 = 2 or 𝑛 = 3, input already computed solution {𝑘−1 (𝑟⃗)} and
can differ in three separate runs provided that Pe and the necessary
{𝜔k } in “𝜔-form” [or {𝛾 k } in “𝛾-form”] with 𝑘 = 2, … , 𝑛;
stability condition 𝑐𝑒 ≤ min𝑟⃗ 𝜙( 𝑟⃗)
are respected.
(g) Precompute 𝑛 (𝑟⃗) with Eq. (A.1) in “𝜔-form” [or Eqs. (A.2)– 𝑑𝛽 𝑒

(A.3) in “𝛾-form”]; check that 𝑟⃗∈{𝑉𝜙 } 𝑛 (𝑟⃗) = 0;
Appendix B. Diffusion in a series of porous blocks
(h) Precompute 𝐽⃗𝑛 (𝑟⃗) with Eq. (A.1) in “𝜔-form” [or Eq. (A.5) in “𝛾-
form”];
B.1. Reference solution
(i) Prescribe a positive distribution [or a constant value exemplified
in Eq. (42)] for collision number Λ𝜙 (𝑟⃗);
Λ𝜙 (𝑟⃗)
In pure diffusion through a series of two porous blocks of length 𝐻 =
Precompute Λ+
𝜙
(𝑟⃗) = Λ− (𝑟⃗)
with Λ−
𝜙
(𝑟⃗) = 𝜙(𝑟⃗)Λ− ; ℎ1 + ℎ2 with the porosity ratio r𝜙 and thickness ratio rh , the effective
𝜙
(j) Prescribe any initial local mass distribution 𝜌(𝑟⃗, 𝑡 = 0); diffusion coefficient 𝑒𝑓 𝑓 from Eq. (14b) and kurtosis coefficient Ku(s)
∑ from Eq. (5) read (see Ginzburg and Vikhansky, 2018):
Precompute global mass 𝑀 (𝑔𝑙𝑜𝑏) = 𝑟⃗∈{𝑉𝜙 } 𝜌(𝑟⃗, 𝑡 = 0);
(k) Prescribe initial population solution {𝑓𝑞 (𝑟⃗, 𝑡 = 0)}, 𝑞 = 0, … , 𝑄𝑚 , (1 + 𝑟ℎ )2 𝑟𝜙 𝜙1 ℎ
e.g., using Eq. (19) with 𝑛 (𝑟⃗) = 0 : 𝑒𝑓 𝑓 = −𝜔2 = 𝑟 0 , 𝑟 = , 𝑟𝜙 = , 𝑟 = 1,
𝑄
(𝑟ℎ + 𝑟𝜙 )(1 + 𝑟ℎ 𝑟𝜙 ) 𝜙2 ℎ ℎ2
𝑓𝑞 (𝑟⃗, 𝑡 = 0) = 𝑒+
𝑞 (𝑟
⃗, 𝑡 = 0) + 𝑒−
𝑞 (𝑟
⃗, 𝑡 = 0), 𝑞 = 1, … , 2𝑚 ; (B.1a)
𝑄
𝑓𝑞̄ (𝑟⃗, 𝑡 = 0) = 𝑒+ 𝑞 (𝑟
⃗, 𝑡 = 0) − 𝑒−
𝑞 (𝑟
⃗, 𝑡 = 0), 𝑞 = 1, … , 2𝑚 , 𝑐⃗𝑞̄ = −𝑐⃗𝑞 ;
∑𝑄𝑚 ∕2
𝑓0 (𝑟⃗, 𝑡 = 0) = 𝜌(𝑟⃗, 𝑡 = 0) − 2 𝑞=1 𝑓𝑞 (𝑟⃗, 𝑡 = 0). 𝑟 ∈]0, 1], 𝑟 |𝑟𝜙 =1 = 𝑟 |𝑟ℎ =0 = 1 ; (B.1b)
3. Computation step:
Perform t(run) collision/streaming steps (7a) using TRT operator (8):
(a) Compute 𝑠± (𝑟⃗) with Eq. (9) from Λ± (𝑟⃗);
𝜙 𝜙 Sk(⋆𝑠) = 0 ; (B.1c)
𝑄𝑚
(b) Compute 𝑒±
𝑞 (𝑟
⃗, 𝑡), 𝑞 = 1, … , 2
, with Eq. (19);
𝑄
(c) Compute {𝑔𝑞 (𝑟⃗, 𝑡)}, 𝑞 = 1, … , 2𝑚 , with Eq. (8);
±

(d) Set (𝑟⃗) = 𝑛 (𝑟⃗) in Eq. (7a) and compute 𝑓0 (𝑟⃗, 𝑡+ 1); 6𝜔4 𝐻2 𝑟2ℎ (𝑟2𝜙 − 1)2
Ku(𝑠) = Ku(𝑡)𝑡 = − = . (B.1d)
(e) Compute 𝑓𝑞 (𝑟⃗ + 𝑐⃗𝑞 , 𝑡 + 1) with Eq. (7a) when 𝑟⃗ + 𝑐⃗𝑞 ∈ {𝑉𝜙 }, 𝑞 = 𝜔22 0 2(1 + 𝑟ℎ )2 𝑟𝜙 (𝑟ℎ + 𝑟𝜙 )(1 + 𝑟ℎ 𝑟𝜙 )
1, … , 𝑄𝑚 ;
(f) Compute 𝑓𝑞̄ (𝑟⃗𝑏 , 𝑡 + 1) in boundary node 𝑟⃗𝑏 with the bounce-back When all the integrals are computed via summation in Eq. (16), Ku(⋆𝑠)
rule using Eq. (29b) when 𝑟⃗𝑏 + 𝑐⃗𝑞 ∉ {𝑉𝜙 }. becomes:
4. Analysis step:
∑ (𝑟𝜙 − 1)2 (𝑟2ℎ (𝐻 2 (1 + 𝑟𝜙 )2 ) + 𝑟ℎ (𝑟ℎ 𝑟𝜙 + 1)(𝑟ℎ + 𝑟𝜙 ))
(a) Compute 𝜌(𝑟⃗, 𝑡) and check whether 𝑀 (𝑔𝑙𝑜𝑏) = 𝑟⃗∈{𝑉𝜙 } 𝜌(𝑟⃗, 𝑡); Ku(⋆𝑠𝑢𝑚) = . (B.2)
b) Compute (𝑛 + 1)th-order moments : 2(1 + 𝑟ℎ )2 𝑟𝜙 (𝑟ℎ + 𝑟𝜙 )(1 + 𝑟ℎ 𝑟𝜙 )0
(a) 𝐵𝑛 (𝑟⃗) with Eq. (19c);
The 𝑟 should be the exact solution of any second-order accurate numer-
(b) 𝑛 (𝑟⃗) with Eq. (20);
ical scheme; the TRT − EMM from Eq. (19) and the direct ADE schemes
(c) 𝑀𝑛+1 (𝑟⃗) and 𝑆𝑛+1 (𝑟⃗) with Eq. (A.1) in “𝜔-form” [or Eqs. (A.2)–
(11) produce 𝑒𝑓 𝑓 exactly for any physical and free parameter choice.
(A.3) in “𝛾-form”];
However, the TRT − EMM with the implicit flux-jump tracking from
(d) 𝜔𝑛+1 with Eq. (16f) [or 𝛾𝑛+1 in “𝛾-form” with Eq. (A.4a)];
Eq. (19) produces Eq. (B.2) exactly only if Λ𝜙 from Eq. (26a) is com-
(e) 𝜔′𝑛+1 [or 𝛾𝑛′ +1 ] with Eq. (13) using, e.g., 0 =  (𝑠𝑢𝑚) ; 𝑠𝑢𝑚)
puted with  = 𝐴(𝐾𝑢 :

32
JID: ADWR
ARTICLE IN PRESS [m5GeSdc;June 29, 2018;16:56]

I. Ginzburg Advances in Water Resources 000 (2018) 1–34

𝑠𝑢𝑚)
𝐴(𝐾𝑢 (𝑟𝜙 , 𝑟ℎ ) Bolle, E.M.L., Quastel, J., Fogg, G.E., Gravner, J., 2000. Diffusion processes in composite
Λ𝜙 = , porous media and their numerical integration by random walks: generalized stochas-
1 − 𝑐𝑒 tic differential equations with discontinuous coefficients. Water Resour. Res. 36 (3),
𝑠𝑢𝑚) 7𝑟𝜙 + 𝑟ℎ (1 + 𝑟𝜙 (12 + 7𝑟ℎ + 𝑟𝜙 )) 𝑠𝑢𝑚)
651–662.
𝐴(𝐾𝑢 (𝑟𝜙 , 𝑟ℎ ) = , lim 𝐴(𝐾𝑢 = ∞. (B.3) Brenner, H., 1980. Dispersion resulting from flow through spatially periodic porous media.
24(1 + 𝑟ℎ )2 𝑟𝜙 𝑟𝜙 →0 Philos. Trans. R. Soc. Lond. Ser. A 297 (1430), 81–133.
Brinkman, H.C., 1947. A calculation of the viscous force exerted by a flowing fluid on a
The parameter-dependent solution (B.3) locates an interface jump in dense swarm of particles.. Appl. Sci. Res. A 1, 27–34.
Buxton, G.A., 2018. Modeling the effects of vegetation on fluid flow through an acid mine
Eq. (41) midway the grid nodes. Eqs. (B.1)–(B.3) are invariant with re- drainage passive remediation system.. Ecol. Eng. 110, 23–37.
spect to a simultaneous transformation 𝑟𝜙 → 𝑟−1
𝜙
and 𝑟ℎ → 𝑟−1

. Cantu-Perez, A., Kee, S.P., Gavriilidis, A., 2008. Mixing and residence time distribution
studies in microchannels with floor herringbone structures. In: Proceedings of the
Comsol Conference. Hannover.
B.2. Semi-explicit diffusion jump-tracking Chai, Z., Huang, C., Shi, B., Guo, Z., 2016a. A comparative study on the Lattice Boltzmann
models for predicting effective diffusivity of porous media. Int. J. Heat Mass Transf.
98, 687–696.
Assume the coordinate discrete-velocity set to be applied in a pure-
Chai, Z., Shi, B., Guo, Z., 2016b. A multiple-relaxation-time Lattice Boltzmann model
diffusion process along the x-axis and the jump values ||⃗ 𝑛 ⋅ 𝐽⃗𝑛 (𝑥 ± for general nonlinear anistropic convection–diffusion equations. J. Sci. Comput. 69,
1
𝑐
⃗ )
2 𝑞 𝐴𝜙
|| from Eq. (17b) to be precomputed midway the grid nodes. The 355–390.
modified TRT update then reads: Chai, Z., Zhao, T.S., 2014. Nonequilibrium scheme for computing the flux of the convec-
tion–diffusion equation in the presence of the Lattice Boltzmann method. Phys. Rev. E
90. 013305–15
𝑔𝑞+ (𝑥) → 𝑔𝑞+ (𝑥) + 𝛿𝑔𝑞+ (𝑥), 𝑔𝑞− (𝑥) → 𝑔𝑞− (𝑥) + 𝛿𝑔𝑞− (𝑥), Chopard, B., Falkone, J.L., Latt, J., 2009. The Lattice Boltzmann advection–diffusion
model revisited. Eur. Phys. J. Spec. Top. 171, 245–249.
1
𝛿𝑔𝑞± (𝑥) = (𝐽 (𝑥) ± 𝐽𝑞̄ (𝑥)), Cozewith, C., Squire, K.R., 2000. Effect of reactor residence time distribution on polymer
2 𝑞 functionalization reactions. Chem. Eng. Sci. 55, 2019–2029.
1
𝐽𝑞 (𝑥) = 𝑡𝑞 ||𝐽⃗𝑛 (𝑥 + 𝑐⃗𝑞 ) ⋅ 𝑛⃗||, 𝐽⃗𝑛 = 𝜙 𝑛−1 1⃗ 𝑥 , 𝑐⃗𝑞 = −𝑐⃗𝑞̄ .
Cui, S., Hong, N., Shi, B., Chai, Z., 2016. Discrete effect on the halfway bounce-back
(B.4) boundary condition of multiple-relaxation-time Lattice Boltzmann model for convec-
2
tion–diffusion equations. Phys. Rev. E 93. 043311–13
In contrast with Eq. (19), Eq. (B.4) copes with the pure diffusion equi- Danckwerts, P.V., 1953. Continuous flow systems. Chem. Eng. Sci. 2, 1–13.
Demuth, C., Mishra, S., Mendes, M.A., Ray, S., Trimis, D., 2016. Application and accuracy
librium function when the mass-source incorporates 𝜕𝑥 𝐽𝑛 (𝑥) = 𝑆𝑛 (𝑥) ac- issues of TRT Lattice Boltzmann method for solving elliptic PDEs commonly encoun-
cording to Eq. (16e): tered in heat transfer and fluid flow problems. Int. J. Therm. Sci. 100, 185–201.
d’Humières, D., Ginzburg, I., 2009. Viscosity independent numerical errors for Lat-
𝑒− ̃ + ̃
𝑞 ≡ 0 ,  𝑛 →  𝑛 =  𝑛 ( 𝑥 ) + 𝑆 𝑛 ( 𝑥 ) , 𝑒 𝑞 ( 𝑥 ) = 𝑡 𝑞 𝑐 𝑒 ( 𝐵 + Λ𝜙  𝑛 ) .
+ tice Boltzmann models: from recurrence equations to “magic” collision numbers.
(B.5)
Comp. Math. Appl. 58 (5), 823–840.
Dubois, F., Lin, C.A., Tekitek, M.M., 2016. Anisotropic thermal Lattice Boltzmann simula-
Eqs. (B.4) and (B.5) allow to verify numerically the symbolic analysis tion of 2d natural convection in a square cavity. Comput. Fluids 124, 278–287.
developed in Section 4. We note that {𝛿𝑔𝑞 = 𝛿𝑔𝑞+ + 𝛿𝑔𝑞− } is non-zero only Gac, J.M., 2014. A large eddy based Lattice-Boltzmann simulation of velocity distribution
for two opposite interface cut links which obtain the same prescribed in an open channel flow with rigid and flexible vegetation. Acta Geophys. 62 (1),
180–198.
flux-jump quantity: Gebäck, T., Marucci, M., Boissier, C., Arnehed, J., Heintz, A., 2015. Investigation of the
effect of the tortuous pore structure on water diffusion through a polymer film using
𝛿𝑔𝑞 |𝑥− =𝑥 1 = 𝑡𝑞 ||𝐽⃗𝑛 ⋅ 𝑛⃗||𝑥0 , Lattice Boltzmann simulations. J. Phys. Chem. B 119 (16), 5220–5227.
0 0 − 2 𝑐𝑞𝑥
Genty, A., Gueddani, S., L. Dymitrowska, M., 2017. Computation of saturation depen-
𝛿𝑔𝑞̄ |𝑥+ =𝑥 1 = 𝑡𝑞 ||𝐽⃗𝑛 ⋅ 𝑛⃗||𝑥0 . (B.6) dence of effective diffusion coefficient in unsaturated argillite micro-fracture by Lat-
0 0 + 2 𝑐𝑞𝑥 tice Boltzmann method. Transp. Porous Media 117 (1), 149–168.
Genty, A., Pot, V., 2014. Numerical calculation of effective diffusion in unsaturated porous
Eq. (B.6) coincides with the flux-jump scheme (Guo et al., 2015) on the media by the TRT Lattice Boltzmann method. Transp. Porous Med. 105, 391–410.
Ginzburg, I., 2005a. Equilibrium-type and link-type Lattice Boltzmann models for generic
midway located straight interface when 𝑡𝑞 = 12 (minimal sets). Continu-
advection and anisotropic-dispersion equation. Adv. Water Res. 28, 1171–1195.
ity relation in Eq. (28a) remains the same but second-order accurate Ginzburg, I., 2005b. Generic boundary conditions for Lattice Boltzmann models and their
Eq. (40) becomes in diffusion case replaced by Eq. (B.7) with the exact application to advection and anisotropic-dispersion equations.. Adv. Water Res. 28,
flux-jump ||𝐽𝑛 ||𝑥=𝑥0 in the RHS: 1196–1216.
Ginzburg, I., 2006. Variably saturated flow described with the anisotropic Lattice Boltz-
mann methods.. J. Comput. Fluids 25, 831–848.
if ||∇̄ 𝑥 ̃ 𝑛 ||𝑥 = 0, or ∀ ̃ 𝑛 if  = 1 ∶ Ginzburg, I., 2007. Lattice Boltzmann modeling with discontinuous collision components.
0 4 Hydrodynamic and advection–diffusion equations.. J. Stat. Phys. 126, 157–203.
̄ 𝑥 𝐵𝑛 )(2) ||𝑥=𝑥 = ||𝐽𝑛 ||𝑥=𝑥 , Φ𝑥 = −𝜙 ∇
||(−𝜙 ∇ ̄ 𝑥 𝐵𝑛 . (B.7) Ginzburg, I., 2012. Truncation errors, exact and heuristic stability analysis of two-relax-
0 0 ation-times Lattice Boltzmann schemes for anisotropic advection–diffusion equation.
Commun. Comput. Phys. 11 (5), 1439–1502.
In pure diffusion in series of two porous blocks at order 𝑛 = 3, 𝑛 (𝑥) Ginzburg, I., 2013. Multiple anisotropic collisions for advection–diffusion Lattice Boltz-
is piecewise-linear and the semi-explicit scheme locates flux-jump con- mann schemes.. Adv. Water. Res. 51, 381–404.
5 Ginzburg, I., 2016. Comment on “an improved gray Lattice Boltzmann model for simu-
dition midway provided that  = 24 with Eq. (41); Eq. (B.2) is then
lating fluid flow in multi-scale porous media”: intrinsic links between LBE brinkman
held using Eq. (A.6) for 𝑆4 (𝑥) = 𝜙 𝜕𝑥 3 (𝑥) in Eq. (18b); second-order schemes. Adv. Water Resour. 88, 241–249.
accurate finite-differences are sufficient at 𝑛 = 1, 2, 3 to precompute Ginzburg, I., 2017. Prediction of the moments in advection–diffusion Lattice Boltzman
method II: attenuation of the boundary layers via double-𝜆 bounce-back flux scheme.
𝑆𝑛 (𝑥) = 𝜙 𝜕𝑥 𝑛−1 and to extrapolate ||𝐽𝑛 ||𝑥0 from the neighbor blocks.
Phys. Rev. E 95. 013305–42
Ginzburg, I., 2017. Prediction of the moments in advection–diffusion Lattice Boltzmann
References method I: truncation dispersion, skewness, and kurtosis. Phys. Rev. E 95. 013304–34
Ginzburg, I., d’Humières, D., 2007. Lattice Boltzmann and analytical modeling of flow
Aris, R., 1956. On the dispersion of a solute in a fluid flowing through a tube.. Proc. R. processes in anisotropic and heterogeneous stratified aquifers. Adv. Water Resour.
Soc. Lond. 235, 67–77. 30, 2202–2234.
Bandyopadhyay, K., Bhattacharya, A.K., Biswas, P., Drabold, A., 2004. Maximum entropy Ginzburg, I., d’Humières, D., Kuzmin, A., 2010. Optimal stability of advection–diffusion
and the problem of moments: a stable algorithm.. Phys. Rev. E 71 (5/2). 057701–4 Lattice Boltzmann models with two relaxation times for positive/negative equilib-
Batôt, G., Talon, L., Peysson, Y., Fleury, M., Bauer, D., 2016. Analytical and numerical rium. J. Stat. Phys. 139 (6), 1090–1143.
investigation of the advective and dispersive transport in Herschel–Bulkley fluids Ginzburg, I., Roux, L., 2015. Truncation effect on Taylor–Aris dispersion in Lattice Boltz-
by means of a Lattice-Boltzmann two-relaxation-time scheme. Chem. Eng. Sci. 141, mann schemes: accuracy towards stability.. J. Comput. Phys. 299, 974–1003.
271–281. Ginzburg, I., Roux, L., Silva, G., 2015a. Local boundary reflections in Lattice Boltzmann
Berkowitz, B., Emmanuel, S., Scher, H., 2008. Non-Fiskian transport and multiple-rate schemes: spurious boundary layers and their impact on the velocity, diffusion and
mass transfer in porous media. Water Resour. Res 44, W03402-16. dispersion. C. R. Mec. 343, 518–532.
Bijeljic, B., Mostaghimi, P., Blunt, M.J., 2011. Signature of non-Fickian solute transport Ginzburg, I., Silva, G., Talon, L., 2015. Analysis and improvement of Brinkman Lat-
in complex heterogeneous porous media. Phys. Rev. Lett. 107 (20), 204502. tice Boltzmann schemes: bulk, boundary, interface. similarity and distinctness with
Bijeljic, B., Mostaghimi, P., Blunt, M.J., 2013. Insights into non-Fickian solute transport finite-elements in heterogeneous porous media.. Phys. Rev. E 91, 023307.
in carbonates. Water Resour. Res. 49, 2714–2728.

33
JID: ADWR
ARTICLE IN PRESS [m5GeSdc;June 29, 2018;16:56]

I. Ginzburg Advances in Water Resources 000 (2018) 1–34

Ginzburg, I., Vikhansky, A., 2018. Determination of the diffusivity, dispersion, skewness Talon, L., Bauer, D., 2013. On the determination of a generalized Darcy equation for
and kurtosis in heterogeneous porous flow part I: analytical solutions with the ex- yield-stress fluid in porous media using a Lattice-Boltzmann TRT scheme. Eur. Phys.
tended method of moments. Adv. Water Res. 115, 60–87. J. E Soft Matter 36 (12), 139.
Golzar, M., 2015. A brief review of pond residence time studies. In: Proceedings of the Talon, L., Bauer, D., Gland, N., Youssef, S., Auradou, H., Ginzburg, I., 2012. Assessment of
Annual Postgraduate Research Student Conference. Sheffield, UK, pp. 32–37. the two relaxation time Lattice-Boltzmann scheme to simulate stokes flow in porous
Guo, K., Li, L., Xiao, G., Yeung, N.A., Mei, R., 2015. Lattice Boltzmann method for conju- media. Water Resour. Res. 48, W04526.
gate heat and mass transfer with interfacial jump conditions. Int. J. Heat Mass Transf. Taylor, G.I., 1953. Dispersion of soluble matter in solvent flowing slowly through a tube.
88, 306–322. Proc. R. Soc. Lond. Ser. A. 219, 186–2003.
Hammou, H., Ginzburg, I., Boulerhcha, M., 2011. Two-relaxation-times Lattice Boltzmann Trinh, S., Arce, P., Locke, B.R., 2000. Effective diffusivities of point-like molecules in
schemes for solute transport in unsaturated water flow, with a focus on stability. Adv. isotropic porous media by monte carlo simulation. Transp. Porous Med. 38, 241–259.
Water Res. 34, 779–793. Valdés-Parada, F.J., Lasseux, D., Bellet, F., 2016. A new formulation of the dispersion in
Hindmarsch, A.C., Grescho, P.M., Griffiths, D.F., 1984. The stability of explicit time-in- homogeneous porous media. Adv. Water Res 90, 70–82.
tegration for certain finite-difference approximation of the multi-dimensional advec- Vikhansky, A., 2008. Effect of diffusion on residence time distribution in chaotic channel
tion–diffusion equation. Int. J. Numer. Methods Fluids 84 (4), 853–897. flow. Chem. Eng. Sci. 63, 1866–1870.
Huang, R., Wu, H., 2014. A modified multiple-relaxation-time Lattice Boltzmann model Vikhansky, A., 2011. Numerical analysis of residence time distribution in microchannels.
for convection-diffusion equation. J. Comput. Phys. 274, 50–63. Chem. Eng. 86, 347–351.
Khirevich, S., Ginzburg, I., Tallarek, U., 2014. Coarse- and fine-grid numerical behavior Vikhansky, A., Ginzburg, I., 2014. Taylor dispersion in heterogeneous porous media: Ex-
of MRT/TRT Lattice-Boltzmann schemes in regular and random sphere packings. J. tended method of moments, theory, and modelling with two-relaxation-times Lattice
Comput. Phys. 281, 708–742. Boltzmann scheme. Phys. Fluids 26. 022104–52
Kim, J.-H., Oschoa, J.A., Whitaker, S., 1987. Diffusion in anisotropic porous media. Vogel, L.E., Makowski, D., Garnier, P., Vieublé-Gonod, L., Coquet, Y., Raynaud, X.,
Transp. Porous Med. 2, 327–356. Nunan, N., Chenu, C., Falconer, R., Pot, V., 2015. Modeling the effect of soil meso- and
Kuzmin, A., Ginzburg, I., Mohamad, A.A., 2011a. The role of the kinetic parameter in macropores topology on the biodegradation of a soluble carbon substrate. Adv. Wa-
the stability of two-relaxation-times advection–diffusion Lattice Boltzmann scheme. ter Res. 83, 123–136.
Comput. Math. Appl. 61 (12), 3417–3442. Wang, L., Shi, B., Chai, Z., 2015. Regularized Lattice Boltzmann model for a class of con-
Kuzmin, A., Guo, Z.L., Mohamad, A.A., 2011b. Simultaneous incorporation of mass and vection-diffusion equations. Phys. Rev. E 92. 043311–13
force terms in the multi-relaxation-time framework for Lattice Boltzmann schemes.. Wassén, S., Bordes, R., Gebäck, T., Bernin, D., Schuster, E., Lorén, N., Hermansson, A.M.,
Phil. Trans. R. Soc. A 369, 2219–2227. 2014. Probe diffusion in phase-separated bicontinuous biopolymer gels. Soft Matter
Mercer, G.N., Roberts, A.J., 1990. A centre manifold description of contaminant dispersion 10 (41). 8276–87
in channels with varying flow properties. SIAM J. Appl. Math 50, 1547. Werner, T.M., Kadlec, R.H., 2000. Wetland residence time distribution modeling. Ecol.
Ngo-Cong, D., Mohammed, F.J., Strunin, D.V., Skvortsov, A.T., Mai-Duy, N., Tran-Cong, T., Eng. 15, 77–90.
2015. Higher-order approximation of contaminant transport equation for turbulent Yan, Z., Yang, X., Li, S., Hilpert, M., 2017. Two-relaxation-time Lattice Boltzmann method
channel flows based on centre manifolds and its numerical solution. J. Hydrol. 525, and its application to advective–diffusive-reactive transport. Adv. Water Res. 109,
87–101. 333–342.
Perko, J., Patel, R.A., 2014. Single-relaxation-time Lattice Boltzmann scheme for advec- Yang, J., Crawshaw, J., Boek, E.S., 2013. Quantitative determination of molecular prop-
tion–diffusion problems with large diffusion-coefficient heterogeneities and high-ad- agator distributions for solute transport in homogeneous and heterogeneous porous
vection transport. Phys. Rev. E 89, 053309. media using Lattice Boltzmann simulations. Water Resour. Res. 49 (12), 8531–8538.
Rasin, I., Saucci, S., Miller, W., 2005. A multi-relaxation Lattice kinetic method for passive Yang, X., Mehmani, Y., Perkins, W.A., Pasquali, A., Schnherr, M., Kim, K., Perego, M.,
scalar diffusion. J. Comput. Phys. 206, 453–462. Parks, M.L., Trask, N., Balhoff, M.T., Richmond, M.C., Geier, M., Krafczyk, M.,
Röding, M., Schuster, E., Logg, K., Lundman, M., Bergström, P., Hanson, C., Gebäck, T., Luo, L.-S., Tartakovsky, A.M., Scheibe, T.D., 2016. Intercomparison of 3d pore-scale
Lorén, N., 2016. Computational high-throughput screening of fluid permeability in flow and solute transport simulation methods. Adv. Water Res. 95, 176–189.
heterogeneous fiber materials. Soft Matter 12 (29). 6293–9 Yang, Z., Bai, F., Huai, W., An, R., Wang, H., 2017. Modelling open-channel flow with
Salles, J., Thovert, J.F., Delannay, R., Prevors, L., Auriault, J.L., Adler, P.M., 1993. Taylor rigid vegetation based on two-dimensional shallow water equations using the Lattice
dispersion in porous media. determination of the dispersion tensor. Phys. Fluids A 5 Boltzmann method. Ecol. Eng. 106, 75–81.
(10), 2348–2377. Yoshida, H., Kobayashi, T., Hayashi, H., Kinjo, T., Washizi, H., Fukuzawa, K., 2014. Bound-
Servan-Camas, B., Tsai, F.T.C., 2008. Lattice Boltzmann method for two relaxation times ary condition at a two-phase interface in the Lattice Boltzmann method for the con-
for advection–diffusion equation: third order analysis and stability analysis.. Adv. Wa- vection–diffusion equation. Phys. Rev. E. 90. 013303–10
ter Res. 31, 1113–1126. Yoshida, H., Nagaoka, M., 2010. Multiple-relaxation-time Lattice Boltzmann model for the
Servan-Camas, B., Tsai, F.T.C., 2010. Two-relaxation time Lattice Boltzmann method for convection and anisotropic diffusion equation. J. Comput. Phys. 229, 7774–7795.
the anisotropic dispersive henry problem. Water Resour. Res. 46. W02515–W025 Zhang, X., Crawford, J.W., Flavel, R.J., Young, I.M., 2016. A multi-scale Lattice Boltzmann
Silva, G., Ginzburg, I., 2016. Stokes–Brinkman–Darcy solutions of bimodal porous flow model for simulating solute transport in 3d X-ray-tomography images of aggregated
across periodic array of permeable cylindrical inclusions: cell model, lubrication the- porous materials. J. Hydrol. 541, 1020.
ory and LBM/FEM numerical simulations. Transp. Porous Med. 111 (3), 795–825. Zhu, J., Ma, J., 2009. An improved gray Lattice Boltzmann model for simulating fluid flow
Silva, G., Talon, L., Ginzburg, I., 2017. Low- and high-order accurate boundary conditions: in multi-scale porous media. Comput. Geosci. 36, 1186.
from stokes to Darcy porous flow modeled with standard and improved Brinkman
Lattice Boltzmann schemes. J. Comput. Phys. 335, 50–83.
Su, T.M., Yang, S.C., Shih, S.S., Lee, H.Y., 2009. Optimal design for hydraulic performance
of free-water-surface constructed wetlands. Ecol. Eng. 35, 1200–1207.

34

You might also like