IPTC 17457

Lattice Boltzmann Flow Simulations With Applications of Reduced Order

Modeling Techniques
Donald L. Brown and Jun Li and Victor M. Calo and Mehdi Ghommem, Center for Numerical Porous Media,
Applied Mathematics and Computational Science, Earth and Environmental Sciences and Engineering, King
Abdullah University of Science and Technology, Thuwal, Saudi Arabia; Yalchin Efendiev, Department of
Mathematics & ISC, Texas A&M University, College Station, Texas

Copyright 2014, International Petroleum Technology Conference

This paper was prepared for presentation at the International Petroleum Technology Conference held in Doha, Qatar, 20–22 January 2014.

This paper was selected for presentation by an IPTC Programme Committee following review of information contained in an abstract submitted by the author(s). Contents of the paper, as
presented, have not been reviewed by the International Petroleum Technology Conference and are subject to correction by the author(s). The material, as presented, does not necessarily
reflect any position of the International Petroleum Technology Conference, its officers, or members. Papers presented at IPTC are subject to publication review by Sponsor Society
Committees of IPTC. Electronic reproduction, distribution, or storage of any part of this paper for commercial purposes without the written consent of the International Petroleum Technology
Conference is prohibited. Permission to reproduce in print is restricted to an abstract of not more than 300 words; illustrations may not be copied. The abstract must contain conspicuous
acknowledgment of where and by whom the paper was presented. Write Librarian, IPTC, P.O. Box 833836, Richardson, TX 75083-3836, U.S.A., fax +1-972-952-9435

Abstract

With the recent interest in shale gas, an understanding of the flow mechanisms at the pore scale and beyond is
necessary, which has attracted a lot of interest from both industry and academia. One of the suggested algorithms to help
understand flow in such reservoirs is the Lattice Boltzmann Method (LBM). The primary advantage of LBM is its ability to
approximate complicated geometries with simple algorithmic modificatoins. In this work, we use LBM to simulate the flow
in a porous medium. More specifically, we use LBM to simulate a Brinkman type flow. The Brinkman law allows us to
integrate fast free-flow and slow-flow porous regions. However, due to the many scales involved and complex
heterogeneities of the rock microstructure, the simulation times can be long, even with the speed advantage of using an
explicit time stepping method. The problem is two-fold, the computational grid must be able to resolve all scales and the
calculation requires a steady state solution implying a large number of timesteps. To help reduce the computational
complexity and total simulation times, we use model reduction techniques to reduce the dimension of the system. In this
approach, we are able to describe the dynamics of the flow by using a lower dimensional subspace. In this work, we utilize
the Proper Orthogonal Decomposition (POD) technique, to compute the dominant modes of the flow and project the solution
onto them (a lower dimensional subspace) to arrive at an approximation of the full system at a lowered computational cost.
We present a few proof-of-concept examples of the flow field and the corresponding reduced model flow field.

Introduction

Understanding the flow mechanics of petroleum reservoirs is critical to economic evaluation of recoverable
resources and reservoir decelopment planning. Laboratory mesurements can be very expensive and may be limited in scope
of scales that can be explored due to inability to accurately describe in-situ conditions. A key tool in understanding flow on
both the pore and reservoirs scales is simulation methods. In reservoir conditions, there are many complex muli-physical
processes that need to be modeled and, complicating issues further, there are a great number of uncertainties in subsurface
parameters. These uncertainties require fast and accurate forward models to update parameters. Finally, due to the scale
disparities involved, simulation of these reservoirs is often very computational expensive. With the recent boom in Shale gas,
understanding fracture modeling and pore-scale flow mechancics has become critical. Computational tools and techniques
must be developed in these areas to reduce simulation cost and time.
There are a wide variety of numerical techniques employed in the oil and gas industry. The primary numerical
tehcniques utilized for flow simulations are based on finite volume methods (FVMs) as they have the property of mass
conservation. A property desired for field simulations as mass balance must be maintained, thus these methods are widely
used in commercial simulators for flow. However, for pore-scale studies, the complicated geometries of the rock
microstructures can make gridding FVMs and finite element methods (FEMs) very tedious and time consuming. One of the
possible candidates for flow simulations to circumvent this difficulty is to use particle methods such as Lattice Boltzman
methods (LBMs). Lattice Boltzmann methods have been used widely in the study of multiphase and multi-component flow
2 IPTC 17457

simulation. Recently, these methods have been adapted to modeling flows in porous media using the force model proposed
in (Guo and Zhao (2002)).
The primary equation used to model fluid flow in petroleum reservoirs is the Darcy equation and its multiphase
counterpart. Much of the industry simulators today are built on this modeling foundation. However, we would like to
incorporate high contrast vugular or channelized regions of the reservoirs common in many carbonate and karst formations
(Popov et al. (2007)). An often utlized equation to incorporate both standard Darcy effects and the high flow regions is the
Brinkman equation. By starting with the Stokes equations of free fluid mechanics and adding a resistive term, we are able to
model effectively both regimes. By tuning the coefficient of this Brinkman term we may enter regimes of high flow in
fissures and channels, or areas of low flow such as the porous rock region or Darcy dominated flow areas. In the work of
(Guo and Zhao (2002)), the authors add a forcing term to the standard LBM formulation that contains the rock
heterogeneities. In this study, we will consider such an LBM for Brinkman flows.
Solving the flow field and pressures for the Brinkman equations can be accomplished via this lattice Boltzmann
model. However, the heterogeneities in the permeability fields and rock microstructures contain many scales and may be high
constrast. Therefore, the grid resolution must be large enough to resolve these heterogeneities and, in addition, since we must
solve to steady state to arrive at an equilibrium velocitiy and pressure field, the number of time steps may be very large. A
way to reduce the complexity of the system is through the use of reduced order models (ROMs). In this methodology, the
dynamics of the system can be captured on a lower dimentional subspace (Akhtar et al. (2009)). By solving the full system
for a short number of timesteps, we are able to capture the underlying dynamics and project onto a lower dimensional
subspace and, subsequently, solve a lower dimensional system. The method we employ in this study is that of Proper
Orthogonal Decomposition (POD) where we project the system onto the dominant modes arising from a singular value
decomposition of the solutions (Akhtar et al. (2012)).
In this paper, we present the Shan-Chan Lattice Boltzmann Model (Chen et al. (1992)) in the first section. This
allows us to understand the basics of the LBM as a numerical method. Then, we present the extension of the LBM to porous
media flow through the addition of a forcing term containing rock permeability information. Once we have the LBM outlined
for porous media, we cover the basic ideas of ROMs and PODs to the corresponding LBM linear system. Finally, we present
a few proof of concept numerical examples that show the POD method can be employed to reduce the LBM model in an
effective way, thereby reducing computational time but still capturing key parts of the dynamics of the flow.

Lattice Boltzmann Methods

Shan-Chen Lattice Boltzmann Model: The continuous Boltzmann equation is a fundamental model in kinetic
theory. The equation is an integro-differential transport equation that models interactions of particle through a complicated
nonlinear collision integral. The connection to fluid mechanics comes from the Chapman-Enskog expansion (Chapman
(1990)) in the Knudsen number 𝜅, relating kinetic theory to the Euler and Navier-Stokes equations. By supposing the BGK
approximation for the collision integral, discretizing both velocity and space (Qian et al. (1992)), and taking Chapman-
Enskog limits, we are able to obtain a method for solving fluid mechanics equations based on the resulting lattice Boltzmann
equations. This idea is the basis for the Shan-Chen model.
In this section, we outline the general LBM methodology (McNamara and Zanetti (1988); Chen et al. (1992); Qian et al.
(1992)) as well as the extension to a Brinkman fluid by adding a forcing term that includes a permeability field as input data
(Guo and Zhao (2002)). This model will be our main focus for this study and we follow the modification and presentation as
in (Li et al. (2013)). As previously stated, the computational domain is discretized into lattice points denoted 𝑥. We let Δ𝑥
and Δ𝑡 be the space and time discretization sizes, respectively. The velocity space at each lattice point can be discretized in a
number of ways resulting in different LBM models. In this work, since our simulations are restricted to two-dimensions, we
use the D2Q9 model (Qian et al. (1992)). In this mode, the velocity space of each particle is discretized into 9 directions. At
each point we have 9 lattice velocities 𝑒𝛼 and weighting values 𝜔𝛼 . We shift to other adjacent lattice point via the relation
Δ𝑥
𝑥 + Δ𝑡𝑒𝛼 . The magnitude of 𝑒𝛼 is proportial to the speed of the particle 𝑐 = . In the D2Q9 model, we have 𝑒0 = (0,0), and
Δ𝑡
4 (𝛼−1)𝜋 1
𝜔0 = , 𝑒𝛼 = 𝑐 (cos(𝜃𝛼 ) , sin(𝜃𝛼 )) where 𝜃𝛼 = , 𝜔𝛼 = , for 𝛼 = 1, … ,4 and 𝑒𝛼 = 𝑐 √2(cos(𝜃𝛼 ) , sin(𝜃𝛼 )) where
9 2 9
(𝛼−5)𝜋 1
𝜃𝛼 = + 𝜋/4, 𝜔𝛼 = , for 𝛼 = 5, … , 8. The evolution of the particles, with this velocity discretization, through the
2 36
lattice is given by

(𝒆𝒒)
𝒇𝜶 (𝒙, 𝒕) − 𝒇𝜶 (𝒙, 𝒕)
𝒇𝜶 (𝒙 + 𝚫𝒕 𝒆𝜶 , 𝒕 + 𝚫𝒕) = 𝒇𝜶 (𝒙, 𝒕) + . (1)
𝝉

Here, 𝑓𝛼 (𝑥, 𝑡) is the number or particles with velocity 𝑒𝛼 at the lattice point 𝑥 at time 𝑡. The initial distribution of particles is
assumed to be dependent on the density 𝜌(x,t) and initial macro-velocity 𝑢(𝑥, 𝑡). More precisely, the initial distribution of
(𝑒𝑞)
particles if given by 𝑓𝛼 (𝑥, 0) = 𝑓𝛼 (𝜌, 𝑢). In the Shan-Chen model, the equilibrium function is given by a BGK type
expansion of the collision integral written as
IPTC 17457 3

(𝑒𝑞) 3 9 2 3
𝑓𝛼 (𝑥, 𝑡) = 𝜔𝛼 𝜌 �1 + 2
𝑒𝛼 ∙ 𝑢(𝑒𝑞) + 4 �𝑒𝛼 ∙ 𝑢(𝑒𝑞) � − 2 𝑢(𝑒𝑞) ∙ 𝑢(𝑒𝑞) � , (2)
𝑐 2𝑐 2𝑐

where the density is 𝜌 = ∑8𝛼=0 𝑓𝛼 . Note that the above relation is nonlinear in equilibrium velocity and thereby making our
system nonlinear. In our calculations we will neglect these nonlinearities and save such complications for future study. We
3𝜈Δ𝑡 1
denote the relaxation time 𝜏 and is related to the kinematic viscosity coefficient 𝜈 by the relation 𝜏 = 2 + . The
Δ𝑥 2
equilibrium velocity is given by

𝜌𝑢′ + 𝜏𝐹
𝑢(𝑒𝑞) = , (3)
𝜌

where 𝐹 is a total volume force. This forcing is how we will connect traditional LBM with the Brinkman equations of porous
media by adding the permeabilities of the rock microstructures. In the above equation we define the velocity 𝑢′ as

∑8𝛼=0 𝑒𝛼 𝑓𝛼
𝑢′ = . (4)
∑8𝛼=0 𝑓𝛼

Brinkman Lattice Boltzmann Model: We will now describe the algorithm used in our simulations. By choosing a
particular forcing in Eq. (3), we are able to incorporate pore structure though a permeability coefficient. This permeability
can be heterogeneous and contain many scales and thereby increasing computational costs needed to capture these features.
In later sections, we will use model reduction to simplify these calculations. We now present the model in (Guo and Zhao
(2002); Guo (2002)) that we use in our simulations. Using the particle stepping of Eq. (1), however, with an equilibrium
function given by

𝜌𝑢′ + 𝜏𝛥𝑡𝜌𝑓𝑚
𝑢(𝑒𝑞) = , (5)
𝜌

where the forcing is given as

𝜀𝜈 𝜀𝐹𝜀
𝑓𝑚 = − 𝑢− |𝑢|𝑢 + 𝜀𝑔. (6)
𝑘 √𝑘

Here the porosity is denoted by 𝜀, 𝑘 is the permeability coefficient, and 𝑔 is an external gravitational force. The term
containing 𝐹𝜀 is the Forcheimer force coefficient that we will be neglecting in our simulations and analysis to avoid
nonlinearities. There are two kinetic viscosities, the one previously defined 𝜈, and the effective viscosity 𝜈𝑒𝑓𝑓 = 𝑐 2 𝛥𝑡(𝜏 −
.5)/3. We obtain explicit formulas for the flow velocities as in (Guo and Zhao (2002)) by the following relations

𝑣
𝑢= ,
𝑎0 + �𝑎02 + 𝑎1 |𝑣|

∑8𝛼=0 𝑒𝛼 𝑓𝛼 + .5 𝛥𝑡𝜀𝜌𝑔
𝑣= , (7)
𝜌

𝜀𝛥𝑡𝜈
𝑎0 = .5 + ,
4𝑘
𝜀𝛥𝑡𝐹𝜀
𝑎1 = .
2√𝑘

The equilibrium velocity, 𝑢(𝑒𝑞) , is computed from the above quantities as

∑8𝛼=0 𝑒𝛼 𝑓𝛼
𝑢(𝑒𝑞) = 2 𝜏𝑢 + (1 − 2𝜏) . (8)
𝜌

In the incompressible limit, the particle velocity is much faster than the velocity of the macro-fluid velocity, more precisely
|𝑢| ≪ 𝑐. We obtain, at the steady state or equilibrium distribution, the Brinkman-Forcheimer equations given by
4 IPTC 17457

−1 𝜀𝜈 𝜀𝐹𝜀
𝛻𝑝 + 𝜈𝑒𝑓𝑓 𝛥𝑢 − 𝑢 − |𝑢|𝑢 + 𝜀𝑔 = 0, (9)
𝜌0 𝑘 √𝑘

𝛻 ∙ 𝑢 = 0. (10)

𝜌𝑐 2
Here, 𝜌0 is the initial uniform density and the pressure is related to the density via the relation 𝑝 = . Depending on the
3
input data and the grid parameter ratios we are able to fix the above steady state. In our analysis, we assume that 𝐹𝜀 = 0 as to
simplify the nonlinearity in the resulting equations. The extension to a nonlinear Forchiemer type steady state as well as a
fully nonlinear BGK type equilibrium distribution (c.f. Eq. (2)) is something left for future investigations.

Model Reduction using Proper Orthogonal Decomposition

Using LBM methods is an effective method to deal with complex pore geometry. However, to deal with physical
problems with many scales and complex physics is computationally expensive. This complexity is exacerbated if we include
convection advection nonlinearities or Forcheimer terms to the system described above. The number of degrees of freedom
may be very large as the number of lattice points must be able to resolve complex heterogeneities in the permeability field. In
addition, combining the stability constraints of using an explicit LBM time stepping method with the fact that we must solve
the LBM algorithm to steady state to yield a proper solution to Eq. (9) and Eq. (10) makes having a large system extremely
computationally costly for real world microstructures or permeabilities as the number of time-steps may be very large.
One approach to reduce the complexity of the system and, in turn, reduce computation times is through the use of
model reduction techniques. It is often possible to describe the fundamental dynamics of a complicated system on a lower
dimensional subspace of the total solution space. By solving a reduced system the computation times can be greatly reduced
only at the one time off-line cost of solving a small time interval the full system. There are many widely used methods for
reducing models in a physically relevant way. The method of model reduction we use in the work is the Proper Orthogonal
Decomposition (POD) (Wang et al. (2011); Hay et al. (2012)). In this section, we give an overview of POD and its
application to LBMs.
To combine model reduction techniques with LBMs, we rewrite the evolution dynamics equation of the distributions
of particles Eq. (1) as a matrix evolution equation

𝐹 𝑛+1 (𝑥, 𝑡 𝑛+1 ) = 𝐴(𝑥)𝐹 𝑛 (𝑥, 𝑡 𝑛 ). (11)

For each lattice point 𝑥 𝑖,𝑗 , time 𝑡 𝑛 , and each velocity direction 𝑒𝛼 , we assign each 𝑓𝛼 (𝑥 𝑖,𝑗 , 𝑡 𝑛 ) → (𝐹 𝑛 )𝑙 for some flattened
index 𝑙. The size of this vector is given by 𝑁 = (𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝐿𝑎𝑡𝑡𝑖𝑐𝑒𝑠) × 9 and so the size of the evolution matrix 𝐴(𝑥)
(assumed time independent and linear) is 𝑁 × 𝑁. Thus, for many particles the above system Eq. (11) can be very large. The
objective is to reduce the size of this system, while capturing the underlying dynamics and maintaining the accuracy of the
fully-resolved solution.
The first step in the POD algorithm is to perform a one-time offline full computation over a short initial time period;
that is, we solve the full system Eq. (11) at 𝑅 initial snapshot times (𝑡 1 , … , 𝑡 𝑅 ) and we collect these solutions as columns of
the 𝑁 × 𝑅 snapshot matrix given by

𝑆 = [𝐹1 , 𝐹 2 , … , 𝐹 𝑅 ]. (12)

Then, we perform a singular value decomposition (SVD) on the above matrix and write 𝑆 = 𝑈𝛴𝑉 𝑇 . We note that a more
compact SVD may be utilized, but that is for future research. Here, the matrix 𝑆 is decomposed into left singular value
vectors in 𝑈, corresponding singular values matrix 𝛴, and right singular values 𝑉 𝑇 . The singular values are ordered in
greatest to least as 𝜎12 ≥ 𝜎22 ≥ ⋯ ≥ 𝜎𝑅2, and corresponding to the column of the matrix 𝑈. Often the first few modes of the
SVD are dominant and contain a large portion of the energy 𝐸 = ∑ 𝜎𝑖2 of the system and in this case the system has a
meaningful reduced subspace. More precisely, the solution may be represented as a linear combination of only a few modes.
By selecting a number of modes that contain most of the energy, we may reduce the system. We consider the first
𝑀 ≪ 𝑅 ≪ 𝑁, modes 𝑈 of the snapshot matrix and construct the column of 𝛹, 𝑁 × 𝑀, as

𝛹(𝑖) = 𝑈(𝑖), 𝑖 = 1, … , 𝑀. (13)

The above matrix allows us to project from the fully-resolved systems into the subspace spanned by the POD modes to
obtaine the reduced system. We are now in the position to write the reduced system of Eq. (11). We denote the 𝑀 × 1,
𝑛 𝑛
reduced unknowns as 𝐹𝑟𝑑 (𝑥, 𝑡 𝑛 ), and we relate this to the fully resolved unknowns by the relation 𝐹 𝑛 ≈ 𝛹𝐹𝑟𝑑 . Applying this
approximation to Eq. (11), we arrive at
IPTC 17457 5

𝑛+1 𝑛
𝛹𝐹𝑟𝑑 = 𝐴𝛹𝐹𝑟𝑑 . (14)

The above system is now reduced but not square. Exploiting the orthogonality property of the POD modes: that is, 𝛹 𝑇 𝛹 =
𝐼𝑑𝑀×𝑀 , we project the above system into the lower dimensional reduced subspace by multiplication of 𝛹 𝑇 , and obtain
𝑛+1 𝑛
𝐹𝑟𝑑 = 𝛹 𝑇 𝐴𝛹𝐹𝑟𝑑 . (15)

The matrix 𝑀 × 𝑀 reduced matrix 𝛹 𝑇 𝐴𝛹 is now significantly smaller than the initial system Eq. (11). By projecting
the governing equations onto the modes, we are able to solve a much smaller system as an approximation to the fully-
resolved system. This is done at a one-time cost of solving the expensive model to construct the snapshot matrix. Given
certain robustness conditions, the above snapshot matrix may be used for different initial conditions, input, and forcing data
as long as the modes capture the underlying dynamics.

Numerical Examples

In this section, we present some demonstration of concept examples comparing the reduced order models to the fully
resolved system. We suppose our computational domain is the unit square [0,1]2 . The boundary conditions that we
implement are periodic, that is, if a particle moves off the right side of the grid it travels to the corresponding lattice on the
1 1
left side etc. The grid size, time step and velocity are taken to be ∆𝑥 = , ∆𝑡 = , 𝑐 = 1000 for all of our simulations.
30 30000
We solve from in the time interval [0,2] and take the gravity forcing to be 𝑔 = (1,0). We also suppose that 𝜌0 = 10, 𝜏 =
.95, 𝜀 = .8. The permeabilitiy field is piecewise defined as 𝑘(𝑥, 𝑦) = .001 if |𝑥 − .25| ≤ .1 and |𝑦 − .25| ≤ .1 or if
|𝑥 − .75| ≤ .1 and |𝑦 − .75| ≤ .1, and 𝑘(𝑥, 𝑦) = 1 otherwise.
For the full resolution solve, we take 200 snapshots at intervals of every 10 timesteps to capture the actual dynamics,
𝑖,𝑗
using the initial particle distribution 𝑓𝛼 = 10 + 𝑠𝑖𝑛(𝑥) to create some initial dynamics that can be captured by the POD
method. Using the resulting snapshot matrix to compute 𝛹, we solve a reduced system Eq. (15) with the initial condition
𝑖,𝑗
𝑓𝛼 = 10, using a different number of modes. Then, we post process the solution to calculate the velocity 𝑢 in the 𝑥
direction. The resulting velocity can be seen in the figures below at time 𝑡 = 1. We plot the fully resolved velocity, then the
velocity from the reduced order model for 5,10, and 15 modes. The resulting relative velocity error from the fully resolved
solution is 67%, 40%, and 15%. Therefore, once a critical number of modes to capture the dynamics is reached, a very good
relative error to the fully resolved system can be achieved in the velocity. These results can be observed in Figures 1-4.

Conclusions

In this work we presented the Shan-Chen model for the Lattice Boltzmann Method and its extensions to porous
media flow by the addition of a forcing term containing the subsurface information. We then covered the proper orthogonal
decomposition method as a reduced order modeling tool for the corresponding linear system. In this framework, we
calculated snapshots of the initial fully resolved system and performed a singular value decomposition on the resulting
matrix. Then, we projected the fully resolved high dimensional system onto a few of dominant modes of the system, while
capturing the underlying physics. Finally, we presented a few proof of concept examples to show that, qualitatively, the fully
resolved flow field can be captured by a few of the POD modes. Future work includes the application of more advanced
ROM techniques and the extension to nonlinear model reduction by incorporating nonlinear flow effects into the ROM.

Acknowledgements

We would like to acknowledge Rio Yakota for his helpful suggestions on parallelization of the code suggestions on how
to fix and better organize the computations. Also, we would like to acknowledge Phil Vignal and Licandro Dalcin for their
help in visualization of the data.

References

I. Akhtar, A. H. Nayfeh, C. J. Ribbens, On the stability and extension of reduced-order Galerkin models in incompressible
flows: A numerical study of vortex shedding, Theoretical and Computational Fluid Dynamics 23 (3) (2009) 213–237.

I. Akhtar, Z. Wang, J. Borggaard, T. Iliescu, A new closure strategy for proper orthogonal decomposition reduced-order
models, Journal of Computational and Nonlinear Dynamics 7 (3) (2012) 034503.

Chen, H., Chen, S. and Matthaeus W. H. Recovery of the Navier-Stokes equations using a lattice-gas Boltzman method,
Physical Review A 45 (1992):5339-5342
6 IPTC 17457

S. Chapman, T. Cowling, The mathematical theory of non-uniform gases: an account of the kinetic theory of viscosity,
thermal conduction, and diffusion in gases, Cambridge University Press, 1990.

Guo Z., and Zhao T. Lattice Boltzmann model for incompressible flows through porous media, Physical Review E. 66.
(2002): 036304.

Guo Z., and Zheng C. and Shi B. Discrete lattice effects on the forcing term in the Lattice Boltzmann method, Physical
Review E. 65. (2002): 046308.

A. Hay, J. Borggaard, I. Akhtar, D. Pelletier, Reduced-order models for parameter dependent geometries based on shape
sensitivity analysis, Journal of Computational Physics 229 (2010) 1327–1352.

A. Hay, I. Akhtar, J. T. Borggaard, On the use of sensitivity analysis in model reduction to predict flows for varying inflow
conditions, International Journal for Numerical Methods in Fluids 68 (1) (2012) 122–134.

J. Li, D.L. Brown, V. Calo, Y. Efendiev, O. Iliev, Multiscale Lattice Boltzmann Method for ow Simulations in Highly
Heterogenous Porous Media.", September 2013, SPE Reservoir Simulation Conference and Exhibition, Abu Dhabi,
UAE.

McNamara G. R., and Zanetti G. Use of the Boltzmann equation to simulate lattice-gas automata, Physical Review Letters
61: (1988) 2332-2335.

P. Popov, G. Qin, L Bi, Y. Efendiev, R. Ewing, Z. Kang and J. Li, Multiscale Methods for Modeling Fluid Flow Through
Naturally Fractured Carbonate Karst Reservoirs, SPE Annual Technical Conference and Exhibition, 11-14 November
2007, Anaheim, California, U.S.A.

Qian Y.H., d’Humieres D., and Lallemand P., Lattice BGK models for Navier-Stokes equations, Europhysics Letter 17:
(1992) 479-484.

Z. Wang, I. Akhtar, J. Borggaard, T. Iliescu, Two-level discretizations of nonlinear closuremodels for proper orthogonal
decomposition, Journal of Computational Physics 230 (1) (2011) 126–146.

Figure 1. The fully resolved x velocity field at time t=1.

IPTC 17457 7

Figure 2. Reduced order x velocity field at t=1, 5 modes used.

Figure 3. Reduced order x velocity field at t=1, 10 modes used.

8 IPTC 17457

Figure 3. Reduced order x velocity field at t=1, 15 modes used.