betaFeedback

WorksMathematics

A lattice Boltzmann method for Biot's consolidation model of linear poroelasticity

Abstract

Biot's consolidation model is a classical model for the evolution of deformable porous media saturated by a fluid and has various interdisciplinary applications. While numerical solution methods to solve poroelasticity by typical schemes such as finite differences, finite volumes or finite elements have been intensely studied, lattice Boltzmann methods for poroelasticity have not been developed yet. In this work, we propose a novel semi-implicit coupling of lattice Boltzmann methods to solve Biot's consolidation model in two dimensions. To this end, we use a single-relaxation-time lattice Boltzmann method for reaction-diffusion equations to solve the Darcy flow and combine it with a recent pseudo-time multi-relaxation-time lattice Boltzmann scheme for quasi-static linear elasticity. We employ a multi-grid method for the latter scheme to achieve quasi-optimal computational cost. For the coupling between the equations, we develop a centered update scheme, that incorporates both explicit and semi-implicit contributions. The numerical results demonstrate that naive (explicit or semi-implicit) coupling schemes lead to instabilities when the poroelastic system is strongly coupled. However, the newly developed centered coupling scheme is stable and accurate in all considered cases, even for the Biot–Willis coefficient being one. Furthermore, the numerical results for Terzaghi's consolidation problem and a two-dimensional extension thereof highlight that the scheme is even able to capture discontinuous solutions arising from instantaneous loading.

Introduction

The evolution of deformable porous media saturated by a fluid is subject to long-lasting scientific investigations due to its various interdisciplinary applications ranging from geological CO2\mathrm{CO}_2 storage [1] and geothermal energy [2], over material design for nano-composites [3] and concrete [4], to the function of biological tissue [5] such as the optic nerve [6] and heart dynamics [7]. The theoretical origin dates back to 1923, when Terzaghi developed a description of one-dimensional consolidation [8]. In the subsequent decades, the theory was extended by Biot [9, 10, 11, 12] leading to the nowadays called Biot’s consolidation model. It describes the two-phase solid-fluid systems based on a linearly elastic, porous solid skeleton that undergoes small deformations, while the fluid flow induced by the deformation of the solid is governed by Darcy’s law. For a general introduction to linear and non-linear poroelasticity, we refer the interested reader to [13, 14].

Since analytical solutions are barely available, there is a need for numerical methods to solve the coupled system. Starting from classical finite difference methods [15, 16, 17] and finite element methods [18, 19], more advanced schemes were developed to improve the approximation of stress and flux and allow for strongly heterogeneous material parameters. This has led to multiple finite volume methods [20, 21, 22, 23], and a large class of tailored finite element methods using, e.g., discontinuous Galerkin [24, 25], mixed finite elements [26, 27], or enriched finite elements [28, 29]. Furthermore, these methods can account for local mass conservation, avoid locking effects and suppress non-physical pressure oscillations. For large deformations, the hybrid usage of Lagrangian particles and Eulerian background mesh has been introduced as material point method and extended to poroelasticity [30, 31].

A substantially different numerical approach are lattice Boltzmann methods (LBMs), which were originally developed for the simulation of fluid dynamics [32, 33]. In contrast to the above mentioned methods, which rely on the approximation of the macroscopic equations, the LBM is based on a mesoscopic description of statistical physics for gas particle dynamics, i.e., the discretization of the Boltzmann equation in physical and velocity space. While the single-relaxation-time collision operator [34] is most popular due to its simplicity, variants based on multiple relaxation times [35] and cumulants [36] are used for more complex problems to improve accuracy, stability and the number of represented physical parameters. For a detailed introduction to LBM, we refer to [37, 38]. A major advantage of LBM is the simple algorithmic structure and the straightforward parallelization, especially on GPUs, which make the method well suited for modern, massively parallel and hybrid computing architectures [39, 40].

Nowadays, various LBMs exist for most of the partial differential equations encountered in practice. In particular, LBMs were quickly adapted for advection-diffusion-reaction equations which have a similar structure as the Navier–Stokes equations. This was originally recognized for pure diffusion in [41], for advection-diffusion problems in [42] and for diffusion-reaction problems in [43]. Based on these works, the method has been extended in several ways. Anisotropic diffusion and dispersion was included by a direction-dependent relaxation in [44] and using a multiple-relaxation-time (MRT) scheme in [45, 46], while high Péclet numbers were treated with an upwind scheme in [47]. The full advection-diffusion-reaction equation was considered in [48] using a source-correction scheme involving the temporal derivative, while a simpler moment update was proposed in [49]. Furthermore, nonlinear coefficients were incorporated in [50] and combined with a MRT scheme for anisotropy in [51]. We refer to [52] for a detailed analysis of LBM for nonlinear and anisotropic advection-diffusion-reaction equations. Finally, we note the connection to relaxation systems for general equilibrium and collision operators established in [53], and the fourth-order accuracy achieved in [54, 55] for a LBM solving diffusion-reaction equations.

Concerning linear elasticity, the first LBM was introduced in [56] using two wave equations of the same speed describing the motion of longitudinal and transverse displacements. This approach was adopted in [57] using extended velocity sets in 2D and 3D combined with finite difference schemes and flux limiters to avoid spurious oscillations. Furthermore, arbitrary Poisson’s ratios, and thus different wave speeds, were achieved by an extension to isotropic linear materials in [58, 59]. The realization of several types of boundary conditions has been developed in [59, 60]. Another approach is discussed in [61], where anti-plane shear deformation is assumed to simplify the equations into a scalar wave equation, that is subsequently solved by LBM. An extension to two wave equations describing dilation and rotation in 2D is presented in [62]. There, the two equations are solved by two separate LBMs, which are coupled via finite difference schemes. Recently, a vectorial LBM has been proposed in [63] for linear elastodynamics in two-dimensional domains with periodic or Dirichlet boundary.

While all these methods solve dynamic elasticity, the approximation of quasi-static elasticity by LBM is more intricate. LBMs are essentially explicit schemes in time, and hence not directly applicable to elliptic equations. To circumvent this, the addition of a time-dependent damping term was proposed in [64], which leads to an iterative LBM that solves the quasi-static equation when approaching the steady-state. In particular, the method in [64] combines a vectorial LBM for the components of the displacement with a finite difference approximation of the divergence. Using the same pseudo-time approach, quasi-static linear elasticity in 2D has been solved in [65] using a MRT LBM on a reduced velocity stencil. While originally posed as an periodic problem, the treatment of Dirichlet and Neumann boundary conditions is discussed in [66].

To accelerate the convergence to a steady-state solution, multi-grid methods are frequently used for all types of discretization methods. The fundamental idea of multi-grid methods is that many iterative methods rapidly smooth high frequency error components on a fine grid, while low frequency error components can be represented on a coarser grid, where they become high frequency error components again. Thus, recursively applying the iterative method on multiple meshes and transferring the corresponding error corrections, this approach achieves a quasi-optimal convergence acceleration. In the context of LBM, multi-grid methods were first introduced in [67, 68] for steady-state solutions of the Navier–Stokes equations. Multiple extensions for this application were developed. In [69], an generalization of MRT LBM was presented, a simplified V-cycle was proposed in [70], and the combination with complex boundaries treated by an immersed-boundary method was discussed in [71]. Moreover, multi-grid LBM for solving second-order elliptic partial differential equations was derived in [72]. However, to the best of the authors’ knowledge, there have not been developed any LBMs for solving poroelasticity yet. In this work, we propose a novel coupled LBM to solve Biot’s consolidation model in two dimensions. To this end, we use the classical LBM for reaction-diffusion equations [49] for the Darcy flow and combine it with the recent pseudo-time LBM for quasi-static linear elasticity [65]. We employ a multi-grid method for the latter scheme and thus achieve quasi-optimal computational cost. To resolve the implicit coupling between the equations, we develop a centered update scheme, that incorporates both explicit and semi-implicit contributions. We demonstrate that naively explicitly or semi-implicitly coupled LBMs lead to instabilities for a strongly coupled problem (Biot–Willis coefficient close to one), while the centered coupling scheme using both contributions is stable and accurate for all Biot–Willis coefficients. The numerical results for Terzaghi’s consolidation problem highlight that the scheme is even able to capture solutions with discontinuities arising from instantaneous loading.

The outline of this paper is as follows. The governing equations of Biot’s consolidation model are presented in § 2. Next, we introduce the LBMs for flow and elasticity in § 3, discussing in detail the different coupling schemes, the multi-grid acceleration and the used initial and boundary conditions. In § 4, we present numerical results for a problem with a manufactured solution, for Terzaghi’s consolidation problem as well as for an two-dimensional extension of the latter. Finally, § 5 summarizes the main conclusions.

Governing equations

In the following two subsections, we first present the system of equations representing Biot’s consolidation model in dimensional form and subsequently derive the dimensionless formulation, which will be used in the LBM.

Biot’s consolidation model

We consider Biot’s model of consolidation for a linearly elastic and porous solid, which is saturated by a slightly compressible viscous fluid [10, 12, 13, 14]. For given final time tf∈(0,∞)t_f \in(0,\infty) and a simply connected, Lipschitz-bounded domain D∈Rd\mathcal{D} \in\mathbb{R}^{d}, we define the time-space cylinder Q=(0,tf)×DQ = (0,t_f) \times\mathcal{D}. Note that we discuss in this section the general model with dimension d∈{1,2,3}d \in\{1,2,3\}, whereas we restrict ourselves to d=2d = 2 thereafter for the LBM. Furthermore, the displacement and stress of the solid are denoted by η:Q→Rd\boldsymbol{\eta}: Q \to\mathbb{R}^{d} and σ:Q→Rd×d\boldsymbol{\sigma}: Q \to\mathbb{R}^{d \times d}, respectively, while the fluid pressure and Darcy velocity are denoted by p:Q→Rp: Q \to\mathbb{R} and u:Q→Rd\boldsymbol{u}: Q \to\mathbb{R}^{d}, respectively. Assuming an isotropic material, the system of partial differential equations reads

−∇⋅σ+α∇p=f(1)-\nabla\cdot\boldsymbol{\sigma} + \alpha\nabla p = \boldsymbol{f} \tag*{(1)}
σ=λ(∇⋅η)I+μ(∇η+∇ηT)(2)\boldsymbol{\sigma} = \lambda(\nabla\cdot\boldsymbol{\eta})\boldsymbol{I} + \mu(\nabla\boldsymbol{\eta} + \nabla\boldsymbol{\eta}^{\mathsf{T}}) \tag*{(2)}
∂t(c0p+α∇⋅η)+∇⋅u=s(3)\partial_t(c_0p + \alpha\nabla\cdot\boldsymbol{\eta}) + \nabla\cdot\boldsymbol{u} = s \tag*{(3)}
u=−κ(∇p−g)(4)\boldsymbol{u} = -\kappa(\nabla p - \boldsymbol{g}) \tag*{(4)}

Here, the first two equations describe the quasi-static force balance in the deformed solid, while the last two equations describe the diffusive Darcy flow of the fluid due to the deformation. The so-called Biot–Willis parameter α∈(0,1]\alpha\in(0,1] represents the interaction strength between solid and fluid ranging from no coupling in the limit α→0\alpha\to0 to strong coupling for α\alpha close to 1. Moreover, λ\lambda and μ\mu denote the first and second Lamé coefficients of the solid, c0c_0 is the specific storage coefficient, and κ\kappa represents the viscosity-reduced permeability. Finally, the right-hand side f\boldsymbol{f} and g\boldsymbol{g} are given volume forces for solid and fluid, respectively, while ss is a volumetric fluid source or sink.

The system has to be closed by appropriate initial and boundary conditions for the primal unknowns η\boldsymbol{\eta} and pp. We omit the details here and will specify the conditions used for the numerical examples in § 4 and their LBM implementations in § 3.6. Note that a derivation of this model by homogenization can be found in [73], while a rigorous mathematical analysis of this problem is presented in [74].

Dimensionless system of equations

For the dimensionless formulation, we introduce a characteristic length LL, time TT, scalar displacement HH and pressure P=c0−1P = c_0^{-1}. With this reference system, the dimensionless quantities are defined by

x^=xL,t^=tT,η^=ηH,σ^=c0LσH,p^=c0p,u^=TuL\hat{\boldsymbol{x}} = \frac{\boldsymbol{x}}{L}, \qquad\hat{t} = \frac{t}{T}, \qquad\hat{\boldsymbol{\eta}} = \frac{\boldsymbol{\eta}}{H}, \qquad\hat{\boldsymbol{\sigma}} = \frac{c_0L\boldsymbol{\sigma}}{H}, \qquad\hat{p} = c_0p, \qquad\hat{\boldsymbol{u}} = \frac{T\boldsymbol{u}}{L}

while the dimensionless derivatives satisfy ∇^=L∇\hat{\nabla}=L\nabla and ∂t^=T∂t\partial_{\hat{t}}=T\partial_t. Accordingly, the scaled time-space cylinder is given by Q^={(t^,x^)∣(t,x)∈Q}\hat{Q}=\{(\hat{t},\hat{x})\mid(t,x)\in Q\}. Inserting the dimensionless quantities into the Biot system Eqs. (1) to (4) yields

−∇^⋅σ^+εα∇^p^=εf^in Q^,(5)-\hat{\nabla}\cdot\hat{\boldsymbol{\sigma}}+\varepsilon\alpha\hat{\nabla}\hat{p}=\varepsilon\hat{\boldsymbol{f}}\qquad\text{in }\hat{Q}, \tag*{(5)}
σ^=λ^(∇^⋅η^)I+μ^(∇^η^+∇^η^T)in Q^,(6)\hat{\boldsymbol{\sigma}}=\hat{\lambda}(\hat{\nabla}\cdot\hat{\boldsymbol{\eta}})I+\hat{\mu}(\hat{\nabla}\hat{\boldsymbol{\eta}}+\hat{\nabla}\hat{\boldsymbol{\eta}}^{\mathsf{T}})\qquad\text{in }\hat{Q}, \tag*{(6)}
∂t^(p^+αε∇^⋅η^)+∇^⋅u^=s^in Q^,(7)\partial_{\hat{t}}\left(\hat{p}+\frac{\alpha}{\varepsilon}\hat{\nabla}\cdot\hat{\boldsymbol{\eta}}\right)+\hat{\nabla}\cdot\hat{\boldsymbol{u}}=\hat{s}\qquad\text{in }\hat{Q}, \tag*{(7)}
u^=−κ^(∇^p^−g^)in Q^,(8)\hat{\boldsymbol{u}}=-\hat{\kappa}(\hat{\nabla}\hat{p}-\hat{\boldsymbol{g}})\qquad\text{in }\hat{Q}, \tag*{(8)}

where the dimensionless parameters are

ε=LH,λ^=c0λ,μ^=c0μ,κ^=Tκc0L2.\varepsilon=\frac{L}{H},\qquad\hat{\lambda}=c_0\lambda,\qquad\hat{\mu}=c_0\mu,\qquad\hat{\kappa}=\frac{T\kappa}{c_0L^2}.

and the dimensionless forces and source are given by

f^=c0Lf,g^=c0Lg,s^=Ts.\hat{\boldsymbol{f}}=c_0L\boldsymbol{f},\qquad\hat{\boldsymbol{g}}=c_0L\boldsymbol{g},\qquad\hat{s}=Ts.

In the following, we drop the hats for readability.

Remark on the incompressible limit cases. For an incompressible fluid (c0=0c_0=0), Biot’s consolidation model is governed by a degenerate parabolic system as the temporal derivative of the pressure vanishes. Then, the above scaling is impossible, but one can use an appropriate scaling, e.g. by replacing c0c_0 by μ\mu for the definition of the dimensionless quantities, leading to well-defined equations. Since the pressure equation is elliptic in that case, standard LBM cannot be used directly, but one also needs a pseudo-timestepping for the pressure equation, similar to the one for linear elasticity presented below. For simplicity, we restrict ourselves to the compressible case (c0>0c_0>0) in this article.

For an incompressible solid (λ→∞\lambda\to\infty), the elastic sub-system must be augmented by the additional constraint ∇⋅η=0\nabla\cdot\boldsymbol{\eta}=0. However, the LBM for linear elasticity presented below becomes unstable in this limit [65]. In the context of finite element and volume methods, remedies such as the total pressure formulation are known and might be applied analogously to the presented LBM. However, this goes beyond the scope of this article, so that we only consider compressible solids from here on.

The lattice Boltzmann method for Biot's consolidation model

In this section, we propose the LBM for Biot’s consolidation model by coupling LBM for the diffusive fluid flow [49] with LBM for two-dimensional linear elasticity [65, 66]. To this end, we introduce notation conventions and basic definitions first and subsequently present the separate LBMs for diffusive flow and elasticity. Then, we discuss the coupling of the two methods, explain the multi-grid method accelerating the pseudo-timestepping, and finally elaborate on the LBM implementation of the used initial and boundary conditions.

Notation and definitions for lattice Boltzmann methods

In the following, we summarize all basics of LBM necessary for this work, for a detailed introduction we refer to [37, 38]. The fundamental principle of LBM is the discretization of the Boltzmann equation in physical and velocity space. For this purpose, the main variable is the distribution function ff, which denotes the probability of material particles at time tt and position xx to have the velocity cc. The velocity space is discretized using a finite set of velocities cic_i and corresponding weights wiw_i, such that fi(t,x)=f(t,x,ci)f_i(t,x)=f(t,x,c_i). These sets are typically derived based on the required mass, momentum and energy conservation using a Hermite series expansion. Typically, the velocity sets are denoted by DdQq\mathrm{D}d\mathrm{Q}q, where dd is the spatial dimension and qq the number of discrete velocities. For the two-dimensional LBMs presented in the following, we utilize the classical D2Q9 velocity set as well as a reduced D2Q8 velocity set without rest velocity c0=0c_0=0, see Figure 1 and Table 1.

Figure 1. D2Q9 velocity set cic_i with linear indices (left) and D2Q8 velocity set cijc_{ij} with 2D Miller indices, i.e., 1ˉ=−1\bar{1}=-1 (right).

012345678
Linear index012345678
Miller index0010011ˉ0\bar{1}001ˉ0\bar{1}111ˉ1\bar{1}11ˉ1ˉ\bar{1}\bar{1}11ˉ1\bar{1}
Velocity(00)\begin{pmatrix}0\\0\end{pmatrix}(10)\begin{pmatrix}1\\0\end{pmatrix}(01)\begin{pmatrix}0\\1\end{pmatrix}(−10)\begin{pmatrix}-1\\0\end{pmatrix}(0−1)\begin{pmatrix}0\\-1\end{pmatrix}(11)\begin{pmatrix}1\\1\end{pmatrix}(−11)\begin{pmatrix}-1\\1\end{pmatrix}(−1−1)\begin{pmatrix}-1\\-1\end{pmatrix}(1−1)\begin{pmatrix}1\\-1\end{pmatrix}
Weight4/94/91/91/91/91/91/91/91/91/91/361/361/361/361/361/361/361/36

Table 1. Discrete velocities cic_i and weights wiw_i with linear and 2D Miller indices, i.e., 1ˉ=−1\bar{1}=-1.

Additionally, the two-dimensional spatial domain D\mathcal{D} is discretized by a uniform lattice of width Δx\Delta x, while the time interval (0,tf)(0,t_f) is discretized by a uniform time step Δt=tf/NT\Delta t=t_f/N_T with NT∈NN_T\in\mathbb{N}. This leads to a discrete lattice speed of sound cs=Δx/(3Δt)c_s=\Delta x/(\sqrt{3}\Delta t). Altogether, the general lattice Boltzmann equation then reads

fi(t+Δt,x+Δtci)−fi(t,x)=Δt Ωi(t,x),(9)f_i(t+\Delta t,\boldsymbol{x}+\Delta t\boldsymbol{c}_i)-f_i(t,\boldsymbol{x})=\Delta t\,\Omega_i(t,\boldsymbol{x}), \tag*{(9)}

where the left-hand side represents an advection along the discrete velocity ci\boldsymbol{c}_i, the so-called streaming step, whereas the right-hand side term, called collision operator, contains all interactions such as relaxation to equilibrium due to particle collision as well as external forces and sources. Note that the lattice Boltzmann equation can be split in two parts: the collision step

fi∗(t,x)=fi(t,x)+Δt Ωi(t,x),(10)f_i^*(t,\boldsymbol{x})=f_i(t,\boldsymbol{x})+\Delta t\,\Omega_i(t,\boldsymbol{x}), \tag*{(10)}

which is local, and the streaming step

fi(t+Δt,x+Δtci)=fi∗(t,x),(11)f_i(t+\Delta t,\boldsymbol{x}+\Delta t\boldsymbol{c}_i)=f_i^*(t,\boldsymbol{x}), \tag*{(11)}

which moves each distribution function one lattice node further in their velocity direction. This is the fundamental reason for the algorithmic simplicity and for the straightforward and well-scaling parallelization.

Even in the absence of external interactions, the collision operator is very complicated as it involves all types of microscopic particle interactions. However, the most simple approximation, which ensures mass and momentum conservation, is given by the Bhatnagar–Gross–Krook collision operator [34]

Ωi=−ω(fi−fieq),(12)\Omega_i=-\omega\left(f_i-f_i^{\mathrm{eq}}\right), \tag*{(12)}

where ω>0\omega>0 is the relaxation rate towards the equilibrium state fieqf_i^{\mathrm{eq}} that depends on the moments of all distribution functions fjf_j according to the equation of interest, e.g., on density ρ=∑ifi\rho=\sum_i f_i and momentum ρu=∑ifici\rho\boldsymbol{u}=\sum_i f_i\boldsymbol{c}_i in case of the Navier–Stokes equations. Note that here and in the following, the arguments (t,x)(t,\boldsymbol{x}) are omitted whenever the meaning is unambiguous. The Bhatnagar–Gross–Krook collision operator will be used below for the diffusive fluid flow. In this application, the relaxation rate is directly related to the diffusion coefficient.

In contrast to this single-relaxation-time scheme, multiple-relaxation-time (MRT) collision operators allow for a larger number of free parameters, increased stability and higher accuracy. This is necessary to formulate a LBM for linear elasticity. For MRT schemes, the distribution functions are transformed into linear independent moments, which are individually relaxed at different rates and then transformed back to the distributions [35]. Denoting the vector of the distribution functions for all discrete velocities by f=(f0,…,f8)T\boldsymbol{f}=(f_0,\ldots,f_8)^{\mathrm{T}}, the transformation into moments can be described by multiplication by an invertible matrix M\mathbf{M}. Analogously, we represent the individual relaxation rates as a diagonal matrix ω\boldsymbol{\omega} and the equilibrium moment vector by meq\mathbf{m}^{\mathrm{eq}}. The vector of MRT collision operators for all discrete velocities is then given by

Ω=−M−1ω(Mf−meq).(13)\boldsymbol{\Omega}=-\mathbf{M}^{-1}\boldsymbol{\omega}\left(\mathbf{M}\mathbf{f}-\mathbf{m}^{\mathrm{eq}}\right). \tag*{(13)}

The specific moments and relaxation rates depend again on the parameters of the considered equations. More details will be given below during the discussion of LBM for linear elasticity. Finally, note that the Bhatnagar–Gross–Krook collision operator is a special case of the MRT one when ω=ωI\boldsymbol{\omega}=\omega\mathbf{I} is chosen.

The lattice Boltzmann method for diffusive fluid flow

In the following, we present the required steps for the LBM solving the diffusive fluid flow Eqs. (7) and (8). Combining these equations and reordering the terms, we obtain the diffusion-reaction equation

∂tp−κΔp=seff=s−κ∇⋅g−αε∂t∇⋅ηin Q.(14)\partial_t p-\kappa\Delta p=s_{\mathrm{eff}}=s-\kappa\nabla\cdot\mathbf{g}-\frac{\alpha}{\varepsilon}\partial_t\nabla\cdot\boldsymbol{\eta}\qquad\text{in }Q. \tag*{(14)}

Subsequently, the LBM for the diffusion-reaction equation proposed in [49] is discussed, while the detailed computation of the effective source term seffs_{\mathrm{eff}} is deferred to § 3.4. Following the presentation in [37 Chpt. 8], this LBM is based on the single-relaxation-time scheme Eqs. (9) and (12) and uses the D2Q9 velocity set. Denoting the distribution functions for the pressure by fif_i with linear index i∈{0,1,…,8}i\in\{0,1,\ldots,8\}, the pressure pp is defined by the zeroth moment of the distribution functions corrected for the effective source seffs_{\mathrm{eff}} as

p=∑i=08fi+Δt2seff.(15)p=\sum_{i=0}^{8}f_i+\frac{\Delta t}{2}s_{\mathrm{eff}}. \tag*{(15)}

Since Eq. (14) is purely diffusive, the equilibrium distribution functions are defined by

fieq=wip.(16)f_i^{\mathrm{eq}}=w_i p. \tag*{(16)}

The single-relaxation-time collision step combining Eqs. (10) and (12) then reads

fi∗=fi−ωΔt(fi−fieq)+ΔtSi,(17)f_i^{*}=f_i-\omega\Delta t\left(f_i-f_i^{\mathrm{eq}}\right)+\Delta t S_i, \tag*{(17)}

where the contribution by the effective source is given by

Si=(1−ω2)wiseff.(18)S_i=\left(1-\frac{\omega}{2}\right)w_i s_{\mathrm{eff}}. \tag*{(18)}

Furthermore, the relaxation rate ω\omega must satisfy ω∈(0,2)\omega\in(0,2) for stability and is related to the permeability via

κ=cs2(1ω−Δt2).(19)\kappa=c_s^2\left(\frac{1}{\omega}-\frac{\Delta t}{2}\right). \tag*{(19)}

Note that this is the standard relation between relaxation rate ω\omega and diffusion coefficient κ\kappa in LBM. In the end of each time step, the streaming step of Eq. (11) is performed. Overall, this LBM yields a pressure solution that is second-order consistent in time and space [49].

The lattice Boltzmann method for quasi-static linear elasticity

Next, we present the necessary steps for the LBM solving the quasi-static linear elasticity Eqs. (5) and (6). The method has been developed in [65, 66] and uses a MRT scheme for the D2Q8 velocity set. Here, we adapt the description to the context of Biot’s consolidation model, where an additional parameter-like time-dependence emerges. In particular, the method solves Eqs. (5) and (6) using an pseudo-timestepping method, where Eq. (5) is replaced by

1ε∂τη−∇⋅σ=εfeff=ε(f−α∇p)for (τ,t,x)∈(0,∞)×Q.(20)\frac{1}{\varepsilon}\partial_\tau\boldsymbol{\eta}-\nabla\cdot\boldsymbol{\sigma}=\varepsilon\mathbf{f}_{\mathrm{eff}}=\varepsilon\left(\mathbf{f}-\alpha\nabla p\right)\qquad\text{for }(\tau,t,\mathbf{x})\in(0,\infty)\times Q. \tag*{(20)}

which is equivalent when the stationary solution is reached (for τ→∞\tau\to\infty). Note that the division by ε\varepsilon in front of the pseudo-time derivative is a result of the viscous scaling leading to a pseudo-time step Δτ=1\Delta\tau= 1. As before, we defer the computation of the effective force fefff_{\mathrm{eff}} to § 3.4, and elaborate on the general method performed within each time step for multiple pseudo-time steps.

The moments for the MRT scheme are linear combinations of the raw countable moments [36]

mab=∑i=−11∑j=−11iajbgij,(21)m_{ab} = \sum_{i=-1}^{1} \sum_{j=-1}^{1} i^{a}j^{b}g_{ij}, \tag*{(21)}

using the two-dimensional Miller index ab∈{10,01,11,20,02,12,21,22}ab \in\{10,01,11,20,02,12,21,22\}, cf. also Table 3 for the inverse transformation. Note that the number of independent moments coincides with the size of the velocity set, i.e., there is no independent zeroth moment m00=∑i,jgijm_{00} = \sum_{i,j}g_{ij} due to the absence of the rest velocity c00\mathbf{c}_{00}. Besides the raw moments mabm_{ab} with ab∈{10,01,11,12,21}ab \in\{10,01,11,12,21\}, we use the combined second-order moments

ij\νij \backslash\nu010110101111ssdd121221212222
101012\frac{1}{2}000014\frac{1}{4}14\frac{1}{4}−12-\frac{1}{2}00−12-\frac{1}{2}
01010012\frac{1}{2}0014\frac{1}{4}−14-\frac{1}{4}00−12-\frac{1}{2}−12-\frac{1}{2}
1ˉ0\bar{1}0−12-\frac{1}{2}000014\frac{1}{4}14\frac{1}{4}12\frac{1}{2}00−12-\frac{1}{2}
01ˉ0\bar{1}00−12-\frac{1}{2}0014\frac{1}{4}−14-\frac{1}{4}0012\frac{1}{2}−12-\frac{1}{2}
1111000014\frac{1}{4}000014\frac{1}{4}14\frac{1}{4}14\frac{1}{4}
1ˉ1\bar{1}10000−14-\frac{1}{4}0000−14-\frac{1}{4}14\frac{1}{4}14\frac{1}{4}
11ˉ1\bar{1}000014\frac{1}{4}0000−14-\frac{1}{4}−14-\frac{1}{4}14\frac{1}{4}
1ˉ1ˉ\bar{1}\bar{1}0000−14-\frac{1}{4}000014\frac{1}{4}−14-\frac{1}{4}14\frac{1}{4}

Table 3. Coefficients cijνc_{ij}^{\nu} of the back-transformation from moments mij∗m_{ij}^{*} to distribution functions gij∗g_{ij}^{*} using 2D Miller indices, i.e., 1ˉ=−1\bar{1}=-1.

ms=m20+m02,md=m20−m02,(22)m_s = m_{20} + m_{02}, \qquad m_d = m_{20} - m_{02}, \tag*{(22)}

which are directly related to the volumetric and deviatoric components of the stress. Additionally, the raw fourth-order moment m22m_{22} is replaced by

mf=m22+112(λ+μ)−4ms,(23)m_f = m_{22} + \frac{1}{12(\lambda+\mu)-4}m_s, \tag*{(23)}

which is required for imposing Neumann boundary conditions [66], see § 3.6.

Interpreting the LBM as a Strang splitting scheme, the effective force is split into two sub-steps of the collision step [75], and then applied as follows

\begin{pmatrix}\bar{m}_{10}\\\bar{m}_{01}\end{pmatrix} = \begin{pmatrix}m_{10}\\m_{01}\end{pmatrix} + \frac{\varepsilon^2}{2}f_{\mathrm{eff}}, \qquad \begin{pmatrix}m_{10}^{*}\\m_{01}^{*}\end{pmatrix} = \begin{pmatrix}\bar{m}_{10}\\\bar{m}_{01}\end{pmatrix} + \frac{\varepsilon^2}{2}f_{\mathrm{eff}}. \tag*{(24)}\end{pmatrix}

Note that this reflects a conservation of the first moments, since no relaxation is applied (ω10=ω01=0\omega_{10} = \omega_{01} = 0). All other moments mνm_\nu, ν∈{s,d,11,12,21,f}\nu\in\{s,d,11,12,21,f\}, are relaxed according to the MRT scheme Eq. (13), viz.,

mν∗=mν−ων(mν−mνeq),(25)m_\nu^{*} = m_\nu- \omega_\nu(m_\nu- m_\nu^{\mathrm{eq}}), \tag*{(25)}

where the relaxation rates ων\omega_\nu and equilibrium moments mνeqm_\nu^{\mathrm{eq}} are given in Table 2. In particular, note that the results after half the collision step are denoted by mˉν=(mν+mν∗)/2\bar{m}_\nu= (m_\nu+ m_\nu^{*})/2.

Index ν\nu1001ssdd111221ff
Relaxation rate ων\omega_\nu002/(3(λ+μ)+1)2/(3(\lambda+\mu)+1)2/(6μ+1)2/(6\mu+1)2/(6μ+1)2/(6\mu+1)111
Equilibrium moment mνeqm_\nu^{\mathrm{eq}}−-−-000mˉ10/3\bar{m}_{10}/3mˉ01/3\bar{m}_{01}/30

Table 2. Relaxation rates ων\omega_\nu and equilibrium moments mνeqm_\nu^{\mathrm{eq}} for quasi-static linear elasticity using 2D Miller indices.

Furthermore, the relaxation rates ωd\omega_d and ω11\omega_{11} are the same in order to obtain isotropic behavior of the deviatoric stress. Indeed, it has been shown in [66] by asymptotic expansion that the displacement η\eta and the stress σ\boldsymbol{\sigma} are given in the limit τ→∞\tau\to\infty by

η=(mˉ10mˉ01),σ=−12(mˉs+mˉd2mˉ112mˉ11mˉs−mˉd),(26)\eta= \begin{pmatrix}\bar{m}_{10}\\\bar{m}_{01}\end{pmatrix}, \qquad \boldsymbol{\sigma} = -\frac{1}{2} \begin{pmatrix} \bar{m}_s+\bar{m}_d & 2\bar{m}_{11}\\ 2\bar{m}_{11} & \bar{m}_s-\bar{m}_d \end{pmatrix}, \tag*{(26)}

respectively, where the involved moments after half the collision can be computed as

mˉs=3(λ+μ)3(λ+μ)+1ms,mˉd=6μ6μ+1md,mˉ11=6μ6μ+1m11.(27)\bar{m}_s = \frac{3(\lambda+\mu)}{3(\lambda+\mu)+1}m_s, \qquad \bar{m}_d = \frac{6\mu}{6\mu+1}m_d, \qquad \bar{m}_{11} = \frac{6\mu}{6\mu+1}m_{11}. \tag*{(27)}

The post-collision distribution functions are obtained from the back-transformation of the post-collision moments by inverting Eqs. (21) to (23). First, note that this leads to

m22∗=−112(λ+μ)−4ms∗=−112(λ+μ)+4ms,(28)m_{22}^{*}=-\frac{1}{12(\lambda+\mu)-4}m_{s}^{*}=-\frac{1}{12(\lambda+\mu)+4}m_{s}, \tag*{(28)}

which is non-singular even for λ+μ=1/3\lambda+\mu=1/3 (alternatively, in this case, the relaxation rate ωf\omega_f can be changed without influencing the leading-order behavior [66]). Then, the post-collision distribution functions are given by

gij∗=∑νcijνmν∗,(29)g_{ij}^{*}=\sum_{\nu}c_{ij}^{\nu}m_{\nu}^{*}, \tag*{(29)}

where the coefficients cijνc_{ij}^{\nu} are stated in Table 3. In the end of each pseudo-time step, the streaming step of Eq. (11) is performed for the distribution functions gijg_{ij}. Overall, this LBM yields displacement and stress solutions which are second-order consistent in space for pseudo time τ→∞\tau\to\infty (in every time step) [65].

The coupling of the lattice Boltzmann methods

To obtain the LBM for Biot’s consolidation model, we couple the above discussed LBMs for diffusive flow and for quasi-static linear elasticity. To this end, in each time step we solve the quasi-static elasticity problem using NE∈NN_E\in\mathbb{N} pseudo-time steps, followed by one regular update of the LBM for diffusive flow.

The effective force term feff=f−α∇pf_{\mathrm{eff}}=f-\alpha\nabla p in the elasticity Eq. (20) involves the gradient of the pressure. Following [76, 77], this gradient can be approximated up to second-order errors in space by the central difference quotient

∇cp(x)=∑i=08wici(p(x+Δtci)−p(x−Δtci))2cs2Δt=∑i=08wicip(x+Δtci)cs2Δt.(30)\nabla^{c}p(x)=\sum_{i=0}^{8}\frac{w_i c_i\left(p(x+\Delta t c_i)-p(x-\Delta t c_i)\right)}{2c_s^2\Delta t}=\sum_{i=0}^{8}\frac{w_i c_i p(x+\Delta t c_i)}{c_s^2\Delta t}. \tag*{(30)}

Note that the second equality only holds for symmetric velocity sets, such as the D2Q9 used here. Alternatively, one can use the first moments of the pressure distribution functions fif_i to obtain the pressure gradient [37 Chpt. 8]. Accounting for the contribution of the source term, this yields

∇∗p=−ωcs2∑i=08fici+12∇cseff.(31)\nabla^{*}p=-\frac{\omega}{c_s^2}\sum_{i=0}^{8}f_i c_i+\frac{1}{2}\nabla^{c}s_{\mathrm{eff}}. \tag*{(31)}

For the numerical examples below, we observe the same order of convergence for both methods, but considerably smaller errors when using Eq. (31), so that we only present results of the latter in the following.

Algorithm 1 Semi-implicit LBM for poroelasticity.

  1. Set t=0t = 0 and initialize the distribution functions fif_i and gijg_{ij} ▷ Eqs. (40) and (41)

  2. for i=1,…,NTi = 1, \ldots, N_T do

  3. ▷ Elasticity update ◁

  4. Set τ=0\tau= 0 and compute the explicit part of seffim(t)s_{\mathrm{eff}}^{\mathrm{im}}(t) ▷ Eq. (34)

  5. for j=1,…,NEj = 1, \ldots, N_E do

  6. Compute seffim(τ,t)s_{\mathrm{eff}}^{\mathrm{im}}(\tau,t) and p(τ,t)p(\tau,t) ▷ Eqs. (15) and (34)

  7. Compute feff(τ,t)f_{\mathrm{eff}}(\tau,t) from ∇p(τ,t)\nabla p(\tau,t) ▷ Eq. (31)

  8. Compute gij(τ+Δτ,t)=ELASTIC_STEP⁡(gij(τ,t),feff(τ,t))g_{ij}(\tau+ \Delta\tau,t) = \operatorname{ELASTIC\_STEP}(g_{ij}(\tau,t), f_{\mathrm{eff}}(\tau,t)) ▷ Algorithm 2

  9. Compute ∇⋅η(τ+Δτ,t)\nabla\cdot\boldsymbol{\eta}(\tau+ \Delta\tau,t) ▷ Eqs. (27) and (33)

  10. Advance pseudo time τ\tau by Δτ\Delta\tau

  11. Compute η(t)\boldsymbol{\eta}(t) and σ(t)\boldsymbol{\sigma}(t) ▷ Eqs. (26) and (27)

  12. ▷ Flow update ◁

  13. Compute seffim(t)s_{\mathrm{eff}}^{\mathrm{im}}(t) and p(τ,t)p(\tau,t) ▷ Eqs. (15) and (34)

  14. Compute fi(t+Δt)=FLOW_STEP⁡(fi(t),seffim(t))f_i(t + \Delta t) = \operatorname{FLOW\_STEP}(f_i(t),s_{\mathrm{eff}}^{\mathrm{im}}(t)) ▷ Algorithm 3

  15. Advance time tt by Δt\Delta t

Irrespective of the used approximation, the gradient depends on the effective source term seff=s−κ∇⋅g−αε∂t∇⋅ηs_{\mathrm{eff}} = s - \kappa\nabla\cdot\boldsymbol{g} - \frac{\alpha}{\varepsilon}\partial_t\nabla\cdot\boldsymbol{\eta} in the flow Eq. (14). While the first and second right-hand side terms are given explicitly, the last one does depend again on the solution of the elastic equation. To eliminate this cyclic self-dependence, one can use a fully explicit scheme

seffex(t)=s(t)−κ∇⋅g(t)−αε∂t∇⋅η(t−Δt)(32)s_{\mathrm{eff}}^{\mathrm{ex}}(t) = s(t) - \kappa\nabla\cdot\boldsymbol{g}(t) - \frac{\alpha}{\varepsilon}\partial_t\nabla\cdot\boldsymbol{\eta}(t-\Delta t) \tag*{(32)}

where ∂tf(t)=(f(t)−f(t−Δt))/Δt\partial_t f(t) = (f(t) - f(t-\Delta t))/\Delta t denotes the temporal backward difference, while the divergence of the displacement is given according to Eqs. (2) and (26) by

∇⋅η=−12(λ+μ)mˉs.(33)\nabla\cdot\boldsymbol{\eta} = -\frac{1}{2(\lambda+\mu)}\bar{m}s. \tag*{(33)}

However, this scheme is unstable for α\alpha close to 1, as shown numerically in § 4.1. Instead, we propose a semi-implicit scheme that utilities the elasticity solution of the current physical time tt, but at the previous pseudo-time step τ−Δτ\tau-\Delta\tau (which is iterated), viz.

seffim(τ,t)=s(t)−κ∇⋅g(t)−αε((1−r)∂t∇⋅η(t−Δt)+r∂t∇⋅η(τ−Δτ,t))(34)s_{\mathrm{eff}}^{\mathrm{im}}(\tau,t) = s(t) - \kappa\nabla\cdot\boldsymbol{g}(t) - \frac{\alpha}{\varepsilon}\left((1-r)\partial_t\nabla\cdot\boldsymbol{\eta}(t-\Delta t) + r\partial_t\nabla\cdot\boldsymbol{\eta}(\tau-\Delta\tau,t)\right) \tag*{(34)}

where r∈[0,1]r \in[0,1] is the ratio between explicit and semi-implicit approximation. Note that only the rightmost term needs to be recomputed in each pseudo-time step, while the others are computed at the beginning of each time step. Therefore, each time step is initialized by setting gij(−1,t)=gij(0,t)=gij(NE,t−Δt)g_{ij}(-1,t) = g_{ij}(0,t) = g_{ij}(N_E,t-\Delta t), and after the NEN_E steps one sets gij(t)=gij(NE,t)g_{ij}(t) = g_{ij}(N_E,t) and ∇⋅η(t)=∇⋅η(NE,t)\nabla\cdot\boldsymbol{\eta}(t) = \nabla\cdot\boldsymbol{\eta}(N_E,t). Altogether, one obtains Algorithm 1 based on the Algorithms 2 and 3 for elasticity and flow updates. While the discussion above only considers the interior of the domain, the algorithms already include the application of initial and boundary conditions, which will be discussed in detail in § 3.6.

Note that an adaptive version of this algorithm using a variable number of pseudo-time steps NEN_E might be preferable for solutions approaching a stationary state. From an algorithmic point of view, such a change is trivial, as one simply replaces Algorithm 1 by a while loop using the stopping criterion. However from the mathematical theory, it remains unclear, what would be an appropriate stopping criterion. Without known convergence order for the elastic sub-system, a simple criterion using the difference between consecutive iterations could easily fail. Therefore, we restrict ourselves here to the study of the algorithm with fixed number of pseudo-time steps.

However, the underlying elasticity problem is elliptic, and thus can be efficiently treated by multi-grid methods, see e.g. [68, 72]. In the following section, we propose such a method for the elasticity LBM discussed above. Thereby, the number of pseudo-time steps becomes essentially independent of the spatial discretization yielding an quasi-optimal algorithm.

Algorithm 2 LBM time step for quasi-static linear elasticity.

  1. function ELASTIC_STEP(gijg_{ij}, fefff_{\mathrm{eff}})

  2. Compute the moments mνm_{\nu} from gijg_{ij} ▷ Eqs. (21) to (23)

  3. Apply half-forcing (first half of Strang splitting) ▷ Eq. (24)

  4. Compute the equilibrium moments mνeqm_{\nu}^{\mathrm{eq}} ▷ Table 2

  5. Apply half-forcing (second half of Strang splitting) ▷ Eq. (24)

  6. Compute the post-collision moments mν∗m_{\nu}^{*} ▷ Eq. (25) and Table 2

  7. Back-transform to distributions gij∗g_{ij}^{*} from mν∗m_{\nu}^{*} ▷ Eqs. (28) and (29)

  8. Pseudo-time streaming of the distributions: gij=STREAM⁡(gij∗)g_{ij} = \operatorname{STREAM}(g_{ij}^{*}) ▷ Eq. (11)

  9. Apply the (nontrivial) boundary conditions for linear elasticity ▷ Implementation details in § 3.6

  10. return gijg_{ij}

Algorithm 3 LBM time step for diffusive fluid flow.

  1. function FLOW_STEP(fif_i, seffs_{\mathrm{eff}})

  2. Compute the equilibrium distributions fieqf_i^{\mathrm{eq}} ▷ Eqs. (15) and (16)

  3. Compute the post-collision distributions fi∗f_i^{*} ▷ Eqs. (17) and (18)

  4. Streaming of the distributions: fi=STREAM⁡(fi∗)f_i = \operatorname{STREAM}(f_i^{*}) ▷ Eq. (11)

  5. Apply the (nontrivial) boundary conditions for the pressure ▷ Implementation details in § 3.6

  6. return fif_i

Furthermore, with respect to the dimensionless formulation discussed in § 2.2, note that in the LBM setting one typically chooses the characteristic scales T∼ΔtphysT \sim\Delta t_{\mathrm{phys}} and L∼ΔxphysL \sim\Delta x_{\mathrm{phys}} to obtain step sizes ΔtLBM=ΔxLBM=1\Delta t_{\mathrm{LBM}} = \Delta x_{\mathrm{LBM}} = 1. Together with a fixed H∼1H \sim1, this also yields a small length ratio ε∼Δxphys\varepsilon\sim\Delta x_{\mathrm{phys}} required for the convergence of the LBM for elasticity [65]. Finally, using the viscous scaling Δtphys∼(Δxphys)2\Delta t_{\mathrm{phys}} \sim(\Delta x_{\mathrm{phys}})^2 yields constant material parameters, but decreasing forces and source.

Acceleration by a multi-grid method

To accelerate the convergence of the LBM for quasi-static linear elasticity, we use a multi-grid method, as discussed in the following. Note that when the steady-state solution is reached, applying one LBM step results in the same populations. Thus, the residual of the LBM for quasi-static linear elasticity is given by

rij=gij−gij∗∗,(35)r_{ij} = g_{ij} - g_{ij}^{**}, \tag*{(35)}

where gij∗∗g_{ij}^{**} is obtained by applying one pseudo-time step (Algorithm 2), viz.

gij∗∗(x+Δτcij)=gij(x)+ΔτΩij(x),(36)g_{ij}^{**}(x + \Delta\tau c_{ij}) = g_{ij}(x) + \Delta\tau\Omega_{ij}(x), \tag*{(36)}

where Ωij\Omega_{ij} denotes the MRT collision operator for elasticity discussed in § 3.3. Hence, our goal is a scheme that rapidly yields rij→0r_{ij} \to0. This is achieved using the multi-grid method via 3 steps: First, one applies a smoother to reduce the high-frequency errors, then one updates the solution with the so-called defect correction on a coarser grid, and finally applies the smoother again to remove high-frequency errors reintroduced by the update. In the following, we first discuss the use of a under-relaxed LBM step as smoother, and afterwards the linear defect correction including the transfer between grids. Note that we apply a standard linear multi-grid method as the considered elasticity equation is linear.

As realized in previous work, see e.g. [68, 72], the application of one LBM step is very similar to one step of the Jacobi iteration used for solving large equation systems. In particular, the convergence of both methods deteriorates for elliptic problems of increasing size. However, it is a suitable smoother when used as under-relaxation with parameter γ∈(0,1)\gamma\in(0, 1), i.e.,

gij(k+1)=(1−γ)gij(k)+γgij(k),∗∗,(37)g_{ij}^{(k+1)} = (1-\gamma)g_{ij}^{(k)} + \gamma g_{ij}^{(k),**}, \tag*{(37)}

where we apply K∈NK \in\mathbb{N} such under-relaxed LBM steps before and after the defect correction. Hence, we start from gij(0)=gij(τ)g_{ij}^{(0)} = g_{ij}(\tau), use Eq. (37) for k=0,…,K−1k = 0, \ldots, K - 1, apply the defect correction to obtain gij(K+1)g_{ij}^{(K+1)}, use Eq. (37) for k=K+1,…,2Kk = K + 1, \ldots, 2K, and finally obtain gij(τ+Δτ)=gij(2K+1)g_{ij}(\tau+ \Delta\tau) = g_{ij}^{(2K+1)}.

For the linear defect correction on a coarser grid, we simply double the lattice width Δx\Delta x to obtain the next coarser one. As we use a cell-centered LBM, each cell of the coarse grid coincides with four cells of the finer one, see Figure 2. First, one computes the residual rij(K)r_{ij}^{(K)} on the current grid using Eq. (35). To transfer the residual to the coarse grid, we use a simple (full) summation as restriction operator R\mathcal{R}, i.e., the values of the four fine cells are added to obtain the value of the corresponding coarse cell, or in stencil notation:

Figure 2. Sketch of a 4 × 4 fine grid (dotted lines) with cell centers (light gray circles) and the corresponding 2 × 2 coarse grid (solid lines) with cell centers (dark gray circles) as used in the multi-grid LBM.

R=[1111].(38)\mathcal{R} = \begin{bmatrix} 1 & 1 \\ 1 & 1 \end{bmatrix}. \tag*{(38)}

This yields the coarse-grid residual rijc=Rrij(K)r_{ij}^{\mathrm{c}} = \mathcal{R}r_{ij}^{(K)}. Vice versa, the coarse-correction values gijcg_{ij}^{\mathrm{c}} are transferred to the current grid via the (full) prolongation P=RT\mathcal{P} = \mathcal{R}^{\mathsf{T}}, i.e., the correction value of all 4 fine cells is set to the value of the corresponding coarse cell. To obtain the coarse-grid correction, one again has to solve Eqs. (35) and (36) on the coarse grid, but with a different collision operator. On the coarse grid, the contribution of the right-hand side force feff\mathbf{f}_{\mathrm{eff}} in the moment update Eq. (24) is removed up to the self-dependence f∗\mathbf{f}_{*} through ∇p\nabla p, viz.

f∗(gij)=α22εΔt∇σ(∇⋅η(gij)),f~eff=feff−f∗,(39)\mathbf{f}_{*}(g_{ij}) = \frac{\alpha^{2}}{2\varepsilon\Delta t}\nabla^{\sigma}(\nabla\cdot\boldsymbol{\eta}(g_{ij})), \qquad\widetilde{\mathbf{f}}_{\mathrm{eff}} = \mathbf{f}_{\mathrm{eff}} - \mathbf{f}_{*}, \tag*{(39)}

where f~eff\widetilde{\mathbf{f}}_{\mathrm{eff}} contains only explicit contributions, i.e., only depends on values of the previous time steps. Moreover, the pre-collision distributions gijg_{ij} must be reduced by the coarse-grid residual rijcr_{ij}^{\mathrm{c}}, leading to Algorithm 4 of Algorithm 4. Finally, the boundary conditions on the coarse grid are the homogeneous versions of those applied on the fine grid. To solve for the coarse-grid correction, one can thus apply the multi-grid method recursively for L∈NL \in\mathbb{N} levels with adapted arguments.

In summary, this leads to Algorithm 4 for the multi-grid LBM method for linear elasticity. To include it into the overall poroelasticity LBM, lines 7 and 8 of Algorithm 1 must be adapted, leading to Algorithm 5.

Initial and boundary conditions

Next, we address the LBM implementation of the used initial and boundary conditions. In particular, we discuss periodic, Dirichlet and Neumann boundary conditions for diffusive flow and elasticity. Since all considered domains are axis-aligned and rectangular, we chose a cell-centered lattice, such that it is sufficient to use simple boundary schemes of (anti-)bounce-back type. For more general boundary conditions and complex boundary shapes, we refer to [37 Sec. 8.5] for diffusive flow and [66] for linear elasticity.

Initial conditions. The initial conditions for all numerical examples given below are

p∣t=0=0,η∣t=0=0.p\big|_{t=0} = 0, \qquad\boldsymbol{\eta}\big|_{t=0} = \mathbf{0}.

Algorithm 4 LBM multi-grid step for quasi-static linear elasticity.

  1. function ELASTIC_MULTIGRID(gijg_{ij}, rijr_{ij}, f~eff\tilde{f}_{\mathrm{eff}}, f∗f_*, ℓ\ell, KK, γ\gamma)

  2. ▷ Pre-smoothing ◁

  3. for k=1,…,Kk = 1, \ldots, K do

  4. Compute gij∗∗g_{ij}^{**} = ELASTIC_STEP(gij−rijg_{ij} - r_{ij}, f~eff+f∗(gij)\tilde{f}_{\mathrm{eff}} + f_*(g_{ij})) ▷ Algorithm 2

  5. Update gij=(1−γ)gij+γgij∗∗g_{ij} = (1 - \gamma)g_{ij} + \gamma g_{ij}^{**} ▷ Eq. (37)

  6. ▷ Recursive defect correction ◁

  7. if ℓ>1\ell> 1 then

  8. Restrict the residual rijc=R(gij−r_{ij}^{c} = \mathcal{R}(g_{ij} - ELASTIC_STEP(gij−rijg_{ij} - r_{ij}, f~eff+f∗(gij))\tilde{f}_{\mathrm{eff}} + f_*(g_{ij}))) ▷ Eq. (38) and Algorithm 2

  9. Compute gijcg_{ij}^{c} = ELASTIC_MULTIGRID(00, rijcr_{ij}^{c}, 00, f∗f_*, ℓ−1\ell- 1, KK, γ\gamma)

  10. Prolongate and update the solution gij=gij+P(gijc)g_{ij} = g_{ij} + \mathcal{P}(g_{ij}^{c}) ▷ Eq. (38)

  11. ▷ Post-smoothing ◁

  12. for k=1,…,Kk = 1, \ldots, K do

  13. Compute gij∗∗g_{ij}^{**} = ELASTIC_STEP(gij−rijg_{ij} - r_{ij}, f~eff+f∗(gij)\tilde{f}_{\mathrm{eff}} + f_*(g_{ij})) ▷ Algorithm 2

  14. Update gij=(1−γ)gij+γgij∗∗g_{ij} = (1 - \gamma)g_{ij} + \gamma g_{ij}^{**} ▷ Eq. (37)

  15. return gijg_{ij}

Algorithm 5 Semi-implicit LBM with multi-grid for poroelasticity.

  1. Set t=0t = 0 and initialize the distribution functions fif_i and gijg_{ij} ▷ Eqs. (40) and (41)

  2. for i=1,…,NTi = 1, \ldots, N_T do

  3. ▷ Elasticity update ◁

  4. Set τ=0\tau= 0 and compute the explicit part of seffim(t)s_{\mathrm{eff}}^{\mathrm{im}}(t) ▷ Eq. (34)

  5. for j=1,…,NEj = 1, \ldots, N_E do

  6. Compute seffim(τ,t)s_{\mathrm{eff}}^{\mathrm{im}}(\tau,t) and p(τ,t)p(\tau,t) ▷ Eqs. (15) and (34)

  7. Compute the forces f~eff(τ,t)\tilde{f}_{\mathrm{eff}}(\tau,t) and f∗f_* using ∇p(τ,t)\nabla p(\tau,t) ▷ Eqs. (31) and (39)

  8. Compute gij(τ+Δτ,t)g_{ij}(\tau+ \Delta\tau,t) = ELASTIC_MULTIGRID(gij(τ,t)g_{ij}(\tau,t), 00, f~eff(τ,t)\tilde{f}_{\mathrm{eff}}(\tau,t), f∗f_*, LL, KK, γ\gamma) ▷ Algorithm 4

  9. Compute ∇⋅η(τ+Δτ,t)\nabla\cdot\eta(\tau+ \Delta\tau,t) ▷ Eqs. (27) and (33)

  10. Advance pseudo time τ\tau by Δτ\Delta\tau

  11. Compute η(t)\eta(t) and σ(t)\sigma(t) ▷ Eq. (26)

  12. ▷ Flow update ◁

  13. Compute seffim(t)s_{\mathrm{eff}}^{\mathrm{im}}(t) and p(τ,t)p(\tau,t) ▷ Eqs. (15) and (34)

  14. Compute fi(t+Δt)f_i(t + \Delta t) = FLOW_STEP(fi(t)f_i(t), seffim(t)s_{\mathrm{eff}}^{\mathrm{im}}(t)) ▷ Algorithm 3

  15. Advance time tt by Δt\Delta t

However, to maintain second-order convergence, the initial conditions in the LBM must be chosen such that the physical initial conditions are satisfied in Eqs. (15) and (26) at t=0t=0 [37, 75]. Hence, the initial distribution functions for pressure and displacement are

fi=−Δt2wi(s−κ∇⋅g)∣t=0,gij=∑νcijνmν0.(40)f_i=-\frac{\Delta t}{2}w_i\left(s-\kappa\nabla\cdot\boldsymbol{g}\right)\bigg|_{t=0},\qquad g_{ij}=\sum_{\nu}c_{ij}^{\nu}m_{\nu}^{0}. \tag*{(40)}

Here, the effective source term at time t=0t=0 disregards the unknown volumetric source −αε∂t∇⋅η-\frac{\alpha}{\varepsilon}\partial_t\nabla\cdot\boldsymbol{\eta}. Moreover, the initial moments at time t=0t=0 are the force-corrected equilibrium moments

\begin{pmatrix}m_{10}^{0}\\m_{01}^{0}\end{pmatrix}=-\frac{\varepsilon^{2}}{2}f_{\mathrm{eff}}\bigg|_{t=0},\qquad m_s^{0}=m_d^{0}=m_{11}^{0}=0,\qquad\begin{pmatrix}m_{12}^{0}\\m_{21}^{0}\end{pmatrix}=-\frac{\varepsilon^{2}}{6}f_{\mathrm{eff}}\bigg|_{t=0},\qquad m_{22}^{0}=0, \tag*{(41)}\end{pmatrix}

where the effective force at time t=0t=0 contains only known terms since feff∣t=0=(f−α∇p)∣t=0f_{\mathrm{eff}}\big|_{t=0}=(f-\alpha\nabla p)\big|_{t=0}. Note that instead of explicitly initializing the distribution functions gijg_{ij} for elasticity, one could just use more iterations in the first time step to achieve the stationary solution.

Periodic boundary conditions. Periodic boundary conditions on an open and connected subset of the boundary Γper⊂∂D\Gamma_{\mathrm{per}}\subset\partial\mathcal{D} describe solutions repeating infinitely in space by using the equality of the solution at opposite boundaries of the finite domain D=(0,a)×(0,b)\mathcal{D}=(0,a)\times(0,b), a,b∈Ra,b\in\mathbb{R}, i.e.,

p∣x1=0=p∣x1=a,p∣x2=0=p∣x2=b,η∣x1=0=η∣x1=a,η∣x2=0=η∣x2=b,on (0,tf)×Γper.p\big|_{x_1=0}=p\big|_{x_1=a},\qquad p\big|_{x_2=0}=p\big|_{x_2=b},\qquad\boldsymbol{\eta}\big|_{x_1=0}=\boldsymbol{\eta}\big|_{x_1=a},\qquad\boldsymbol{\eta}\big|_{x_2=0}=\boldsymbol{\eta}\big|_{x_2=b},\qquad\text{on }(0,t_f)\times\Gamma_{\mathrm{per}}.

This requires that distribution functions leaving the domain on one side, simultaneously, re-enter at the opposite side. Since the lattice is cell-centered, this can be realized by simply copying the outgoing distribution functions during streaming to the respective cells on the opposite side. Furthermore, this scheme maintains the convergence order of the underlying LBM. Note that this scheme is automatically included in the streaming step, when a rolling shift implementation is used, e.g., by the function roll in the libraries numpy or jax of Python.

In case of other boundary conditions than periodic ones, there are missing distribution functions after the streaming which would come from outside the domain to the boundary nodes, see Figure 3 for a sketch at the bottom boundary. These missing distributions can be reconstructed from the post-collision distribution functions taking into account the respective boundary condition. To this end, the basic idea of bounce-back schemes is that distribution functions hitting the boundary are reflected, and thus the missing distributions after streaming are computed from the leaving post-collision distributions in opposite directions.

Figure 3. Missing incoming distribution functions (blue) at the bottom boundary of the rectangular domain D\mathcal{D} and known incoming distribution functions (black) from inside the domain. Black dots: interior nodes, blue dots: boundary nodes, empty gray nodes: ghost nodes.

Dirichlet boundary conditions. First, we consider a Dirichlet boundary condition for the pressure pp to be prescribed on an open and connected subset of the boundary ΓD,p⊂∂D\Gamma_{D,p}\subset\partial\mathcal{D}, viz,

p=pDon (0,tf)×ΓD,p,p=p_D\qquad\text{on }(0,t_f)\times\Gamma_{D,p},

for some given continuous function pD:(0,tf)×ΓD,p→Rp_D:(0,t_f)\times\Gamma_{D,p}\to\mathbb{R}. In this case, we apply the second-order consistent, halfway anti-bounce-back scheme following [45]. When a boundary node with position x\boldsymbol{x} connects via the velocity direction ci\boldsymbol{c}_i to the Dirichlet boundary ΓD,p\Gamma_{D,p}, then the distribution function at the next time step is given by

fi∗(t+Δt,x)=−fi∗∗(t,x)+2wipD(t+Δt,x+12Δtci).(42)f_i^*(t+\Delta t,\boldsymbol{x})=-f_{i^*}^*(t,\boldsymbol{x})+2w_i p_D\left(t+\Delta t,\boldsymbol{x}+\frac{1}{2}\Delta t\boldsymbol{c}_i\right). \tag*{(42)}

Here, iˉ\bar{i} denotes the index of the opposite direction of index ii, e.g. 4ˉ=2\bar{4}=2 and 5ˉ=7\bar{5}=7, cf. Figure 1.

Similarly, when a Dirichlet boundary condition for the displacement η\boldsymbol{\eta} at an open and connected subset of the boundary ΓD,η⊂∂D\Gamma_{D,\eta} \subset\partial D is given by

η=ηDon (0,tf)×ΓD,η,\boldsymbol{\eta}=\boldsymbol{\eta}_{D}\quad\text{on }(0,t_f)\times\Gamma_{D,\eta},

for some given continuous function ηD:(0,tf)×ΓD,η→R\boldsymbol{\eta}_{D}:(0,t_f)\times\Gamma_{D,\eta}\to\mathbb{R}, then we use the fullway bounce-back scheme developed in [66]. For a boundary node with position x\boldsymbol{x} connecting to the Dirichlet boundary ΓD,η\Gamma_{D,\eta} via the velocity direction cij\boldsymbol{c}_{ij}, the distribution function at the next pseudo-time step is given by

giˉjˉ(τ+Δτ,t,x)=gij(τ−Δτ,t,x)−6wijcij⋅ηD(t,x+12Δtcij).(43)g_{\bar{i}\bar{j}}(\tau+\Delta\tau,t,\boldsymbol{x})=g_{ij}(\tau-\Delta\tau,t,\boldsymbol{x})-6w_{ij}\boldsymbol{c}_{ij}\cdot\boldsymbol{\eta}_{D}\left(t,\boldsymbol{x}+\frac{1}{2}\Delta t\boldsymbol{c}_{ij}\right). \tag*{(43)}

As before, we use the 2D Miller index notation here, i.e., 1ˉ=−1\bar{1}=-1, cf. Figure 1. Note that the delay of one pseudo-time step is equivalent to two full streaming steps into the respective ghost node and back, hence the term “fullway”. Furthermore, this delay is necessary to suppress grid-scale oscillations in the displacement solution according to [66]. As shown therein, this boundary scheme yields a second-order consistent displacement solution and a first-order consistent stress solution for pseudo time τ→∞\tau\to\infty.

Neumann boundary conditions. For the pressure pp, we consider homogeneous Neumann boundary conditions (no-flow conditions) posed on an open and connected subset of the boundary ΓN,p⊂∂D\Gamma_{N,p}\subset\partial D, viz.

κ∇p⋅n=0on (0,tf)×ΓN,p,\kappa\nabla p\cdot\boldsymbol{n}=0\quad\text{on }(0,t_f)\times\Gamma_{N,p},

where n\boldsymbol{n} denotes the outward unit normal at the boundary ΓN,p\Gamma_{N,p}. This condition can be implemented using the simple halfway bounce-back scheme [37 Sec. 8.5.3]. The new distribution function at a boundary node connecting via the velocity direction ci\boldsymbol{c}_{i} to the Neumann boundary ΓN,p\Gamma_{N,p} is thus given by

fiˉ(t+Δt,x)=fi∗(t,x),(44)f_{\bar{i}}(t+\Delta t,\boldsymbol{x})=f_{i}^{*}(t,\boldsymbol{x}), \tag*{(44)}

where iˉ\bar{i} denotes the opposite direction index introduced above for Dirichlet conditions. Note that this boundary scheme is first-order consistent in general, but second-order consistent if the flux tangential to the boundary vanishes as considered in the numerical examples below.

Finally, a Neumann boundary condition for the displacement η\boldsymbol{\eta} (traction condition) can be imposed at an open and connected subset of the boundary ΓN,η⊂∂D\Gamma_{N,\eta}\subset\partial D, i.e.,

σn=ton (0,tf)×ΓN,η,\boldsymbol{\sigma}\boldsymbol{n}=\boldsymbol{t}\quad\text{on }(0,t_f)\times\Gamma_{N,\eta},

with the outward unit normal n\boldsymbol{n} and some given continuous traction function t:(0,tf)×ΓN,η→R2\boldsymbol{t}:(0,t_f)\times\Gamma_{N,\eta}\to\mathbb{R}^{2}. Following [66], we apply a second-order consistent, halfway anti-bounce-back scheme to obtain the missing distribution function at a boundary node connecting via the velocity direction ci\boldsymbol{c}_{i} to the Neumann boundary ΓN,η\Gamma_{N,\eta}. This yields

giˉjˉ(τ+Δτ,t,x)=−gij(τ,t,x)+ε Sij(t,x+12Δtcij),(45)g_{\bar{i}\bar{j}}(\tau+\Delta\tau,t,\boldsymbol{x})=-g_{ij}(\tau,t,\boldsymbol{x})+\varepsilon\,\mathcal{S}_{ij}\left(t,\boldsymbol{x}+\frac{1}{2}\Delta t\boldsymbol{c}_{ij}\right), \tag*{(45)}

where the traction correction terms Sij\mathcal{S}_{ij} are obtained via the stress Eq. (26) from the traction t\boldsymbol{t} and the normal direction n\boldsymbol{n}. For the left boundary, where n=(−1,0)T\boldsymbol{n}=(-1,0)^{\mathsf{T}}, one gets

S10=−S11=t1,S11=−12S12=12t2,S1ˉ1ˉ=12S12=−12t2.\mathcal{S}_{10}=-\mathcal{S}_{11}=t_{1},\qquad\mathcal{S}_{11}=-\frac{1}{2}\mathcal{S}_{12}=\frac{1}{2}t_{2},\qquad\mathcal{S}_{\bar{1}\bar{1}}=\frac{1}{2}\mathcal{S}_{12}=-\frac{1}{2}t_{2}.

Analogous formulas hold for the right, bottom and top boundaries, but are not explicitly shown here.

Numerical Experiments

This section presents several numerical examples to study the numerical accuracy and stability. To this end, we consider three problem settings: the first one is a periodic problem with manufactured solution to assess the performance of the proposed LBM in the bulk without interference of boundary, while the second problem is the two-dimensional version of the classical (quasi-1D) Terzaghi consolidation problem, which even has a discontinuous solution at the initial time due to the instantaneous response to a traction boundary. Finally, the third example extends the latter setting to a truly two-dimensional problem. We present detailed parameter studies concerning the coupling strength, i.e., the Biot–Willis coefficient α∈(0,1]\alpha\in(0,1], as well as the convergence with respect to the lattice width Δx\Delta x and the number of pseudo-time steps NEN_E, while the time step Δt=Δx2\Delta t = \Delta x^2 and the length ratio ε=Δx\varepsilon= \Delta x follow the typical viscous and elastic scaling, respectively. When applying NEN_E steps of the multi-grid LBM of Section 3.5 using KK smoothing steps with under-relaxation rate γ=3/4\gamma= 3/4, we fix the maximal grid level L=log⁡2(Nx)L = \log_2(N_x), and abbreviate this as MG(NE,KN_E,K) for simplicity. Our implementation is based on XLB [40], a jax-based Python library allowing for parallel LBM simulations. The code is available in the GitHub repository https://github.com/s-lunowa/xlb-biot. All presented simulations were run on an ordinary workstation with Intel i7-11700 CPU (16 × 2.50GHz).

For the convergence studies, the errors of the numerical solutions of the poroelasticity Eqs. (5) to (8) at the times tk=kΔtt_k = k\Delta t, k∈{0,1,…,NT}k \in\{0,1,\ldots,N_T\}, are measured based on the discrete L2(D)L^2(\mathcal{D})-norm, i.e.,

epk=∥p(tk)−pLBM(tk)∥L2(D)=Δx(∑m=1NxNy∣p(tk,xm)−pLBM(tk,xm)∣2)1/2,e_p^k = \left\|p(t_k)-p^{\mathrm{LBM}}(t_k)\right\|_{L^2(\mathcal{D})} = \Delta x\left(\sum_{m=1}^{N_xN_y}\left|p(t_k,\mathbf{x}_m)-p^{\mathrm{LBM}}(t_k,\mathbf{x}_m)\right|^2\right)^{1/2},
eηk=∥η(tk)−ηLBM(tk)∥L2(D)=(∑i=12∥ηi(tk)−ηiLBM(tk)∥L2(D)2)1/2,e_\eta^k = \left\|\boldsymbol{\eta}(t_k)-\boldsymbol{\eta}^{\mathrm{LBM}}(t_k)\right\|_{L^2(\mathcal{D})} = \left(\sum_{i=1}^{2}\left\|\eta_i(t_k)-\eta_i^{\mathrm{LBM}}(t_k)\right\|_{L^2(\mathcal{D})}^2\right)^{1/2},
eσk=∥σ(tk)−σLBM(tk)∥L2(D)=(∑ab∈{11,12,22}∥σab(tk)−σabLBM(tk)∥L2(D)2)1/2.e_\sigma^k = \left\|\boldsymbol{\sigma}(t_k)-\boldsymbol{\sigma}^{\mathrm{LBM}}(t_k)\right\|_{L^2(\mathcal{D})} = \left(\sum_{ab\in\{11,12,22\}}\left\|\sigma_{ab}(t_k)-\sigma_{ab}^{\mathrm{LBM}}(t_k)\right\|_{L^2(\mathcal{D})}^2\right)^{1/2}.

where NxN_x and NyN_y denote the number of lattice cells in x1x_1- and x2x_2-direction, respectively. Furthermore, the time-space L2(Q)L^2(Q)-error of the pressure is given by

ep=∥p−pLBM∥L2(Q)=(Δt∑k=0NT∥epk∥L2(D)2)1/2,e_p = \left\|p-p^{\mathrm{LBM}}\right\|_{L^2(Q)} = \left(\Delta t\sum_{k=0}^{N_T}\left\|e_p^k\right\|_{L^2(\mathcal{D})}^2\right)^{1/2},

and analogously for the time-space L2(Q)L^2(Q)-error of displacement eηe_\eta and of stress eσe_\sigma. To have scaling-independent, easily interpretable results, we will primarily report relative errors such as eprel=ep/∥p∥L2(Q)e_p^{\mathrm{rel}} = e_p/\left\|p\right\|_{L^2(Q)}.

Periodic problem with manufactured analytical solution

First, a manufactured solution to Eqs. (5) to (8) is considered for numerical validation of accuracy and stability. To avoid the influence of boundary conditions, a periodic problem (Γper=∂D\Gamma_{\mathrm{per}} = \partial\mathcal{D}) is posed in the domain D=(0,1)2\mathcal{D} = (0,1)^2 up to the final time tf=1t_f = 1. Inspired by [65], the manufactured solution used here is given by

η(t,x)=1−exp⁡(−8π2κt)2(9cos⁡(2πx1)sin⁡(2πx2)7sin⁡(2πx1)cos⁡(2πx2)),\boldsymbol{\eta}(t,\mathbf{x}) = \frac{1-\exp(-8\pi^2\kappa t)}{2} \begin{pmatrix} 9\cos(2\pi x_1)\sin(2\pi x_2) \\ 7\sin(2\pi x_1)\cos(2\pi x_2) \end{pmatrix},
p(t,x)=−(16λ+30μ)πα(1−exp⁡(−8π2κt))sin⁡(2πx1)sin⁡(2πx2),p(t,\mathbf{x}) = -\frac{(16\lambda+30\mu)\pi}{\alpha}\left(1-\exp(-8\pi^2\kappa t)\right)\sin(2\pi x_1)\sin(2\pi x_2),

which leads to the following forces and source term

f(t,x)=1−exp⁡(−8π2κt)2(16π2μcos⁡(2πx1)sin⁡(2πx2)0),g(t,x)=0,\mathbf{f}(t,\mathbf{x}) = \frac{1-\exp(-8\pi^2\kappa t)}{2} \begin{pmatrix} 16\pi^2\mu\cos(2\pi x_1)\sin(2\pi x_2) \\ 0 \end{pmatrix}, \qquad \mathbf{g}(t,\mathbf{x}) = 0,
s(t,x)=−8π3κ(16αexp⁡(−8π2κt)+16λ+30μα)sin⁡(2πx1)sin⁡(2πx2).s(t,\mathbf{x}) = -8\pi^3\kappa\left(16\alpha\exp(-8\pi^2\kappa t)+\frac{16\lambda+30\mu}{\alpha}\right)\sin(2\pi x_1)\sin(2\pi x_2).

Setting the (dimensionless) Young’s modulus E=0.11E = 0.11 and Poisson’s ratio ν=0.8\nu= 0.8 as in [65] (with plane-stress assumption), the physical parameters are

λ=Eν1−ν2=1145,μ=E2(1+ν)=11360,κ=110,\lambda= \frac{E\nu}{1-\nu^2} = \frac{11}{45}, \qquad \mu= \frac{E}{2(1+\nu)} = \frac{11}{360}, \qquad \kappa= \frac{1}{10},

while we consider varying Biot–Willis coefficients α∈[0.5,1.0]\alpha\in[0.5,1.0]. In the following, we discuss the influence of this coefficient α\alpha, of the number of pseudo-time steps NEN_E and of the ratio r∈{0,1/2,1}r \in\{0,1/2,1\} in the semi-implicit discretization.

Weak coupling. First, we consider rather low Biot–Willis coefficients α≤0.5\alpha\le0.5, which means that the poroelastic system is weakly coupled. In the following, we exemplary present the case α=0.5\alpha= 0.5, but observe the same behavior also in the cases α∈{0.3,0.4}\alpha\in\{0.3, 0.4\}. Using a fixed number NEN_E of multi-grid steps MG(NEN_E,2), the algorithmic complexity of O(NENx2)O(N_E N_x^2) per time step is optimal. For NE=3N_E = 3, we observe second-order convergence in ε=Δx=1/Nx\varepsilon= \Delta x = 1/N_x for the explicit (r=0r = 0), centered (r=1/2r = 1/2) and implicit (r=1r = 1) coupling scheme, see Figure 4. In particular, we observe that the errors in pressure and stress are similar, while the displacement errors are about two times larger than the others. A detailed convergence study for varying number NEN_E of multi-grid steps is displayed in Figure 5 for α=0.5\alpha= 0.5 and all three schemes. While the numerical errors might be dominated by the elasticity approximation for NE≤2N_E \le2 multi-grid steps, as visible by the slightly slanted contour lines, they follow the typical second-order convergence rate for NE≥3N_E \ge3, where the contour lines are vertically aligned. In general, there are virtually no differences between the results of all three coupling schemes.

Figure 4. Convergence results for the periodic problem with α=0.5\alpha= 0.5 using the explicit, centered and implicit coupling schemes with MG(3,2).

Figure 5. Convergence results for the periodic problem with α=0.5\alpha= 0.5 using the explicit (r=0r = 0), centered (r=1/2r = 1/2) and implicit (r=1r = 1) coupling schemes (top to bottom) with MG⁡(NE,2)\operatorname{MG}(N_E,2).

Using the simpler pressure gradient formulation of Eq. (30), we also observe second-order convergence, but an about two times increased elasticity error (in displacement and stress), as depicted in Figure 6 for the centered scheme (all three schemes produce practically indistinguishable results). Note however, that in case of strong coupling discussed below, the onset of instability for the explicit coupling scheme (r=0r = 0) occurs slightly later with that formulation, while the implicit coupling scheme (r=1r = 1) remains stable.

Figure 6. Convergence results for the periodic problem with α=0.5\alpha= 0.5 using the simpler pressure gradient formulation of Eq. (30) in the centered coupling scheme (r=1/2r = 1/2) with MG⁡(NE,2)\operatorname{MG}(N_E,2).

Furthermore, we compare this to the simple Algorithm 1 using pseudo-timestepping without multi-grid method for the elasticity. Fixing the number of pseudo-time steps to NE=1.5NxN_E = 1.5N_x, the numerical results indicate almost second-order convergence, as depicted exemplary for α=0.5\alpha= 0.5 in Figure 7 for all three coupling schemes. This linearly growing number of pseudo-time steps is not sufficient, as a detailed convergence study for varying NEN_E shows, cf. Figure 8 for the centered scheme. While the numerical errors are dominated by the elasticity approximation for small NEN_E as visible by the rather horizontal contour lines, they follow the typical second-order convergence rate for large NEN_E, where the contour lines are vertically aligned. In particular, the transition between these two regimes seems to occur around NE≈0.015Nx2N_E \approx0.015N_x^2, explaining the degradation of the convergence order visible for small Δx\Delta x in Figure 7. Note that this quadratic scaling is characteristic for point-wise iterative solvers when applied to second-order elliptic problems. Hence, the multi-grid method achieves second-order convergence at significantly reduced computational cost.

Convergence results for the periodic problem with α = 0.5 using the explicit, centered and implicit coupling schemes

Figure 7. Convergence results for the periodic problem with α=0.5\alpha= 0.5 using the explicit, centered and implicit coupling schemes with NE=1.5NxN_E = 1.5N_x pseudo-time steps (no multi-grid).

Figure 8. Convergence results for the periodic problem with α=0.5\alpha= 0.5 using the centered coupling scheme (r=1/2r = 1/2) with NEN_E pseudo-time steps (no multi-grid). The orange dashed line NE=0.015Nx2N_E = 0.015N_x^2 indicates the transition between dominant error due to spatial discretization (above the line) and due to insufficient pseudo-timestepping (below the line).

Strong coupling. Next, we consider large Biot–Willis coefficients α∈{0.6,0.7,0.8,0.9,1.0}\alpha\in\{0.6, 0.7, 0.8, 0.9, 1.0\}, which means that the poroelastic system is strongly coupled. Here, both the explicit scheme and (slightly later) the implicit scheme show instabilities for a sufficiently large number of multi-grid steps NEN_E, which decreases with increasing Biot–Willis coefficient α\alpha. In particular, when using NE=4N_E = 4 multi-grid steps, we observe divergence of the errors for the explicit scheme when α≥0.6\alpha\ge0.6, and for the implicit scheme when α≥0.8\alpha\ge0.8, as depicted in Figure 9. Hence, the implicit scheme is slightly better than the explicit one, but ultimately also unstable for α\alpha close to 1. Similar behavior has also been observed for other numerical methods due to spurious oscillations between stress and pressure, cf. [17, 23, 24].

Figure 9. Divergence results for the periodic problem using the explicit coupling scheme (r=0r = 0) for α=0.6\alpha= 0.6 and the implicit coupling scheme (r=1r = 1) for α=0.8\alpha= 0.8, but convergence for the centered coupling scheme (r=1/2r = 1/2) for α=1.0\alpha= 1.0, each using MG(4,2).

On the other hand, the centered scheme still gives second-order convergence for all α\alpha up to one, see Figure 9 for α=1\alpha= 1. Probably, this is due to averaging over two time step which leads to the cancellation of those spurious oscillations. Furthermore, a numerical convergence study varying the number NEN_E of multi-grid steps yields again a

clear second-order convergence for the centered scheme when NE≥3N_E \ge3, cf. Figure 10 for $\alpha = 1.0. Note that we observe analogous results for α∈{0.6,0.7,0.8,0.9}\alpha\in\{0.6, 0.7, 0.8, 0.9\}.

Figure 10. Convergence results for the periodic problem with α=1.0\alpha= 1.0 using the centered coupling scheme (r=1/2r = 1/2) with MG(NEN_E,2).

Terzaghi's consolidation problem

The second poroelastic example originates from Terzaghi’s consolidation problem [8], which describes a fluid-filled soil layer which becomes loaded from above. This leads to a so-called overpressurization of the fluid which continually vanishes due to flow through the surface. Consequently, the elastic soil bears more and more of the loading and subsides. This problem has an analytical solution and, hence, has been considered as benchmark for many numerical methods, see e.g. [19, 24, 28, 29] for finite elements, [21, 22] for finite volumes, and [78] for a direct numerical simulations of the averaged Navier–Stokes equation coupled to discrete particles.

While the original problem considers the quasi-1D solution to the three-dimensional poroelastic equations (d=3d = 3), we adapt the quasi-1D formulation for d=2d = 2. Following the derivation in [14 Secs. 5.1 & 5.2.2], we consider Eqs. (5) to (8) in a domain D=(0,l)×(0,1)\mathcal{D} = (0, l) \times(0, 1) until the final time tf=1t_f = 1 with fixed pressure and traction at the top

p=0andσn=(0−ϖ)on (0,tf)×{x∈∂D∣x2=1},p = 0 \qquad\text{and} \qquad\boldsymbol{\sigma}\mathbf{n} = \begin{pmatrix} 0 \\ -\varpi\end{pmatrix} \qquad\text{on } (0,t_f) \times\{x \in\partial\mathcal{D} \mid x_2 = 1\},

while no-flow condition and fixed displacement are prescribed at the bottom

κ∇p⋅n=0andη=0on (0,tf)×{x∈∂D∣x2=0},\kappa\nabla p \cdot\mathbf{n} = 0 \qquad\text{and} \qquad\boldsymbol{\eta} = 0 \qquad\text{on } (0,t_f) \times\{x \in\partial\mathcal{D} \mid x_2 = 0\},

and periodic boundary conditions are posed on the left and right at x1=0x_1 = 0 and x1=lx_1 = l. Note that the length l>0l > 0 is irrelevant since the solution is independent of the first spatial coordinate and we choose l=1l = 1 in the numerical experiments. No external forces or source are present, i.e., f=g=0\boldsymbol{f} = \boldsymbol{g} = \boldsymbol{0} and s=0s = 0. Moreover, the initial conditions are given by

p∣t=0=0andη∣t=0=0in D.p\big|_{t=0} = 0 \qquad\text{and} \qquad\boldsymbol{\eta}\big|_{t=0} = 0 \qquad\text{in } \mathcal{D}.

Due to the incompatibility of the initial and boundary conditions, the solution has a discontinuity at time t=0t = 0 with an instantaneous response

p(0+,x)=p0=αϖλ+2μ+α2andη(0+,x)=(0−p0x2/α)for x∈D.p(0^+,\boldsymbol{x}) = p_0 = \frac{\alpha\varpi}{\lambda+ 2\mu+ \alpha^2} \qquad\text{and} \qquad \boldsymbol{\eta}(0^+,\boldsymbol{x}) = \begin{pmatrix} 0 \\ -p_0x_2/\alpha \end{pmatrix} \qquad\text{for } \boldsymbol{x} \in\mathcal{D}.

Hereafter, the solution evolves continuously. In particular, the pressure and the surface subsidence S=−η2∣x2=1S = -\eta_2|_{x_2=1} are explicitly given by

p(t,x)=p0∑k=0∞2(−1)k(k+12)πexp⁡(−(k+12)2π2cft)cos⁡((k+12)πx2)for (t,x)∈(0,∞)×D,p(t,\boldsymbol{x}) = p_0 \sum_{k=0}^{\infty} \frac{2(-1)^k}{(k+\frac{1}{2})\pi} \exp\left(-\left(k+\frac{1}{2}\right)^2\pi^2 c_f t\right)\cos\left(\left(k+\frac{1}{2}\right)\pi x_2\right) \qquad\text{for } (t,\boldsymbol{x}) \in(0,\infty)\times\mathcal{D},
S(t)=S∞+(S0−S∞)∑k=0∞2(k+12)2π2exp⁡(−(k+12)2π2cft)for t∈(0,∞),S(t) = S_\infty+ (S_0-S_\infty)\sum_{k=0}^{\infty}\frac{2}{(k+\frac{1}{2})^2\pi^2}\exp\left(-\left(k+\frac{1}{2}\right)^2\pi^2c_ft\right) \qquad\text{for } t \in(0,\infty),

where the effective diffusivity cfc_f as well as the initial and final surface subsidence are defined as

cf=λ+2μλ+2μ+α2κ,S0=ϖλ+2μ+α2,S∞=ϖλ+2μ.c_f = \frac{\lambda+2\mu}{\lambda+2\mu+\alpha^2}\kappa,\qquad S_0 = \frac{\varpi}{\lambda+2\mu+\alpha^2},\qquad S_\infty= \frac{\varpi}{\lambda+2\mu}.

Setting the (dimensionless) Young’s modulus E=1E = 1 and Poisson’s ratio ν=0.8\nu= 0.8 (with plane-stress assumption), we choose the physical parameters

λ=Eν1−ν2=209,μ=E2(1+ν)=518,κ=λ+2μ+α2λ+2μ=25+9α225,ϖ=α+λ+2μα.\lambda= \frac{E\nu}{1-\nu^2} = \frac{20}{9},\qquad \mu= \frac{E}{2(1+\nu)} = \frac{5}{18},\qquad \kappa= \frac{\lambda+2\mu+\alpha^2}{\lambda+2\mu} = \frac{25+9\alpha^2}{25},\qquad \varpi= \alpha+\frac{\lambda+2\mu}{\alpha}.

which implies p0=1p_0 = 1. Note that due to the permeability κ>1\kappa> 1, we use a scaling Δt=Δx2/4\Delta t = \Delta x^2/4 to reduce the numerical errors in the pressure solution. Since only the pressure and the surface subsidence are given analytically, we report relative errors of pressure eprele_p^{\mathrm{rel}} and of subsidence esrel=∥S−SLBM∥L2(0,T)/∥S∥L2(0,T)e_s^{\mathrm{rel}} = \lVert S-S^{\mathrm{LBM}}\rVert_{L^2(0,T)}/\lVert S\rVert_{L^2(0,T)}, where the numerical subsidence is defined as average of −η2-\eta_2 over the last row of cells next to the boundary at x2=1x_2 = 1.

The numerical solution of the centered coupling scheme (r=1/2r = 1/2) for Nx=128N_x = 128 using MG(3,4) and the analytical solution are depicted in detail for α=1\alpha= 1 in Figure 11. The results excellently match already at time t=10−4t = 10^{-4} (after only 8 time steps), thus illustrating that the LBM scheme is well able to reproduce the solution due to instantaneous loading. Particularly, we observe first-order convergence of the centered coupling scheme with MG(3,4) for varying Biot–Willis coefficient α∈{0.8,0.9,1.0}\alpha\in\{0.8, 0.9, 1.0\}, as shown in Figure 12. Note that convergence order is reduced due to the implementation of the boundary conditions and the discontinuity at time t=0t = 0. In contrast to the previous numerical example, also the explicit scheme (r=0r = 0) and the implicit scheme (r=1r = 1) feature such a convergence behavior for all α∈{0.8,0.9,1.0}\alpha\in\{0.8, 0.9, 1.0\}. Concerning the influence of the number of multi-grid steps NEN_E, a convergence study for the centered coupling scheme (r=1/2r = 1/2) is displayed in Figure 13 for α=1\alpha= 1. Analogously to the periodic example, the numerical errors are dominated by the elasticity approximation for small NE≤2N_E \leq2, while they are dominated by the first-order spatio-temporal error for larger NEN_E. Here however, the transition region seems slightly wider. Note that we observe very similar results for α∈{0.8,0.9}\alpha\in\{0.8, 0.9\} and all three coupling schemes.

Figure 11. Exact and numerical solutions for Terzaghi’s problem with α=1.0\alpha= 1.0 using the centered scheme (r=1/2r = 1/2) with MG(3,4) for Nx=128N_x = 128. In the left sub-figure, dotted lines represent the exact solution, while solid lines depict the numerical solution.

Convergence plots for $\alpha = 0.8$, $\alpha = 0.9$, and $\alpha = 1.0$

Figure 12. Convergence results for Terzaghi’s problem with α=0.8\alpha= 0.8, 0.9 and 1.0 using the centered coupling scheme (r=1/2r = 1/2) with MG(3,4).

Figure 13. Convergence results for Terzaghi’s problem with α=1.0\alpha= 1.0 using the centered coupling scheme (r=1/2r = 1/2) with MG(NEN_E,4).

Two-dimensional loading problem

The third example extends Terzaghi’s consolidation problem to two dimensions by introducing a position-dependent load. Again, this describes a fluid-filled soil layer which becomes locally loaded from above. The resulting overpressurized fluid slowly flows out through the surface, so that the elastic soil bears more and more of the loading.

and subsides. While Terzaghi’s problem considers a quasi-1D solution, here we also observe horizontal displacement due to the nonuniform load.

To this end, we consider the dimensional Eqs. (1) to (4) in the domain D=(0,100)2\mathcal{D} = (0, 100)^2 until the final time tf=2⋅106t_f = 2 \cdot10^{6}. No external forces or source are present, i.e., f=g=0\boldsymbol{f} = \boldsymbol{g} = \boldsymbol{0} and s≡0s \equiv0, and the initial conditions are

p∣t=0=0andη∣t=0=0in D.p|_{t=0} = 0 \qquad\text{and} \qquad\boldsymbol{\eta}|_{t=0} = \boldsymbol{0} \qquad\text{in } \mathcal{D}.

At the top boundary, pressure and traction are imposed

p=0andσn=(0−104(1−cos⁡(x1+2550π)))on (0,tf)×{x∈∂D∣x2=100},p = 0 \qquad\text{and} \qquad\boldsymbol{\sigma}\boldsymbol{n} = \begin{pmatrix} 0\\ -10^{4}\left(1-\cos\left(\frac{x_1+25}{50}\pi\right)\right) \end{pmatrix} \qquad\text{on } (0,t_f) \times\{\boldsymbol{x} \in\partial\mathcal{D} \mid x_2 = 100\},

whereas no-flow condition and fixed displacement are given at the bottom

κ∇p⋅n=0andη=0on (0,tf)×{x∈∂D∣x2=0}.\kappa\nabla p \cdot\boldsymbol{n} = 0 \qquad\text{and} \qquad\boldsymbol{\eta} = \boldsymbol{0} \qquad\text{on } (0,t_f) \times\{\boldsymbol{x} \in\partial\mathcal{D} \mid x_2 = 0\}.

Furthermore, periodic boundary conditions are prescribed on the left and right boundaries at x1∈{0,100}x_1 \in\{0, 100\}. Setting the (dimensionless) Young’s modulus E=105E = 10^{5} and Poisson’s ratio ν=0.9\nu= 0.9 (with plane-stress assumption), we choose the physical parameters

λ=Eν1−ν2=9⋅10619,μ=E2(1+ν)=5⋅10519,c0=10−6,κ=10−9.\lambda= \frac{E\nu}{1-\nu^2} = \frac{9 \cdot10^{6}}{19}, \qquad \mu= \frac{E}{2(1+\nu)} = \frac{5 \cdot10^{5}}{19}, \qquad c_0 = 10^{-6}, \qquad \kappa= 10^{-9}.

and consider the strongly coupled situation where α=1\alpha= 1. Note that the incompatibility of the initial and boundary conditions again yields a solution which is discontinuous at time t=0t = 0.

The numerical solution of the centered coupling scheme (r=1/2r = 1/2) using MG(3,4) for Nx=128N_x = 128 and NT=Nx2=16384N_T = N_x^2 = 16384 is depicted at the times t∈{2⋅104,2⋅105,2⋅106}t \in\{2 \cdot10^{4}, 2 \cdot10^{5}, 2 \cdot10^{6}\} in Figure 14. Since the load is considerably larger on the left half, the subsidence is significant there, while the surface on the right is even lifted. Furthermore, this leads to a horizontal expansion from the left half to the right one. In the beginning, the fluid pressure is very large close to the left half of the surface, while it is low on the right half. As expected, the pressure slowly vanishes over time resulting in a further subsidence of the whole surface.

Numerical solutions at the times $t \in \{2 \cdot 10^{4}, 2 \cdot 10^{5}, 2 \cdot 10^{6}\}$ for the 2D loading problem

Figure 14. Numerical solutions at the times t∈{2⋅104,2⋅105,2⋅106}t \in\{2 \cdot10^{4}, 2 \cdot10^{5}, 2 \cdot10^{6}\} (left to right) for the 2D loading problem using the centered coupling scheme (r=1/2r = 1/2) with MG(3,4) for Nx=128N_x = 128. The pressure is represented in color, while the shown displacement is amplified by a factor of 3 and additionally represented by the arrows. The original outline is given by the solid orange line.

Conclusion

In this work, we developed a novel coupling of LBMs to solve Biot’s consolidation model in two dimensions. To this end, we combined the classical LBM for reaction-diffusion equations [49] with the recent pseudo-time LBM for quasi-static linear elasticity [65]. Therefore, the overall method uses only two distribution functions with (reduced) standard velocity sets for the solution fields, so that 17 unknowns per cell are required to solve for the 6 components of pressure, displacement and stress field at second-order consistency. We addressed the coupling between the equations by a newly developed, centered update scheme which uses half of both explicit and semi-implicit contributions.

Numerical results obtained with the method of manufactured solutions confirm the analytical properties of the coupled methods. This concerns the second-order convergence with respect to time-step size and lattice width. Using a multi-grid method to accelerate the explicit pseudo-timestepping of the LBM for quasi-static elasticity, the required number of pseudo-time steps reduces from quadratic (NE∼Nx2N_E \sim N_x^2) to constant (NE≈3N_E \approx3 in the presented examples). We have demonstrated that naively coupled LBMs lead to instabilities for strong coupling with Biot–Willis coefficient α\alpha close to 1, as both the fully explicit scheme and the semi-implicit scheme yield diverging solutions. However, the centered coupling scheme is stable and accurate for all Biot–Willis coefficients. Moreover, the presented numerical results for Terzaghi’s (quasi-1D) consolidation problem and for a two-dimensional expansion thereof, both incorporate an instantaneous response due to external loading, and hence demonstrate that the coupled LBM is even applicable to discontinuous problems.

The method has several advantages, as it allows for a simple and scalable parallelization and is second-order consistent in space and time for all three solution fields. The main drawback of the naive implementation lies in its pseudo-timestepping for quasi-static elasticity. However, the presented multi-grid method eliminates this shortcoming since only few multi-grid steps are necessary. Note that an adaptive number of multi-grid steps depending on the residual might reduce the computational cost even further. Moreover, advanced multi-grid cycles, such as W- and F-cycle, as well as a full multi-grid approach are left for future research. Considering complex geometries and the resulting implementation of boundary conditions in LBMs, the application of algebraic rather than geometric multi-grid might be preferable. Lastly, also the Darcy flow subsystem can be accelerated analogously to allow for arbitrary time-step sizes.

In view of the fact, that this is the initial application of LBM to poroelasticity, there are ample opportunities for further development. Besides possible extensions to three-dimensional and nonlinear poroelastic equations, different LBMs might be used to solve the flow and/or elasticity sub-problems. In particular, solving the flow equations by a MRT scheme like in [45] could allow for anisotropic permeability, while accounting for elastodynamics solved by LBMs as e.g. in [59, 63] would enable the study of seismic waves in aquifers and reservoirs. Furthermore, the limiting regimes of incompressible fluid and/or solid require appropriate treatment. Finally, Biot’s consolidation theory strongly simplifies real-world behavior, so that extensions towards nonlinear poro-(visco-)elastoplasticity including rate-dependency, as well as coupling to further processes, such as thermal exchange or chemical reactions, would open the methods applicability to many more contemporary applications.

CRediT authorship contribution statement

Stephan B. Lunowa: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Software, Validation, Visualization, Writing – original draft, Writing – review and editing. Barbara Wohlmuth: Conceptualization, Funding acquisition, Methodology, Resources, Supervision, Writing – review and editing.

Acknowledgments

The authors thank Laura De Lorenzis, Oliver Bolakee and Martin Geier for the discussions about multi-grid methods for LBM.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Data availability

The python implementation and scripts required to reproduce the numerical results of this article are available in the following GitHub repository: https://github.com/s-lunowa/xlb-biot.

References

  1. [1]J. M. Nordbotten, M. A. Celia, Geological Storage of CO2: Modeling Approaches for Large-Scale Simulation, Wiley, 2011. doi:10.1002/9781118137086.DOI
  2. [2]T. Sharmin, N. R. Khan, M. S. Akram, M. M. Ehsan, A state-of-the-art review on geothermal energy extraction, utilization, and improvement strategies: conventional, hybridized, and enhanced geothermal systems, Internat. J. Thermofluids 18 (2023) 100323. doi:10.1016/j.ijft.2023.100323.DOI
  3. [3]H. Alkhateb, A. Al-Ostaz, A. H.-D. Cheng, X. Li, Materials genome for graphene-cement nanocomposites, J. Nanomech. Micromech. 3 (3) (2013) 67–77. doi:10.1061/(ASCE)NM.2153-5477.0000055.
  4. [4]M. Vandamme, F.-J. Ulm, Nanoindentation investigation of creep properties of calcium silicate hydrates, Cem. Concr. Res. 52 (2013) 38–52. doi:10.1016/j.cemconres.2013.05.006.DOI
  5. [5]A.-R. A. Khaled, K. Vafai, The role of porous media in modeling flow and heat transfer in biological tissues, Internat. J. Heat Mass Transfer 46 (26) (2003) 4989–5003. doi:10.1016/S0017-9310(03)00301-6.DOI
  6. [6]P. Causin, G. Guidoboni, A. Harris, D. Prada, R. Sacco, S. Terragni, A poroelastic model for the perfusion of the lamina cribrosa in the optic nerve head, Math. Biosci. 257 (2014) 33–41. doi:10.1016/j.mbs.2014.08.002.DOI
  7. [7]D. Chapelle, J.-F. Gerbeau, J. Sainte-Marie, I. E. Vignon-Clementel, A poroelastic model valid in large strains with applications to perfusion in cardiac modeling, Comput. Mech. 46 (2010) 91–101. doi:10.1007/s00466-009-0452-x.
  8. [8]K. Terzaghi, Die Berechnung der Durchlässigkeitsziffer des Tones aus dem Verlauf der hydrodynamischen Spannungsercheinungen, Sitzungsberichte Akademie der Wissenschaften, Wien Mathematisch-Naturwissenschaftliche Klasse, Abteilung IIa 132 (1923) 105–124.
  9. [9]M. A. Biot, Consolidation settlement under a rectangular load distribution, J. Appl. Phys. 12 (5) (1941) 426–430. doi:10.1063/1.1712921.DOI
  10. [10]M. A. Biot, General theory of three-dimensional consolidation, J. Appl. Phys. 12 (2) (1941) 155–164. doi:10.1063/1.1712886.DOI
  11. [11]M. A. Biot, Theory of elasticity and consolidation for a porous anisotropic solid, J. Appl. Phys. 26 (2) (1955) 182–185. doi:10.1063/1.1721956.DOI
  12. [12]M. A. Biot, D. G. Willis, The elastic coefficients of the theory of consolidation, J. Appl. Mech. 15 (1957) 594–601. doi:10.1115/1.4011606.DOI
  13. [13]E. Detournay, A. H.-D. Cheng, Fundamentals of poroelasticity, in: C. Fairhurst (Ed.), Analysis and Design Methods, Pergamon, 1993, pp. 113–171. doi:10.1016/B978-0-08-040615-2.50011-3.DOI
  14. [14]O. Coussy, Poromechanics, Wiley, 2004. doi:10.1002/0470092718.DOI
  15. [15]G. N. Mercer, S. I. Barry, Flow and deformation in poroelasticity – II Numerical method, Math. Comput. Model. 30 (9) (1999) 31–38. doi:10.1016/S0895-7177(99)00178-8.DOI
  16. [16]F. J. Gaspar, F. J. Lisbona, P. N. Vabishchevich, Finite difference schemes for poro-elastic problems, Comput. Meth. Appl. Math. 2 (2) (2002) 132–142. doi:10.2242/cmam.2002.0008.DOI
  17. [17]M. Bean, S.-Y. Yi, An immersed interface method for a 1D poroelasticity problem with discontinuous coefficients, J. Comput. Appl. Math. 272 (2014) 81–96. doi:10.1016/j.cam.2014.05.009.DOI
  18. [18]R. W. Lewis, B. A. Schrefler, The Finite Element Method in the Deformation and Consolidation of Porous Media, Wiley, 1987.
  19. [19]M. A. Murad, A. F. D. Loula, On stability and convergence of finite element approximations of Biot’s consolidation problem, Internat. J. Numer. Meth. Engrg. 37 (4) (1994) 645–667. doi:10.1002/nme.1620370407.DOI
  20. [20]J. M. Nordbotten, Stable cell-centered finite volume discretization for Biot equations, SIAM J. Numer. Anal. 54 (2) (2016) 942–968. doi:10.1137/15M1014280.DOI
  21. [21]H. T. Hohm, C. R. Maliska, M. Ferronato, C. Janna, A stabilized element-based finite volume method for poroelastic problems, J. Comput. Phys. 364 (2018) 49–72. doi:10.1016/j.jcp.2018.03.010.
  22. [22]I. Sokolova, M. G. Bastysia, H. Hajibeygi, Multiscale finite volume method for finite-volume-based simulation of poroelasticity, J. Comput. Phys. 379 (2019) 309–324. doi:10.1016/j.jcp.2018.11.039.DOI
  23. [23]K. M. Terekhov, Cell-centered finite-volume method for heterogeneous anisotropic poromechanics problem, J. Comput. Appl. Math. 365 (2020) 112357. doi:10.1016/j.cam.2019.112357.DOI
  24. [24]P. J. Phillips, M. F. Wheeler, A coupling of mixed and discontinuous Galerkin finite-element methods for poroelasticity, Comput. Geosci. 12 (2008) 417–435. doi:10.1007/s10596-008-9082-1.DOI
  25. [25]R. Liu, M. F. Wheeler, C. N. Dawson, R. H. Dean, A coupled discontinuous/continuous Galerkin framework and an adaptive penalty scheme for poroelasticity problems, Comput. Meth. Appl. Mech. Engrg. 198 (41) (2009) 3499–3510. doi:10.1016/j.cma.2009.07.005.DOI
  26. [26]M. F. Wheeler, G. Xue, I. Yotov, Coupling multipoint flux mixed finite element methods with continuous Galerkin methods for poroelasticity, Comput. Geosci. 18 (2014) 57–75. doi:10.1007/s10596-013-9382-y.DOI
  27. [27]I. Ambartsumyan, E. Khattatov, I. Yotov, A coupled multipoint stress–multipoint flux mixed finite element method for the Biot system of poroelasticity, Comput. Meth. Appl. Mech. Engrg. 372 (2020) 113407. doi:10.1016/j.cma.2020.113407.DOI
  28. [28]J. Choo, S. Lee, Enriched Galerkin finite elements for coupled poromechanics with local mass conservation, Comput. Meth. Appl. Mech. Engrg. 341 (2018) 311–332. doi:10.1016/j.cma.2018.06.022.DOI
  29. [29]T. Kadeethum, H. M. Nick, S. Lee, F. Ballarin, Enriched Galerkin discretization for modeling poroelasticity and permeability alteration in heterogeneous porous media, J. Comput. Phys. 427 (2021) 110030. doi:10.1016/j.jcp.2020.110030.DOI
  30. [30]D. Sulsky, Z. Shi-Jian, H. L. Schreyer, Application of a particle-in-cell method to solid mechanics, Comput. Phys. Commun. 87 (1) (1995) 236–252, particle Simulation Methods. doi:10.1016/0010-4655(94)00107-7.DOI
  31. [31]J. Yu, J. Zhao, W. Liang, S. Zhao, A semi-implicit material point method for coupled thermo-hydro-mechanical simulation of saturated porous media in large deformation, Comput. Meth. Appl. Mech. Engrg. 418 (2024) 116462. doi:10.1016/j.cma.2023.116462.DOI
  32. [32]G. R. McNamara, G. Zanetti, Use of the Boltzmann equation to simulate lattice-gas automata, Phys. Rev. Lett. 61 (1988) 2332–2335. doi:10.1103/PhysRevLett.61.2332.DOI
  33. [33]S. Chen, G. D. Doolen, Lattice Boltzmann method for fluid flows, Annu. Rev. Fluid Mech. 30 (1998) 329–364. doi:10.1146/annurev.fluid.30.1.329.DOI
  34. [34]P. L. Bhatnagar, E. P. Gross, M. Krook, A model for collision processes in gases. I. Small amplitude processes in charged and neutral one-component systems, Phys. Rev. 94 (1954) 511–525. doi:10.1103/PhysRev.94.511.DOI
  35. [35]D. d’Humieres, Multiple-relaxation-time lattice Boltzmann models in three dimensions, Phil. Trans. R. Soc. A. 360 (2002) 437–451. doi:10.1098/rsta.2001.0955.DOI
  36. [36]M. Geier, M. Schönherr, A. Pasquali, M. Krafczyk, The cumulant lattice Boltzmann equation in three dimensions: Theory and validation, Comput. Math. Appl. 70 (4) (2015) 507–547. doi:10.1016/j.camwa.2015.05.001.DOI
  37. [37]T. Krüger, H. Kusumaatmaja, A. Kuzmin, O. Shardt, G. Silva, E. M. Viggen, The Lattice Boltzmann Method, Springer, 2017. doi:10.1007/978-3-319-44649-3.DOI
  38. [38]S. Succi, The Lattice Boltzmann Equation: For Complex States of Flowing Matter, Oxford University Press, 2018. doi:10.1093/oso/9780199592357.001.0001.
  39. [39]M. Schönherr, K. Kucher, M. Geier, M. Stiebler, S. Freudiger, M. Krafczyk, Multi-thread implementations of the lattice Boltzmann method on non-uniform grids for CPUs and GPUs, Comput. Math. Appl. 61 (12) (2011) 3730–3743. doi:10.1016/j.camwa.2011.04.012.DOI
  40. [40]M. Ataei, H. Salehipour, XLB: A differentiable massively parallel lattice Boltzmann library in Python, Comput. Phys. Commun. 300 (2024) 109187. doi:10.1016/j.cpc.2024.109187.DOI
  41. [41]D. Wolf-Gladrow, A lattice Boltzmann equation for diffusion, J. Stat. Phys. 79 (1995) 1023–1032. doi:10.1007/BF02181215.
  42. [42]E. G. Flekkøy, Lattice Bhatnagar-Gross-Krook models for miscible fluids, Phys. Rev. E 47 (1993) 4247–4257. doi:10.1103/PhysRevE.47.4247.DOI
  43. [43]S. Ponce Dawson, S. Chen, G. D. Doolen, Lattice Boltzmann computations for reaction-diffusion equations, J. Chem. Phys. 98 (2) (1993) 1514–1523. doi:10.1063/1.464316.DOI
  44. [44]W. Zhang, A. G. Bengough, J. W. Crawford, I. M. Young, A lattice BGK model for advection and anisotropic dispersion equation, Adv. Water Resour. 25 (1) (2002) 1–8. doi:10.1016/S0309-1708(01)00047-1.DOI
  45. [45]I. Ginzbourg, Equilibrium-type and link-type lattice Boltzmann models for generic advection and anisotropic-dispersion equation, Adv. Water Resour. 28 (11) (2005) 1171–1195. doi:10.1016/j.advwatres.2005.03.004.
  46. [46]I. Rasin, S. Succi, W. Miller, A multi-relaxation lattice kinetic method for passive scalar diffusion, J. Comput. Phys. 206 (2) (2005) 453–462. doi:10.1016/j.jcp.2004.12.010.DOI
  47. [47]D. Dapelo, S. Simonis, M. J. Krause, J. Bridgeman, Lattice-Boltzmann coupled models for advection–diffusion flow on a wide range of Péclet numbers, J. Comput. Sci. 51 (2021) 101363. doi:10.1016/j.jocs.2021.101363.
  48. [48]B. Shi, B. Deng, R. Du, X. Chen, A new scheme for source term in LBGK model for convection–diffusion equation, Comput. Math. Appl. 55 (7) (2008) 1568–1575. doi:10.1016/j.camwa.2007.08.016.DOI
  49. [49]T. Seta, Implicit temperature-correction-based immersed-boundary thermal lattice Boltzmann method for the simulation of natural convection, Phys. Rev. E 87 (2013) 063304. doi:10.1103/PhysRevE.87.063304.DOI
  50. [50]Q. Li, Z. Chai, B. Shi, Lattice Boltzmann model for a class of convection–diffusion equations with variable coefficients, Comput. Math. Appl. 70 (4) (2015) 548–561. doi:10.1016/j.camwa.2015.05.008.DOI
  51. [51]Z. Chai, B. Shi, Z. Guo, A multiple-relaxation-time lattice Boltzmann model for general nonlinear anisotropic convection-diffusion equations, J. Sci. Comput. 69 (2016) 355–390. doi:10.1007/s10915-016-0198-5.DOI
  52. [52]Z. Chai, B. Shi, Multiple-relaxation-time lattice Boltzmann method for the Navier-Stokes and nonlinear convection-diffusion equations: Modeling, analysis, and elements, Phys. Rev. E 102 (2020) 023306. doi:10.1103/PhysRevE.102.023306.DOI
  53. [53]S. Simonis, M. Frank, M. J. Krause, On relaxation systems and their relation to discrete velocity Boltzmann models for scalar advection–diffusion equations, Phil. Trans. R. Soc. A. 378 (2020) 20190400. doi:10.1098/rsta.2019.0400.DOI
  54. [54]M. Geier, A. Pasquali, M. Schönherr, Parameterization of the cumulant lattice Boltzmann method for fourth order accurate diffusion part I: Derivation and validation, J. Comput. Phys. 348 (2017) 862–888. doi:10.1016/j.jcp.2017.05.040.DOI
  55. [55]Y. Chen, Z. Chai, B. Shi, A general fourth-order mesoscopic multiple-relaxation-time lattice Boltzmann model and its macroscopic finite-difference scheme for two-dimensional diffusion equations, J. Comput. Phys. 509 (2024) 113045. doi:10.1016/j.jcp.2024.113045.DOI
  56. [56]S. Marconi, B. Chopard, A lattice Boltzmann model for a solid body, Internat. J. Modern Phys. B 17 (1–2) (2003) 153–156. doi:10.1142/S0217979203021754.DOI
  57. [57]G. S. O’Brien, T. Nissen-Meyer, C. J. Bean, A lattice Boltzmann method for elastic wave propagation in a Poisson solid, Bull. Seismol. Soc. Am. 102 (3) (2012) 1224–1234. doi:10.1785/0120110191.DOI
  58. [58]J. S. N. Murthy, P. K. Kolluru, V. Kumaran, S. Ansumal, Lattice Boltzmann method for wave propagation in elastic solids, Commun. Comput. Phys. 23 (4) (2018) 1223–1240. doi:10.4208/cicp.OA-2016-0259.
  59. [59]M. Escande, P. K. Kolluru, L. M. Cléon, P. Sagaut, Lattice Boltzmann method for wave propagation in elastic solids with a regular lattice: Theoretical analysis and validation (2020). arXiv:2009.06404.arxiv.org/abs/2009.06404
  60. [60]E. Faust, A. Schlüter, H. Müller, F. Steinmetz, R. Müller, Dirichlet and Neumann boundary conditions in a lattice Boltzmann method for elastodynamics, Comput. Mech. 73 (2024) 317–339. doi:10.1007/s00466-023-02369-w.DOI
  61. [61]A. Schlüter, C. Kuhn, R. Müller, Multiple crack propagation in a lattice Boltzmann method for plane shear loading of a stationary crack, Comput. Mech. 62 (2018) 1059–1069. doi:10.1007/s00466-018-1550-4.
  62. [62]A. Schlüter, H. Müller, R. Müller, Boundary conditions in a lattice Boltzmann method for plane strain problems, PAMM 21 (1) (2021) e202100085. doi:10.1002/pamm.202100085.DOI
  63. [63]O. Boolakee, M. Geier, L. De Lorenzis, Lattice Boltzmann for linear elastodynamics: Periodic problems and Dirichlet boundary conditions, Comput. Meth. Appl. Mech. Engrg. 433 (2025) 117469. doi:10.1016/j.cma.2024.117469.DOI
  64. [64]X. Yin, G. Yan, T. Li, Direct simulations of the linear elastic displacements field based on a lattice Boltzmann model, Internat. J. Numer. Meth. Engrg. 107 (3) (2016) 234–251. doi:https://doi.org/10.1002/nme.5167.
  65. [65]O. Boolakee, M. Geier, L. De Lorenzis, A new lattice Boltzmann scheme for linear elastic solids: periodic problems, Comput. Math. Appl. Mech. Engrg. 404 (2023) 115756. doi:10.1016/j.cma.2022.115756.DOI
  66. [66]O. Boolakee, M. Geier, L. De Lorenzis, Dirichlet and Neumann boundary conditions for a lattice Boltzmann scheme for linear elastic solids on arbitrary domains, Comput. Meth. Appl. Mech. Engrg. 415 (2023) 116225. doi:10.1016/j.cma.2023.116225.DOI
  67. [67]J. Tölke, M. Krafczyk, E. Rank, A multigrid-solver for the discrete Boltzmann equation, J. Statist. Phys. 107 (2002) 573–591. doi:10.1023/A:1014551813787.DOI
  68. [68]D. J. Mavriplis, Multigrid solution of the steady-state lattice Boltzmann equation, Comput. Fluids 35 (2006) 793–804. doi:10.1016/j.compfluid.2005.07.020.DOI
  69. [69]C. Armstrong, Y. Peng, An MRT extension to the multigrid lattice Boltzmann method, Commun. Comput. Phys. 26 (2019) 1178–1195. doi:10.4208/cicp.OA-2018-0166.DOI
  70. [70]B. An, J. M. Bergadà, W. Sang, A simplified new multigrid algorithm of lattice Boltzmann method for steady states, Comput. Math. Appl. 135 (2023) 102–110. doi:10.1016/j.camwa.2023.01.013.DOI
  71. [71]S. Gsell, U. D’Ortona, J. Favier, Multigrid dual-time-stepping lattice Boltzmann method, Phys. Rev. E 101 (2020) 023309. doi:10.1103/PhysRevE.101.023309.DOI
  72. [72]D. V. Patil, K. N. Premnath, S. Banerjee, Multigrid lattice Boltzmann method for accelerated solution of elliptic equations, J. Comput. Phys. 265 (2014) 172–194. doi:10.1016/j.jcp.2014.01.049.DOI
  73. [73]J.-L. Auriault, E. Sanchez-Palencia, Etude du comportement macroscopique d’un milieu poreux saturé deformable, J. Mecanique 16 (4) (1977) 575–603.
  74. [74]R. E. Showalter, Diffusion in poro-elastic media, J. Math. Anal. Appl. 251 (1) (2000) 310–340. doi:10.1006/jmaa.2000.7048.DOI
  75. [75]P. J. Dellar, An interpretation and derivation of the lattice Boltzmann method using Strang splitting, Comput. Math. Appl. 65 (2) (2013) 129–141. doi:10.1016/j.camwa.2011.08.047.
  76. [76]T. Lee, C.-L. Lin, A stable discretization of the lattice Boltzmann equation for simulation of incompressible two-phase flows at high density ratio, J. Comput. Phys. 206 (1) (2005) 16–47. doi:10.1016/j.jcp.2004.12.001.DOI
  77. [77]T. Lee, P. F. Fischer, Eliminating parasitic currents in the lattice Boltzmann equation method for nonideal gases, Phys. Rev. E 74 (2006) 046709. doi:10.1103/PhysRevE.74.046709.DOI
  78. [78]J. Zhao, T. Shan, Coupled CFD–DEM simulation of fluid–particle interaction in geomechanics, Powder Technol. 239 (2013) 248–258. doi:10.1016/j.powtec.2013.02.003.DOI

Paper details

vlnc:W_tywsd9zt
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents