betaFeedback

WorksMathematics

MARS : a Method for the Adaptive Removal of Stiffness in PDEs

Abstract

The E(xplicit)I(implicit)N(null) method was developed recently to remove numerical instability from PDEs, adding and subtracting an operator D\mathcal{D} of arbitrary structure, treating the operator implicitly in one case, and explicitly in the other. Here we extend this idea by devising an adaptive procedure to find an optimal approximation for D\mathcal{D}. We propose a measure of the numerical error which detects numerical instabilities across all wavelengths, and adjust each Fourier component of D\mathcal{D} to the smallest value such that numerical instability is suppressed. We show that for a number of nonlinear and non-local PDEs, in one and two dimensions, the spectrum of D\mathcal{D} adapts automatically and dynamically to the theoretical result for marginal stability. Our method thus has the same stability properties as a fully implicit method, while only requiring the computational cost comparable to an explicit solver. The adaptive implicit part is diagonal in Fourier space, and thus leads to minimal overhead compared to the explicit method.

Introduction

Our ability to model many key physical processes is limited by the stability of the numerical schemes we use to simulate the partial differential equations (PDEs) describing them. The reason is that the maximal stable time step of an explicit numerical integration scheme is of the order of the shortest time-scale in the system. In a stable physical system these are typically exponentially damped modes which relax back to equilibrium; the smaller the length scale, the faster the relaxation. This makes it particularly hard to simulate systems at large values of the viscosity or of the surface tension. For instance, surface tension driven flows in the open source fluid dynamics code Gerris [1] (followed by Basilisk : http://basilisk.fr ) require a time step proportional to Δ3\Delta^3 [2], where Δ\Delta is the grid spacing, which this program adapts dynamically in order to ensure a sufficient spatial accuracy [3]. As a result, for small geometries Δ\Delta can be very small, resulting in time steps which are prohibitively small. This constraint is more restrictive than the CFL constraint, related to advection, for which the time step depends on scale like Δ\Delta. Another example is the numerical computation of solidification/fusion fronts, which uses a non-linear heat equation [4] : the corresponding time step constraint is Δ2\Delta^2.

If for example relaxation toward equilibrium is controlled by a differential operator of order mm (m=2m = 2 for ordinary diffusion, m=3m = 3 for the Hele-Shaw flow to be described below), then the required maximum time step δt\delta t scales as δt=Cδxm\delta t = C\delta x^m, where δx\delta x is the smallest grid spacing or the size of the smallest sub-division. In a well-resolved numerical simulation, this should be considerably smaller than the smallest relevant physical feature. Rapid exponential decay implies that the amplitude of perturbations on the grid scale is very small, and contributes negligible to the numerical solution. Thus one arrives at the paradoxical situation that the stability of the numerical scheme is controlled by a part of the solution which contributes negligibly, and which is actually the most stable from a physical perspective. This property is sometimes referred to as the stiffness of the PDE [5], which becomes worse with increasing spatial order mm of the operator.

To deal with this constraint on the time step, which often is so severe that it makes the exploration of important physical parameter regimes impractical, one has to resort to implicit methods. This means that the right hand side of the equation (or at least the stiffest parts of it) has to be evaluated at a future time step, making it necessary to solve an implicit equation at each time step [6, 7]. This makes the numerical code both complicated to write and time-consuming to solve. This is true in particular if the operator is non-local (as is the case for example of integral operators, as they appear in boundary integral type codes [8, 9]). Indeed, in this particular case, when writing an implicit scheme, each element of the discretized solution depends on all the others, requiring a large number of operations to solve the implicit equation.

To address this problem, it has long been realized that not the whole of the right hand side of an equation has to be treated implicitly, as long as the “stiffest” part of the operator is dealt with implicitly. This gives rise to the so-called “implicit-explicit methods” [10, 11], which divide up the problem between explicit and implicit parts, such that hopefully the implicit contribution is sufficiently simple to invert. If this is not clear, as is typically the case for an integral operator, the problem can be solved by judiciously slicing off the stiffest part, which can be local [9]. However, this has to be done on a case-by-case basis, and will not always be possible. Recently, we have presented a much more general method to stabilize stiff equations, which makes use of the arbitrariness in which splitting between explicit and implicit parts can take place [12, 13]. We consider a partial differential equation of the form

∂u∂t=f(u,t),(1)\frac{\partial u}{\partial t}=f(u,t), \tag*{(1)}

where u(x,t)u(x,t) is a function of space and time or a vector of functions of space and time, and f(u,t)f(u,t) generally is a non-linear operator involving spatial derivatives of u(x,t)u(x,t). In the present article, after explaining the adaptive stabilization procedure in section 2, we shall treat the following three examples :

  • A non-linear operator with a fourth-order spatial derivative, related to the thin film flow equation (section 3) :

f(u,t)=−∂∂x(u3∂3u∂x3+1u∂u∂x),f(u,t)=-\frac{\partial}{\partial x}\left(u^3\frac{\partial^3u}{\partial x^3}+\frac{1}{u}\frac{\partial u}{\partial x}\right),
  • A two-dimensional example with a fourth-order derivative (section 4) :

f(u,t)=−N(u)−Δu−νΔ2u,f(u,t)=-\mathcal{N}(u)-\Delta u-\nu\Delta^2u,

where u(x,y)u(x,y), N(u)\mathcal{N}(u) is a non-linear operator, Δ\Delta the Laplacian, and ν\nu a constant,

  • A boundary integral equation (section 5) :

f(u,t)=∫g(v)K(u,v) dv,f(u,t)=\int g(v)K(u,v)\,dv,

where uu and vv are a two-dimensional vectors, g(v)g(v) is a function of vv involving second-order derivatives in space, K(u,v)K(u,v) is a singular kernel, and the integral is performed along a curve.

As explained in our previous article [13], in order to stabilize the stiff terms in f(u,t)f(u,t), we add two terms on the right-hand-side of the discretized version of equation (1) :

ujn+1−ujnδt=fj(un,tn)−Dj[un]+Dj[un+1],(2)\frac{u_j^{n+1}-u_j^n}{\delta t}=f_j(u^n,t^n)-\mathcal{D}_j[u^n]+\mathcal{D}_j[u^{n+1}], \tag*{(2)}

where nn denotes the time variable (tn=nδtt^n= n\delta t) and D\mathcal{D} is an arbitrary operator. The variable uu as well as ff are defined on a spatial grid xj=jδxx_j=j\delta x, where δx\delta x is the grid spacing. Clearly, the added terms are exactly zero apart from the first-order error that comes from the fact that D\mathcal{D} is evaluated at different time levels, which motivates the name “Explicit-Implicit-Null” method or “EIN”. If D\mathcal{D} is the same as the original operator f(u,t)f(u,t), this is a purely implicit method, if D=0\mathcal{D}=0, it is explicit. Similar ideas have been implemented to stabilize the motion of a surface in the diffuse interface and level-set methods [14, 15, 16], and for the solution of PDEs on surfaces [17]. We also show that by a simple step-halving procedure [18], (2) can always be turned into a scheme which is second order accurate in time [13].

When applying (2), we want D\mathcal{D} to be a reasonable approximation to the stiff part of ff. If D\mathcal{D} were much larger, it would stabilize the scheme, but would introduce an additional time truncation error. We thus want to choose D\mathcal{D} for optimal effectiveness, in the sense that it adds the perfect amount of damping, without adding a supplementary error to the numerical scheme.

This paper presents a numerical scheme to achieve this goal automatically, adjusting D\mathcal{D} to the threshold value. Restricting ourselves to periodic boundary conditions, we choose D\mathcal{D} to be diagonal in Fourier space, which renders it both simple to handle and sufficiently flexible. Indeed, the implicit step becomes almost trivial to perform:

u^kn+1−u^knδt=f^k(un,tn)+λ(k)u^kn−λ(k)u^kn+1,(3)\frac{\hat{u}^{n+1}_k-\hat{u}^{n}_k}{\delta t}=\hat{f}_k(u^n,t^n)+\lambda(k)\hat{u}^{n}_k-\lambda(k)\hat{u}^{n+1}_k, \tag*{(3)}

where  ^\hat{\ } denotes the Fourier transform and the damping spectrum λ(k)≥0\lambda(k)\ge0 is an arbitrary function. Adjustment of D\mathcal{D} is now down to finding the scalar damping spectrum λ(k)\lambda(k) for a discrete sequence of kk’s.

The Fourier transform f^k\hat{f}_k can be calculated effectively from the spatial discretization fjf_j using the fast Fourier transform (FFT) [19]. From (3), we find

u^kn+1=u^kn+f^k(un,tn)δt−1+λ(k),(4)\hat{u}^{n+1}_k=\hat{u}^{n}_k+\frac{\hat{f}_k(u^n,t^n)}{\delta t^{-1}+\lambda(k)}, \tag*{(4)}

so we obtain the desired solution ujn+1u_j^{n+1} at the new time step from the inverse transform. The scheme (4) (as well as any other first order scheme) can be turned into a second order scheme by Richardson extrapolation [18]. Namely, let u1n+1u_1^{n+1} be the solution for one step δt\delta t, u2n+1u_2^{n+1} the solution for two half steps δt/2\delta t/2. Then

un+1=2u2n+1−u1n+1+O(δt3),(5)u^{n+1}=2u_2^{n+1}-u_1^{n+1}+O(\delta t^3), \tag*{(5)}

is second order accurate in time, and

E=u1n+1−u2n+1(6)E=u_1^{n+1}-u_2^{n+1} \tag*{(6)}

can be used as an error estimator [20].

To analyze (3) further, we adopt a “frozen-coefficient” hypothesis, that the solution is essentially constant over the time scale on which numerical instability is developing. Then assuming small perturbations δu^kn\delta\hat{u}_k^n, the problem is turned into a linear equation for δu^kn\delta\hat{u}_k^n, with constant coefficients. At least on a small scale (i.e. in the large kk limit), much smaller than any externally imposed scale, f^k(un,tn)\hat{f}_k(u^n,t^n) is expected to be translationally invariant, making the operator diagonal in Fourier space. Namely, let us assume a more general non-local operator

f(x)=∫−∞∞D(x,y)u(y) dy.f(x)=\int_{-\infty}^{\infty}D(x,y)u(y)\,dy.

Translational invariance implies that D(x,y)=D(x−y)D(x,y)=D(x-y); taking the Fourier transform, we arrive at f^(k)=D^(k)u^(k)\hat{f}(k)=\hat{D}(k)\hat{u}(k). This means in the large kk limit we expect

f^k(un+δu,tn)∼−e(k)δu^kn,(7)\hat{f}_k(u^n+\delta u,t^n)\sim-e(k)\delta\hat{u}_k^n, \tag*{(7)}

where δu\delta u is a small perturbation around the solution at tnt^n. Here we assume that the eigenvalues e(k)e(k) are real, as it is typically the case for physical problems, where the dominant process on a small scale is dissipative.

We have shown in [13] that, as long as λ(k)>e(k)/2\lambda(k)>e(k)/2, the system (3) with the approximation (7) is unconditionally stable. This is a generalization of a method first presented, for the case of the diffusion equation in two dimensions, in [21]. If (4) is turned into a second order scheme using (5), this condition is [22]:

λ(k)>λc(k)=23e(k),(8)\lambda(k)>\lambda_c(k)=\frac{2}{3}e(k), \tag*{(8)}

with λc(k)\lambda_c(k) the theoretical stability limit. Thus for sufficiently large values of λ(k)\lambda(k), there is always stability; however, the time truncation error of the second order scheme is O(λδt2)O(\lambda\delta t^2). This is to be compared to the time truncation error O(ϵδt2)O(\epsilon\delta t^2) of the original scheme, and should therefore be kept as small as possible, consistent with the stability constraint. There is a certain similarity here with the preconditioning of matrices, where a matrix is approximated by a simple diagonal matrix [23, 24].

We also showed in [13] that the explicit scheme (i.e. for which λ(k)=0\lambda(k)=0) is stable as long as:

ϵ(k)<2δt(9)\epsilon(k)<\frac{2}{\delta t} \tag*{(9)}

Equation (9) defines a threshold value of wave numbers kek_e, below which the scheme is stable even without stabilization. As a result, for k<kek<k_e, λ(k)\lambda(k) can be chosen to vanish, without affecting stability.

In [13] we have tested the ideas underlying the EIN method, calculating the spectrum ϵ(k)\epsilon(k) for a variety of operators, including nonlocal operators treated previously in [9]. We approximated λ(k)\lambda(k) as a power law, derived from the low wavenumber limit of the exact discrete spectrum. As predicted by the above analysis, we find the scheme (3) unconditionally stable, and performing with the same accuracy as that proposed in [9]. Obviously, this still requires one to obtain a good estimate for the spectrum.

In the present paper, we aim to remove this analytical step, and to make the calculation of λ(k)\lambda(k) self-consistent. The idea is to determine λ(k)\lambda(k) iteratively, by detecting numerical instability. If, for a given wave number kk, random perturbations due to numerical instability of the time-stepping grow in time, then the damping is increased, while λ(k)\lambda(k) can be reduced if the code is stable. In the simplest version of our procedure, we focus on the high wave number limit, where most of the stiffness is coming from, and approximate λ(k)\lambda(k) by a power law, determined by one or two parameters, depending on whether the exponent is to be prescribed. While we found this approach to work, it introduces arbitrary assumptions into the procedure, and assumes a separation between a high and low wave number regimes. Instead, here we present the results of a scheme which adjusts each Fourier mode individually, based on noise detected in the same Fourier mode. This models the original operator in much greater detail, and leads to a spectrum λ(k)\lambda(k) which corresponds closely to the theoretical stability limit.

In the next section we develop and describe our procedure for automatic stabilization. The following three sections are each dedicated to a particular example, to illustrate how the method is implemented, and to demonstrate its effectiveness.

Adaptive stabilization

Our method is based on the formulation (3), which together with (5) is an unconditionally stable second order scheme, as long as λ(k)\lambda(k) is sufficiently large. We would like to find an adaptive procedure which refines λ(k)\lambda(k) at each time step, so as to keep it as small as possible, consistent with stability. To achieve this, we have to address two issues: (i) find a measure ϵ(k)\epsilon(k) of the noise, or of numerical instability, for each Fourier mode kk; (ii) specify the evolution of λ(k)\lambda(k) for a given noise.

Finding a suitable measure of the error is the crucial question, to be discussed in more detail below. As for (ii), we aim to adjust each Fourier component λ(k)\lambda(k) individually, although we have also explored representing λ(k)\lambda(k) by a finite number of parameters. We adopt a simple approach, taking a local relation between ϵ(k)\epsilon(k) and λ(k)\lambda(k), which is shown to be sufficient for the examples to be presented below. For each Fourier mode, if ϵ(k)\epsilon(k) is larger than an upper bound ϵu\epsilon_u, the corresponding λ(k)\lambda(k) is increased by a factor of 1.2. If on the other hand ϵ(k)<ϵu\epsilon(k)<\epsilon_u, λ(k)\lambda(k) is decreased slowly by a factor of 1/1.021/1.02 at each time step, in order to avoid a sudden onset of instability. We have used the same rates in all examples, but confirmed that the method is robust against change of parameters.

As to a measure of noise, a first guess might be to take ϵ(k)\epsilon(k) as the Fourier transform of the error estimator (6) E^k\hat{E}_k. We tested this idea using the interface dynamics discussed in more detail in section 5, and illustrated in Figure 6. Figure 1 shows the evolution of the Fourier transform E^k\hat{E}_k of this error estimator, for the first four time steps, without using the EIN method. The time step is chosen to be δt=3.125×10−5\delta t=3.125\times10^{-5}, the number of points N=1024N=1024, and we use a purely explicit scheme (no stabilization), so that the modes with the largest wavenumbers are unstable.

Figure 1. The evolution of the error estimator E^k\hat{E}_k (black curve) for the Hele-Shaw flow (27), (28), discussed in more detail in Sect. 5. Initially the error is uniformly small for a flat interface with a white noise. The vertical dashed line is the stability boundary k=kek = k_e.

Figure 6. A simulation of the Hele-Shaw problem (27), (28) shown on the top row. On the lower row, the corresponding spectrum of ϵ(k)\epsilon(k) defined by equation (37) (green), as well as λ(k)\lambda(k) (red). The dotted line is the stability limit λc(k)=2e(k)/3\lambda_c(k)=2e(k)/3, with e(k)e(k) given by (33), the top horizontal blue line the explicit stability boundary 2/δt2/\delta t. The vertical orange line is k=kek=k_e given by equation (35), and the vertical dashed line its initial position.

Indeed, as explained in the next sections, there exists a region k>kek > k_e in kk-space which is stable with an explicit scheme (on the left of the vertical dashed line), and an unstable region (on the right), where we would like to detect numerical instability. As a result, the noise level grows very rapidly for the right-hand side of the spectrum, and for the first two time steps there is little power in the k<kek < k_e modes. Thus E^k\hat{E}_k could be used to detect correctly the numerical instability for large kk.

However, the left part of the spectrum is soon invaded through non-linear mode-coupling, and there grows a considerable component of the error at small kk (corresponding to large scales), which would not be damped away if λ(k)\lambda(k) was increased. The problem is clear: in the proposed scheme, there is no clean distinction between noise resulting from numerical instability, and the broad spectrum of unstable modes which is part of the physical solution. The crucial problem of defining the numerical noise ϵ\epsilon lies in this distinction.

A successful procedure came from the idea of spatial smoothing, taking the truncation error as the starting point. To compute ϵ\epsilon at the jj’th gridpoint, we consider the error estimator EjE_j, and compare it to a smoothed version Eˉj\bar{E}_j at the same point. The reasoning is that Eˉj\bar{E}_j contains the full spectrum coming from the deterministic nonlinear dynamics, so Ej−EˉjE_j - \bar{E}_j only contains the random noise produced by numerical instability. There are many possible choices for the smoothed-out error. We chose a polynomial approximation over 2n2n gridpoints, but excluding jj itself, otherwise ϵ\epsilon would be identically zero. In other words,

Eˉj=P(Ej−n,…,Ej−1,Ej+1,…,Ej+n),(10)\bar{E}_j = \mathcal{P}\left(E_{j-n},\ldots,E_{j-1},E_{j+1},\ldots,E_{j+n}\right), \tag*{(10)}

where P\mathcal{P} is the (2n−1)(2n - 1) degree polynomial, passing through (Ej−n,…,Ej−1,Ej+1,…,Ej+n)\left(E_{j-n},\ldots,E_{j-1},E_{j+1},\ldots,E_{j+n}\right). Taking the Fourier transform of the difference between EjE_j and Eˉj\bar{E}_j, we define the noise measure ϵ(k)\epsilon(k) as

ϵ(k)=E^k−Eˉ^k.(11)\epsilon(k) = \hat{E}_k - \hat{\bar{E}}_k. \tag*{(11)}

The difference between the naive error measure E^k\hat{E}_k and ϵ(k)\epsilon(k) based on smoothing is illustrated in Figure 2. We ran the same computation as in figure 1, but using our adaptive procedure. Equation (11) is used as a measure of the noise (with n=2n = 2) to adapt λ(k)\lambda(k) at each time step, with the threshold ϵu=10−10\epsilon_u = 10^{-10}. The left curve shows the Fourier transform E^k\hat{E}_k as a function of kk: clearly, this error alone is ill-suited to detect instability, since it has significant components for k<kek < k_e, where the explicit scheme is stable, i.e. where there is no instability even for λ(k)=0\lambda(k) = 0. The right curve shows the noise measure ϵ(k)\epsilon(k) given by (11) used to adapt λ(k)\lambda(k) as a function of kk: the instability is correctly detected at large values of kk and this error remains low for k<kek < k_e, i.e. does not require any damping, in the region where an explicit scheme is stable.

Figure 2. The effect of smoothing on the solution of the Hele-Shaw flow (27),(28) at t=0.04t = 0.04, as shown in the third panel of Figure 6 below. On the left, we show the spectrum of the error estimator E^k\widehat{E}_k, which is broad in the nonlinear regime. On the right, we show the noise measure ϵ(k)\epsilon(k) as defined by (11), which is substantial only in a high wave number region where noise is detected. The vertical dashed line is k=kek = k_e. The horizontal red line is the threshold ϵu\epsilon_u used to adapt λ(k)\lambda(k).

Example: thin film flow with van der Waals forces

Equation of motion

As an example of a non-linear equation in one dimension, we first consider a thin liquid film on a horizontal solid substrate. Assuming lubrication theory and taking into account van der Waals forces, which can destabilize the film, the 1D evolution equation for the height of the film h(x,t)h(x,t) reads [25, 26]:

∂h∂t=−∂∂x[h33η∂∂x(γ∂2h∂x2−A6πh3)](12)\frac{\partial h}{\partial t}=-\frac{\partial}{\partial x}\left[\frac{h^3}{3\eta}\frac{\partial}{\partial x}\left(\gamma\frac{\partial^2 h}{\partial x^2}-\frac{A}{6\pi h^3}\right)\right] \tag*{(12)}

where η\eta is the dynamic viscosity of the fluid, γ\gamma its surface tension coefficient and AA the Hamaker constant. Using L=A/2πγL=\sqrt{A/2\pi\gamma} as the lengthscale and T=3ηL/γT=3\eta L/\gamma as the timescale, the dimensionless equation reads:

∂h∂t=−∂∂x(h3∂3h∂x3+1h∂h∂x).(13)\frac{\partial h}{\partial t}=-\frac{\partial}{\partial x}\left(h^3\frac{\partial^3 h}{\partial x^3}+\frac{1}{h}\frac{\partial h}{\partial x}\right). \tag*{(13)}

Considering the linear stability of (13), we study the growth of a small-amplitude single mode added to an initially flat interface:

h(x,t)=h0+εeikx+ωt,(14)h(x,t)=h_0+\varepsilon e^{ikx+\omega t}, \tag*{(14)}

where x∈[0,1]x\in[0,1]. Linearizing equation (13) for ε≪1\varepsilon\ll1 gives the dispersion relation:

ω=−h03k4+k2h0.(15)\omega=-h_0^3k^4+\frac{k^2}{h_0}. \tag*{(15)}

As the initial condition, we start from a flat film with a sinusoidal perturbation added to it, corresponding to the most unstable (or Rayleigh) mode. In order to initiate an instability on the scale of the entire computational domain, we fix k=2πk=2\pi and set the initial height h0h_0 that corresponds to the maximum growth rate:

dωdk∣k=2π=0⟺h0=121/4(2π)1/2≃0.34\left.\frac{d\omega}{dk}\right|_{k=2\pi}=0\Longleftrightarrow h_0=\frac{1}{2^{1/4}(2\pi)^{1/2}}\simeq0.34

Using this initial thickness, we choose as the initial condition:

h(x,0)=h0+Acos⁡(2πx),(16)h(x,0)=h_0+A\cos(2\pi x), \tag*{(16)}

where A=0.01A=0.01.

In order to compute the right-hand-side fj(hn,tn)f_j(h^n,t^n) of equation (13), we use second-order centered finite differences on a regular grid xj=j/Nx_j=j/N, where j∈[0,N]j\in[0,N] and N=128N=128 is the number of grid points:

fj(hn,tn)=−hj3hj−2−4hj−1+6hj−4hj+1+hj+2δx4−3hj2hj+1−hj−12δxhj−2−2hj−1+2hj+1+hj+22δx3−1hjhj+1−2hj+hj−1δx2+1hj2(hj+1−hj−12δx)2.(17)\begin{aligned} f_j(h^n,t^n)={}&-h_j^3\frac{h_{j-2}-4h_{j-1}+6h_j-4h_{j+1}+h_{j+2}}{\delta x^4}\\ &-3h_j^2\frac{h_{j+1}-h_{j-1}}{2\delta x}\frac{h_{j-2}-2h_{j-1}+2h_{j+1}+h_{j+2}}{2\delta x^3}\\ &-\frac{1}{h_j}\frac{h_{j+1}-2h_j+h_{j-1}}{\delta x^2} +\frac{1}{h_j^2}\left(\frac{h_{j+1}-h_{j-1}}{2\delta x}\right)^2. \tag*{(17)} \end{aligned}

Using the Fourier transform of (17) in (4), we obtain h^kn+1\hat{h}_k^{n+1}, from which the new points hjn+1h_j^{n+1} are obtained from the inverse Fourier transform. The Richardson scheme (5), based on each grid point, then leads to a second-order accurate result for hjn+1h_j^{n+1}.

Stabilization

Before proceeding to the automatic stabilization of the numerical scheme, we adopt a von Neumann stability analysis, in order to predict the theoretical value of λ(k)\lambda(k) for the scheme to be stable. For this purpose, we only need to consider the fourth-order derivative in equation (13), which is the stiff term to be stabilized. Using a “frozen coefficient” hypothesis, we look for perturbations to the mean profile hˉ\bar{h} in the form of a single Fourier mode:

hjn=hˉ(j/N,nδt)+δh^kn=hˉ(j/N,nδt)+ξneikδα,(18)h_j^n=\bar{h}(j/N,n\delta t)+\delta\hat{h}_k^n=\bar{h}(j/N,n\delta t)+\xi^n e^{ik\delta\alpha}, \tag*{(18)}

where ξ(δt,k)\xi(\delta t,k) is the amplification factor [27], hˉ\bar{h} is assumed constant over the time step, δα=2πδx=2π/N\delta\alpha=2\pi\delta x=2\pi/N, and k∈[0,N−1]k\in[0,N-1]. Inserting this expression into (17), retaining only −hˉ3hxxxx-\bar{h}^3h_{xxxx} in the original equation, the linearization (7) of the right-hand-side of (13) gives:

e(k)=2hˉ3δx4(cos⁡(2kδα)−4cos⁡(kδα)+3).(19)e(k)=2\frac{\bar{h}^3}{\delta x^4}\left(\cos(2k\delta\alpha)-4\cos(k\delta\alpha)+3\right). \tag*{(19)}

The modified numerical scheme (4) will be stable as long as λ(k)\lambda(k) meets the stability criterion (8) with e(k)e(k) given by (19). Instead of this λ(k)\lambda(k), for simplicity we initialize λ(k)\lambda(k) with an expression of the form λ0k4\lambda_0k^4, where λ0\lambda_0 must be chosen as the maximum of 2e(k)/3k42e(k)/3k^4 over all kk, in order to satisfy the stability criterion everywhere. It is easily seen that this maximum is attained in the limit k→0k\to0, from which we obtain

λ0≥323π4hˉ3.(20)\lambda_0\ge\frac{32}{3}\pi^4\bar{h}^3. \tag*{(20)}

Note that (20) depends on the local height hˉ\bar{h}, which varies slightly for the initial condition; we choose it to correspond to the maximum of the surface elevation, for which the stability requirement is most stringent. In addition, the condition (9) defines a threshold value kek_e, below which the explicit time step is stable:

2δt=e(k)≃hˉ3(2πke)4⇒ke≃12π(2δthˉ3)1/4≃4.25,(21)\frac{2}{\delta t}=e(k)\simeq\bar{h}^3(2\pi k_e)^4\Rightarrow k_e\simeq\frac{1}{2\pi}\left(\frac{2}{\delta t\bar{h}^3}\right)^{1/4}\simeq4.25, \tag*{(21)}

where we have used hˉ=h0=0.34\bar{h}=h_0=0.34 and δt=10−4\delta t=10^{-4}.

Using the procedure described in § 2, for each Fourier mode e(k)e(k) we adjust λ(k)\lambda(k) depending on whether it is larger or smaller than an upper bound ϵu=10−8\epsilon_u=10^{-8}. The value of this bound is subject to some experimentation to make the scheme work, but can be varied by several orders of magnitude without affecting the functioning of the scheme.

However, εu\varepsilon_u must be chosen in accordance with the typical value of ε(k)\varepsilon(k). Indeed, in the current case, the absolute error (6) is O(δt2)O(\delta t^2), and the noise measure (11) introduces a spatial error O(δx2)O(\delta x^2), leading to a global absolute error O(δx2δt2)≈6×10−13O(\delta x^2\delta t^2)\approx6\times10^{-13}, consistent with the green curve in the first panel of figure 3. As a consequence εu\varepsilon_u has to be larger than this value for the adaptive procedure to work.

Figure 3 shows our adaptive scheme at work, as the interface (shown on the top row) deforms; the noise measure ε(k)\varepsilon(k) is defined as in (11) (See supplementary movie TF_movie.mpeg). We initialized λ(k)\lambda(k) to the asymptotic power-law λ0k4\lambda_0k^4, which is seen as the red line in the lower panel of the first row (which shows the system after the first time step). The true stability boundary, based on the full expression (19), is shown as the dotted line; for small kk, it is slightly lower than the more stringent power law approximation chosen as the initial condition. The noise measure ε(k)\varepsilon(k) after the first time step is very small as expected.

Figure 3. A simulation of the thin film equation with van der Waals forces (13), with the interface shown on the top row. On the lower row, the corresponding spectrum of ε(k)\varepsilon(k) defined by equation (11) (green), as well as λ(k)\lambda(k) (red). The dotted line is the stability limit λc(k)=2ε(k)/3\lambda_c(k)=2\varepsilon(k)/3, with ε(k)\varepsilon(k) given by (19) and hˉ≈h0\bar{h}\approx h_0, the top horizontal blue line the explicit stability boundary 2/δt2/\delta t. The vertical dashed line is k=kek=k_e given by equation (21), with hˉ=h0\bar{h}=h_0, whereas the vertical orange line is k=kek=k_e computed for hˉ=hmax⁡\bar{h}=h_{\max}.

As seen in the second panel, after some time λ(k)\lambda(k) has converged onto the theoretical stability limit (8) for small kk, with ε(k)\varepsilon(k) given by (19). Here we have assumed hˉ≈h0\bar{h}\approx h_0 for the initial stages of the dynamics, an approximation that will no longer be valid near the end of the computation (fourth row of Figure 3), where hh varies considerably in space, and (18) can only be applied locally. However, adjustment of λ(k)\lambda(k) toward the stability boundary only occurs for k>kek>k_e, since there is no numerical instability below k=kek=k_e. As a result, for k<kek<k_e the stabilizing spectrum λ(k)\lambda(k) is reduced at every time step, and in the second panel has already fallen by orders of magnitude below the stability limit of the EIN scheme.

The only source for concern is seen in the 4th panel, when the film thickness has become very non-uniform. In that case there is a region just above kek_e where λ(k)\lambda(k) is quite elevated relative to the theoretical limit (but which is based on the assumption of a uniform thickness), as well as noisy. This comes from the fact that the explicit stability boundary k=kek=k_e given by (19) moves to the left when values of hh in space become significantly higher than h0h_0. The vertical orange line in figure 3 corresponds to k=kek=k_e computed using the maximum height of the interface hmax⁡h_{\max} instead of h0h_0. As long as this value does not reach the next smaller integer value of kk, λ(k)\lambda(k) is progressively decreased on the left of this boundary. As soon as kek_e crosses an integer value k′k', exponential growth is observed for k′k' and λ(k′)\lambda(k') cannot be adapted quickly enough, resulting in increasing values of λ\lambda for the adjacent values of kk. This problem could probably be solved by changing the way λ(k)\lambda(k) is adapted at each timestep, by testing the stability of each mode and rejecting the timestep in case of instability.

2D Kuramoto–Sivashinsky equation

Equation of motion

To demonstrate that our method works in higher dimensions, we consider the example of the 2D Kuramoto–Sivashinsky equation [28], which is known to exhibit spatio-temporal chaos [28, 29]:

∂u∂t=−N(u)−Δu−νΔ2u.(22)\frac{\partial u}{\partial t} = -\mathcal{N}(u) - \Delta u - \nu\Delta^2 u. \tag*{(22)}

The single Laplacian on the right has a minus sign in front of it, leading to instability on the smallest scale; this is stabilized by the last term, which is of forth order, making the problem very stiff. Nonlinearity is introduced through the term

N(u)=12(∣∇u∣2−14π2∫02π∫02π∣∇u∣2 dx dy);(23)\mathcal{N}(u) = \frac{1}{2}\left(|\nabla u|^2 - \frac{1}{4\pi^2}\int_0^{2\pi}\int_0^{2\pi}|\nabla u|^2\,dx\,dy\right); \tag*{(23)}

following [29], the spatially constant integral term has been introduced for convenience only, to make sure that uu always has zero spatial mean. The variable u(x,y,t)u(x,y,t) is defined on a two-dimensional square domain, which we can rescale to ensure that (x,y)∈[0,2π](x,y) \in[0,2\pi].

Equation (22) is discretized on a regular grid using centered finite differences in order to find fjx,jy(un,tn)f_{j_x,j_y}(u^n,t^n), whose two-dimensional Fourier transform is f^kx,ky(un,tn)\hat{f}_{k_x,k_y}(u^n,t^n). We can then use the modified time step (4) to find u^kx,kyn+1\hat{u}^{n+1}_{k_x,k_y} and thus ujx,jyn+1u^{n+1}_{j_x,j_y}. As usual, the scheme is then turned into a second order method using Richardson extrapolation (5). In [29], (22) is treated implicitly using a Fourier pseudospectral method. The purpose of our treatment is to demonstrate the effectiveness of our general scheme, which does not pay attention to the specifics of the operator, in spatial dimensions greater than one.

Stabilization

In order to study the numerical stability, we only consider the bi-Laplacian, which is the stiffest term to be stabilized. Inserting the single Fourier mode

ujx,jyn=ξnei(kxjxδx+kyjyδy)u_{j_x,j_y}^{n} = \xi^{n}e^{i(k_xj_x\delta x+k_yj_y\delta y)}

into the discretized version of equation (22), and retaining only the bi-Laplacian term, we obtain:

e(kx,ky)=ν{2cos⁡2kxδx−8cos⁡kxδx+6δx4+2cos⁡2kyδy−8cos⁡kyδy+6δy4+22cos⁡(kxδx+kyδy)−4cos⁡kyδy+2cos⁡(kxδx−kyδy)−4cos⁡kxδx+4δx2δy2}.(24)e(k_x,k_y) = \nu\left\{ \frac{2\cos2k_x\delta x - 8\cos k_x\delta x + 6}{\delta x^{4}} + \frac{2\cos2k_y\delta y - 8\cos k_y\delta y + 6}{\delta y^{4}} + 2\frac{2\cos(k_x\delta x+k_y\delta y)-4\cos k_y\delta y+2\cos(k_x\delta x-k_y\delta y)-4\cos k_x\delta x+4}{\delta x^{2}\delta y^{2}} \right\}. \tag*{(24)}

The modified numerical scheme (4) will be stable as long as λ(kx,ky)\lambda(k_x,k_y) meets the stability criterion (8): λ(kx,ky)>2e(kx,ky)/3\lambda(k_x,k_y) > 2e(k_x,k_y)/3, and the stability of the explicit scheme is given by (9): e(kx,ky)<2/δte(k_x,k_y) < 2/\delta t. For the explicit stability boundary, one can take the small-kk limit of (24):

es(kx,ky)∼ν(kx2+ky2)2.(25)e_s(k_x,k_y) \sim\nu\left(k_x^2+k_y^2\right)^2. \tag*{(25)}

As in the previous example, this approximation overpredicts the true value of e(kx,ky)e(k_x,k_y) for large wavenumbers. For simplicity, we will use this conservative approximation to set the initial value of λ(kx,ky)\lambda(k_x,k_y). The adjustment of λ\lambda is based on an upper bound ϵu=10−5\epsilon_u=10^{-5} for ϵ\epsilon. This value is significantly higher than in the previous example, since the typical absolute error in the present case is O(δx2δt2)≈2×10−7O(\delta x^2\delta t^2)\approx2\times10^{-7}, as seen on the green curve in the first panel of figure 4.

In the definition (6) of ϵ(kx,ky)\epsilon(k_x,k_y), we now have to interpolate the error estimator E(jx,jy)E(j_x,j_y) on a two-dimensional grid; we find Eˉ(jx,jy)\bar{E}(j_x,j_y) using a bilinear interpolation of the error estimator:

Eˉ(jx,jy)=14[E(jx,jy−1)+E(jx+1,jy)+E(jx,jy+1)+E(jx−1,jy)](26)\bar{E}(j_x,j_y)=\frac{1}{4}\left[E(j_x,j_y-1)+E(j_x+1,j_y)+E(j_x,j_y+1)+E(j_x-1,j_y)\right] \tag*{(26)}

Our first attempt was to use E(jx+1,jy−1)E(j_x+1,j_y-1), E(jx+1,jy+1)E(j_x+1,j_y+1), E(jx−1,jy+1)E(j_x-1,j_y+1) and E(jx−1,jy−1)E(j_x-1,j_y-1) to interpolate EE in (jx,jy)(j_x,j_y), but it turned out that the interpolation values for two adjacent nodes where decoupled, making the process of estimating the error unstable. This issue is addressed by using the four neighbors E(jx,jy−1)E(j_x,j_y-1), E(jx+1,jy)E(j_x+1,j_y), E(jx,jy+1)E(j_x,j_y+1), and E(jx−1,jy)E(j_x-1,j_y), as seen in figure 5.

Figure 4 shows a computation of the Kuramoto–Sivashinsky equation (22), for ν=0.2\nu=0.2, in the chaotic regime (See supplementary movie KS_movie.mpeg). Thus the interface u(x,y)u(x,y) deforms in an irregular, unpredictable fashion on many scales. On the lower row we show the corresponding spectrum of ϵ(kx,0)\epsilon(k_x,0) (green), as well as λ(kx,0)\lambda(k_x,0) (red). For plotting purpose, we chose to show only a slice of the spectrum (ky=0k_y=0), but the adaption procedure works for the whole 2D spectrum.

Figure 4. A simulation of the Kuramoto–Sivashinsky equation (22) for ν=0.2\nu= 0.2, with the interface shown on the top row. Nx=Ny=128N_x = N_y = 128 points are used to discretize (22) in both directions, and the time step is δt=0.01\delta t = 0.01. On the lower row, the corresponding spectrum of es(kx,0)e_s(k_x,0) defined by equation (11) (green), as well as λ(kx,0)\lambda(k_x,0) (red). The dotted line is the stability limit λc(kx,0)=2es(kx,0)/3\lambda_c(k_x,0) = 2e_s(k_x,0)/3 (cf. (24)); the top horizontal blue line the explicit stability boundary 2/δt2/\delta t. The vertical dashed line is k=kek = k_e given by 2/δt=es(kx,0)2/\delta t = e_s(k_x,0) (cf. (25)); λ(kx,ky)\lambda(k_x,k_y) is initialized using 2es(kx,ky)/32e_s(k_x,k_y)/3. Half of the spectrum in kxk_x is shown, since it is symmetric around Nx/2N_x/2 (the other half corresponding to negative wave-numbers kx∈[−Nx/2+1:−1]k_x \in[-N_x/2 + 1:-1]).

We have initialized λ(kx,ky)\lambda(k_x,k_y) to 2es(kx,ky)/32e_s(k_x,k_y)/3 (see (25)). As in the previous example, the initial condition for λ\lambda (red line) is slightly above the theoretical stability limit (dotted line) for kk values corresponding to small scales. This is still true after the first time step, while e(kx,0)e(k_x,0) is very small as expected. As seen in the second panel, λ(kx,0)\lambda(k_x,0) has converged onto the theoretical stability limit λc(kx,0)=2e(kx,0)/3\lambda_c(k_x,0)=2e(k_x,0)/3, with e(kx,0)e(k_x,0) given by (24), since the initial condition overpredicts the stability boundary.

However, convergence only occurs for k>kek>k_e, since below k=kek=k_e no numerical instability occurs. As a result, for k<kek<k_e the stabilizing spectrum λ(kx,0)\lambda(k_x,0) is reduced at every time step, and has already fallen by orders of magnitude below the stability limit of the EIN scheme. Correspondingly, by adjusting λ(kx,0)\lambda(k_x,0) the error ϵ(kx,0)\epsilon(k_x,0) is kept close to the threshold ϵu=10−5\epsilon_u=10^{-5} for k>kek>k_e.

Figure 5. Bilinear interpolation Eˉ(jx,jy)\bar{E}(j_x,j_y) using the four neighboring points in red.

Hele–Shaw flow

Equations of motion

As an example of a non-local, but stiff operator, we consider an interface in a vertical Hele-Shaw cell, separating two viscous fluids with the same dynamic viscosity, with the heavier fluid on top [9]. As heavy fluid falls, small perturbations on the interface grow exponentially: this is known as the Rayleigh-Taylor instability [30]. However, surface tension assures regularity on small scales. For simplicity, we assume the flow to be periodic in the horizontal direction. We briefly recall the dynamics of the interface here; for more details, see [9, 13].

The interface is discretized using marker points labeled with α\alpha, which represents the motion of a fluid particle. They are advected according to:

∂X(α)∂t=Un+Ts.(27)\frac{\partial\mathbf{X}(\alpha)}{\partial t} = U\mathbf{n} + T\mathbf{s}. \tag*{(27)}

Here X(α)=(x,y)\mathbf{X}(\alpha)=(x,y) is the position vector, n=(−yα/sα,xα/sα)\mathbf{n}=(-y_\alpha/s_\alpha,x_\alpha/s_\alpha) and s=(xα/sα,yα/sα)\mathbf{s}=(x_\alpha/s_\alpha,y_\alpha/s_\alpha) are the normal and tangential unit vectors, respectively, and sα=(xα2+yα2)1/2s_\alpha=(x_\alpha^2+y_\alpha^2)^{1/2}. Hence U=(u,v)⋅nU=(u,v)\cdot\mathbf{n} and T=(u,v)⋅sT=(u,v)\cdot\mathbf{s} are the normal and tangential velocities, respectively. The tangential velocity does not affect the motion, but is chosen so as to maintain a reasonably uniform distribution of points [9, 13]. If z(α,t)=x+iyz(\alpha,t)=x+iy is the complex position of the interface, which is assumed periodic with period 1 (z(α+2π)=z(α)+1z(\alpha+2\pi)=z(\alpha)+1), the complex velocity becomes:

u(α)−iv(α)=12iPV⁡∫02πγ(α′,t)cot⁡[π(z(α,t)−z(α′,t))] dα′.(28)u(\alpha)-iv(\alpha)=\frac{1}{2i}\operatorname{PV}\int_{0}^{2\pi}\gamma(\alpha',t)\cot\left[\pi\left(z(\alpha,t)-z(\alpha',t)\right)\right]\,d\alpha'. \tag*{(28)}

where γ\gamma is the vortex sheet strength. For two fluids of equal viscosities [31],

γ=Sκα−Ryα,(29)\gamma=S\kappa_\alpha-Ry_\alpha, \tag*{(29)}

where κ\kappa is the mean curvature of the interface:

κ(α)=xαyαα−yαxααsα3,recalling thatsα=(xα2+yα2)1/2.(30)\kappa(\alpha)=\frac{x_\alpha y_{\alpha\alpha}-y_\alpha x_{\alpha\alpha}}{s_\alpha^3}, \qquad\text{recalling that}\qquad s_\alpha=(x_\alpha^2+y_\alpha^2)^{1/2}. \tag*{(30)}

Here SS is the non-dimensional surface tension coefficient and RR is the non-dimensional gravity force, chosen to be 0.1 and 50, respectively, in the following example. To compute the complex Lagrangian velocity of the interface (28), we use the spectrally accurate alternate point discretization [32]:

uj−ivj≃−2πiN∑l=0j+l oddN−1γlcoth⁡[π(zj−zl)].(31)u_j-\mathrm{i}v_j \simeq-\frac{2\pi\mathrm{i}}{N}\sum_{\substack{l=0\\j+l\ \text{odd}}}^{N-1}\gamma_l\coth\left[\pi(z_j-z_l)\right]. \tag*{(31)}

Derivatives κα\kappa_\alpha and yαy_\alpha are computed at each time step using second-order centered finite differences, and α\alpha is defined by α(j)=2πj/N\alpha(j)=2\pi j/N, where j∈[0,N]j\in[0,N] and N=1024N=1024 is the number of points describing the periodic surface. Note that the numerical effort of evaluating (31) requires O(N2)O(N^2) operations, and thus will be the limiting factor of our algorithm.

Stabilization

Although the character of the non-local operator in this example is very different from previous equations, our numerical stabilization works in a fashion that is remarkably similar. The modified scheme (4) now becomes

x^kn+1=x^kn+u^knδt−1+λ(k),y^kn+1=y^kn+v^knδt−1+λ(k),(32)\hat{x}_k^{n+1}=\hat{x}_k^n+\frac{\hat{u}_k^n}{\delta t^{-1}+\lambda(k)},\qquad\hat{y}_k^{n+1}=\hat{y}_k^n+\frac{\hat{v}_k^n}{\delta t^{-1}+\lambda(k)}, \tag*{(32)}

where u^n\hat{u}^n and v^n\hat{v}^n are calculated from the Fourier transform of (31). The new grid points xjn+1,yjn+1x_j^{n+1},y_j^{n+1} are obtained from the inverse Fourier transform of x^jn+1,y^jn+1\hat{x}_j^{n+1},\hat{y}_j^{n+1}, and for each component (32) is turned into a second-order scheme using (5).

In [13], we performed a linear analysis (7) of the discrete modes of (31) about a flat interface. We found that

e(k)=SN3L3(1−cos⁡2πkN)sin⁡2πkN≡e~(x)=(1−cos⁡x)sin⁡x,x=2πkN,(33)e(k)=\frac{SN^3}{L^3}\left(1-\cos\frac{2\pi k}{N}\right)\sin\frac{2\pi k}{N}\equiv\tilde{e}(x)=(1-\cos x)\sin x,\qquad x=\frac{2\pi k}{N}, \tag*{(33)}

where LL is the length of the interface, and NN the number of gridpoints. As before, we use the long-wavelength approximation to e(k)e(k):

e(k)≈S2L3(2πk)3≡es(x)=x32,(34)e(k)\approx\frac{S}{2L^3}(2\pi k)^3\equiv e_s(x)=\frac{x^3}{2}, \tag*{(34)}

to find the explicit stability boundary (9) as

ke≈L2π(4Sδt)1/3.(35)k_e\approx\frac{L}{2\pi}\left(\frac{4}{S\delta t}\right)^{1/3}. \tag*{(35)}

Using the same approximation (34), which overpredicts the critical value λ(k)\lambda(k), one finds

λc(k)=S3(2πkL)3(36)\lambda_c(k)=\frac{S}{3}\left(\frac{2\pi k}{L}\right)^3 \tag*{(36)}

as a sufficient condition for stability. In [13], we used a fixed spectrum λ(k)\lambda(k), slightly larger than (36), to stabilize the Hele-Shaw dynamics.

We now use the same procedure as before, with the upper bound ϵu=10−10\epsilon_u=10^{-10} for ϵ(k)\epsilon(k). Figure 6 shows our adaptive scheme at work, as the interface (shown on the top row) deforms, and the length LL of the interface increases (See supplementary movie HS_movie.mpeg). The error ϵ(k)\epsilon(k) is defined as the maximum of (11) over the two components:

ϵ(k)=MAX⁡(E^kx−Ekx,E^ky−Eky).(37)\epsilon(k)=\operatorname{MAX}\left(\hat{E}_k^x-E_k^x,\hat{E}_k^y-E_k^y\right). \tag*{(37)}

Initializing λ(k)\lambda(k) to the approximation (36) (red line), we observe the same convergence toward the theoretical stability boundary as before (dotted line). A new feature is that on account of the length LL of the boundary increasing in time, the explicit stability boundary kek_e (vertical dashed and orange lines) increases in time, and the theoretical stability boundary for λ(k)\lambda(k) (dotted line) comes down. As seen in the second panel of Fig. 6, our adaptive scheme for $k>k_e has converged toward the theoretical prediction, and then continues to trace it as LL increases. The region where no stabilization is required increases as well, and λ(k)\lambda(k) decreases to very low values on an increasingly large domain. The results described above are not changed significantly as ϵu\epsilon_u is varied over several orders of magnitude up or down from 10−1010^{-10}, but of course the value must be significantly over the rounding error, and below the expected truncation error.

In our earlier EIN scheme [13], we used λ(k)\lambda(k) based on the simplified stability boundary (36) to stabilize the Hele-Shaw interface motion shown in Figure 6. However, this overpredicts the necessary damping for large kk. In addition, for k<kek<k_e, no damping is necessary, and our adaptive scheme reflects that by decreasing λ(k)\lambda(k) more and more. As a result, the damping in the adaptive scheme is significantly smaller than in our previous EIN scheme. In Figure 7 we show a comparison of the numerical results to those of the earlier scheme, and find very good agreement. The major advance is of course that λ(k)\lambda(k) no longer needs to be prescribed, but is found self-consistently as part of the algorithm which ensures stability. Only in the last panel is there a significant discrepancy between the two results. This occurs in places where two sides of the interface have come in close proximity, comparable to the spacing between grid points. But this means our evaluation of the velocity integral is no longer sufficiently accurate to be reliable.

Figure 7. A comparison of the interface as obtained from our current adaptive scheme (red curves) and our earlier EIN scheme (black curves) [13], which used the theoretical stability boundary (36).

Outlook and conclusions

We have demonstrated the feasibility of our method using three different model problems, highlighting different aspects of physical problems containing a wide spectrum of time scales, making them stiff. Clearly, there are many ways in which to extend and improve the present approach. Firstly, we estimated the damping spectrum by analyzing the current solution in Fourier space, which is particularly easy for the periodic domain considered by us. However, this may be circumvented by periodically continuing a solution defined over a finite domain only. In addition, one could formulate the entire method in real space, as done in some cases described in [13].

A second, more important issue is our assumption of the spectrum e(k)e(k) in (7) being real. This assumption is well founded, since the ultimate physical damping process is dissipative, leading to real eigenvalues. However, as demonstrated by the example of an inertial vortex sheet considered in [9], even problems lacking dissipation can display significant stiffness. This case leads to a system of PDEs, with pairs of complex eigenvalues e(k)e(k) on the right-hand-side of (7), corresponding to traveling waves. In that case the damping spectrum λ(k)\lambda(k) would also have to be complex to ensure stability [13], a case we have not yet considered.

Finally, a problem we still need to address is how to choose an initial condition for the damping spectrum λ(k)\lambda(k). In the present work we choose a power-law spectrum which can be inferred from a simple analysis of the continuum version of the equations of motion, which then adapts to an optimal spectrum. It would be ideal if no input whatsoever was necessary, choosing for example λ(k)=0\lambda(k) = 0 initially. At present, this is not possible, as the quality of the numerical solution deteriorates before λ(k)\lambda(k) can adapt. We suspect that in order for such a scheme to be successful, one needs to implement a variable time step, such that initial steps during which λ(k)\lambda(k) is found are very small.

In conclusion, following our previous study on this subject, we propose a new method to remove the stiffness of PDEs containing non-linear stiff terms, i.e. high spatial derivatives embedded into non-linear terms. This method allows for the self-consistent estimation of a stabilizing term on the right-hand-side of the PDE, that ensures absolute stability for the numerical scheme. Analyzing the spectrum of the solution at each time step, we adapt automatically the stabilizing term such that each unstable Fourier mode is damped optimally. The computational cost of this method is essentially the same as that of the explicit method.

References

  1. [1]S. Popinet, An accurate adaptive solver for surface-tension-driven interfacial flows, J. Comp. Phys. 228 (2009) 5838–5866.DOI
  2. [2]J. Brackbill, D. B. Kothe, C. Zemach, A continuum method for modeling surface tension, Journal of computational physics 100 (1992) 335–354.DOI
  3. [3]S. Popinet, Numerical models of surface tension, Annual Review of Fluid Mechanics 50 (2018) 49–75.DOI
  4. [4]M. Ulvřová, S. Labrosse, N. Coltice, P. Raback, P. Tackley, Numerical modelling of convection interacting with a melting and solidification front: Application to the thermal evolution of the basal magma ocean, Physics of the Earth and Planetary Interiors 206–207 (2012) 51 – 66.DOI
  5. [5]A.-K. Kassam, L. Trefethen, Fourth-order time-stepping for stiff pdes, SIAM J. Sci. Comput. 26 (2005) 1214–1233.DOI
  6. [6]A. Iserles, A first course in the numerical analysis of differential equations, 44, Cambridge university press, 2009.
  7. [7]W. F. Ames, Numerical methods for partial differential equations, Academic press, 2014.
  8. [8]C. Pozrikidis, Boundary Integral and singularity methods for linearized flow, Cambridge University Press, Cambridge, 1992.DOI
  9. [9]T. Hou, J. Lowengrub, M. Shelley, Removing the stiffness from interfacial flows with surface tension, J. Comp. Physics 114 (1994) 312–338.DOI
  10. [10]U. M. Ascher, S. J. Ruuth, B. T. R. Wetton, Implicit-explicit methods for time-dependent partial differential equations, SIAM J. Numer. Anal. 32 (1995) 797.
  11. [11]D. R. D. Rann, P. N. Blossey, Implicit-explicit multistep methods for fast-wave-slow-wave problems, Monthly Weather Rev. 140 (2012) 1307.
  12. [12]J. Eggers, J. R. Lister, H. A. Stone, Coalescence of liquid drops, J. Fluid Mech. 401 (1999) 293–310.DOI
  13. [13]L. Duchemin, J. Eggers, The explicit-implicit-null method: removing the numerical instability of pdes, Journal of Computational Physics 263 (2014) 37–52.DOI
  14. [14]P. Smereka, Semi-implicit level set methods for curvature and surface diffusion motion, J. Sci. Comput. 19 (2002) 439–456.DOI
  15. [15]K. Glasner, A diffuse interface approach to hele-shaw flow, Nonlinearity 16 (2003) 49–66.DOI
  16. [16]D. Salac, W. Lu, A local semi-implicit level-set method for interface motion, J. Sci. Comput. 35 (2008) 330–349.DOI
  17. [17]C. B. MacDonald, S. J. Ruuth, The implicit closest point method for the numerical solution of partial differential equations on surfaces, SIAM J. Sci. Comput. 31 (2009) 4330–4350.DOI
  18. [18]B. P. Ayati, T. F. Dupont, Convergence of a step-doubling galerkin method for parabolic problems, Math. Comput. 74 (2004) 1053–1065.DOI
  19. [19]M. Frigo, S. G. Johnson, The design and implementation of FFTW3, Proceedings of the IEEE 93 (2005) 216–231. Special issue on “Program Generation, Optimization, and Platform Adaptation”.DOI
  20. [20]E. Hairer, S. P. Nørsett, G. Wanner, Solving ordinary differential equations I: nonstiff problems, volume 8, Springer Science & Business Media, 2008.
  21. [21]J. Douglas Jr., T. F. Dupont, Alternating-direction Galerkin methods on rectangles, in: B. Hubbard (Ed.), Numerical Solution of Partial Differential Equations II, Academic Press, 1971, pp. 133–214.DOI
  22. [22]P. Concus, G. H. Golub, Use of fast direct methods for the efficient numerical solution of nonseparable elliptic equations, SIAM J. Numer. Anal. 10 (1973) 1103–1120.DOI
  23. [23]A. Wathen, D. Silvester, Fast iterative solution of stabilised stokes systems. part i: Using simple diagonal preconditioners, SIAM Journal on Numerical Analysis 30 (1993) 630–649.DOI
  24. [24]H. C. Elman, D. J. Silvester, A. J. Wathen, Finite elements and fast iterative solvers: with applications in incompressible fluid dynamics, Numerical Mathematics and Scientific Computation, 2014.
  25. [25]M. B. Williams, S. H. Davis, Nonlinear theory of film rupture, Journal of Colloid and Interface Science 90 (1982) 220–228.DOI
  26. [26]W. W. Zhang, J. R. Lister, Similarity solutions for van der waals rupture of a thin film on a solid substrate, Physics of Fluids 11 (1999) 2454–2462.DOI
  27. [27]W. Press, Numerical recipes : the art of scientific computing, Cambridge University Press, Cambridge, UK New York, 2007.
  28. [28]M. C. Cross, P. C. Hohenberg, Pattern formation outside of equilibrium, Rev. Mod. Phys. 65 (1993) 851–1112.DOI
  29. [29]A. Kalogirou, E. E. Keaveny, D. T. Papageorgiou, An in-depth numerical study of the two-dimensional kuramoto–sivashinsky equation, Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences 471 (2015) 20140932.pubmed.ncbi.nlm.nih.gov/26345218
  30. [30]P. G. Drazin, W. H. Reid, Hydrodynamic stability, Cambridge University Press, Cambridge, 1981.
  31. [31]A. J. Majda, A. L. Bertozzi, Vorticity and Incompressible Flow, Cambridge University Press, Cambridge, 2002.DOI
  32. [32]M. Shelley, A study of singularity formation in vortex sheet motion by a spectrally accurate vortex method, J. Fluid Mech. 244 (1992) 493.DOI

Paper details

vlnc:W_2dpx7x8s
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

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