betaFeedback

WorksPhysics and Astronomy

A Fast second-order solver for stiff multifluid dust and gas hydrodynamics

Abstract

We present MDIRK: a Multifluid second-order Diagonally-Implicit Runge-Kutta method to study momentum transfer between gas and an arbitrary number (NN) of dust species. The method integrates the equations of hydrodynamics with an Implicit Explicit (IMEX) scheme and solves the stiff source term in the momentum equation with a diagonally-implicit asymptotically stable Runge-Kutta method (DIRK). In particular, DIRK admits a simple analytical solution that can be evaluated with O(N)\mathcal{O}(N) operations, instead of standard matrix inversion, which is O(N)3\mathcal{O}(N)^3. Therefore the analytical solution significantly reduces the computational cost of the multifluid method, making it suitable for studying the dynamics of systems with particle-size distributions. We demonstrate that the method conserves momentum to machine precision and converges to the correct equilibrium solution with constant external acceleration. To validate our numerical method we present a series of simple hydrodynamic tests, including damping of sound waves, dusty shocks, a multi-fluid dusty Jeans instability, and a steady-state gas-dust drift calculation. The simplicity of MDIRK lays the groundwork to build fast high-order asymptotically stable multifluid methods.

Introduction

Theories of star and planet formation are grounded in the study of astrophysical fluids. Consequently, many decades of research have been devoted to the development of algorithms to capture their non-linear evolution [33, 52]. For a subset of systems, however, it is insufficient to consider only the evolution of a single species. For example, in cold systems, dust is often the dominant opacity source [43, 13, 10], and thus dictates the thermal evolution of the gas. However, for a range of gas densities and dust particle sizes, the coupling between molecular gas and dust is imperfect. The density ratio between gas and different solids may not be spatially or temporally constant, leading to non-negligible relative velocities, and thus momentum exchange [14, 54, 1]. These

multi-species systems are susceptible to a wide range of multi-fluid instabilities that can substantially alter the evolutionary pathway of a system [57, 48, 26, 22]. A corollary of the fact that dust controls the thermodynamic behavior of these systems is that it also dramatically impacts the observational signatures of these objects, especially in continuum emission at millimeter - infrared wavelengths, and scattered light at optical wavelengths. Thus accurately predicting the non-uniform distribution of dust grains of various sizes is also vital for predicting and interpreting data from ALMA (e.g., [3]), SPHERE (e.g., [5]) and forthcoming JWST observations.

To fully capture the impact of dust on astrophysical systems, we require numerical methods that can tackle the problem of momentum exchange between NN species in an efficient manner. Several approaches have been developed to study the interaction between dust and gas numerically. So-called “Hybrid” integration schemes utilize Eulerian grids for the gas, but separate Lagrangian super particles for dust (e.g., Youdin & Johansen [56]; Balsara et al. [7]; Bai & Stone [6]; Miniati [39]; Yang & Johansen [55]; Mignone et al. [38]). In smoothed particle hydrodynamics (SPH), both gas and dust are treated as Lagrangian particles, with careful choice of the kernel for the drag terms (e.g., Laibe & Price [29, 30]; Booth et al. [11]; Price et al. [44]; Stoyanovskaya et al. [50]). Staggered Semianalytic Methods (e.g., Fung & Muley [16]), are best applied to tightly coupled distributions, where a staggered time step allows for accurate calculation of dust grain terminal velocity. Finally, there exists a range of Euler-Euler methods where all species utilize the same grids, (e.g., Johansen et al. [25]; Paardekooper & Mellema [41]; Hanasz et al. [19]) including first-order multifluid (Benítez-Llambay et al. [8]), second-order multifluid (Huang & Bai [23]), and single-fluid with modified energy equation (Chen & Lin [12]).

Solving for the momentum exchange between gas and particles requires special focus since it introduces source terms (and time scales) that modify the stiffness ratio to the system of equations of motion. These source terms are usually modeled as drag forces that depend on the relative velocity between gas and dust (in both the linear and non-linear regime), the dust-to-gas density ratio, and the characteristic stopping time (Whipple [54]; Adachi et al. [1]). In the limit of strong coupling (small stopping time relative to a characteristic time scale) and / or strong feedback (large dust-to-gas density ratio) the time-explicit numerical integration of drag forces requires an excessively small time step in order to preserve stability. Such small timesteps dramatically increase the computational cost of a given calculation.

Alternative methods that circumvent the reduction of the time step are time-implicit methods. However it is not straightforward to couple a time-implicit drag-force scheme with a time-explicit numerical method for hydrodynamics, (see e.g. Huang & Bai [23]). In general, this coupling requires a splitting of the hydrodynamics equations which, if not done properly, downgrades the precision of the numerical solution. Moreover, combined time-implicit and time-explicit methods must still convergence to the correct equilibrium solution. One example of a time-implicit method was presented in Benítez-Llambay et al. [8]. The method coupled a first-order time-implicit scheme for the drag force with an explicit first-order hydrodynamics solver. This multi-species numerical solver is unconditionally and asymptotically stable and conserves momentum to machine precision. Moreover, the complexity of the scheme is order O(N)O(N), with NN the number of fluids (Krapp & Benítez-Llambay [26]).

In this work, we present MDIRK: a novel multi-fluid second-order method that is asymptotically stable, conserves momentum to machine precision, and has complexity O(N)O(N). We solve the hydrodynamics equations a two-stage Runge-Kutta Implicit-Explicit (IMEX) method. In addition, we highlight the benefits of a stiff-accurate multifluid solver, and lay the groundwork to develop even high-order implicit schemes that retain complexity O(N)O(N).

This work is organized as follows. In § 2 we introduce the fluid equations for a multifluid mixture of gas and dust and showcase the numerical method. In § 3 we present the results from a test-suite and validate the properties of our second-order method. In § 4 we summarize the main results of our work and suggest possible extensions to achieve higher order accuracy.

Numerical Method

We now describe the numerical method for a 1D system. All of its important properties such as momentum conservation, convergence to the correct equilibrium, and order of accuracy are independent of the number of dimensions. Extensions to 2D and 3D are straightforward, because the drag force does not couple momentum in orthogonal directions.

Consider a system of N−1N-1 dust species that exchange momentum with an inviscid gas through drag forces. Each dust species is characterized by a density, velocity, and a characteristic stopping time tst_s, or equivalent a rate of momentum exchange with the gas, αd\alpha_d. Gas density and velocity are defined as ρg\rho_g and vgv_g, respectively. Equivalently, for the ii-th dust species we define the density and velocity as ρd,i\rho_{d,i} and vd,iv_{d,i}. We neglect the momentum exchange between dust species. We assume an isothermal equation of state and the gas pressure is defined as P=cs2ρgP = c_s^2\rho_g, with csc_s as the sound speed. Under these assumptions, the continuity and momentum equations for gas and dust species are

∂tρg+∂x(ρgvg)=0,(1)\partial_t \rho_g + \partial_x \left(\rho_g v_g\right) = 0, \tag*{(1)}
∂tρd,i+∂x(ρd,ivd,i)=0,(2)\partial_t \rho_{d,i} + \partial_x \left(\rho_{d,i}v_{d,i}\right) = 0, \tag*{(2)}
∂t(ρgvg)+∂x(ρgvg2+P)=−∑i=1N−1ρgαg,i(vg−vd,i)+Gg,(3)\partial_t\left(\rho_gv_g\right) + \partial_x\left(\rho_gv_g^2 + P\right) = -\sum_{i=1}^{N-1}\rho_g\alpha_{g,i}\left(v_g-v_{d,i}\right) + G_g, \tag*{(3)}
∂t(ρd,ivd,i)+∂x(ρd,ivd,i2)=−ρd,iαd,i(vd,i−vg)+Gd,i,(4)\partial_t\left(\rho_{d,i}v_{d,i}\right) + \partial_x\left(\rho_{d,i}v_{d,i}^2\right) = -\rho_{d,i}\alpha_{d,i}\left(v_{d,i}-v_g\right) + G_{d,i}, \tag*{(4)}

with i=1,…,N−1i = 1,\ldots,N-1. The right-hand sides of Equations (3) and (4) contain the drag force terms and any external forces GgG_g and Gd,iG_{d,i}, respectively. The rates of momentum exchange are αg,i\alpha_{g,i} and αd,i\alpha_{d,i}. Momentum conservation implies

ρgαg,i=ρd,iαd,i,(5)\rho_g\alpha_{g,i}=\rho_{d,i}\alpha_{d,i}, \tag*{(5)}

therefore αg,i=ϵiαd,i\alpha_{g,i}=\epsilon_i\alpha_{d,i}, with ϵi=ρd,i/ρg\epsilon_i=\rho_{d,i}/\rho_g. The collision rates may depend on the density, sound speed, and relative velocity between species. To rewrite the momentum equations in matrix form, we define

u=(ρgvg,ρd,1vd,1,…,ρd,N−1vd,N−1),(6)\mathbf{u}=(\rho_gv_g,\rho_{d,1}v_{d,1},\ldots,\rho_{d,N-1}v_{d,N-1}), \tag*{(6)}
F=(ρgvg2+P,ρd,1vd,12,…,ρd,N−1vd,N−12),(7)\mathcal{F}=(\rho_gv_g^2+P,\rho_{d,1}v_{d,1}^2,\ldots,\rho_{d,N-1}v_{d,N-1}^2), \tag*{(7)}
G=(Gg,Gd,1,…,Gd,N−1),(8)\mathbf{G}=(G_g,G_{d,1},\ldots,G_{d,N-1}), \tag*{(8)}

now, Eqs. (3)-(4) can be cast into the form

∂tu+∂xF=Mu+G,(9)\partial_t\mathbf{u}+\partial_x\mathcal{F}=\mathbf{M}\mathbf{u}+\mathbf{G}, \tag*{(9)}

where M\mathbf{M} is the N×NN\times N matrix.

M=(−∑k=1N−1ϵkαd,kαd,1αd,2⋯αd,N−1ϵ1αd,1−αd,10⋯0ϵ2αd,20−αd,2⋯0⋮⋮⋮⋱⋮ϵN−1αd,N−100⋯−αd,N−1).(10)\mathbf{M}= \begin{pmatrix} -\sum_{k=1}^{N-1}\epsilon_k\alpha_{d,k} & \alpha_{d,1} & \alpha_{d,2} & \cdots& \alpha_{d,N-1}\\ \epsilon_1\alpha_{d,1} & -\alpha_{d,1} & 0 & \cdots& 0\\ \epsilon_2\alpha_{d,2} & 0 & -\alpha_{d,2} & \cdots& 0\\ \vdots& \vdots& \vdots& \ddots& \vdots\\ \epsilon_{N-1}\alpha_{d,N-1} & 0 & 0 & \cdots& -\alpha_{d,N-1} \end{pmatrix}. \tag*{(10)}

Note that in our case the matrix M\mathbf{M} is defined from the RHS of the momentum equation in terms of conserved variables, not primitives. This is a subtle distinction with the similar coupling matrix MBKP19\mathbf{M}_{\mathrm{BKP19}}, defined in equation (11) of Benítez-Llambay et al. [8] (see Appendix § 11). In this work the matrix M=−RMBKP19R−1\mathbf{M}=-\mathbf{R}\mathbf{M}_{\mathrm{BKP19}}\mathbf{R}^{-1}, where R\mathbf{R} a diagonal matrix such that Rii=ρiR_{ii}=\rho_i. The relation between these matrices, their eigenvectors, and eigenvalues will be utilized to explain the properties of the numerical method. Both matrices are similar to a real symmetric matrix and therefore diagonalizable. Moreover, if λk\lambda_k is an eigenvalue of M\mathbf{M}, then −λk-\lambda_k is an eigenvalue of MBKP19\mathbf{M}_{\mathrm{BKP19}}. Note that the eigenvalues of MBKP19\mathbf{M}_{\mathrm{BKP19}} are all real and non-negative (there is only one zero eigenvalue) (see Benítez-Llambay et al. [8]). It is also straightforward to show that if x\mathbf{x} is an eigenvector of MBKP19\mathbf{M}_{\mathrm{BKP19}}, then x~=−Rx\tilde{\mathbf{x}}=-\mathbf{R}\mathbf{x} is an eigenvector of the matrix M\mathbf{M}. Since 1T=(1,1,…,1)\mathbf{1}^{T}=(1,1,\ldots,1) is eigenvector of MBKP19\mathbf{M}_{\mathrm{BKP19}}, it follows that 1~T=−(ρg,ρd,1,…,ρd,N−1)\tilde{\mathbf{1}}^{T}=-(\rho_g,\rho_{d,1},\ldots,\rho_{d,N-1}) is eigenvector of M\mathbf{M}.

Implicit solutions of the drag-force term

Before introducing a numerical method to solve Eqs. (1)-(4), we will discuss three different alternatives to solve the momentum equation (9) neglecting all terms, except for the drag force. The three methods correspond to the first-order fully implicit method of Benítez-Llambay et al. [8], a second-order fully implicit method (see e.g., Huang & Bai [23]), and the diagonally implicit Runge-Kutta method (DIRK) utilized in this work. Our goal is to highlight the major differences between the three methods when solving the momentum equation

∂tu=Mu.(11)\partial_t\mathbf{u}=\mathbf{M}\mathbf{u}. \tag*{(11)}

We define the numerical solution at t=tn+Δtt=t_n+\Delta t of Eq. (5) as un+1\mathbf{u}^{n+1}. A first-order fully implicit solution can be obtained after solving

(I−ΔtM)un+1=un,(12)\left(\mathbf{I}-\Delta t\mathbf{M}\right)\mathbf{u}^{n+1}=\mathbf{u}^{n}, \tag*{(12)}

whereas a second-order fully implicit method can be cast as

(I−ΔtM+Δt22M2)un+1=un.(13)\left(\mathbf{I}-\Delta t\mathbf{M}+\frac{\Delta t^2}{2}\mathbf{M}^2\right)\mathbf{u}^{n+1}=\mathbf{u}^{n}. \tag*{(13)}

These two fully implicit methods converge to the correct equilibrium solution and are asymptotically stable, that is,

lim⁡Δt→∞∣un+1−u(Δt)∣=0.(14)\lim_{\Delta t\to\infty}\left|\mathbf{u}^{n+1}-\mathbf{u}(\Delta t)\right|=0. \tag*{(14)}

In particular, utilizing the properties of the matrix M\mathbf{M} it can be shown that the second-order method asymptotically converges as lim⁡Δt→∞∣un+1−u(Δt)∣∼1/Δt2\lim_{\Delta t\to\infty}\left|\mathbf{u}^{n+1}-\mathbf{u}(\Delta t)\right|\sim1/\Delta t^2, whereas the first-order method converges as lim⁡Δt→∞∣un+1−u(Δt)∣∼1/Δt\lim_{\Delta t\to\infty}\left|\mathbf{u}^{n+1}-\mathbf{u}(\Delta t)\right|\sim1/\Delta t. Therefore, the fully implicit second-order method provides a more accurate solution for all Δt\Delta t.

The advantage of the first-order method is that no matrix inversion is required to obtain the solution (Krapp & Benítez-Llambay [26]). The matrix inversion can add a complexity O(N3)\mathcal{O}(N^3) to the problem, making the second-order method ∼×N2\sim\times N^2 more expensive than the first-order method. The extension of the analytical solution of Krapp & Benítez-Llambay [26] to the second-order method is challenging since (I−ΔtM+Δt22M2)\left(\mathbf{I}-\Delta t\mathbf{M}+\frac{\Delta t^2}{2}\mathbf{M}^2\right) is not a sparse matrix.

Ideally, we would like a numerical solution that takes advantage of the reduced complexity of the first-order method and preserves the accuracy of the fully implicit second-order method. Thus, in this work, we study and test the properties of a second-order diagonally implicit Runge-Kutta (DIRK) method of the form

(I−γΔtM)k1=Mun(15)(\mathbf{I}-\gamma\Delta t\mathbf{M})\mathbf{k}_1=\mathbf{M}\mathbf{u}^n \tag*{(15)}
(I−γΔtM)k2=M(un+(1−γ)Δtk1),(16)(\mathbf{I}-\gamma\Delta t\mathbf{M})\mathbf{k}_2=\mathbf{M}\left(\mathbf{u}^n+(1-\gamma)\Delta t\mathbf{k}_1\right), \tag*{(16)}
un+1=un+(1−γ)Δtk1+γΔtk2.(17)\mathbf{u}^{n+1}=\mathbf{u}^n+(1-\gamma)\Delta t\mathbf{k}_1+\gamma\Delta t\mathbf{k}_2. \tag*{(17)}

where γ\gamma is known as the singular value of the DIRK method Alexander [2]. This two-stage method is second-order accurate (see Appendix § 6). Moreover, the solution of each stage requires the inversion of the matrix (I−γΔtM)(\mathbf{I}-\gamma\Delta t\mathbf{M}) and therefore an extension of the analytical solution of Krapp & Benítez-Llambay [26] can be utilized to reduce the complexity to order O(N)\mathcal{O}(N). We will show the analytical solution of the DIRK method in § 2.2. We also emphasize that coupling the solution (17) with other numerical methods for hydrodynamics is not a trivial task, as standard splitting may fail to converge to the correct equilibrium. In § 2.2 we describe the Implicit-Explicit Runge-Kutta method from Ascher et al. [4] to solve the full system of momentum equations (3)-(4) utilizing the solution (17). Continuity equations are then included in the MDIRK method described in § 2.3.

On the value of γ\gamma and stiff condition

Alternative versions of the method of Alexander [2] can be obtained by varying the coefficients of the Runge-Kutta. Depending on the coefficients, the value of γ\gamma is obtained either by the order conditions (e.g., second-order accuracy as Δt→0\Delta t \to0 and/or as Δt→∞\Delta t \to\infty) or by the asymptotic stability condition lim⁡Δt→∞∣un+1−u(Δt)∣→0\lim_{\Delta t\to\infty}\lvert\mathbf{u}^{n+1}-\mathbf{u}(\Delta t)\rvert\to0. Since the form of Eqs. (15)–(17) provides an asymptotically stable solution for any value of γ\gamma (see Appendix § 7), we will use the order conditions to set its value.

When Δt→0\Delta t \to0, the Taylor series of un+1\mathbf{u}^{n+1} should match the Taylor series of the solution of Eq. (11) up to second-order, which implies (2−γ)γ=1/2(2-\gamma)\gamma=1/2, i.e., γ=1±1/2\gamma=1\pm1/\sqrt{2} (see Appendix § 6 for details). With this choice of γ\gamma, the leading term in the difference between the two Taylor series goes as γ2(2γ−3)\gamma^2(2\gamma-3). Therefore, the smallest error is obtained for γ=1−1/2\gamma=1-1/\sqrt{2}. However, we will show in § 3.1 that the choice of γ=1−1/2\gamma=1-1/\sqrt{2} can undershot the solution. While γ=1+1/2\gamma=1+1/\sqrt{2} provides a less accurate (but still second-order) solution, it ensures a monotonic convergence, which is a numerically desirable property.

Instead of utilizing the Taylor series at Δt→0\Delta t\to0, one could also set the value of γ\gamma by matching the series expansion of un+1\mathbf{u}^{n+1} and u(Δt)\mathbf{u}(\Delta t) at Δt→∞\Delta t\to\infty. It can be shown that un+1\mathbf{u}^{n+1} asymptotically converges to the correct solution as ∼(2γ−1)/γ2λk−1Δt−1+(2−3γ)/γ3λk−2Δt−2\sim(2\gamma-1)/\gamma^2\lambda_k^{-1}\Delta t^{-1}+(2-3\gamma)/\gamma^3\lambda_k^{-2}\Delta t^{-2}, for some λk\lambda_k (non-zero) eigenvalue of the matrix M\mathbf{M}. In this

case, second-order accuracy is obtained with γ=1/2\gamma=1/2, otherwise, the method will converge as a first-order method.

We note that the two order conditions can not be simultaneously satisfied by the same value of γ\gamma. Therefore, there must be a criterion condition that sets the value of γ\gamma for a given time step Δt\Delta t. We define this criterion as,

γ={1/2if Δt>max⁡(αd,i−1),1±1/2otherwise,(18)\gamma= \begin{cases} 1/2 & \text{if }\Delta t>\max(\alpha_{d,i}^{-1}),\\ 1\pm1/\sqrt{2} & \text{otherwise}, \end{cases} \tag*{(18)}

where max⁡(αd,i−1)\max(\alpha_{d,i}^{-1}) is the largest stopping time among the dust species. The choice of Δt>max⁡(αd,i−1)\Delta t>\max(\alpha_{d,i}^{-1}) is only approximate since formally one should look into the non-zero eigenvalues of the matrix M\mathbf{M}. For instance, the timescale for the condition (18) could be better defined in terms of the largest eigenvalue (non-zero and smallest in absolute value) of the matrix M\mathbf{M}, e.g., Δt>∣max⁡(λk)∣−1\Delta t>\lvert\max(\lambda_k)\rvert^{-1}. However, it can be shown that ∣max⁡(λk)∣−1\lvert\max(\lambda_k)\rvert^{-1} is bound between the two largest stopping times of the system (see Appendix D Benítez-Llambay et al. [8]). This justify our choice of Δt>max⁡(αd,i−1)\Delta t>\max(\alpha_{d,i}^{-1}). As we describe below, when Δt∼max⁡(αd,i−1)\Delta t\sim\max(\alpha_{d,i}^{-1}) the truncation error is in the asymptotic regime. We therefore define the system of Eqs. (1)–(4) as a “stiff system” when Δt≥max⁡(αd,i−1)\Delta t\geq\max(\alpha_{d,i}^{-1}).

To highlight the definition of the stiff condition and further compare the methods described in this section, we solve Eq. (11) for a system with N=17N=17 fluids (16 dust species). For each dust fluid, the dust-to-gas mass ratios are obtained as ϵk=ϵ(ts,k+11/2−ts,k1/2)/(ts,min⁡1/2−ts,max⁡1/2)\epsilon_k=\epsilon(t_{s,k+1}^{1/2}-t_{s,k}^{1/2})/(t_{s,\min}^{1/2}-t_{s,\max}^{1/2}), and ϵ=1.0\epsilon=1.0. The stopping time ts,k=αd,k−1t_{s,k}=\alpha_{d,k}^{-1} is uniformly spaced in logarithmic space, with ts,min⁡=10−3t_{s,\min}=10^{-3} and ts,max⁡=10t_{s,\max}=10. Note that ts,max⁡=max⁡(αd,i−1)t_{s,\max}=\max(\alpha_{d,i}^{-1}) in Eq. (18). In Figure 1 we show the truncation error, e(Δt)e(\Delta t), for the three different methods described in this section. The error is obtained as

Image not extracted

Figure 1. Comparison of the (local) truncation error for different methods described in § 2.1. The blue solid line corresponds to the first-order implicit method of Eq. (12) whereas the red solid line shows the error for the second-order implicit method of Eq. (13). The solutions obtained with the DIRK method are shown in blue and red dashed lines. The black solid line indicates the transition to a stiff regime where γ\gamma is set to γ=1/2\gamma= 1/2 for the DIRK method. For Δt≳0.002\Delta t \gtrsim0.002 (vertical dotted line) time explicit methods will be unstable. The analytical solution, uk(Δt)u_k(\Delta t), is obtained by solving the eigenvalue problem of the linear equation (11).

e(Δt)=1N∑k∣ukn+1−uk(Δt)∣.(19)e(\Delta t)=\frac{1}{N}\sum_k\left\lvert u_k^{n+1}-u_k(\Delta t)\right\rvert. \tag*{(19)}

As Δt→0\Delta t\to0, the difference between the Taylor series implies that e(Δt)∼Δt2e(\Delta t)\sim\Delta t^2 and e(Δt)∼Δt3e(\Delta t)\sim\Delta t^3 for the first and second-order methods, respectively. These scaling are shown in Figure 1 for Δt≲0.001\Delta t\lesssim0.001. As already mentioned in this section, the DIRK method is more accurate for γ=1−1/2\gamma=1-1/\sqrt{2} than γ=1+1/2\gamma=1+1/\sqrt{2}. In the regime where ts,min⁡≲Δt≲ts,max⁡t_{s,\min}\lesssim\Delta t\lesssim t_{s,\max}, the DIRK method shows an error comparable to that obtained with the first-order implicit method. Similarly, the second-order fully implicit method is also more accurate in this regime.

As the time-step increases the system reaches the stiff regime where Δt≥ts,max⁡=10\Delta t \geq t_{s,\max}=10. In this regime, the numerical solutions show the expected asymptotic convergences e(Δt)∼Δt−1e(\Delta t) \sim\Delta t^{-1} and e(Δt)∼Δt−2e(\Delta t) \sim\Delta t^{-2} for the first and second-order methods, respectively. If condition (18) is not applied to the DIRK method, i.e. γ=1±1/2\gamma= 1 \pm1/\sqrt{2} for all Δt\Delta t, the asymptotic convergence will be that of the first-order method. Therefore, without condition (18) for γ\gamma there is not a clear advantage of utilizing a DIRK method in the stiff regime.

Second-order Implicit-Explicit Method (IMEX)

In this section, we describe the Runge-Kutta IMEX numerical method that will be utilized to solve the hydrodynamics equations. We begin by focusing on the implementation of the more complex momentum equation; the continuity equation is shown in § 2.3. The novelty of this second-order Runge-Kutta scheme, from Ascher et al. [4], is that it incorporates an explicit term and implicit term together in a single time update. As we will demonstrate, the method conserves momentum to machine precision, is asymptotically stable (L-stable), and converges to the correct equilibrium. We can write Eq. (9) as,

∂tu=[H(u)]exp+[Mu]imp(20)\partial_t \mathbf{u} = [\mathcal{H}(\mathbf{u})]_{\mathrm{exp}} + [\mathbf{M}\mathbf{u}]_{\mathrm{imp}} \tag*{(20)}

with H(u)=−∂xF+G\mathcal{H}(\mathbf{u})=-\partial_x\mathbf{F}+\mathbf{G}. The brackets [ ]exp[\ ]_{\mathrm{exp}} and [ ]imp[\ ]_{\mathrm{imp}} indicate which terms are time-explicit and time-implicit, respectively. For our particular system of equations, the term solved with a diagonally implicit method admits an analytical solution [26], which transforms the complexity of the method from O(N3)O(N^3) to O(N)O(N), making it suitable for a system with a large number of species. Consider the momenta vector u(tn)=un\mathbf{u}(t_n)=\mathbf{u}^n. The full update to a new time t=tn+Δtt=t_n+\Delta t is obtained as follows:

un+1=un+δΔtH(un)+(1−δ)ΔtH(w1)+b1ΔtMw1+b2ΔtMw2(21)\mathbf{u}^{n+1}=\mathbf{u}^n+\delta\Delta t\mathcal{H}(\mathbf{u}^n)+(1-\delta)\Delta t\mathcal{H}(\mathbf{w}_1) +b_1\Delta t\mathbf{M}\mathbf{w}_1+b_2\Delta t\mathbf{M}\mathbf{w}_2 \tag*{(21)}

where the coefficients γ~\widetilde{\gamma}, δ\delta and β\beta are defined in Eq. (30). The vectors w1\mathbf{w}_1 and w2\mathbf{w}_2 are the solution of the following equations:

w1=un+γ~ΔtH(un)+γΔtMw1(22)\mathbf{w}_1=\mathbf{u}^n+\widetilde{\gamma}\Delta t\mathcal{H}(\mathbf{u}^n)+\gamma\Delta t\mathbf{M}\mathbf{w}_1 \tag*{(22)}
w2=un+δΔtH(un)+(1−δ)ΔtH(w1)+βΔtMw1+γΔtMw2(23)\mathbf{w}_2=\mathbf{u}^n+\delta\Delta t\mathcal{H}(\mathbf{u}^n)+(1-\delta)\Delta t\mathcal{H}(\mathbf{w}_1) +\beta\Delta t\mathbf{M}\mathbf{w}_1+\gamma\Delta t\mathbf{M}\mathbf{w}_2 \tag*{(23)}

Alternatively, the solution can be also obtained in terms of k1=Mw1\mathbf{k}_1=\mathbf{M}\mathbf{w}_1 and k2=Mw2\mathbf{k}_2=\mathbf{M}\mathbf{w}_2, by applying the M\mathbf{M} matrix to Eqs. (22)–(23).

(I−γΔtM)k1=M(un+γ~ΔtH(un))(24)(\mathbf{I}-\gamma\Delta t\mathbf{M})\mathbf{k}_1=\mathbf{M}\left(\mathbf{u}^n+\widetilde{\gamma}\Delta t\mathcal{H}(\mathbf{u}^n)\right) \tag*{(24)}
(I−γΔtM)k2=M(un+δΔtH(un)+βΔtk1)+M((1−δ)ΔtH(w1))(25)(\mathbf{I}-\gamma\Delta t\mathbf{M})\mathbf{k}_2=\mathbf{M}\left(\mathbf{u}^n+\delta\Delta t\mathcal{H}(\mathbf{u}^n)+\beta\Delta t\mathbf{k}_1\right) +\mathbf{M}\left((1-\delta)\Delta t\mathcal{H}(\mathbf{w}_1)\right) \tag*{(25)}

The major advantage of the diagonally implicit method is that one first solves for k1\mathbf{k}_1 and then for k2\mathbf{k}_2, which simplifies the calculations. Note that this is not the case in fully implicit methods since Eq. (24) will also depend on k2\mathbf{k}_2. Moreover, we can find the analytical solution of the linear system, which has a complexity O(N)O(N) (c.f. [26]). For instance, consider the arbitrary vector qn=(qgn,qd,1n,…,qd,N−1n)\mathbf{q}^n=(q^n_g,q^n_{d,1},\ldots,q^n_{d,N-1}) and the equation

(I−γΔtM)kn=Mqn(26)(\mathbf{I}-\gamma\Delta t\mathbf{M})\mathbf{k}^n=\mathbf{M}\mathbf{q}^n \tag*{(26)}

the solution kn=(kgn,kd,1n,…,kd,N−1n)\mathbf{k}^n=(k^n_g,k^n_{d,1},\ldots,k^n_{d,N-1}) has the components

kgn=A−qgnB1+γΔtB(27)k^n_g=\frac{\mathcal{A}-q^n_g\mathcal{B}}{1+\gamma\Delta t\mathcal{B}} \tag*{(27)}
kd,in=αin1+γΔtαin(ϵinqgn−qd,in+γΔtϵinkgn)(28)k^n_{d,i}=\frac{\alpha^n_i}{1+\gamma\Delta t\alpha^n_i}\left(\epsilon^n_iq^n_g-q^n_{d,i}+\gamma\Delta t\epsilon^n_i k^n_g\right) \tag*{(28)}

for i=1,…,N−1i=1,\ldots,N-1, with

A=∑k=1N−1αd,knqd,kn1+γΔtαd,kn,B=∑k=1N−1ϵknαd,kn1+γΔtαd,kn.(29)\mathcal{A}=\sum_{k=1}^{N-1}\frac{\alpha_{d,k}^{n}q_{d,k}^{n}}{1+\gamma\Delta t\alpha_{d,k}^{n}},\qquad\mathcal{B}=\sum_{k=1}^{N-1}\frac{\epsilon_{k}^{n}\alpha_{d,k}^{n}}{1+\gamma\Delta t\alpha_{d,k}^{n}}. \tag*{(29)}

It can be shown that γ=1±1/2\gamma=1\pm1/\sqrt{2} are the only two possible values for γ\gamma that provide an asymptotically stable and second-order accurate solution (for Δt→0\Delta t\to0) with two-stages [45, 2]. The rest of the coefficients are chosen to satisfy the order condition and asymptotic behavior of the solution. In this work, we adopt the values from [4] (see also Appendix § 6)

γ~=γ,δ=1−12γ,β=1−γ,\tilde{\gamma}=\gamma,\qquad\delta=1-\frac{1}{2\gamma},\qquad\beta=1-\gamma,
b1=1−γ,b2=γ.(30)b_{1}=1-\gamma,\qquad b_{2}=\gamma. \tag*{(30)}

An alternative version of the second-order Implicit-Explicit Runge-Kutta can be found in Pareschi & Russo [42] with coefficients γ~=1\tilde{\gamma}=1, δ=1/2\delta=1/2, β=1−2γ\beta=1-2\gamma, b1=1/2b_{1}=1/2 and b2=1/2b_{2}=1/2. This alternative version has an explicit part that matches the optimal second-order Runge-Kutta derived in Shu & Osher [47], with important Total-Variation-Diminishing (TVD) properties. This TVD property also holds for the Runge-Kutta utilized in this work when γ=1+1/2\gamma=1+1/\sqrt{2}, provided a Courant number C≲0.58C\lesssim0.58.

MDIRK

To solve the fluid equations we have reduced the Runge-Kutta described in § 2.2 into a more convenient form, taking into account the continuity equation as well. To describe the algorithm we define the discrete vectors at time tnt_n as Ugn=(ρgn,ρgnvgn)=(Ug0n,Ug1n)\mathbf{U}_{g}^{n}=(\rho_{g}^{n},\rho_{g}^{n}\mathbf{v}_{g}^{n})=(U_{g0}^{n},U_{g1}^{n}) and Kgn=(0,kgn)\mathbf{K}_{g}^{n}=(0,\mathbf{k}_{g}^{n}) for the gas, and Ud,in=(ρd,in,ρd,invd,in)=(Ud0,in,Ud1,in)\mathbf{U}_{d,i}^{n}=(\rho_{d,i}^{n},\rho_{d,i}^{n}\mathbf{v}_{d,i}^{n})=(U_{d0,i}^{n},U_{d1,i}^{n}), Kd,in=(0,kd,in)\mathbf{K}_{d,i}^{n}=(0,\mathbf{k}_{d,i}^{n}) for the ii-th dust-species, with kgn\mathbf{k}_{g}^{n} and kd,in\mathbf{k}_{d,i}^{n} defined in Eq. (27) and Eq. (28), respectively. Therefore, Kg\mathbf{K}_{g} and Kg,i\mathbf{K}_{g,i} depend on the density and momentum of all the fluids. The zero components of these vectors ensure that the drag force is only considered in the momentum equation.

The flux of mass and momentum are included in the L\mathcal{L} operator, as well as external forces, except of course the drag force. To calculate the fluxes we utilize a standard Riemann solver (see Appendix § 9).

For each fluid2, a full update of the conserved variables from time step tnt_n to time step tn+1=tn+Δtt_{n+1}=t_n+\Delta t is obtained as follows

U∗=Un+γ~ΔtL(Un)\mathbf{U}^{*}=\mathbf{U}^{n}+\tilde{\gamma}\Delta t\mathcal{L}(\mathbf{U}^{n})
U∗∗=Un+(1−δ)ΔtL(U∗+γΔtK∗)\mathbf{U}^{**}=\mathbf{U}^{n}+(1-\delta)\Delta t\mathcal{L}(\mathbf{U}^{*}+\gamma\Delta t\mathbf{K}^{*})
Un+1=Up+(b1−β)ΔtK∗+b2ΔtKp,(31)\mathbf{U}^{n+1}=\mathbf{U}^{p}+(b_{1}-\beta)\Delta t\mathbf{K}^{*}+b_{2}\Delta t\mathbf{K}^{p}, \tag*{(31)}

with

Up=U∗∗+δγ~(U∗−Un)+βΔtK∗.(32)\mathbf{U}^{p}=\mathbf{U}^{**}+\frac{\delta}{\tilde{\gamma}}(\mathbf{U}^{*}-\mathbf{U}^{n})+\beta\Delta t\mathbf{K}^{*}. \tag*{(32)}

The coefficients γ~\tilde{\gamma}, δ\delta, β\beta, b1b_{1} and b2b_{2} are given in Eq. (30). The vectors K∗\mathbf{K}^{*} and Kp\mathbf{K}^{p} are obtained from the analytical solution given in Eqs. (27)–(28). When utilizing these analytical solutions, the densities and momenta must be taken from the partial updates U∗\mathbf{U}^{*} and Up\mathbf{U}^{p}, respectively. For example,

Kg∗=[0A−Ug1∗B1+γΔtB],(33)\mathbf{K}_{g}^{*}= \begin{bmatrix} 0\\ \frac{\mathcal{A}-U_{g1}^{*}\mathcal{B}}{1+\gamma\Delta t\mathcal{B}} \end{bmatrix}, \tag*{(33)}
Kd,i∗=αi∗1+γΔtαi∗([0ϵd1,i∗Ug1∗−Ud1,i∗]+γΔtϵi∗Kg∗).(34)\mathbf{K}_{d,i}^{*}=\frac{\alpha_{i}^{*}}{1+\gamma\Delta t\alpha_{i}^{*}} \left( \begin{bmatrix} 0\\ \epsilon_{d1,i}^{*}U_{g1}^{*}-U_{d1,i}^{*} \end{bmatrix} +\gamma\Delta t\epsilon_{i}^{*}\mathbf{K}_{g}^{*} \right). \tag*{(34)}

with

A=∑k=1N−1αd,k∗Ud1,k∗1+γΔtαd,k∗,B=∑k=1N−1ϵk∗αd,k∗1+γΔtαd,k∗.(35)\mathcal{A}=\sum_{k=1}^{N-1}\frac{\alpha_{d,k}^{*}U_{d1,k}^{*}}{1+\gamma\Delta t\alpha_{d,k}^{*}},\qquad \mathcal{B}=\sum_{k=1}^{N-1}\frac{\epsilon_{k}^{*}\alpha_{d,k}^{*}}{1+\gamma\Delta t\alpha_{d,k}^{*}}. \tag*{(35)}

Note that when the vector K\mathbf{K} is set to zero (no drag forces), second-order accuracy can be obtained by setting γ~=1\tilde{\gamma}=1 and δ=1/2\delta=1/2.

Momentum conservation

The drag-force between species is an internal force that cannot change the total momentum of the system. Our second-order solver has this property as well. Momentum conservation follows from

Ug1n+1+∑kUd1,kn+1=Ug1n+∑kUd1,kn+b1Δt(kg∗+∑kkd,k∗)+b2Δt(kgp+∑kkd,kp)=Ug1n+∑kUd1,kn.(36)U_{g1}^{n+1}+\sum_{k}U_{d1,k}^{n+1} = U_{g1}^{n}+\sum_{k}U_{d1,k}^{n} +b_{1}\Delta t\left(k_{g}^{*}+\sum_{k}k_{d,k}^{*}\right) +b_{2}\Delta t\left(k_{g}^{p}+\sum_{k}k_{d,k}^{p}\right) = U_{g1}^{n}+\sum_{k}U_{d1,k}^{n}. \tag*{(36)}

where for any arbitrary q\mathbf{q} (which for example can be taken as u∗\mathbf{u}^{*} or up\mathbf{u}^{p}), the vector k\mathbf{k} satisfies

kg+∑kkd,k=kg+∑kαk1+γΔtαk(ϵkqg−qd,k+γΔtϵkkg)=kg(1+γΔtB)−(A−qgB)=A−qgB1+γΔtB(1+γΔtB)−(A−qgB)=0.(37)\begin{aligned} k_{g}+\sum_{k}k_{d,k} = k_{g}+\sum_{k}\frac{\alpha_{k}}{1+\gamma\Delta t\alpha_{k}} \left(\epsilon_{k}q_{g}-q_{d,k}+\gamma\Delta t\epsilon_{k}k_{g}\right) \\ &= k_{g}(1+\gamma\Delta t\mathcal{B})-(\mathcal{A}-q_{g}\mathcal{B}) \\ &= \frac{\mathcal{A}-q_{g}\mathcal{B}}{1+\gamma\Delta t\mathcal{B}}(1+\gamma\Delta t\mathcal{B})-(\mathcal{A}-q_{g}\mathcal{B}) \\ &=0. \tag*{(37)} \end{aligned}

recalling A\mathcal{A} and B\mathcal{B} as defined in Eq. (29). We have also utilized the analytical solutions provided in Eqs. (27)–(28) (see § 2.2). The solution of Eq. (39) can be expressed as u(t)=∑k=0N−1cˉkxkeλkt\mathbf{u}(t)=\sum_{k=0}^{N-1}\bar{c}_k\mathbf{x}_k e^{\lambda_k t}. To compare with the work of Huang & Bai [23], we perform their tests B and C. Both tests consider one gas and two dust species with constant stopping times and homogeneous density and velocity backgrounds. The cases differ by a factor of 1000 in damping timescale. In terms of the velocity, the analytical solution corresponds to

vanalytical=vCM+c1eλ1t+c2eλ2t,(38)v_{\mathrm{analytical}}=v_{\mathrm{CM}}+c_1e^{\lambda_1t}+c_2e^{\lambda_2t}, \tag*{(38)}

Method Test

In this section, we present a test-suite that validates our second-order method, and illustrates that the computational costs scales linearly with number of species. We first compare our implicit second-order drag force solver with the first-order implicit solver of Benítez-Llambay et al. [8]. Next, we present four test cases using MDIRK method: damped sound waves, a multifluid shock, an idealized, multi-fluid Jeans instability, and steady-state drift in a 2D shearing box.

Damping

As we show in Appendix A, our drag force solver provides solutions that converge asymptotically to the velocity of the center of mass. We show this property and study the accuracy of the method by solving the momentum exchange between gas and two dust species. The problem reduces to solving the following equation for the momenta of all fluids

∂tu=Mu.(39)\partial_t\mathbf{u}=\mathbf{M}\mathbf{u}. \tag*{(39)}

for a given initial condition u0\mathbf{u}_0. After solving the eigenvalue problem Mxˉ=λxˉ\mathbf{M}\bar{\mathbf{x}}=\lambda\bar{\mathbf{x}}, the solution of Eq. (39) can be expressed as u(t)=∑k=0N−1cˉkxˉkeλkt\mathbf{u}(t)=\sum_{k=0}^{N-1}\bar{c}_k\bar{\mathbf{x}}_k e^{\lambda_k t}. To compare with the work of Huang & Bai [23], we perform their tests B and C. Both tests consider one gas and two dust species with constant stopping times and homogeneous density and velocity backgrounds. The cases differ by a factor of 1000 in damping timescale. In terms of the velocity, the analytical solution corresponds to

vanalytical=vCM+c1eλ1t+c2eλ2t,(40)v_{\mathrm{analytical}}=v_{\mathrm{CM}}+c_1e^{\lambda_1t}+c_2e^{\lambda_2t}, \tag*{(40)}

where vCMv_{\mathrm{CM}} is the velocity of the center of mass. The values of the eigenvalues λ1\lambda_1, λ2\lambda_2 and the coefficients c1c_1 and c2c_2 can be obtained from table 1 of Huang & Bai [23].

In Figure 2 we show a comparison between the numerical and the analytical solutions, together with the numerical error. We include the first-order solution obtained with the method of Benítez-Llambay et al. [8] (green circle curve). In addition, the relative error with respect to the analytic solution as a function of the time step is displayed in the rightmost panel of Figure 2. The error is obtained as

Image not extracted

Figure 2. From left to right we plot the evolution of the velocity for the gas and two dust species. The dashed line corresponds to the analytical solution, given by Eq. (39). The circles indicate the numerical solutions obtained with different integrators. Green color shows the first-order method, orange and dark lines the second-order method with γ=γ+=1+1/2\gamma=\gamma^{+}=1+1/\sqrt{2} and γ=γ−=1−1/2\gamma=\gamma^{-}=1-1/\sqrt{2}, respectively. The upper and lower panels show the results for the tests 1 and 2, respectively. In the top and bottom panels, the time step is Δt1=0.005\Delta t_1=0.005, Δt2=0.05\Delta t_2=0.05, respectively (see § 3.1). In the rightmost panel, we plot, the error by Eq. (41) as a function of the time step Δt\Delta t, for the first-order method (green color) and the DIRK method with γ=1±1/2\gamma=1\pm1/\sqrt{2}. The vertical solid line corresponds to Δt=ts,max\Delta t=t_{\mathrm{s,max}}, where condition (18) is applied for the DIRK method.

e1=1M∑k=02∑i=0M−1∣vnumeric,k(Δti)−vanalytic,k(Δti)vanalytic,k(Δti)∣,(41)e_1=\frac{1}{M}\sum_{k=0}^{2}\sum_{i=0}^{M-1}\left|\frac{v_{\mathrm{numeric},k}(\Delta t_i)-v_{\mathrm{analytic},k}(\Delta t_i)}{v_{\mathrm{analytic},k}(\Delta t_i)}\right|, \tag*{(41)}

where Δti\Delta t_i is the time step and MM is the number of time steps during the integration. The index k=0,1,2k = 0, 1, 2 is the index of gas, and dust species, respectively. We calculate the numerical solution for different time steps starting at Δt=10−4\Delta t = 10^{-4} until a final time tfinal>tdampt_{\mathrm{final}} > t_{\mathrm{damp}}, where tdamp≃∣max⁡(λk)∣−1t_{\mathrm{damp}} \simeq|\max(\lambda_k)|^{-1} is a characteristic damping time to reach the equilibrium velocity. When Δt≤tdamp\Delta t \leq t_{\mathrm{damp}}, the number of times steps MM is given by the ratio tdamp/Δtt_{\mathrm{damp}}/\Delta t. When Δt≥tdamp\Delta t \geq t_{\mathrm{damp}} we only have one timestep, i.e., M=1M = 1. For the case of test B, tfinal=2t_{\mathrm{final}} = 2 and tdamp≃0.007t_{\mathrm{damp}} \simeq0.007, whereas in test C, tfinal=2×103t_{\mathrm{final}} = 2 \times10^3 and tdamp≃1.9t_{\mathrm{damp}} \simeq1.9.

In Figure 2 we show that the error as a function of Δt\Delta t. In the regime of Δt<tdamp\Delta t < t_{\mathrm{damp}}, we obtain e1∼Δt2e_1 \sim\Delta t^2 for both values of γ\gamma of the DIRK method. When γ=1+1/2\gamma= 1 + 1/\sqrt{2}, the absolute error is significantly larger in comparison with the solution obtained for γ=1−1/2\gamma= 1 - 1/\sqrt{2}. However, for γ=1−1/2\gamma= 1 - 1/\sqrt{2} the numerical solution may introduce small oscillations about the correct equilibrium. Fortunately, since the method is asymptotically stable, any initial oscillations are damped to the correct equilibrium. This is shown in the left panel of Figure 2 where the gas velocity initially undershoots the analytical solution (see Appendix D for additional comments on the value of γ\gamma).

When Δt≫tdamp\Delta t \gg t_{\mathrm{damp}}, the error obtained from Eq. (41) should agree with the characteristic asymptotic convergence rate of the numerical method. For large time steps, the error displayed in Figure 2 converges as ∼1/Δt\sim1/\Delta t for the first-order method, whereas for the DIRK method, the error goes as ∼1/Δt2\sim1/\Delta t^2. Therefore, the DIRK method provides a second-order accurate solution in the stiff regime. We significantly improve the convergence rate of the DIRK method utilizing condition for γ\gamma given in Eq. (18). The condition sets γ=1/2\gamma= 1/2 if Δt≥ts,max\Delta t \geq t_{\mathrm{s,max}}. In the top panel, ts,max=0.01t_{\mathrm{s,max}} = 0.01 whereas in the bottom panel ts,max=2t_{\mathrm{s,max}} = 2. Note that these stopping times are upper bounds for tdampt_{\mathrm{damp}} as expected from the discussion in § 2.3 and the results on Appendix D of Benítez-Llambay et al. [8].

Damping with an external force

In this section, we extend the test of § 3.1 by including an external force. We illustrate the convergence properties of our method when the time-explicit and time-implicit integrators are coupled together as described in § 2.3.

We again consider one gas and two dust species and solve the momentum equation using Eq. (31), neglecting the continuity equation as well as the divergence of fluxes. Under these assumptions, the momentum equation becomes

∂tu=Mu+G(42)\partial_t \mathbf{u} = \mathbf{M}\mathbf{u} + \mathbf{G} \tag*{(42)}

with G=(G0,0,0)\mathbf{G} = (G_0, 0, 0). The analytical solution of Eq. (42) corresponds to,

uanalytical=C[t,eλ1t,eλ2t,1]T,(43)\mathbf{u}_{\mathrm{analytical}} = \mathbf{C}\left[t, e^{\lambda_1 t}, e^{\lambda_2 t}, 1\right]^{\mathrm{T}}, \tag*{(43)}

where λ1\lambda_1 and λ2\lambda_2 are the non-zero eigenvalues of M\mathbf{M} and C\mathbf{C} is a coefficient matrix that depends on the initial condition, external force, and eigenvalues of M\mathbf{M}. We set as initial condition ρg=1.0\rho_g = 1.0, ρd1=0.1\rho_{d1} = 0.1, ρd2=0.1\rho_{d2} = 0.1 for the density and vg=2.0v_g = 2.0, vd1=0.1v_{d1} = 0.1, vd2=−0.5v_{d2} = -0.5 for the velocity. The external force is set to G0=1G_0 = 1 whereas the drag force coefficients are set to αd,1=1.0\alpha_{d,1} = 1.0, αd,2=0.75\alpha_{d,2} = 0.75, respectively. With these values we obtain λ1=−1.125\lambda_1 = -1.125 and λ2=−0.8\lambda_2 = -0.8 and the coefficient matrix C\mathbf{C} is

C=(0.8333330.1450140.0596151.795370.083333−0.1160110.0298080.0962040.083333−0.029003−0.0894230.068426).(44)\mathbf{C} = \begin{pmatrix} 0.833333 & 0.145014 & 0.059615 & 1.79537 \\ 0.083333 & -0.116011 & 0.029808 & 0.096204 \\ 0.083333 & -0.029003 & -0.089423 & 0.068426 \end{pmatrix}. \tag*{(44)}

In Figure 3 we show the relative velocity between the gas and the dust species. The dashed line corresponds to the analytical solution, the agreement between the analytical and numerical solutions is excellent. In the limit of Δt→∞\Delta t \rightarrow\infty, the momentum of the jj-th species (see Eq. (43)) converges to

Image not extracted

Figure 3. Relative velocity between the gas and dust species for the damping test with a constant external force (see § 3.2). The numerical solution is shown with orange and blue circles. The dashed line corresponds to the analytical solution.

uj=Pj0((P−1G)0t+(P−1u0))−∑k=1N−1Pjk(P−1G)kλk(45)u_j = P_{j0}\left((\mathbf{P}^{-1}\mathbf{G})_0 t + (\mathbf{P}^{-1}\mathbf{u}_0)\right) - \sum_{k=1}^{N-1} P_{jk}\frac{(\mathbf{P}^{-1}\mathbf{G})_k}{\lambda_k} \tag*{(45)}

where we assume λ0=0\lambda_0 = 0 and PjkP_{jk} denotes the jkjk entry of the matrix P\mathbf{P}, with P\mathbf{P} and P−1\mathbf{P}^{-1} the matrices such that D=P−1MP\mathbf{D} = \mathbf{P}^{-1}\mathbf{M}\mathbf{P} is diagonal, that is, λk=Dkk\lambda_k = D_{kk}.

After dividing Eq. (45) by the corresponding density we can obtain the relative velocity between species. First, note that Pj0/ρj−Pi0/ρi=0P_{j0}/\rho_{j} - P_{i0}/\rho_{i} = 0, for any ii and jj components. This follows from the properties of the matrix M\mathbf{M} described below Eq. (125) and using that the first (zero) column of the matrix P\mathbf{P} is the eigenvector 1~T\tilde{\mathbf{1}}^{T}. Therefore, the relative velocity between two species is obtained as

lim⁡Δt→∞vj−vi=1ρi∑k=1N−1Pik(P−1G)kλk−1ρj∑k=1N−1Pjk(P−1G)kλk(46)\lim_{\Delta t \to\infty} v_{j} - v_{i} = \frac{1}{\rho_{i}}\sum_{k=1}^{N-1}P_{ik}\frac{\left(\mathbf{P}^{-1}\mathbf{G}\right)_{k}}{\lambda_{k}}-\frac{1}{\rho_{j}}\sum_{k=1}^{N-1}P_{jk}\frac{\left(\mathbf{P}^{-1}\mathbf{G}\right)_{k}}{\lambda_{k}} \tag*{(46)}

In this case we obtain lim⁡Δt→∞∣vg−vd1∣=0.8333\lim_{\Delta t \to\infty} |v_{g} - v_{d1}| = 0.8333 and lim⁡Δt→∞∣vg−vd2∣=1.111\lim_{\Delta t \to\infty} |v_{g} - v_{d2}| = 1.111, which are the asymptotic values displayed in Figure 3. We conclude that the numerical method converges to the correct equilibrium for the case with external constant force. We study this convergence property further in Appendix § 7.

Damping of a soundwave

In this section we reproduce the test described in Section 3.2 of Benítez-Llambay et al. [8]. For this purpose, we have developed a simple code that solves the Equations (1)–(4) in a periodic domain with the second-order Runge-Kutta method of Eq. (31). We utilize a uniform mesh in the xx-coordinate with NxN_{x} grid cells. After obtaining the numerical solution, we compare it with the analytical one and show that our method de-

scribed in § 2.3 achieves the expected second-order accuracy.

In Figure 4, we show the analytic (solid lines) and numerical (open circles) solutions for the normalized density, defined as δρ^/(ρA)\delta\hat{\rho}/(\rho A). The test includes five species represented by different colors. Numerical and analytical solutions are in very good agreement. We study the convergence of the numerical scheme computing the numerical error as

Image not extracted

Figure 4. Numerical (open circles) and analytical (solid lines) solution of the five-fluid test described in Sec 3.3 in Benítez-Llambay et al. [8]. We plot the time evolution of the normalized density. The blue circles correspond to the gas, while the other color corresponds to the dust species. All solutions were obtained at x=0x = 0. The normalized density is defined as δρ^/(Aρ0)\delta\hat{\rho}/(A\rho^{0}), with A=10−4A = 10^{-4}. We include in the plot the L1L_{1} error composed of the sum errors of density and velocity (not shown) components of the five fluids. Individual errors are calculated over the whole grid after t=2.0t = 2.0.

e1=1Nx∑k=15∑i=12∣δfi,numeric,k−δfi,analytic,k∣,(47)e_{1} = \frac{1}{N_{x}}\sum_{k=1}^{5}\sum_{i=1}^{2}\left|\delta f_{i,\mathrm{numeric},k} - \delta f_{i,\mathrm{analytic},k}\right|, \tag*{(47)}

with kk the index of the fluid and ii the index of the density and velocity field, respectively. In the bottom panel of Figure 4 we show that the error effectively scales as e1∼Nx−2e_{1} \sim N_{x}^{-2}.

Shock test

In this section we test our numerical method by reproducing the shock test from Lehmann & Wardle [32], extended to multiple fluids in (see § 3.3 Benítez-Llambay et al. [8]). We solve Eqs. (1)–(4) in an evenly spaced mesh along the xx-coordinate with 400 grid cells. In this shock test, we assume that Eq. (5) is constant, that is, ρgαg,i=ρd,iαd,i≡Ki\rho_{\mathrm{g}}\alpha_{\mathrm{g},i}=\rho_{\mathrm{d},i}\alpha_{\mathrm{d},i}\equiv K_i. Since the drag-force solver has been written in terms of the parameter ϵi\epsilon_i, the drag-force coefficient αd,i\alpha_{\mathrm{d},i} must be replaced by αd,i=Ki/ρd,i\alpha_{\mathrm{d},i}=K_i/\rho_{\mathrm{d},i} in the matrix M\mathbf{M}. The details of the test together with initial and boundary conditions are in section 3.3.2 from Benítez-Llambay et al. [8].

In Figure 5, we plot the exact (solid line) and numerical (open circles) solutions for the density of the four fluid shock test. To compare the numerical solution with the exact one, we shift the exact solution 2.97 in a positive xx-direction until the shock position as explained in (Benítez-Llambay et al. [8]). The unfilled circles are a sub-sampling (1:4) of the grid points. Inside, we plot a zoomed-in region containing the discontinuity to show the quality of the numerical solution across the shock in the actual grid (all the points are shown). The Riemann solver captures the shock within one cell. For the dust fluid we utilize the solver described in Huang & Bai [23].

Image not extracted

Figure 5. Numerical (open circles) and analytical (solid lines) solution of the five-fluid test described in Sec 3.2 in Benítez-Llambay et al. [8]. The numerical solution was obtained at time t=500t = 500, starting from the initial jump condition. The plot shows the density of the gas (blue) and dust (orange,green,red) species, respectively. In the large panel, to allow the quality of the asymptotic behavior far away from the shock to be assessed, the sampling rate for the open circles was reduced to 1:4 of the original data. We additionally plot as an inset a zoomed region about the shock showing the full sampling, where the unfilled circles correspond to the actual grid points. The code resolves the shock within a single cell, even when an increasing number of fluids is considered. The overall agreement between the numerical and analytical solutions is excellent.

Jeans Instability

In order to expand our benchmarking to include codes with self-gravity, we simulate a multifluid Jeans instability in an infinite uniform medium. To our knowledge this is the first implementation of such a test, and so for completeness we present the analytic solution in detail below.

We add the gravitational potential term into Equations (3) and (4), now written as

ρgDvgDt=ρg∇Φ−∇P−∑i=1N−1ρgαg,i(vg−vd,i)(48)\rho_{\mathrm{g}}\frac{D\boldsymbol{v}_{\mathrm{g}}}{Dt}=\rho_{\mathrm{g}}\nabla\Phi-\nabla P-\sum_{i=1}^{N-1}\rho_{\mathrm{g}}\alpha_{\mathrm{g},i}(\boldsymbol{v}_{\mathrm{g}}-\boldsymbol{v}_{\mathrm{d},i}) \tag*{(48)}
ρd,iDvd,iDt=ρd,i∇Φ−ρd,iαd,i(vd,i−vg)(49)\rho_{\mathrm{d},i}\frac{D\boldsymbol{v}_{\mathrm{d},i}}{Dt}=\rho_{\mathrm{d},i}\nabla\Phi-\rho_{\mathrm{d},i}\alpha_{\mathrm{d},i}(\boldsymbol{v}_{\mathrm{d},i}-\boldsymbol{v}_{\mathrm{g}}) \tag*{(49)}

with D/Dt=∂/∂t+v⋅∇D/Dt=\partial/\partial t+\boldsymbol{v}\cdot\nabla the material derivate. The Poisson equation that governs the gravitational potential of the system is

∇2Φ=4πG(ρg+∑i=1N−1ρd,i)(50)\nabla^2\Phi=4\pi G\left(\rho_{\mathrm{g}}+\sum_{i=1}^{N-1}\rho_{\mathrm{d},i}\right) \tag*{(50)}

In order to study the Jeans instability, we linearize Equations (1)–(2), and (48)–(50) with respect to an equilibrium solution obtained after neglecting the contribution of the background gravitational field in the Poisson equation (Jeans Swindle) (Jeans [24]; Binney & Tremaine [9]). The equilibrium gas density is constant and denoted as ρg0\rho_{\mathrm{g}}^{0}. The equilibrium dust density of the ii-th species is also a constant, equal to ρd,i0=ϵi0ρg0\rho_{\mathrm{d},i}^{0}=\epsilon_i^{0}\rho_{\mathrm{g}}^{0}. Finally, the equilibrium velocities are set to zero. We define small density perturbations δρg\delta\rho_{\mathrm{g}} and δρd,i\delta\rho_{\mathrm{d},i}, whereas the velocity perturbations are δvg=δvgx^\delta\boldsymbol{v}_{\mathrm{g}}=\delta v_{\mathrm{g}}\hat{x}, δvd,i=δvd,ix^\delta\boldsymbol{v}_{\mathrm{d},i}=\delta v_{\mathrm{d},i}\hat{x}. We set ρg0=1.0\rho_{\mathrm{g}}^{0}=1.0 and the total dust-to-gas mass ratio ϵ0=∑i=1N−1ϵi0=0.01\epsilon^{0}=\sum_{i=1}^{N-1}\epsilon_i^{0}=0.01, with ϵi0=ϵ0/N\epsilon_i^{0}=\epsilon^{0}/N. We consider N=128N=128 fluids with a uniform distribution of stopping times between ts=10−4−10t_{\mathrm{s}}=10^{-4}-10.

The linearized continuity, momentum, and Poisson equation are

∂δρg∂t+ρg0∂δvg∂x=0(51)\frac{\partial\delta\rho_{\mathrm{g}}}{\partial t}+\rho_{\mathrm{g}}^{0}\frac{\partial\delta v_{\mathrm{g}}}{\partial x}=0 \tag*{(51)}
∂δρd,i∂t+ρd,i0∂δvd,i∂x=0(52)\frac{\partial\delta\rho_{\mathrm{d},i}}{\partial t}+\rho_{\mathrm{d},i}^{0}\frac{\partial\delta v_{\mathrm{d},i}}{\partial x}=0 \tag*{(52)}
∂δvg∂t+∑m=1N−1ϵm0ts,m(δvg−δvd,m)+cs2ρg0∂δρg∂x=−∂δΦ∂x(53)\frac{\partial\delta v_{\mathrm{g}}}{\partial t}+\sum_{m=1}^{N-1}\frac{\epsilon_m^{0}}{t_{\mathrm{s},m}}(\delta v_{\mathrm{g}}-\delta v_{\mathrm{d},m})+\frac{c_{\mathrm{s}}^{2}}{\rho_{\mathrm{g}}^{0}}\frac{\partial\delta\rho_{\mathrm{g}}}{\partial x}=-\frac{\partial\delta\Phi}{\partial x} \tag*{(53)}
∂δvd,i∂t+1ts,i(δvd,i−δvg)=−∂δΦ∂x(54)\frac{\partial\delta v_{\mathrm{d},i}}{\partial t}+\frac{1}{t_{\mathrm{s},i}}(\delta v_{\mathrm{d},i}-\delta v_{\mathrm{g}})=-\frac{\partial\delta\Phi}{\partial x} \tag*{(54)}
4πG(δρg+∑m=1N−1δρd,m)=∇2δΦ(55)4\pi G\left(\delta\rho_{\mathrm{g}}+\sum_{m=1}^{N-1}\delta\rho_{\mathrm{d},m}\right)=\nabla^{2}\delta\Phi \tag*{(55)}

For NN fluids, these 2N+12N+1 equations are coupled, due to the gravitational field and the drag-force.

After assuming plane-wave perturbations δf=δf~eik⋅x−ωt\delta f=\delta\tilde{f}e^{i\boldsymbol{k}\cdot\boldsymbol{x}-\omega t} with k\boldsymbol{k} a real wave-number, we obtain the eigenvalues and eigenvectors of the system. We normalize the set of linearized equations defining

ω~=ω/(cskJ),k~=k/kJ(56)\tilde{\omega}=\omega/(c_{\mathrm{s}}k_{\mathrm{J}}),\qquad\tilde{k}=k/k_{\mathrm{J}} \tag*{(56)}
ρ~g=δρg/ρg0,ρ~d,i=δρd,i/ρg0(57)\tilde{\rho}_{\mathrm{g}}=\delta\rho_{\mathrm{g}}/\rho_{\mathrm{g}}^{0},\qquad\tilde{\rho}_{\mathrm{d},i}=\delta\rho_{\mathrm{d},i}/\rho_{\mathrm{g}}^{0} \tag*{(57)}
v~g=δvg/cs,v~d,i=δvd,i/cs(58)\tilde{v}_{\mathrm{g}}=\delta v_{\mathrm{g}}/c_{\mathrm{s}},\qquad\tilde{v}_{\mathrm{d},i}=\delta v_{\mathrm{d},i}/c_{\mathrm{s}} \tag*{(58)}
βi=1/(ts,icskJ)(59)\beta_i=1/(t_{\mathrm{s},i}c_{\mathrm{s}}k_{\mathrm{J}}) \tag*{(59)}

where

kJ=4πGρg0(1+ϵ0)/cs,ϵ0=∑m=1N−1ϵm0(60)k_{\mathrm{J}}=\sqrt{4\pi G\rho_{\mathrm{g}}^{0}(1+\epsilon^{0})}/c_{\mathrm{s}},\qquad\epsilon^{0}=\sum_{m=1}^{N-1}\epsilon_m^{0} \tag*{(60)}

After normalization, (51)–(56) become

−ω~ρ~g+ik~v~g=0(61)-\tilde{\omega}\tilde{\rho}_{\mathrm{g}}+i\tilde{k}\tilde{v}_{\mathrm{g}}=0 \tag*{(61)}
−ω~ρ~d,i+iϵi0k~v~d,i=0(62)-\tilde{\omega}\tilde{\rho}_{\mathrm{d},i}+i\epsilon_i^{0}\tilde{k}\tilde{v}_{\mathrm{d},i}=0 \tag*{(62)}
−ω~v~g+∑m=1N−1ϵm0βm(v~g−v~d,i)+iρ~gk~=−ik~δΦ~cs2(63)-\tilde{\omega}\tilde{v}_{\mathrm{g}}+\sum_{m=1}^{N-1}\epsilon_m^{0}\beta_m(\tilde{v}_{\mathrm{g}}-\tilde{v}_{\mathrm{d},i})+i\tilde{\rho}_{\mathrm{g}}\tilde{k}=-i\tilde{k}\frac{\delta\tilde{\Phi}}{c_{\mathrm{s}}^{2}} \tag*{(63)}
−ω~v~d,i+βi(v~d,i−v~g)=−ik~δΦ~cs2(64)-\tilde{\omega}\tilde{v}_{\mathrm{d},i}+\beta_i(\tilde{v}_{\mathrm{d},i}-\tilde{v}_{\mathrm{g}})=-i\tilde{k}\frac{\delta\tilde{\Phi}}{c_{\mathrm{s}}^{2}} \tag*{(64)}
δΦ~cs2=−1k~2(1+ϵ0)(ρ~g+∑mρ~d,m)(65)\frac{\delta\tilde{\Phi}}{c_{\mathrm{s}}^{2}}=-\frac{1}{\tilde{k}^{2}(1+\epsilon^{0})}\left(\tilde{\rho}_{\mathrm{g}}+\sum_m\tilde{\rho}_{\mathrm{d},m}\right) \tag*{(65)}

and can be reduced to a 2N2N equation system, written as

ik~v~g=ω~ρ~g,(66)i\tilde{k}\tilde{v}_{g}=\tilde{\omega}\tilde{\rho}_{g}, \tag*{(66)}
ik~ϵiv~d,i=ω~ρ~d,i,(67)i\tilde{k}\epsilon_{i}\tilde{v}_{d,i}=\tilde{\omega}\tilde{\rho}_{d,i}, \tag*{(67)}
ω~v~g=(ik~−ik~(1+ϵ0))ρ~g−ik~(1+ϵ0)∑m=1N−1ρ~d,m+∑m=1N−1ϵm0βmv~g−∑m=1N−1ϵm0βmv~d,m(68)\tilde{\omega}\tilde{v}_{g}=\left(i\tilde{k}-\frac{i}{\tilde{k}(1+\epsilon^{0})}\right)\tilde{\rho}_{g}-\frac{i}{\tilde{k}(1+\epsilon^{0})}\sum_{m=1}^{N-1}\tilde{\rho}_{d,m} +\sum_{m=1}^{N-1}\epsilon_{m}^{0}\beta_{m}\tilde{v}_{g}-\sum_{m=1}^{N-1}\epsilon_{m}^{0}\beta_{m}\tilde{v}_{d,m} \tag*{(68)}
ω~v~d,i=−ik~(1+ϵ0)ρ~g−ik~(1+ϵ0)∑m=1N−1ρ~d,m−βiv~g+βiv~d,i.(69)\tilde{\omega}\tilde{v}_{d,i}=-\frac{i}{\tilde{k}(1+\epsilon^{0})}\tilde{\rho}_{g}-\frac{i}{\tilde{k}(1+\epsilon^{0})}\sum_{m=1}^{N-1}\tilde{\rho}_{d,m} -\beta_{i}\tilde{v}_{g}+\beta_{i}\tilde{v}_{d,i}. \tag*{(69)}

The equivalent matrix equation is AV=ω~V\mathbf{A}\mathbf{V}=\tilde{\omega}\mathbf{V}, V=(ρ~g,ρ~d,i,…,ρ~d,N−1,v~g,v~d,i,…,v~d,N−1)\mathbf{V}=(\tilde{\rho}_{g},\tilde{\rho}_{d,i},\ldots,\tilde{\rho}_{d,N-1},\tilde{v}_{g},\tilde{v}_{d,i},\ldots,\tilde{v}_{d,N-1}). The eigenvalues and eigenvectors of the problem, in our tests, have been obtained with the function eig() of the Numpy Python package.

Figure 6 displays the grow rate of the dusty Jeans instability σ=−min⁡{R(ω~)}1+ϵ0\sigma=-\min\{\mathfrak{R}(\tilde{\omega})\}\sqrt{1+\epsilon^{0}} as a function of k/kJk/k_J. Note that when ϵ0=0\epsilon_{0}=0, σ\sigma reduces to the normalized growth rate of the gas-only problem. Thus, the inclusion of dust enables higher growth rates as shown in Figure 6 (see solid line vs dashed line (dust-free case). Comparing the gas only and dusty cases illustrates that while both feature a growth rate that is larger for lower wave numbers, the critical k/kJk/k_J for stability is shifted to larger values as dust density increases. In Figure 6 we also the numerical solution obtained with MDIRK described in § 2.3. We calculate the growth rate varying the number of cells from Ncells=16N_{\mathrm{cells}}=16 to Ncells=128N_{\mathrm{cells}}=128. The growth rate converges to the analytic solution for all unstable modes considered in this work.3

Image not extracted

Figure 6. Test resolution convergence for Jeans instability including drag forces. The dashed line denotes the analytic solution for only gas. The solid line shows the analytic solution for 128 fluids including gas a dust species with a linear distribution from ts=10−4t_{\mathrm{s}}=10^{-4} to 10 and uniform density distribution with a total dust-to-gas ratio ϵ0=0.01\epsilon^{0}=0.01. The open circles indicate that the numerical solutions for different numbers of cells converge to the solid line. The vertical dotted line indicates k/kJ=1/1+ϵ0k/k_J=1/\sqrt{1+\epsilon^{0}}.

Axisymmetric steady-state drift solution

To demonstrate the scaling of our method with the number of species, we recover the steady-state gas and dust drift solution in a shearing box, as in section 3.5.2 of Benítez-Llambay et al. [8]. This test consists of a 1D shearing box along the xx direction, with size L=1L=1 and with constant frequency Ω0=1\Omega_{0}=1. The shear parameter is set to q=3/2q=3/2. We solve the continuity and xx-yy momentum equations for each fluid. Gas and dust momentum equations include the centrifugal force (fce=ρ2qΩ02xx^f_{\mathrm{ce}}=\rho2q\Omega_{0}^{2}x\hat{x}) and the Coriolis force fc=−2ρΩ0z^×(vx,vy)f_{\mathrm{c}}=-2\rho\Omega_{0}\hat{z}\times(v_{x},v_{y}). In addition, we add an external constant force on the gas fluid that mimics the global pressure gradient in a protoplanetary disk, fg=χ0Ω02x^\mathbf{f}_{g}=\chi_{0}\Omega_{0}^{2}\hat{x} with χ0=0.005\chi_{0}=0.005. We add these external forces to the time-explicit hydrodynamic solver in an unsplit fashion (see e.g. Gressel & Ziegler [18]). We apply periodic boundary conditions in ρ,ρvx\rho,\rho v_{x}. Whereas for vyv_{y} we add the relative shear velocity after applying periodic boundary conditions, that is ρvy(x)=ρvy(x±L)±qΩ0Lρ\rho v_{y}(x)=\rho v_{y}(x\pm L)\pm q\Omega_{0}L\rho (Hawley et al. [21]). The initial condition is set to the exact equilibrium following Eqs. (81–84) of Benítez-Llambay et al. [8]. We integrate the system until t=20t=20 and vary the number of grid points in xx from Nx=32N_{x}=32 to Nx=1024N_{x}=1024. In all cases, our numerical method conserves the initial equilibrium to machine precision.

In Figure 7, we show the normalized CPU time obtained for this test, where we vary the number of dust species from 1 to 63. As expected, the computational cost is proportional to the number of fluids (complexity O(N)O(N)). We moreover compare our result with the second-order multifluid method RK2Im of Huang & Bai [23], which is found to be close to O(N3)O(N^{3}) for a large number of fluids (N>10N>10). It is interesting to note that the complexity of Huang & Bai [23] method can be reduced to O(N)O(N) by utilizing our analytical solution in Appendix A.

Image not extracted

Figure 7. Normalized CPU time vs. the number of fluids for the test described in § 3.6. As expected, the complexity of MDIRK is O(N)\mathcal{O}(N). For comparison, we also include the results of the multifluid second-order fully implicit method (RK2Im) from [23] which has complexity O(N3)\mathcal{O}(N^3).

Discussion and conclusions

In this work, we have presented MDIRK, a multifluid asymptotically stable numerical method to solve the momentum exchange between gas and dust species. We have shown that a hydrodynamical solver based on a Runge-Kutta implicit-explicit (IMEX) algorithm converges to the correct equilibrium solution with second-order accuracy. The time-implicit part of the IMEX utilizes a diagonally implicit Runge–Kutta method (DIRK) [2]. The major advantage of a diagonally implicit method for the drag force is that it admits an analytical solution that reduces the complexity of the problem from O(N3)\mathcal{O}(N^3) (needed to invert the matrix numerically) to O(N)\mathcal{O}(N) (needed to evaluate the solution). Our second-order implicit-explicit solver can be easily implemented in state-of-the-art hydrodynamics codes (e.g., PLUTO [37], ATHENA++ [49], among others) and it allows for efficient parallelization by fluids through the utilization of weighted momenta and density (c.f. [26]). Extension to different second-order methods can be also obtained utilizing the analytical solution of Appendix A.

Our results suggest that IMEX Runge-Kutta methods with diagonally implicit drag-force schemes are an interesting alternative for developing high-order multifluid codes. Examples of third-order and four-order IMEX methods can be found in [4, 42]. Note however that a high-order integration will require monotonic preserving spatial reconstruction techniques with adequate accuracy [36, 15]. In addition, we stress that only a subset of high-order Runge-Kutta methods are consistent with a total-variation-diminishing property (TVD) [47, 17] and therefore special care must be taken with high-order methods to prevent the presence of spurious oscillation.

While we focus on 1D tests, our method can be utilized for 2D and 3D with static and adaptive mesh refinement with the appropriate extension of the hydrodynamics [31, 23]. Also, in this work we did not take into account mass transfer between species, however, future extensions could explore this feature by combining our method with a dust-growth solution based on discretized dust bins method as described in [35]. Another important future extension of our method would be the inclusion of energy exchange between dust and gas [40], since collisional cooling could play an important role in the thermodynamics of protoplanetary disks. In radiation hydrodynamics, multi-fluid dynamics can also impact the opacity distribution and observational signatures of accreting, embedded planets [28]. We emphasize that the complexity O(N)\mathcal{O}(N) combined with the stability properties will expedite the study of particle-size distribution, especially in problems where hundreds and/or thousands of bins are required to properly model the dust component (see e.g., [27, 58]).

We thank the reviewer for a valuable clear report. L. K. and K. M. K acknowledge support from the Heising-Simons 51 Pegasi b postdoctoral fellowship and TCAN grant 80NSSC19K0639. P. B. L. acknowledges support from ANID, QUIMAL fund ASTRO21-0039 and FONDECYT project 1231205. J.G. acknowledge support by ANID, – Millennium Science Initiative Program – NCN19_171, CCJRFI2105, FONDECYT Regular grant 1210425 and CATA-BASAL project FB210003.

All codes and data utilized in this work are available in [bitbucket.org/secondorderri...]{.} (bitbucket.org/secondorderri... truncated?)

Asymptotic stability

In this section, we show that MDIRK is asymptotically stable. As in § 2.2, we will only focus on the system of momentum equations. Moreover, we only consider the implicit part of the method. In this case, MDIRK reduces to a two-stage diagonally implicit method of the form

(1−γΔtM) k1=Mun,(70)(1-\gamma\Delta t\mathbf{M})\,\mathbf{k}_1=\mathbf{M}\mathbf{u}^n, \tag*{(70)}
(1−γΔtM) k2=M(un+βΔtk1)(71)(1-\gamma\Delta t\mathbf{M})\,\mathbf{k}_2=\mathbf{M}(\mathbf{u}^n+\beta\Delta t\mathbf{k}_1) \tag*{(71)}
un+1=un+b1Δtk1+b2Δtk2,(72)\mathbf{u}^{n+1}=\mathbf{u}^n+b_1\Delta t\mathbf{k}_1+b_2\Delta t\mathbf{k}_2, \tag*{(72)}

with β=1−γ\beta=1-\gamma, b1=1−γb_1=1-\gamma and b2=γb_2=\gamma. The analytical solution of Eq. (72) can be obtained after solving Eqs. (70)–(71) for k1\mathbf{k}_1 and k2\mathbf{k}_2 utilizing the following summations

A=∑k=1N−1αd,kud,k1+γΔtαd,k,B=∑k=1N−1ϵkαd,k1+γΔtαd,k,C=∑k=1N−1αd,k2ud,k(1+γΔtαd,k)2,D=∑k=1N−1ϵkαd,k2(1+γΔtαd,k)2,(73)\begin{aligned} A&=\sum_{k=1}^{N-1}\frac{\alpha_{d,k}u_{d,k}}{1+\gamma\Delta t\alpha_{d,k}},& B&=\sum_{k=1}^{N-1}\frac{\epsilon_k\alpha_{d,k}}{1+\gamma\Delta t\alpha_{d,k}},\\ C&=\sum_{k=1}^{N-1}\frac{\alpha_{d,k}^{2}u_{d,k}}{(1+\gamma\Delta t\alpha_{d,k})^{2}},& D&=\sum_{k=1}^{N-1}\frac{\epsilon_k\alpha_{d,k}^{2}}{(1+\gamma\Delta t\alpha_{d,k})^{2}}, \tag*{(73)} \end{aligned}

with

k1g=A−ugnB1+γΔtB(74)k_{1g}=\frac{A-u_g^nB}{1+\gamma\Delta tB} \tag*{(74)}
k2g=k1g(1+(γ−β)ΔtB+γβΔt2D1+γΔtB)+βΔt1+γΔtB(ugnD−C)(75)k_{2g}=k_{1g}\left(\frac{1+(\gamma-\beta)\Delta tB+\gamma\beta\Delta t^2D}{1+\gamma\Delta tB}\right)+\frac{\beta\Delta t}{1+\gamma\Delta tB}(u_g^nD-C) \tag*{(75)}

and

k1d,i=αi1+γΔtαi(ϵiug−ud,i+γΔtϵik1g)(76)k_{1d,i}=\frac{\alpha_i}{1+\gamma\Delta t\alpha_i}\left(\epsilon_i u_g-u_{d,i}+\gamma\Delta t\epsilon_i k_{1g}\right) \tag*{(76)}
k2d,i=(αi+(γ−β)Δtαi2)(1+γΔtαi)2(ϵiug−ud,i)+αiβΔtϵi1+γΔtαik1g+αiγΔtϵi1+γΔtαik2g.(77)k_{2d,i}=\frac{\left(\alpha_i+(\gamma-\beta)\Delta t\alpha_i^2\right)}{(1+\gamma\Delta t\alpha_i)^2}(\epsilon_i u_g-u_{d,i})+\frac{\alpha_i\beta\Delta t\epsilon_i}{1+\gamma\Delta t\alpha_i}k_{1g}+\frac{\alpha_i\gamma\Delta t\epsilon_i}{1+\gamma\Delta t\alpha_i}k_{2g}. \tag*{(77)}

MDIRK does not utilize the full analytical solutions given in Eqs. (74)–(76) since the calculation of k1\mathbf{k}_1 and k2\mathbf{k}_2 are combined with the explicit updates (see § 2.3). However, this analytical solution could be implemented in the method of Huang & Bai [23] reducing the from O(N3)O(N^3) complexity to O(N)O(N) without shifting to a mixed implicit explicit scheme. Another alternative could be a method based on Strang-Splitting [51], however, we anticipate that strang-splitting may not converge to the correct equilibrium (see Appendix § 7.§ 1).

Asymptotic stability

We will show that this solution is asymptotically stable, that is,

lim⁡Δt→∞un+1−c=0,(78)\lim_{\Delta t\to\infty}\mathbf{u}^{n+1}-\mathbf{c}=0, \tag*{(78)}

where c\mathbf{c} is a constant vector. In fact, the terminal velocity of all fluids converges to a steady state given by the velocity of the center of mass, vcmv_{\rm cm}, defined as

vcm=ug+∑kud,kρg+∑kρd,k=vg+ϵkvd,k1+∑kϵk.(79)v_{\rm cm}=\frac{u_g+\sum_k u_{d,k}}{\rho_g+\sum_k\rho_{d,k}}=\frac{v_g+\epsilon_kv_{d,k}}{1+\sum_k\epsilon_k}. \tag*{(79)}

To demonstrate the asymptotic convergence we define ud=∑kud,k\mathbf{u}_{\mathrm{d}}=\sum_k \mathbf{u}_{\mathrm{d},k}, ϵ=∑kϵk\epsilon=\sum_k\epsilon_k, and X=(ud−ϵug)X=(\mathbf{u}_{\mathrm{d}}-\epsilon\mathbf{u}_{\mathrm{g}}). With these definitions the summations from Eq. (73) can be reduced to

lim⁡Δt→∞ΔtA=γ−1ud,lim⁡Δt→∞ΔtB=γ−1ϵ,lim⁡Δt→∞Δt2C=γ−2ud,lim⁡Δt→∞Δt2D=γ−2ϵ.(80)\begin{aligned} \lim_{\Delta t\to\infty}\Delta t A&=\gamma^{-1}\mathbf{u}_{\mathrm{d}},& \lim_{\Delta t\to\infty}\Delta t B&=\gamma^{-1}\epsilon,\\ \lim_{\Delta t\to\infty}\Delta t^2 C&=\gamma^{-2}\mathbf{u}_{\mathrm{d}},& \lim_{\Delta t\to\infty}\Delta t^2 D&=\gamma^{-2}\epsilon. \tag*{(80)} \end{aligned}

Using the limits from Eq. (80) and the coefficients β=b1=1−γ\beta=b_1=1-\gamma and b2=γb_2=\gamma it can be shown that

lim⁡Δt→∞Δtk1g=γ−1X1+ϵ,lim⁡Δt→∞Δtk2g=γ−1X1+ϵ−γ−2(1−γ)1+ϵX,lim⁡Δt→∞Δtk1d,k=γ−1(ϵkug−ud,k)+γ−1ϵk1+ϵX,lim⁡Δt→∞Δtk2d,k=γ−2(2γ−1)(ϵkug−ud,k)+ϵk1+ϵX(γ−1−γ−2(1−γ)).(81)\begin{aligned} \lim_{\Delta t\to\infty}\Delta t k_{1g}&=\gamma^{-1}\frac{X}{1+\epsilon},\\ \lim_{\Delta t\to\infty}\Delta t k_{2g}&=\gamma^{-1}\frac{X}{1+\epsilon}-\gamma^{-2}\frac{(1-\gamma)}{1+\epsilon}X,\\ \lim_{\Delta t\to\infty}\Delta t k_{1d,k}&=\gamma^{-1}(\epsilon_k\mathbf{u}_{\mathrm{g}}-\mathbf{u}_{\mathrm{d},k})+\gamma^{-1}\frac{\epsilon_k}{1+\epsilon}X,\\ \lim_{\Delta t\to\infty}\Delta t k_{2d,k}&=\gamma^{-2}(2\gamma-1)(\epsilon_k\mathbf{u}_{\mathrm{g}}-\mathbf{u}_{\mathrm{d},k})+\frac{\epsilon_k}{1+\epsilon}X\left(\gamma^{-1}-\gamma^{-2}(1-\gamma)\right). \tag*{(81)} \end{aligned}

Taking the limit of Δt→∞\Delta t\to\infty to Eq. (72) using the limits from Eq. (81) we obtain

lim⁡Δt→∞ugn+1=ugn+((1−γ)γ−1+1−γ−1(1−γ))X1+ϵ=ud+ug1+ϵ.(82)\begin{aligned} \lim_{\Delta t\to\infty}\mathbf{u}_{\mathrm{g}}^{n+1} &=\mathbf{u}_{\mathrm{g}}^n+\left((1-\gamma)\gamma^{-1}+1-\gamma^{-1}(1-\gamma)\right)\frac{X}{1+\epsilon}\\ &=\frac{\mathbf{u}_{\mathrm{d}}+\mathbf{u}_{\mathrm{g}}}{1+\epsilon}. \tag*{(82)} \end{aligned}
lim⁡Δt→∞ud,kn+1=ud,kn(1−γ−1(1−γ)−γ−1(2γ−1))+ugϵkγ−1((1−γ)+(2γ−1))+ϵk1+ϵX((1−γ)γ−1+γ(γ−1−γ−2(1−γ)))=ugϵk+ϵk1+ϵX=ϵkud+ug1+ϵ.(83)\begin{aligned} \lim_{\Delta t\to\infty}\mathbf{u}_{\mathrm{d},k}^{n+1} &=\mathbf{u}_{\mathrm{d},k}^n\left(1-\gamma^{-1}(1-\gamma)-\gamma^{-1}(2\gamma-1)\right)+\mathbf{u}_{\mathrm{g}}\epsilon_k\gamma^{-1}\left((1-\gamma)+(2\gamma-1)\right)\\ &\quad+\frac{\epsilon_k}{1+\epsilon}X\left((1-\gamma)\gamma^{-1}+\gamma\left(\gamma^{-1}-\gamma^{-2}(1-\gamma)\right)\right)\\ &=\mathbf{u}_{\mathrm{g}}\epsilon_k+\frac{\epsilon_k}{1+\epsilon}X\\ &=\epsilon_k\frac{\mathbf{u}_{\mathrm{d}}+\mathbf{u}_{\mathrm{g}}}{1+\epsilon}. \tag*{(83)} \end{aligned}

In this derivation, it seems that the result does not depend on the value of γ\gamma. However, this is only due to the particular choice of the coefficients β\beta, b1b_1 and b2b_2. A different set of coefficients will lead to the condition γ=1±1/2\gamma=1\pm1/\sqrt{2}, together with constraint for second-order accuracy.

Second-order time accuracy

To demonstrate the second-order accuracy in time of the method described in § 2.2 we consider an equation of the form

dudt=G(u)+Mu(84)\frac{\mathrm{d}\mathbf{u}}{\mathrm{d}t}=\mathbf{G}(\mathbf{u})+\mathbf{M}\mathbf{u} \tag*{(84)}

where G(u)\mathbf{G}(\mathbf{u}) is a vector of any external forces i.e. those that do not involve coupling between different fluids, that is G(u)i=Gi(ui)\mathbf{G}(\mathbf{u})_i=G_i(u_i). Forces due to coupling enter only through M\mathbf{M}. To proceed we define u(tn)=un\mathbf{u}(t_n)=\mathbf{u}^n and emphasise that the notation qn\mathbf{q}^n, (q)n(\mathbf{q})^n, and [q]n[\mathbf{q}]^n indicates the function q(t)q(t) evaluated when t=tnt=t_n, for any arbitrary function q(t)q(t). The analytical solution of Eq. (84) is obtained as

u(t)=un+∫tntG(u) dt+∫tntMu dt.(85)\mathbf{u}(t)=\mathbf{u}^n+\int_{t_n}^{t}\mathbf{G}(\mathbf{u})\,\mathrm{d}t+\int_{t_n}^{t}\mathbf{M}\mathbf{u}\,\mathrm{d}t. \tag*{(85)}

We now approximate the solution (B16) to second-order in time. First, the conserved variable u(t)\mathbf{u}(t) and the force, G(u)\mathbf{G}(\mathbf{u}), can be approximated (to first order in time) as

u(t)≃un+(t−tn)(dudt)n(86)\mathbf{u}(t)\simeq\mathbf{u}^n+(t-t_n)\left(\frac{\mathrm{d}\mathbf{u}}{\mathrm{d}t}\right)^n \tag*{(86)}
G(u)≃G(un)+(t−tn)[∇uG(u)]n(dudt)n\mathbf{G}(\mathbf{u})\simeq\mathbf{G}(\mathbf{u}^n)+(t-t_n)\left[\nabla_{\mathbf{u}}\mathbf{G}(\mathbf{u})\right]^n\left(\frac{\mathrm{d}\mathbf{u}}{\mathrm{d}t}\right)^n

where [∇uG(u)]n[\nabla_{\mathbf{u}}G(\mathbf{u})]^{n} is a diagonal operator with diagonal elements [∇uG(u)]iin=(dGi(ui)/dui)n[\nabla_{\mathbf{u}}G(\mathbf{u})]^{n}_{ii} = (dG_i(u_i)/du_i)^{n}. We emphasize again that GiG_i is only a function of uiu_i. With this consideration, the solution (B16) can be approximated as follows

u(t)≃un+G(un)(t−tn)+[∇uG(u)]n(dudt)n(t−tn)22+M(un(t−tn)+(dudt)n(t−tn)22).(87)\mathbf{u}(t) \simeq\mathbf{u}^{n} + G(\mathbf{u}^{n})(t-t_n) + [\nabla_{\mathbf{u}}G(\mathbf{u})]^{n}\left(\frac{d\mathbf{u}}{dt}\right)^{n}\frac{(t-t_n)^2}{2} + \mathbf{M}\left(\mathbf{u}^{n}(t-t_n) + \left(\frac{d\mathbf{u}}{dt}\right)^{n}\frac{(t-t_n)^2}{2}\right). \tag*{(87)}

Eq. (B15) implies that

(dudt)n=G(un)+Mun,(88)\left(\frac{d\mathbf{u}}{dt}\right)^{n} = G(\mathbf{u}^{n}) + \mathbf{M}\mathbf{u}^{n}, \tag*{(88)}

we finally obtain

u(t)≃un+(t−tn)G(un)+(t−tn)Mun+(t−tn)22([∇uG(u)]n+M)G(un)+(t−tn)22([∇uG(u)]nM+M2)un.(89)\mathbf{u}(t) \simeq\mathbf{u}^{n} + (t-t_n)G(\mathbf{u}^{n}) + (t-t_n)\mathbf{M}\mathbf{u}^{n} + \frac{(t-t_n)^2}{2}\left([\nabla_{\mathbf{u}}G(\mathbf{u})]^{n}+\mathbf{M}\right)G(\mathbf{u}^{n}) + \frac{(t-t_n)^2}{2}\left([\nabla_{\mathbf{u}}G(\mathbf{u})]^{n}\mathbf{M}+\mathbf{M}^2\right)\mathbf{u}^{n}. \tag*{(89)}

We now apply the numerical scheme from § 2.3 with the coefficients β=b1=1−γ\beta=b_1=1-\gamma and b2=γb_2=\gamma. A full update is obtained as follows4.

(I−γΔtM)k1=M(un+γΔtG(un))(90)(I-\gamma\Delta t\mathbf{M})\mathbf{k}_1=\mathbf{M}\left(\mathbf{u}^{n}+\gamma\Delta tG(\mathbf{u}^{n})\right) \tag*{(90)}
u∗=un+γΔtk1+ΔtγG(un)(91)\mathbf{u}^{*}=\mathbf{u}^{n}+\gamma\Delta t\mathbf{k}_1+\Delta t\gamma G(\mathbf{u}^{n}) \tag*{(91)}
(I−γΔtM)k2=M(un+(1−γ)Δtk1+δΔtG(un)+(1−δ)ΔtG(u∗))un+1=un+(1−γ)Δtk1+γΔtk2+δΔtG(un)+(1−δ)ΔtG(u∗)(92)\begin{aligned} (I-\gamma\Delta t\mathbf{M})\mathbf{k}_2&=\mathbf{M}\left(\mathbf{u}^{n}+(1-\gamma)\Delta t\mathbf{k}_1+\delta\Delta tG(\mathbf{u}^{n})+(1-\delta)\Delta tG(\mathbf{u}^{*})\right)\\ \mathbf{u}^{n+1}&=\mathbf{u}^{n}+(1-\gamma)\Delta t\mathbf{k}_1+\gamma\Delta t\mathbf{k}_2+\delta\Delta tG(\mathbf{u}^{n})+(1-\delta)\Delta tG(\mathbf{u}^{*}) \tag*{(92)} \end{aligned}

Defining

A=(I−γΔtM)−1M,(93)\mathbf{A}=(I-\gamma\Delta t\mathbf{M})^{-1}\mathbf{M}, \tag*{(93)}

the solution follows as

un+1=(I+ΔtA+γ(1−γ)Δt2A2)un+((γ(1−γ)+γδ)Δt2A+γ2(1−γ)Δt3A2+δΔtI)G(un)+(γ(1−δ)Δt2A+(1−δ)ΔtI)G(u∗).(94)\begin{aligned} \mathbf{u}^{n+1}={}&\left(I+\Delta t\mathbf{A}+\gamma(1-\gamma)\Delta t^2\mathbf{A}^2\right)\mathbf{u}^{n}\\ &+\left((\gamma(1-\gamma)+\gamma\delta)\Delta t^2\mathbf{A}+\gamma^2(1-\gamma)\Delta t^3\mathbf{A}^2+\delta\Delta t I\right)G(\mathbf{u}^{n})\\ &+\left(\gamma(1-\delta)\Delta t^2\mathbf{A}+(1-\delta)\Delta t I\right)G(\mathbf{u}^{*}). \tag*{(94)} \end{aligned}

To find the second order approximation of the numerical solution (B26) we first note that

A≃M+ΔtM2,(95)\mathbf{A}\simeq\mathbf{M}+\Delta t\mathbf{M}^{2}, \tag*{(95)}
G(u∗)≃G(un)+γΔt([∇uG(u)]nG(un)+[∇uG(u)]nMun),(96)G(\mathbf{u}^{*})\simeq G(\mathbf{u}^{n})+\gamma\Delta t\left([\nabla_{\mathbf{u}}G(\mathbf{u})]^{n}G(\mathbf{u}^{n})+[\nabla_{\mathbf{u}}G(\mathbf{u})]^{n}\mathbf{M}\mathbf{u}^{n}\right), \tag*{(96)}

where we have used that k1≃ΔtMun\mathbf{k}_1\simeq\Delta t\mathbf{M}\mathbf{u}^{n} to first order. Replacing these two expansions in Eq. (B26), and neglecting terms with order higher than (Δt)2(\Delta t)^2, we obtain

un+1≃(I+Δt(M+ΔtM2)+γ(1−γ)Δt2M2)un+((γ(1−γ)+γδ)Δt2M+δΔtI)G(un)+(γ(1−δ)Δt2M+(1−δ)ΔtI)×[G(un)+γΔt([∇uG(u)]nG(un)+[∇uG(u)]nMun)].(97)\begin{aligned} \mathbf{u}^{n+1}\simeq{}&\left(I+\Delta t(\mathbf{M}+\Delta t\mathbf{M}^{2})+\gamma(1-\gamma)\Delta t^2\mathbf{M}^{2}\right)\mathbf{u}^{n}\\ &+\left((\gamma(1-\gamma)+\gamma\delta)\Delta t^2\mathbf{M}+\delta\Delta t I\right)G(\mathbf{u}^{n})\\ &+\left(\gamma(1-\delta)\Delta t^2\mathbf{M}+(1-\delta)\Delta t I\right)\\ &\times\left[G(\mathbf{u}^{n})+\gamma\Delta t\left([\nabla_{\mathbf{u}}G(\mathbf{u})]^{n}G(\mathbf{u}^{n})+[\nabla_{\mathbf{u}}G(\mathbf{u})]^{n}\mathbf{M}\mathbf{u}^{n}\right)\right]. \tag*{(97)} \end{aligned}

which can be cast as

un+1≃un+Δt(M+ΔtG(un)+Δt2(γ(2−γ)M2+γ(1−δ)[∇uG(u)]nM)un+Δt2(γ(1−δ)[∇uG(u)]n+γ(2−γ)M)G(un).(98)\begin{aligned} \mathbf{u}^{n+1}\simeq{}&\mathbf{u}^{n}+\Delta t(\mathbf{M}+\Delta tG(\mathbf{u}^{n})\\ &+\Delta t^2\left(\gamma(2-\gamma)\mathbf{M}^{2}+\gamma(1-\delta)[\nabla_{\mathbf{u}}G(\mathbf{u})]^{n}\mathbf{M}\right)\mathbf{u}^{n}\\ &+\Delta t^2\left(\gamma(1-\delta)[\nabla_{\mathbf{u}}G(\mathbf{u})]^{n}+\gamma(2-\gamma)\mathbf{M}\right)G(\mathbf{u}^{n}). \tag*{(98)} \end{aligned}

In order to match the Taylor series (B20) for t=tn+Δtt=t_n+\Delta t, we must have

γ(2−γ)=12,γ(1−δ)=12(99)\gamma(2-\gamma)=\frac{1}{2}, \qquad\gamma(1-\delta)=\frac{1}{2} \tag*{(99)}

These two conditions give

γ=1±12,δ=1−12γ,(100)\gamma=1\pm\frac{1}{\sqrt{2}}, \qquad\delta=1-\frac{1}{2\gamma}, \tag*{(100)}

which is the solution from Ascher et al. [4].

Convergence to the correct equilibrium

To study the convergence properties of the numerical method we consider the case of Eq. (84) with constant external forces g(u)=g0\mathbf{g}(\mathbf{u})=\mathbf{g}_0. In this case, the second-order Runge–Kutta IMEX method described in § 2.2 reduces to

un+1=(I+ΔtA+γ(1−γ)Δt2A2)un+(γ(2−γ)Δt2A+γ2(1−γ)Δt3A2+ΔtI)G0.(101)\mathbf{u}^{n+1}=\left(\mathbf{I}+\Delta t\mathbf{A}+\gamma(1-\gamma)\Delta t^2\mathbf{A}^2\right)\mathbf{u}^n +\left(\gamma(2-\gamma)\Delta t^2\mathbf{A}+\gamma^2(1-\gamma)\Delta t^3\mathbf{A}^2+\Delta t\mathbf{I}\right)\mathbf{G}_0. \tag*{(101)}

Since A\mathbf{A} and A2\mathbf{A}^2 are diagonalizable under the same transformation (see Section C.2), we apply the transformation matrix P−1\mathbf{P}^{-1} to Eq. (105) and obtain

u^kn+1=u^kn(1+Δtλk1−γΔtλk+γ(1−γ)Δt2(λk1−γΔtλk)2)+G^k[γ(2−γ)Δt2λk1−γΔtλk+γ2(1−γ)Δt3(λk1−γΔtλk)2+Δt],(102)\hat{u}_k^{n+1}=\hat{u}_k^n\left(1+\Delta t\frac{\lambda_k}{1-\gamma\Delta t\lambda_k}+\gamma(1-\gamma)\Delta t^2\left(\frac{\lambda_k}{1-\gamma\Delta t\lambda_k}\right)^2\right) +\hat{G}_k\left[\gamma(2-\gamma)\Delta t^2\frac{\lambda_k}{1-\gamma\Delta t\lambda_k}+\gamma^2(1-\gamma)\Delta t^3\left(\frac{\lambda_k}{1-\gamma\Delta t\lambda_k}\right)^2+\Delta t\right], \tag*{(102)}

where u^=P−1u\hat{\mathbf{u}}=\mathbf{P}^{-1}\mathbf{u} and u^k\hat{u}_k corresponds to the entry of u^\hat{\mathbf{u}} associated to the eigenvalue λk\lambda_k of the matrix M\mathbf{M} (equivalently for G^k\hat{G}_k). As mentioned in § 2, all eigenvalues of the matrix M\mathbf{M} are negative, except for a unique eigenvalue λj=0\lambda_j=0. We define x=∣λk∣Δtx=|\lambda_k|\Delta t (with ∣λk∣=−λk|\lambda_k|=-\lambda_k) and write Eq. (102) as

u^kn+1=u^kn(x(2γ−1)+1(1+γx)2)+G^k∣λk∣x(γ2x+1)(1+γx)2.(103)\hat{u}_k^{n+1}=\hat{u}_k^n\left(\frac{x(2\gamma-1)+1}{(1+\gamma x)^2}\right)+\frac{\hat{G}_k}{|\lambda_k|}\frac{x(\gamma^2x+1)}{(1+\gamma x)^2}. \tag*{(103)}

For the case of the eigenvalue λj=0\lambda_j=0, u^jn+1=u^jn+ΔtG^j\hat{u}_j^{n+1}=\hat{u}_j^n+\Delta t\hat{G}_j, which corresponds to a uniformly accelerated center of mass of the system. For the case with λk≠0\lambda_k\ne0, the limit of x→∞x\to\infty gives

lim⁡x→∞u^kn+1=lim⁡x→∞(u^kn2γ−1γ2x+G^k∣λk∣(1+1−2γγ2x))=G^k∣λk∣.(104)\lim_{x\to\infty}\hat{u}_k^{n+1}=\lim_{x\to\infty}\left(\hat{u}_k^n\frac{2\gamma-1}{\gamma^2x}+\frac{\hat{G}_k}{|\lambda_k|}\left(1+\frac{1-2\gamma}{\gamma^2x}\right)\right)=\frac{\hat{G}_k}{|\lambda_k|}. \tag*{(104)}

This limit is identical to that obtained after applying the transformation P−1\mathbf{P}^{-1} to the analytical solution (B16) with a constant external force G=G0\mathbf{G}=\mathbf{G}_0 (see Eq. (45)). Thus, the solution corresponds to a system with constant velocities relative to the center of mass. Note as well that in cases with no external acceleration lim⁡x→∞u^kn+1∝lim⁡x→∞1/x\lim_{x\to\infty}\hat{u}_k^{n+1}\propto\lim_{x\to\infty}1/x, which explains the asymptotic convergence of the error as e∼1/Δte\sim1/\Delta t obtained in § 3.1.

Strang-splitting

We now show that a strang-splitting method does not converge to the correct equilibrium solution. Consider a strang-splitting method where the drag force is integrated on Δt/2\Delta t/2 before and after a full integration of the hydrodynamics. To study the convergence of the strang-splitting method we assume un=0\mathbf{u}^n=0 since the initial condition should not affect the convergence to the correct equilibrium. Thus, we skip the first implicit update on the half step. The solution of the strang-splitting method is obtained as follows

un+1=(I+12ΔtA+γ(1−γ)14Δt2A2)ΔtG0.(105)\mathbf{u}^{n+1}=\left(\mathbf{I}+\frac{1}{2}\Delta t\mathbf{A}+\gamma(1-\gamma)\frac{1}{4}\Delta t^2\mathbf{A}^2\right)\Delta t\mathbf{G}_0. \tag*{(105)}

Following the diagonalization steps and normalization described for Eq. (102) we obtain

u^kn+1=G^k∣λk∣x(1−x/2(1−γ)1+γx/2−x/2γ(1+γx/2)2(1+x/2(2γ−1)))(106)\hat{u}_k^{n+1}=\frac{\hat{G}_k}{\lvert\lambda_k\rvert}x\left(1-\frac{x/2(1-\gamma)}{1+\gamma x/2}-\frac{x/2\gamma}{(1+\gamma x/2)^2}\left(1+x/2(2\gamma-1)\right)\right) \tag*{(106)}

for some eigenvalue λk≠0\lambda_k \ne0. Now we take the limit of x→∞x \to\infty and obtain

lim⁡x→∞u^kn+1=(G^k∣λk∣)lim⁡x→∞x(1−x/2(1−γ)1+γx/2−x/2γ(1+γx/2)2(1+x/2(2γ−1)))=(G^k∣λk∣)4γ−2γ2.(107)\lim_{x\to\infty}\hat{u}_k^{n+1}=\left(\frac{\hat{G}_k}{\lvert\lambda_k\rvert}\right)\lim_{x\to\infty}x\left(1-\frac{x/2(1-\gamma)}{1+\gamma x/2}-\frac{x/2\gamma}{(1+\gamma x/2)^2}\left(1+x/2(2\gamma-1)\right)\right)=\left(\frac{\hat{G}_k}{\lvert\lambda_k\rvert}\right)\frac{4\gamma-2}{\gamma^2}. \tag*{(107)}

The convergence to the correct equilibrium implies 4γ−2/γ2=14\gamma-2/\gamma^2=1 which gives γ=2±2\gamma=2\pm\sqrt{2}. However, this value is not consistent with the second-order accuracy requirement discussed in § 2.3. Therefore, the strang-splitting fashion will not converge to the correct steady-state solution.

The matrix A is diagonalizable

First note that M\mathbf{M} is diagonalizable since it is similar to a real symmetric matrix as shown in Benítez-Llambay et al. [8]. Now, we define P\mathbf{P} as the transformation matrix such that P−1MP=Λ\mathbf{P}^{-1}\mathbf{M}\mathbf{P}=\mathbf{\Lambda}, with Λ\mathbf{\Lambda} a diagonal matrix. The transformation defined by P\mathbf{P} also diagonalize I−γΔtM\mathbf{I}-\gamma\Delta t\mathbf{M}, together with the inverse (I−γΔtM)−1(\mathbf{I}-\gamma\Delta t\mathbf{M})^{-1}. Therefore,

P−1AP=P−1((I−γΔtM)−1M)P=(P−1(I−γΔtM)−1P)(P−1MP)=DΛ=AD(108)\mathbf{P}^{-1}\mathbf{A}\mathbf{P}=\mathbf{P}^{-1}\left((\mathbf{I}-\gamma\Delta t\mathbf{M})^{-1}\mathbf{M}\right)\mathbf{P} =\left(\mathbf{P}^{-1}(\mathbf{I}-\gamma\Delta t\mathbf{M})^{-1}\mathbf{P}\right)\left(\mathbf{P}^{-1}\mathbf{M}\mathbf{P}\right)=\mathbf{D}\mathbf{\Lambda}=\mathbf{A}_{\mathbf{D}} \tag*{(108)}

where D\mathbf{D} and AD\mathbf{A}_{\mathbf{D}} are diagonal matrices.

On the value of γ\gamma

We stress that γ=1−1/2\gamma=1-1/\sqrt{2} (instead of γ=1+1/2\gamma=1+1/\sqrt{2}) will provide the most accurate solution since the truncation error is minimized for this value. However, depending on the problem at hand, the value of γ=1−1/2\gamma=1-1/\sqrt{2} may not be consistent with a TVD Runge-Kutta in the form of Shu & Osher [47]. For example, the explicit part of the solver in § 2.3 can be written as

u(0)=unu(1)=α10u0+β10ΔtH(u0)u(2)=α20u(0)+α21u(1)+β20ΔtH(u(0))+β21ΔtH(u(1))un+1=u(2)(109)\begin{aligned} \mathbf{u}^{(0)}&=\mathbf{u}^n\\ \mathbf{u}^{(1)}&=\alpha_{10}\mathbf{u}^{0}+\beta_{10}\Delta t\mathcal{H}(\mathbf{u}^{0})\\ \mathbf{u}^{(2)}&=\alpha_{20}\mathbf{u}^{(0)}+\alpha_{21}\mathbf{u}^{(1)}+\beta_{20}\Delta t\mathcal{H}(\mathbf{u}^{(0)})+\beta_{21}\Delta t\mathcal{H}(\mathbf{u}^{(1)})\\ \mathbf{u}^{n+1}&=\mathbf{u}^{(2)} \tag*{(109)} \end{aligned}

with α10=1\alpha_{10}=1, α20=1\alpha_{20}=1, α21=0\alpha_{21}=0, β10=γ\beta_{10}=\gamma, β20=δ\beta_{20}=\delta, β21=1−δ\beta_{21}=1-\delta. These values are consistent with TVD second-order Runge-Kutta provided γ=1+1/2\gamma=1+1/\sqrt{2}, since all the coefficients of the explicit update are non-negative (Gottlieb & Shu [17]). However, when γ=1−1/2\gamma=1-1/\sqrt{2} the coefficient β20=δ<0\beta_{20}=\delta<0.

There is yet another concern about the value γ=1−1/2\gamma=1-1/\sqrt{2}. To showcase the problem, we consider a reduced system where only drag forces are taken into account. In this case, the numerical solution of the system of equations (3)-(4) reduces to the form

un+1=(I+ΔtA+γ(1−γ)Δt2A2)un≡T(Δt,γ)un(110)\mathbf{u}^{n+1}=\left(\mathbf{I}+\Delta t\mathbf{A}+\gamma(1-\gamma)\Delta t^2\mathbf{A}^2\right)\mathbf{u}^n\equiv\mathbf{T}(\Delta t,\gamma)\mathbf{u}^n \tag*{(110)}

where A\mathbf{A} is defined in Eq. (93). If for a given time step the det⁡(T(Δt,γ))=0\det(\mathbf{T}(\Delta t,\gamma))=0, the solution un+1\mathbf{u}^{n+1} may deviate from the correct instantaneous equilibrium. First, note that the eigenvalues, aka_k, of A=(I−γΔtM)−1M\mathbf{A}=(\mathbf{I}-\gamma\Delta t\mathbf{M})^{-1}\mathbf{M} are

ak=λk1−γΔtλk,(111)a_k=\frac{\lambda_k}{1-\gamma\Delta t\lambda_k}, \tag*{(111)}

with λk\lambda_k eigenvalue of M\mathbf{M}. Now, we define μk\mu_k the eigenvalues of T\mathbf{T} and note that

μk=1+Δtλk1−γΔtλk+γ(1−γ)Δt2(λk1−γΔtλk)2.(112)\mu_k=1+\Delta t\frac{\lambda_k}{1-\gamma\Delta t\lambda_k}+\gamma(1-\gamma)\Delta t^2\left(\frac{\lambda_k}{1-\gamma\Delta t\lambda_k}\right)^2. \tag*{(112)}

We introduce the variable x=Δt∣λk∣x = \Delta t\lvert\lambda_k\rvert as in Section C with ∣λk∣=−λk\lvert\lambda_k\rvert= -\lambda_k. Thus,

μk(x)=1−x1+γx+γ(1−γ)x2(1+γx)2.(113)\mu_k(x) = 1 - \frac{x}{1+\gamma x} + \gamma(1-\gamma)\frac{x^2}{(1+\gamma x)^2}. \tag*{(113)}

which can be reduced to

μk(x)=x(2γ−1)+1(1+γx)2.(114)\mu_k(x) = \frac{x(2\gamma-1)+1}{(1+\gamma x)^2}. \tag*{(114)}

We conclude that μk(x)=0\mu_k(x)=0 for x=1/(1−2γ)x=1/(1-2\gamma) and therefore

det⁡(T)=∏kμk=0⟺x=1/(1−2γ).(115)\det(\mathbf{T}) = \prod_k \mu_k = 0 \Longleftrightarrow x = 1/(1-2\gamma). \tag*{(115)}

In terms of Δt\Delta t, the condition (D49) implies

det⁡(T)=0⟺Δt=∣λk∣−11−2γ,(116)\det(\mathbf{T}) = 0 \Longleftrightarrow\Delta t = \frac{\lvert\lambda_k\rvert^{-1}}{1-2\gamma}, \tag*{(116)}

for any k∈[0,N−1]k \in[0,N-1]. In the case of γ=1−1/2\gamma=1-1/\sqrt{2}, 1−2γ>01-2\gamma>0 and therefore there is Δt>0\Delta t>0 such that det⁡(T)=0\det(\mathbf{T})=0. In the other hand, if γ=1+1/2\gamma=1+1/\sqrt{2}, 1−2γ<01-2\gamma<0 and therefore det⁡(T)≠0\det(\mathbf{T})\ne0 for all Δt>0\Delta t>0. The same result also holds for the second-order Runge-Kutta of Pareschi & Russo [42]. Finally, the same analysis can be applied to a case with constant external force, however, in that case, the condition that could affect the solution corresponds to x=−1/γ2x=-1/\gamma^2, which is never satisfied for Δt>0\Delta t>0.

Flux calculation

To calculate the mass and momentum flux (L\mathcal{L} operator neglecting external forces) we consider a 1D Cartesian evenly spaced mesh along the xx-direction. For each fluid, the operator L\mathcal{L} in Eq. (31) returns the divergence of the flux as follows

L(Uin)=−1Δx(Fi+1/2(Un)−Fi−1/2(Un)),(117)\mathcal{L}(\mathbf{U}_i^n)=-\frac{1}{\Delta x}\left(\mathbf{F}_{i+1/2}(\mathbf{U}^n)-\mathbf{F}_{i-1/2}(\mathbf{U}^n)\right), \tag*{(117)}

where the index ii denotes the spatial position in the discretized mesh and i±1/2i\pm1/2 indicates the position of the interfaces where fluxes are calculated. U=(ρ,ρv)\mathbf{U}=(\rho,\rho\mathbf{v}) correspond to the vector of conserved variables. The flux, Fi+1/2\mathbf{F}_{i+1/2}, is obtained by solving the Riemann problem with initial condition

U(x,0)={UL,i+1/2if x<xi+1/2,UR,i+1/2if x>xi+1/2,(118)\mathbf{U}(x,0)= \begin{cases} \mathbf{U}_{\mathrm{L},i+1/2} & \text{if } x<x_{i+1/2},\\ \mathbf{U}_{\mathrm{R},i+1/2} & \text{if } x>x_{i+1/2}, \end{cases} \tag*{(118)}

where UL,i+1/2\mathbf{U}_{\mathrm{L},i+1/2} and UR,i+1/2\mathbf{U}_{\mathrm{R},i+1/2} correspond to the left and right states with respect to the interface at x=xi+1/2x=x_{i+1/2}. In this work, we use a piece-wise linear reconstruction with a Van-Leer slope limiter (van Leer [53]) on the primitive variables V=(ρ,v)\mathbf{V}=(\rho,v). We approximate the left and right states as follows

VL,i+1/2=Vi+12ΔxVi,VR,i+1/2=Vi+1−12ΔxVi+1,(119)\mathbf{V}_{\mathrm{L},i+1/2}=\mathbf{V}_i+\frac{1}{2}\Delta_x\mathbf{V}_i,\qquad\mathbf{V}_{\mathrm{R},i+1/2}=\mathbf{V}_{i+1}-\frac{1}{2}\Delta_x\mathbf{V}_{i+1}, \tag*{(119)}

with

ΔxVi=max⁡{(Vi+1−Vi)(Vi−Vi−1),0}Vi+1−Vi−1.(120)\Delta_x\mathbf{V}_i=\frac{\max\{(\mathbf{V}_{i+1}-\mathbf{V}_i)(\mathbf{V}_i-\mathbf{V}_{i-1}),0\}}{\mathbf{V}_{i+1}-\mathbf{V}_{i-1}}. \tag*{(120)}

After the reconstruction of the left and right states calculate the fluxes utilizing the approximate solution from Harten, Lax and van Leer (HLL) (Harten et al. [20]),

Fi+1/2={FLif 0<SL,SRFL−SLFR+SRSL(UR−UL)SR−SLif SL≤0≤SR,FRif 0>SR.(121)\mathbf{F}_{i+1/2}= \begin{cases} \mathbf{F}_{\mathrm{L}} & \text{if } 0<S_{\mathrm{L}},\\ \dfrac{S_{\mathrm{R}}\mathbf{F}_{\mathrm{L}}-S_{\mathrm{L}}\mathbf{F}_{\mathrm{R}}+S_{\mathrm{R}}S_{\mathrm{L}}(\mathbf{U}_{\mathrm{R}}-\mathbf{U}_{\mathrm{L}})}{S_{\mathrm{R}}-S_{\mathrm{L}}} & \text{if } S_{\mathrm{L}}\le0\le S_{\mathrm{R}},\\ \mathbf{F}_{\mathrm{R}} & \text{if } 0>S_{\mathrm{R}}. \end{cases} \tag*{(121)}

where FL\mathbf{F}_{\mathrm{L}} and FR\mathbf{F}_{\mathrm{R}} are fluxes computed with VL,i+1/2\mathbf{V}_{\mathrm{L},i+1/2} and VR,i+1/2\mathbf{V}_{\mathrm{R},i+1/2}, respectively. Thus, in terms of primitive variables, we define the vector fluxes as Fg=(ρgvg,ρgvg2+P)\mathbf{F}_{\mathrm{g}}=(\rho_{\mathrm{g}}v_{\mathrm{g}},\rho_{\mathrm{g}}v_{\mathrm{g}}^{2}+P), Fd=(ρdvd,ρdvd2)\mathbf{F}_{\mathrm{d}}=(\rho_{\mathrm{d}}v_{\mathrm{d}},\rho_{\mathrm{d}}v_{\mathrm{d}}^{2}), for gas and dust, respectively. For the gas species, the speeds SRS_{\mathrm{R}} and SLS_{\mathrm{L}} are estimated as

SL=min⁡{vL−cs,vR−cs},SR=max⁡{vL+cs,vR+cs},(122)S_{\mathrm{L}}=\min\{v_{\mathrm{L}}-c_{\mathrm{s}},v_{\mathrm{R}}-c_{\mathrm{s}}\},\qquad S_{\mathrm{R}}=\max\{v_{\mathrm{L}}+c_{\mathrm{s}},v_{\mathrm{R}}+c_{\mathrm{s}}\}, \tag*{(122)}

where vLv_{\mathrm{L}} and vRv_{\mathrm{R}} are obtained after transforming the conservative left and right states into primitive variables. For pressureless fluids (dust), the sound speed is set to zero, SL=−SRS_{\mathrm{L}}=-S_{\mathrm{R}}, and SR=max⁡{∣vL∣,∣vR∣}S_{\mathrm{R}}=\max\{|v_{\mathrm{L}}|,|v_{\mathrm{R}}|\}. Therefore, for the dust species the flux in Equation (121) reduces to the Rusanov flux Rusanov [46]. We have also implemented the Riemann solver for the dust described in Huang & Bai [23]. We utilize this method in the shock test and steady-state gas and dust drift calculation. While in this work we adopted the (perhaps) simpler flux estimation, there are different alternatives for the gas (see e.g., Toro [52]), and the dust as a pressureless fluid Leveque [34]; Paardekooper & Mellema [41].

Gravitational potential

The gravitational potential is included as a source term in Eqs. (3)-(4). To incorporate the potential into our numerical method we solve the discretized form of the Poisson equation using a sparse linear solver with periodic boundary conditions,

Φi+1−2Φi−Φi−1Δx2=4πGρin,(123)\frac{\Phi_{i+1}-2\Phi_i-\Phi_{i-1}}{\Delta x^2}=4\pi G\rho_i^n, \tag*{(123)}

and obtain Φi\Phi_i and include the potential into the L\mathcal{L} operator as

L(Uin)=−ΔtΔx(Fi+1/2(Un)−Fi−1/2(Un)−ρin2(Φi+1−Φi−1)).(124)\mathcal{L}(\mathbf{U}_i^n)=-\frac{\Delta t}{\Delta x}\left(\mathbf{F}_{i+1/2}(\mathbf{U}^n)-\mathbf{F}_{i-1/2}(\mathbf{U}^n)-\frac{\rho_i^n}{2}(\Phi_{i+1}-\Phi_{i-1})\right). \tag*{(124)}

Matrix MBKP19

MBKP19=(∑k=1N−1ϵkαd,k−ϵ1αd,1−ϵ2αd,2⋯−ϵN−1αd,N−1−αd,1αd,10⋯0−αd,20αd,2⋯0⋯⋯⋯⋯⋯−αd,N−100⋯αd,N−1).(125)\mathbf{M}_{\mathrm{BKP19}}= \begin{pmatrix} \sum_{k=1}^{N-1}\epsilon_k\alpha_{\mathrm{d},k} & -\epsilon_1\alpha_{\mathrm{d},1} & -\epsilon_2\alpha_{\mathrm{d},2} & \cdots& -\epsilon_{N-1}\alpha_{\mathrm{d},N-1}\\ -\alpha_{\mathrm{d},1} & \alpha_{\mathrm{d},1} & 0 & \cdots& 0\\ -\alpha_{\mathrm{d},2} & 0 & \alpha_{\mathrm{d},2} & \cdots& 0\\ \cdots& \cdots& \cdots& \cdots& \cdots\\ -\alpha_{\mathrm{d},N-1} & 0 & 0 & \cdots& \alpha_{\mathrm{d},N-1} \end{pmatrix}. \tag*{(125)}

References

  1. [1]Adachi, I., Hayashi, C., & Nakazawa, K. 1976, Progress of Theoretical Physics, 56, 1756, doi: 10.1143/PTP.56.1756
  2. [2]Alexander, R. 1977, SIAM Journal on Numerical Analysis, 14, 1006, doi: 10.1137/0714068
  3. [3]Andrews, S. M., Huang, J., Pérez, L. M., et al. 2018, ApJL, 869, L41, doi: 10.3847/2041-8213/aaf741DOI
  4. [4]Ascher, U. M., Ruuth, S. J., & Spiteri, R. J. 1997, Applied Numerical Mathematics, 25, 151, doi: https://doi.org/10.1016/S0168-9274(97)00056-1
  5. [5]Avenhaus, H., Quanz, S. P., Garufi, A., et al. 2018, ApJ, 863, 44, doi: 10.3847/1538-4357/aab846DOI
  6. [6]Bai, X.-N., & Stone, J. M. 2010, ApJS, 190, 297, doi: 10.1088/0067-0049/190/2/297DOI
  7. [7]Balsara, D. S., Tilley, D. A., Rettig, T., & Brittain, S. D. 2009, Monthly Notices of the Royal Astronomical Society, 397, 24, doi: 10.1111/j.1365-2966.2009.14606.xDOI
  8. [8]Benítez-Llambay, P., Krapp, L., & Pessah, M. E. 2019, ApJS, 241, 25, doi: 10.3847/1538-4357/aba0ae
  9. [9]Binney, J., & Tremaine, S. 2008, Galactic Dynamics: Second Edition, rev - revised, 2 edn. (Princeton University Press). http://www.jstor.org/stable/j.ctvc778ff
  10. [10]Birnstiel, T., Dullemond, C. P., Zhu, Z., et al. 2018, ApJL, 869, L45, doi: 10.3847/2041-8213/aaf743DOI
  11. [11]Booth, R. A., Sijacki, D., & Clarke, C. J. 2015, Monthly Notices of the Royal Astronomical Society, 452, 3932, doi: 10.1093/mnras/stv1486DOI
  12. [12]Chen, J.-W., & Lin, M.-K. 2018, Monthly Notices of the Royal Astronomical Society, 478, 2737, doi: 10.1093/mnras/sty1166DOI
  13. [13]Draine, B. T. 2003, ARA&A, 41, 241, doi: 10.1146/annurev.astro.41.011802.094840DOI
  14. [14]Epstein, P. S. 1924, Phys. Rev., 23, 710, doi: 10.1103/PhysRev.23.710DOI
  15. [15]Felker, K. G., & Stone, J. M. 2018, Journal of Computational Physics, 375, 1365, doi: 10.1016/j.jcp.2018.08.025DOI
  16. [16]Fung, J., & Muley, D. 2019, ApJS, 244, 42, doi: 10.3847/1538-4365/ab45f3DOI
  17. [17]Gottlieb, S., & Shu, C.-W. 1998, Mathematics of Computation, 67, 73. http://www.jstor.org/stable/2584973
  18. [18]Gressel, O., & Ziegler, U. 2007, Computer Physics Communications, 176, 652, doi: 10.1016/j.cpc.2007.01.010
  19. [19]Hanasz, M., Kowalik, K., Wóltański, D., Pawłaszck, R., & Kornet, K. 2010, in EAS Publications Series, Vol. 42, EAS Publications Series, ed. K. Goździewski, A. Niedzielski, & J. J. Schneider, 281–285, doi: 10.1051/eas/1042030DOI
  20. [20]Harten, A., Lax, P. D., & Leer, B. v. 1983, SIAM Review, 25, 35, doi: 10.1137/1025002DOI
  21. [21]Hawley, J. F., Gammie, C. F., & Balbus, S. A. 1995, ApJ, 440, 742, doi: 10.1086/175311DOI
  22. [22]Hopkins, P. F., Squire, J., & Seligman, D. 2020, MNRAS, 496, 2123, doi: 10.1093/mnras/staa1046DOI
  23. [23]Huang, P., & Bai, X.-N. 2022, ApJS, 262, 11, doi: 10.3847/1538-4365/ac76cbDOI
  24. [24]Jeans, J. H. 1902, Philosophical Transactions of the Royal Society of London Series A, 199, 1, doi: 10.1098/rsta.1902.0012DOI
  25. [25]Johansen, A., Andersen, A. C., & Brandenburg, A. 2004, A&A, 417, 361, doi: 10.1051/0004-6361:20034417DOI
  26. [26]Krapp, L., & Benítez-Llambay, P. 2020, Research Notes of the American Astronomical Society, 4, 198, doi: 10.3847/2515-5172/abc7beDOI
  27. [27]Krapp, L., Benítez-Llambay, P., Gressel, O., & Pessah, M. E. 2019, ApJL, 878, L30, doi: 10.3847/2041-8213/ab2596DOI
  28. [28]Krapp, L., Kratter, K. M., & Youdin, A. N. 2022, ApJ, 928, 156, doi: 10.3847/1538-4357/ac5899DOI
  29. [29]Laibe, G., & Price, D. J. 2012, MNRAS, 420, 2345, doi: 10.1111/j.1365-2966.2011.20202.xDOI
  30. [30]Laibe, G. 2014, MNRAS, 444, 1940, doi: 10.1093/mnras/stu1367DOI
  31. [31]Lebreuilly, U., Commerçon, B., & Laibe, G. 2019, A&A, 626, A96, doi: 10.1051/0004-6361/201834147DOI
  32. [32]Lehmann, A., & Wardle, M. 2018, MNRAS, 476, 3185, doi: 10.1093/mnras/sty450DOI
  33. [33]LeVeque, R. J. 1992, Numerical methods for conservation laws (2. ed.)., Lectures in mathematics (Birkhäuser), 1–214
  34. [34]Leveque, R. J. 2004, Journal of Hyperbolic Differential Equations, 01, 315, doi: 10.1142/S0219891604000135
  35. [35]Lombart, M., & Laibe, G. 2021, MNRAS, 501, 4298, doi: 10.1093/mnras/staa3682DOI
  36. [36]Mignone, A. 2014, Journal of Computational Physics, 270, 784, doi: 10.1016/j.jcp.2014.04.001
  37. [37]Mignone, A., Bodo, G., Massaglia, S., et al. 2007, ApJS, 170, 228, doi: 10.1086/513316DOI
  38. [38]Mignone, A., Flock, M., & Vaidya, B. 2019, ApJS, 244, 38, doi: 10.3847/1538-4365/ab4356DOI
  39. [39]Miniati, F. 2010, Journal of Computational Physics, 229, 3916, doi: https://doi.org/10.1016/j.jcp.2010.01.034
  40. [40]Muley, D., Melon Fuksman, J. D., & Klahr, H. 2023, arXiv e-prints, arXiv:2308.03504, doi: 10.48550/arXiv.2308.03504DOI
  41. [41]Paardekooper, S. J., & Mellema, G. 2006, A&A, 453, 1129, doi: 10.1051/0004-6361:20054449DOI
  42. [42]Pareschi, L., & Russo, G. 2005, Journal of scientific computing, 25, 129–155, doi: 10.1007/s10915-004-4636-4
  43. [43]Pollack, J. B., Hollenbach, D., Beckwith, S., et al. 1994, ApJ, 421, 615, doi: 10.1086/173677DOI
  44. [44]Price, D. J., Wurster, J., Tricco, T. S., et al. 2018, Publications of the Astronomical Society of Australia, 35, e031, doi: 10.1017/pasa.2018.25
  45. [45]Rosenbrock, H. H. 1963, The Computer Journal, 5, 329, doi: 10.1093/comjnl/5.4.329DOI
  46. [46]Rusanov, V. 1970, Journal of Computational Physics, 5, 507, doi: https://doi.org/10.1016/0021-9991(70)90077-X
  47. [47]Shu, C.-W., & Osher, S. 1988, Journal of Computational Physics, 77, 439, doi: https://doi.org/10.1016/0021-9991(88)90177-5
  48. [48]Squire, J., & Hopkins, P. F. 2018, MNRAS, 477, 5011, doi: 10.1093/mnras/sty854DOI
  49. [49]Stone, J. M., Tomida, K., White, C. J., & Felker, K. G. 2020, ApJS, 249, 4, doi: 10.3847/1538-4357/ab929b
  50. [50]Stoyanovskaya, O. P., Glushko, T. A., Syntnikov, N. V., & Snytnikov, V. N. 2018, Astronomy and Computing, 25, 25, doi: 10.1016/j.ascom.2018.08.004DOI
  51. [51]Strang, G. 1968, SIAM Journal on Numerical Analysis, 5, 506, doi: 10.1137/0705041DOI
  52. [52]Toro, E. F. 2009, Riemann Solvers and Numerical Methods for Fluid Dynamics, 3rd edn. (Berlin, Heidelberg: Springer-Verlag), doi: 10.1007/b79761DOI
  53. [53]van Leer, B. 1977, Journal of Computational Physics, 23, 276, doi: 10.1016/0021-9991(77)90095-XDOI
  54. [54]Whipple, F. L. 1972, in From Plasma to Planet, ed. A. Elvius, 211
  55. [55]Yang, C.-C., & Johansen, A. 2016, ApJS, 224, 39, doi: 10.3847/0067-0049/224/2/39DOI
  56. [56]Youdin, A., & Johansen, A. 2007, ApJ, 662, 613, doi: 10.1086/516729DOI
  57. [57]Youdin, A. N., & Goodman, J. 2005, ApJ, 620, 459, doi: 10.1086/426895DOI
  58. [58]Zhu, Z., & Yang, C.-C. 2021, MNRAS, 501, 467, doi: 10.1093/mnras/staa3628DOI

Paper details

vlnc:W_v7bfurqs
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 11

downloads · 0

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

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents