betaFeedback

WorksPhysics and Astronomy

Pseudo-spectral solver versus grid-based solver: A quantitative accuracy test using GMHD3D and PLUTO4.4

Abstract

We provide a thorough comparison of the GMHD3D code and the PLUTO4.4 code for both two and three-dimensional hydrodynamic and magnetohydrodynamic problems. The open-source finite-volume solver PLUTO4.4 and the in-house developed pseudo-spectral multi-GPU solver GMHD3D both can be used to model the dynamics and turbulent motions of astrophysical plasmas. Although GMHD3D and PLUTO4.4 utilize different implementations, it is found that simulation results for hydrodynamic and magnetohydrodynamic problems, such as the rate of instability growth, 3-dimensional turbulent dynamics, oscillation of kinetic & magnetic energy, and recurrence dynamics, are remarkably similar. However, it is shown that the pseudo spectral solver GMHD3D is significantly more superior than the grid based solver PLUTO4.4 for certain category of physics problems.

Introduction

Plasma is a collection of charged particles that is frequently represented as a fluid on which electromagnetic body forces operate. It has been observed that a spatially averaged model termed the “fluid model” is particularly efficient in predicting the behavior of the plasma when a large number of charged particles is present. Therefore, Maxwell’s equations coupled with the equations of hydrodynamics become the primary governing equations for the motion of the charged-fluid element in the presence of an self-generated electric and magnetic field. The subject that studies the self-consistent evolution of such a magnetized plasma fluid is known as MagnetoHydroDynamics (MHD). In order to better understand the behavior of astrophysical plasmas like those found in the Sun or other young stars, the theories of HydroDynamics (HD)[1, 2, 3] and MagnetoHydroDynamics (MHD)[4, 5, 6] are frequently employed to investigate HD turbulence and magnetized plasma turbulence, respectively. The theory of MHD is useful for a wide variety of purposes, including the analysis of three-dimensional magnetized plasma turbulence, which is essential for understanding the fundamental behavior of astro-plasmas present in the Sun and other young stars[7], the physics of magnetic reconnection, and the careful operation of complex fusion reactors like Stellarators and Tokamaks[8].

Plasmas are inherently turbulent, whether they are found in burning stars or fusion devices. Because of nonlinear interactions across different scales of length, energy can cascade through different modes in a fully formed turbulent plasma medium. It has long been a problem in fluid dynamics to characterize the nature of the cascade of kinetic energy for a given initial spectrum. Almost all plasma physics problems - from relativistic jets[9] to angular momentum transport in accretion disks[10] - require an in-depth understanding of the turbulence in the plasma on a variety of length scales. Understanding plasma turbulence is essential for controlling the disruption[11] of plasma in such experimental devices, thereby enhancing plasma confinement for fusion plasmas and enabling the prediction of extreme events in astrophysical objects and stellar matter. Additionally, the magnetic field-lines combined with such a plasma flow provide mechanism for the transfer of energy and dynamics within the plasma. Due to the strong coupling, the dynamics of a completely developed turbulent plasma present in stellar objects need to be treated appropriately. Understanding and accurately analyzing the defining characteristics of this plasma turbulence is essential for forecasting events in our nearest star, our “Sun”, or in the fusion reactors operating in the many laboratories.

An important challenge in astrophysical plasmas is the creation of multi-scale magnetic fields, which occurs in the Sun, newborn stars, accretion disks, and other astronomical entities. The “Dynamo Theory” of Parker[12] is one of the first explanations for the generation of such magnetic multi-scale fields. Such large or intermediate-scale magnetic field is generated at the expense of the plasma kinetic energy, which primarily governs the dynamics of the charged fluid (plasma) via a time-dependent Lorentz force (back-reaction) term added to the Navier-Stokes equation, thereby self-consistently influencing the dynamics of the fluid flow. Consequently, the turbulence in an MHD plasma is fundamentally different from hydrodynamic turbulence. Search for a rapid growth of magnetic fields in astrophysical objects remains one of the most fascinating areas of study[13, 14, 15, 16, 17, 18, 19, 20].

In general, one needs to solve the set of coupled partial differential equations in order to deal with the complex astrophysical MHD phenomena outlined above. It is extremely challenging to solve the set of coupled nonlinear MHD equations analytically. For this reason, high-performance numerical solvers are required to accurately represent the physics problems occurring in plasmas. In order to simulate the MHD systems on a wide scale, including astrophysical entities and laboratory scenarios, it is necessary to develop highly scalable codes.

For the purpose of simulating plasma flows in astrophysics and in the lab, a variety of numerical MHD solvers have been developed. Popular examples of such codes include ZEUS[21], BATS-R-US[22], FLASH[23, 24], PLUTO[25], ATHENA[26], NIRVANA[27], M3DC1[28], GHOST[29, 30], BIFORST[31], PEGASUS[32], ENZO[33], CAFEQ[34], GKEYLL[35, 36], HMHD[37], CANS+[38], DEDALUS[39], PENCIL[40], CLT[41] to mention some.

At the Institute for Plasma Research [IPR], INDIA, we have recently upgraded an already existing three-dimensional compressible single GPU MHD solver[42, 43] to multi-node, multi-card GPU architecture [GMHD3D][44] using OpenAcc & MPI and achieved considerable speed increases across 32 P100 GPU cards[45]. The continuity, momentum, and energy equations for fluid and magnetic variables, with a thermodynamic closure for pressure, are solved using the GMHD3D solver using a pseudo-spectral technique. The solver currently employs OpenMPI/4.0.1 for its multi-node communication and the AccFFT library46 for FFT operations. PyEVTK[47], a data converter (ASCII to BINARY) written in Python, is designed to dump data in VTK binary format for the sake of visualization. After dumping the data file to binary an open source visualization softwares like, VisIt 3.1.2[48] and Paraview[49] are used for visualization.

In this study, we provide a comprehensive comparison between the aforementioned in-house pseudo-spectral MHD solver (GMHD3D)[44] and an open source grid based MHD solver PLUTO4.4[25] for some specific physics problems. A number of earlier works[50, 51] have reported comparative analyses between several numerical codes. The primary goal of current investigation is to validate the precision of the recently developed GPU solver and to compare the superiority of a pseudo-spectral solver to that of a grid-based solver atleast for certain class of physics problems.

The organization of the paper is as follows. In Sec. II we present about the dynamic equations. About our numerical solver and simulation details of the solver are described in Sec. III. Section IV is dedicated to the simulation results that we obtained from both the codes. Finally the summary and conclusions are listed in Sec. V.

Governing Equations

The governing equations for the single fluid MHD plasma are as follows,

∂ρ∂t+∇⃗⋅(ρu⃗)=0(1)\frac{\partial\rho}{\partial t}+\vec{\nabla}\cdot(\rho\vec{u})=0 \tag*{(1)}
∂(ρu⃗)∂t+∇⃗⋅[ρu⃗⊗u⃗+(P+B22)I−B⃗⊗B⃗]=1Re∇2u⃗+f⃗(2)\frac{\partial(\rho\vec{u})}{\partial t}+\vec{\nabla}\cdot\left[\rho\vec{u}\otimes\vec{u}+\left(P+\frac{B^2}{2}\right)\mathbf{I}-\vec{B}\otimes\vec{B}\right]=\frac{1}{R_e}\nabla^2\vec{u}+\vec{f} \tag*{(2)}
P=Cs2ρ(3)P=C_s^2\rho \tag*{(3)}
∂B⃗∂t+∇⃗⋅(u⃗⊗B⃗−B⃗⊗u⃗)=1Rm∇2B⃗(4)\frac{\partial\vec{B}}{\partial t}+\vec{\nabla}\cdot\left(\vec{u}\otimes\vec{B}-\vec{B}\otimes\vec{u}\right)=\frac{1}{R_m}\nabla^2\vec{B} \tag*{(4)}

for the above said system of equations, ρ\rho, u⃗\vec{u}, PP and BB represent the density, velocity, kinetic pressure and magnetic fields respectively. f⃗\vec{f} is the external driver available in the system. All quantities are appropriately normalised as discussed below. GMHD3D suite also provides a choice between utilizing the energy equation or the equation of state. For all cases studied here, we have used equation of state (see Eq. (3)).

We define Alfven speed as, VA=u0MAV_A=\frac{u_0}{M_A}, here MAM_A is the Alfven Mach number of the plasma flow and u0u_0 is a typical velocity scale. Sound speed of the fluid is defined as Cs=u0MsC_s=\frac{u_0}{M_s}, where MsM_s is the sonic Mach number of the fluid flow and the dynamic sound speed CsC_s contains the inherent information regarding the temperature of the system. The initial magnetic field present in the plasma is calculated from relation B0=VAρ0B_0=V_A\sqrt{\rho_0}, ρ0\rho_0 is the initial density of the flow. The time is normalized to Alfven times as t=t0×t′t=t_0\times t', t0=LVAt_0=\frac{L}{V_A} and length to a typical characteristic length scale L.

The dimensionless numbers are defined as, Re=u0LμR_e=\frac{u_0L}{\mu}, Rm=u0LηR_m=\frac{u_0L}{\eta}, here ReR_e and RmR_m are the kinetic Reynolds number and magnetic Reynolds number, μ\mu & η\eta are the kinematic viscosity and magnetic diffusivity. Magnetic Prandtl number is also defined as, PM=RmReP_M=\frac{R_m}{R_e}. The symbol ’⊗\otimes’ represents the dyadic between two vector quantities.

All magnetic variables, including Equation (4), are disabled for addressing the physics of hydrodynamic systems.

About numerical solver & Simulation Details

Before comparing the open-source grid-based MHD solver PLUTO4.4 with the in-house developed pseudo-spectral MHD solver GMHD3D, we provide additional details on both numerical solvers.

Details of GMHD3D suite

In order to study the plasma dynamics governed by MHD equations described above, we have recently upgraded an already existing well bench-marked single GPU MHD solver, developed in house at Institute For Plasma Research to multi-node, multi-card (multi-GPU) architecture. After multi-GPU upgrade, we obtain a 675.5x speedup across 32 P100 GPU cards in comparison to the MPI version, and a 32x speedup in comparison to the single-GPU version (See Table I & Figure 1)[44]. This GPU based magnetohydrodynamic solver (GMHD3D) is capable of handling very large grid sizes.

ArchitectureWall time (in Hrs)
8 Core (Intel Xeon 6148)60.80
1 NVIDIA P100 Card2.88
2 NVIDIA P100 Cards1.45
32 NVIDIA P100 Cards0.09

Table I. Time taken for hundred iteration at 5123512^3 grid resolution using GMHD3D.

Speed Up of GMHD3D code

Figure 1. Speed Up of GMHD3D code. A 675.5x speedup is obtained across 32 P100 GPU cards in comparison to the MPI version, and a 32x speedup in comparison to a single GPU version. Vertical axis set in log for plotting purpose only.

As outlined before, GMHD3DGMHD3D is a multi-node, multi-card, three dimensional (3D), weakly compressible, pseudo-spectral, visco-resistive solver[44]. This suite (GMHD3D) includes both 2-dimensional and 3-dimensional HydroDynamic (HD) and MagnetoHydrodynamic (MHD) solvers. It uses pseudo-spectral technique to simulate the dynamics of 3D magnetohydrodynamic plasma in a cartesian box with periodic boundary condition. By this technique one calculates the spatial derivative to evaluate non-linear term in governing equations with a standard 23\frac{2}{3} de-aliasing rule[52]. OpenACC FFT library (AccFFT library46) is used to perform Fourier transform and Adams-bashforth time solver, for time integration.

For 3D iso-surface visualization, an open source Python based data converter to VTK (Visualization Tool kit) by “PyEVTK”[47] is developed, which converts ASCII data to VTK binary format. After dumping the state data files to VTK, an open source visualization softwares, VisIt 3.1.2[48] and Paraview[49] is used to visualize the data. The further details of GMHD3D suite are given in Table II.

EquationsCompressible Navier-Stokes + Maxwell’s Equations
Dimension2D & 3D
Physics Modules2D Hydrodynamics, 3D Hydrodynamics, 2D MHD, 3D MHD
Spatial Derivative SolverPseudo-Spectral
Time integrationAdams-Bashforth, Runge-Kutta 4
ArchitectureSingle GPU, Multiple GPU
ParallelizationOpenACC, OpenMPI
LibrariescuFFT, AccFFT
PrecisionDouble
LanguageFortran 95
VisualizationGnuplot, Python, VisIt, Paraview

Table II. Features of GMHD3D suite.

As we mentioned above, we have upgraded a well bench-marked single GPU MHD solver to multi-GPU architecture [GMHD3D][44], we only crosschecked the upgraded solver accuracy with the existing one and observe that results match upto machine precision. Few more benchmark details can be found in some earlier works[3, 19, 44].

Details of PLUTO4.4 code

PLUTO4.4 is a multi-physics, multi-algorithm, high resolution code that can solve hyper sonic flows in one, two, and three spatial dimensions[25]. In order to solve the system of non-linear equations, a finite volume/finite difference approach is employed. PLUTO4.4 is parallelized with the help of the MPI Library via global domain decomposition.

Different Reconstruction algorithms are available in PLUTO4.4; however, we employ PARABOLIC reconstruction, which employs the piece wise parabolic method (PPM) to determine the spatial order of integration. The stencil requires either 3 or 5 zones and was implemented by Migone et al.[53]. We consider the RK3 scheme for time stepping because it is compatible with PARABOLIC reconstruction[25].

Since the governing equation requires an Isothermal equation of state, we set the EOS module in PLUTO4.4 to be in the ISOTHERMAL state[25]. For the isothermal equation of state, Migone et al. demonstrated that the hlld solver provides the highest precision of any Riemann solver[54]. For this reason, we employ the hlld Riemann solver for PLUTO4.4 throughout our study. Table III contains a complete listing of all input modules for PLUTO4.4.

ReconstructionPARABOLIC
Time SteepingRK3
EOSISOTHERMAL
Div. B ControlCONSTRAINED TRANSPORT
ResistivityEXPLICIT
ViscosityEXPLICIT
LimiterMC LIM

Table III. Initial input modules for PLUTO4.4.

Cost metric comparison

All simulations using the PLUTO4.4 and GMHD3D codes were executed on the 1 Petaflop ANTYA[45] super-computer located at the Institute for Plasma Research in India. PLUTO4.4 utilizes a dual configuration of 20 CPU cores, namely the Intel Xeon 6148 model, operating at a clock speed of 2.4 GHz. The system is equipped with a total of 384 GB DDR4 RAM. For simulations utilizing GMHD3D code we have used GPU nodes of ANTYA cluster with similar specification along with two NVIDIA tesla P100 GPU cards in a single node with 16 GB RAM each.

To conduct a cost metric comparison between the two solvers, a series of simulation runs have been performed, varying the number of resources (CPUs and GPUs) and grid points. We have plotted the normalized computational costs for the GPUs and CPUs in relation to the grid resolutions (refer to Figure 2a & Figure 2b). From Figure 2a & Figure 2b it is readily understood that the computational expenses increases linearly for both CPUs and GPUs as grid points increases. It can be observed from Figure 2a & Figure 2b that the normalized computational time of 16 GPUs is nearly similar to the computational time of 400 CPUs in cases when the computational workload is significant.

Cost metric comparison plots

Figure 2. Cost metric comparison in terms of GPU/CPU hours (normalized) versus number of grid points for GMHD3D and PLUTO4.4. For both algorithms, the time stepping is kept constant, and the wall time is calculated for a total of 10410^{4} iterations. For this investigation, we have utilized a total of 16 GPU cards (P100) for GMHD3D and (a) 16 CPU cores & (b) 400 CPU cores (Intel Xeon 6148) for PLUTO4.4. (c) Comparison of cost vs accuracy as a function of the number of grid points (NN) for GMHD3D and PLUTO4.4. The accuracy is measured by the value of the error percentage (Error(%)Error(\%)) as explained in the text.

It is widely acknowledged that CPUs exhibit a higher power consumption, as compared to CPUs, GPUs which have a lower power consumption because of their shared memory architecture. Therefore, the maintenance of 400 CPUs would result in higher computational expenses, including electrical power consumption, cooling, and rack space, as compared to the maintenance of 16 GPU cards. This, in turn, signifies the cost-effectiveness of the GPU solver.

It is also important to note that, the GPUs performances would also depend on the architectures. The computational efficiency of P100 cards, for instance, is lower than that of V100 and A100 cards. Consequently, additional investigation is required to determine the efficacy of the GMHD3D solver on the A100 architecture. Furthermore, comparing two GPU solvers is more meaningful than comparing the computational expenses of a CPU solver and a GPU solver, and to the best of our knowledge, the GPU version of the PLUTO4.4 code is currently being developed. Hence, such comparisons are what we plan to include in future communications.

The calculation of the accuracy (value of error %) is determined by the following formula:

Error(%)=100∗W1−W2W1(5)Error(\%) = 100 \ast\frac{W_1 - W_2}{W_1} \tag*{(5)}

where, W1W_1 represents the original expected value, while W2W_2 represents the value observed in the numerical simulation. In order to determine the accuracy, expressed as a percentage of error, we have considered an individual test problem that was utilized for the purpose of conducting a cost comparison analysis. The difference in peak values are calculated for various grid points using both codes. We have plotted the value of percentage of error (Error(%)Error(\%)) for the GPU solver (GMHD3D) and CPU solver (PLUTO 4.4) in relation to the grid resolutions (see Figure 2c). It is evident from Figure 2c that the GPU solver (GMHD3D) converged rapidly to errors of the order of less than 1%, whereas PLUTO 4.4 requires a more precise grid (higher grid resolution) in order to converge. Figure 2c demonstrates that the GMHD3D code fulfills the accuracy criteria at the lowest cost, whereas the PLUTO 4.4 code is the most expensive. This indicates the cost-effectiveness and accuracy of the GPU solver compared to the CPU solver being discussed.

Numerical Tests

It has been evident from the preceding discussion that we intend to provide a detailed comparison between two alternative framework solvers, one of which is an in-house developed pseudo-spectral solver and the other being an open source grid based solver. We have considered some well-known test problems in two- and three-dimensional hydrodynamics and magnetohydrodynamics to accomplish this. For example,

  • 2-dimensional Kelvin-Helmholtz instability (Details are described in IV A).

  • Dynamics of 3-dimensional Taylor-Green (TG) vortex (Details are described in IV B).

  • Coherent nonlinear oscillations using 2D Orszag-Tang (OT) Flow (Details are described in IV C).

  • Coherent nonlinear oscillations using 2D Cats Eye (CE) Flow (Details are described in IV D).

  • Coherent nonlinear oscillations using 3-dimensional astrophysical Flows (Details are described in IV E).

  • Coherent nonlinear oscillations for driven Flows (Details are described in Appendix B).

  • Recurrence dynamics in 3D MHD plasma (Details are described in IV F).

Parameter information for each individual test problem is provided in their respective subsections.

Test 1 [Hydrodynamics]: 2-dimensional Kelvin-Helmholtz instability

Using the GMHD3D solver and the PLUTO4.4 solver, we have investigated the 2-dimensional Kelvin-Helmholtz (KH) instability for hydrodynamic systems. We assume a simulation box with dimensions Lx=1L_x = 1 and Ly=2L_y = 2, and that the initial pressure (p0p_0) and initial density (ρ0\rho_0) are each to be unity. We apply a shear velocity along the xx-direction of the form,

ux=U0[tanh⁡(y−Ly3a)−tanh⁡(y−2Ly3a)−1](6)u_x = U_0 \left[\tanh\left(\frac{y-\frac{L_y}{3}}{a}\right) - \tanh\left(\frac{y-\frac{2L_y}{3}}{a}\right) - 1\right] \tag*{(6)}

where U0=0.645U_0 = 0.645 is the shear flow strength and a=0.05a = 0.05 is the shear width. We introduce a sinusoidal perturbation in the direction perpendicular to the initial flow velocity, of the form,

uy=uy0sin⁡(kxx)exp⁡[−(y−Ly3)2σ2]+uy0sin⁡(kxx)exp⁡[−(y−2Ly3)2σ2](7)\begin{aligned} u_y &= u_{y0}\sin(k_xx)\exp\left[-\frac{\left(y-\frac{L_y}{3}\right)^2}{\sigma^2}\right] \\ &\quad+u_{y0}\sin(k_xx)\exp\left[-\frac{\left(y-\frac{2L_y}{3}\right)^2}{\sigma^2}\right] \tag*{(7)} \end{aligned}

with σ=4a\sigma= 4a, to this initial configuration. Here, uy0=10−4u_{y0} = 10^{-4} represents the amplitude of the velocity perturbation, and kxk_x represents the mode of the velocity perturbation. For our system, the Sonic Mach number is defined as Ms=U0CsM_s = \frac{U_0}{C_s}, where CsC_s is the sound speed. We have investigated KH instability in the compressible limit using these parameter spaces. While maintaining the Sonic Mach number constant (i.e. Ms=0.5M_s = 0.5), we run simulations of the KH instability with both the solver (GMHD3D & PLUTO4.4) at different modes of perturbation (kxk_x) [See Figure 3a]. Also, by tracking the yy-direction kinetic energy (Ey=∫12ρvy2dxdyE_y = \int\frac{1}{2}\rho v_y^2 dxdy), we are able to calculate the growth rate (γ\gamma) of the KH instability in both codes [See Figure 3a]. The relationship between the KH growth rate (γ\gamma) and the perturbation wave number (kxk_x) is also calculated. As can be seen in Figure 3b, a perfect inverted parabola fits the predicted growth rates for varied mode numbers from both the GMHD3D and PLUTO4.4 codes. From Figure 3b, we can further conclude that the KH instability is stabilized for both small and large wave numbers, with the maximum growth rate occurring at kxa=0.4k_xa = 0.4. This finding agrees well with that of Keppens et al[55].

Comparison of kinetic energy and KH instability growth rate

Figure 3. (a) Comparison of kinetic energy from both GMHD3D and PLUTO4.4 code at different mode number of perturbation in the direction perpendicular to the flow direction is evaluated with time. (b) The growth rate (γaU‾0\frac{\gamma_a}{\overline{U}_0}) of KH instability at different normalized mode of perturbation (kxak_xa) with sonic Mach number Ms=0.5M_s = 0.5, where aa is the shear width. The calculated growth rates from both the codes are exactly identical and it is fitted perfectly by an inverted parabola which is identical to Keppens et al.[55]. Simulation Details: Grid resolution N=5122N = 512^2, Time stepping dt=10−4dt = 10^{-4}.

Next, by holding the mode number of excitation at kx=2πk_x = 2\pi, we investigate the impact of compressibility on KH instability [See Fig. 4a]. Similar growth rates are obtained from both codes for various Sonic Mach numbers [See Fig. 4b]. Also, the KH growth rates as a function of MsM_s exhibit an inverted parabolic nature [See Fig. 4b], as determined by Keppens et al.[55].

We have also shown the evolution of the vorticity profile of KH instability to conduct a more in-depth comparison between the GMHD3D and PLUTO4.4 codes. It is evident from Figure 5 that the outcomes of both solvers are identical.

Time evolution of vorticity for two oppositely directed KH unstable jets from GMHD3D and PLUTO4.4

Figure 5. Time evolution of vorticity for two oppositely directed KH unstable jets (broken jet) from GMHD3D code [upper panel (a–d)] and PLUTO4.4 code [lower panel (e–h)]. Both the solver has captured identical dynamics. Simulation Details: Grid resolution N=5122N = 512^{2}, Time stepping dt=10−4dt = 10^{-4}.

Test 2 [Hydrodynamics]: Dynamics of 3-dimensional Taylor-Green (TG) vortex

In this subsection, we investigate the conventional 3-dimensional Taylor-Green (TG) vortex problem in the incompressible. For the purpose of testing the accuracy of numerical solvers and algorithms, Taylor-Green vortex flow is frequently employed as a usual benchmark problem. At time t=0t = 0, the components of velocity are as follows,

ux=cos⁡xsin⁡ycos⁡zuy=−sin⁡xcos⁡ycos⁡zuz=0(8)\begin{aligned} u_x &= \cos x \sin y \cos z \\ u_y &= -\sin x \cos y \cos z \\ u_z &= 0 \tag*{(8)} \end{aligned}

Even though the initial zz-component of the velocity field is zero, the flow evaluated over time is three-dimensional. At time t=3.5t = 3.5 and in the z=π4z = \frac{\pi}{4} plane, we have calculated the xx-component (uxu_x) and the zz-component (uzu_z) of velocity using the GMHD3D code and the PLUTO4.4 code, respectively [See Fig. 6]. As can be seen in Figure 6, our findings from both codes are identical. When compared to Orszag[56] and Sharma et al.[57], our numerical observation using the GMHD3D code and the PLUTO4.4 code shows good agreement.

Figure 6. Contour plot of xx-component of velocity (uxu_x) and zz-component of velocity (uzu_z) at z=π4z = \frac{\pi}{4} plane from GMHD3D code (a & b) and PLUTO4.4 code (c & d). Our numerical observation from both the codes match well with the observation of Orszag[56] and Sharma et al[57]. Simulation Details: Reynolds number Re=100Re = 100, Grid resolution N=1283N = 128^3, Time stepping dt=10−4dt = 10^{-4}.

We compute the growth of normalized mean square vorticity, denoted by (Ω(t)Ω(0))\left(\frac{\Omega(t)}{\Omega(0)}\right) [where Ω(t)=1N∭ω2(x,y,z,t)dxdydz\Omega(t) = \frac{1}{N}\iiint\omega^2(x,y,z,t)\mathrm{d}x\mathrm{d}y\mathrm{d}z; NN is total number of grid points], using the GMHD3D code and the PLUTO4.4 code for two distinct values of Reynolds numbers [See Fig. 7a & 7b]. The results are in agreement with the earlier observation by Sharma et al.[57]. From Figure 7a & Figure 7b, we observe that the GMHD3D code effectively reproduces the growth of normalized men square vorticity at a grid resolution of 64364^3, whereas the PLUTO4.4 code requires at least 2563256^3. The accuracy of the pseudo-spectral solver over a grid-based solution is evident from this finding.

Growth of normalized mean square vorticity for Reynolds numbers $R_e=300$ and $R_e=400$

Figure 7. Growth of normalized mean square vorticity Ω(t)Ω(0)\frac{\Omega(t)}{\Omega(0)} calculated as function of time for Reynolds number Re=300R_e=300 & 400400. We observe a secondary peak for Re=400R_e=400, at around t=9t=9 from both the solver. Simulation Details: Time stepping dt=10−4dt=10^{-4}.

For Reynolds number Re=2000Re = 2000, we plot ωx\omega_x contours in the x=0.03πx = 0.03\pi plane. Based on Figure 8, we see that the vortices start stretching at an early time (t=0.03t = 0.03), and that this vortex stretching causes a period doubling bifurcation at a later time (t≥1.25t \ge1.25). Each Taylor-Green vortex cell bifurcates into four daughter cells due to period doubling bifurcation. We report identical observation from both the numerical solver (GMHD3D & PLUTO4.4) at grid resolution 5123512^3 [See Fig. 8].

Contours of the x-component of vorticity at five times from GMHD3D and PLUTO4.4

Figure 8. Contours of the xx-component of vorticity (ωx\omega_x) at the x=0.03πx=0.03\pi plane for Reynolds number Re=2000R_e=2000 from GMHD3D code [upper panel (a–e)] and PLUTO4.4 code [lower panel (f–j)]. Each Taylor-Green vortex cell forms into four new smaller cells due to period doubling bifurcation. Simulation Details: Reynolds number Re=2000R_e=2000, Grid resolution N=5123N=512^3, Time stepping dt=10−4dt=10^{-4}.

We have created a 3D iso-contour visualization of ωz\omega_z using data from both GMHD3D and PLUTO4.4 to obtain a more comprehensive look at the flow evolution. Figure 9 shows that the Taylor-Green vortex is laminar and anisotropic at early stages and that energy is transported to smaller scales as a result of vortex stretching. The impact of the vortex stretching term is clearly seen in Figure 9a & Figure 9f. As it is seen in Figure 9b & Figure 9g, the vortices roll up at an intermediate time, followed by a coherent breakdown [See Figure 9c & Figure 9h] and the creation of small-scale structures. Later on, as shown in Figure 9d & Figure 9i, the flow becomes completely turbulent, eventually ending in a turbulent decay [See Figure 9e & Figure 9j]. Using the GMHD3D code and the PLUTO4.4 code, we are able to capture the whole dynamics of a Taylor-Green vortex, including its stretching, rolling up, dividing, and reconnecting (vortex break down), turbulence, turbulent decays, etc. We find that our numerical observation is consistent with other, earlier studies[57].

3-dimensional iso-surfaces of the z-component of vorticity from GMHD3D and PLUTO4.4 at five times

Figure 9. 3-dimensional Iso-surfaces of the zz-component of vorticity (ωz\omega_z) from GMHD3D code [upper panel (a–e)] and PLUTO4.4 code [lower panel (f–j)]. The flow evolution dynamics consists of several process like vortex stretching (a & f), vortex roll up (b & g), vortex break down (c & h), turbulence (d & i) and finally turbulent decay (e & j). Simulation Details: Reynolds number Re=400Re = 400, Grid resolution N=2563N = 256^3, Time stepping dt=10−4dt = 10^{-4}.

Comparison of kinetic energy and growth rate at different sonic Mach numbers

Figure 4. (a) Comparison of kinetic energy from both GMHD3D and PLUTO4.4 code at different sonic Mach number (MsM_s) in the direction perpendicular to the flow direction is evaluated with time. (b) The growth rate (γaU0\frac{\gamma_a}{U_0}) of KH instability at different sonic Mach number (MsM_s) with kx=2πk_x = 2\pi. The calculated growth rates from both the codes are exactly identical and it is fitted perfectly by an inverted parabola which is identical to Keppens et al[55]. Simulation Details: Grid resolution N=5122N = 512^2, Time stepping dt=10−4dt = 10^{-4}.

In the following, some standard magnetohydrodynamic problems are considered to test the MHD module. We present the same in the coming subsection.

Test 3 [Magnetohydrodynamics]: Coherent nonlinear oscillations using 2D Orszag-Tang (OT) Flow

Orszag and Tang were the first to examine the Orszag-Tang flow[58]. Since then, it has been tested and compared a lot in numerical MHD simulation models. 2D Orszag-Tang [OT] Flow is known as the divergence free flow. For 2D Orszag-Tang [OT] flow, the velocity profile takes the form,

ux=−u0[Asin⁡(k0y)]uy=u0[Asin⁡(k0x)](9)\begin{aligned} u_x &= -u_0[A\sin(k_0y)] \\ u_y &= u_0[A\sin(k_0x)] \tag*{(9)} \end{aligned}

with A=1.0A = 1.0 and k0=1.0k_0 = 1.0. The initial magnetic field is assumed to be homogeneous and ambient, with a value determined by the Alfven Mach number (MAM_A) and the initial fluid velocity (u0u_0). Sonic mach (MsM_s) for our model is equal to 0.01. In this simulation, we consider a grid resolution of 1282128^2 for both codes, and we find that kinetic energy is converted to magnetic energy and vise-versa at regular intervals, in the form of coherent nonlinear oscillations [See Figure 10a]. Now, a close review of Figure 10a reveals that, substantial differences between the PLUTO4.4 and GMHD3D data do exist. Even though both sets of code have identical parameters, the data does not match very well. Using the appropriate electric field reconstruction techniques [CT_EMF_AVG] available in the PLUTO4.4 code, we are able to improve the results obtained from the code. We utilize the CT_CONTACT scheme, which is the least dissipative EMF AVG scheme available in PLUTO4.4[59]. Using this method, it is seen that the kinetic and magnetic energy oscillations from the PLUTO4.4 code are accurately recreated, and that it matches the GMHD3D data perfectly [See Figure 10b]. The period of oscillation is seen to be T=2.971T = 2.971 from both the code.

Shifted kinetic and magnetic energies for 2D Orszag–Tang Flow from GMHD3D and PLUTO4.4

Figure 10. The shifted kinetic and magnetic energies for 2D Orszag-Tang Flow from GMHD3D and PLUTO4.4 code (a) with out CT_CONTACT scheme (b) with CT_CONTACT scheme at grid resolution 1282128^2. PLUTO4.4 data is perfectly matched with GMHD3D data by employing appropriate electric field reconstruction algorithms. Simulation Details: Time stepping dt=10−4dt = 10^{-4}.

We also visualize the kinetic energy contour [See Figure 11a & Figure 11c] and the magnetic energy contour [See Figure 11b & Figure 11d] using GMHD3D and PLUTO4.4 data, and we find that they are similar.

Figure 11. 2D Orszag-Tang Flow kinetic energy contour and magnetic energy contour from GMHD3D code (a & b) and PLUTO4.4 code (c & d). Simulation Details: Reynolds number Re=Rm=105R_e = R_m = 10^5, Grid resolution N=5122N = 512^2, Time stepping dt=10−4dt = 10^{-4}, initial fluid velocity u0=1.0u_0 = 1.0, Alfven Mach number MA=1.0M_A = 1.0.

Test 4 [Magnetohydrodynamics]: Coherent nonlinear oscillations using 2D Cats Eye (CE) Flow

The velocity profile for 2D Cats Eye [CE] flow is given by,

ux=u0[sin⁡(k0x)cos⁡(k0y)−Acos⁡(k0x)sin⁡(k0y)]u_x = u_0[\sin(k_0x)\cos(k_0y) - A\cos(k_0x)\sin(k_0y)]
uy=−u0[cos⁡(k0x)sin⁡(k0y)+Asin⁡(k0x)cos⁡(k0y)](10)u_y = -u_0[\cos(k_0x)\sin(k_0y) + A\sin(k_0x)\cos(k_0y)] \tag*{(10)}

with AA equal to 0.5 and k0k_0 equal to 1. Here, we consider that the Alfven Mach number MA=1.0M_A = 1.0 and the initial speed of the fluid, u0=1.0u_0 = 1.0, from which the initial magnetic field strength is determined. For this simulation, we use a grid resolution of 1282128^{2} and set the sonic Mach number (MsM_s) to 0.01. The conversion of energy from the kinetic to the magnetic mode is shown clearly in Figure 12a. It is also evident that, the oscillations are significantly dampened for the PLUTO4.4 solver [See Figure 12a].

Two-panel plots of shifted kinetic, magnetic, and total energies for 2D Cats Eye Flow

Figure 12. (a) The shifted kinetic and magnetic energies for 2D Cats Eye Flow from GMHD3D and PLUTO4.4 at grid resolution 1282128^2. (b) The shifted kinetic energy for 2D Cats Eye Flow from GMHD2D and PLUTO4.4 (with all CT_EMF_AVG schemes) at grid resolution 1282128^2. Among all the CT_EMF_AVG schemes UCT_HLLD (shown by the cyan line) and CT_CONTACT (represented by the orange line) shows the least amount of dissipation. Simulation Details: Time stepping dt=10−4dt = 10^{-4}

We employ various electric field averaging approach (CT EMF AVG) in PLUTO4.4, similar to the previously stated instance, in order to improve the precision of the results. If we adopt the UCT HLL[60, 61] technique, we find that the kinetic and magnetic energy oscillations are similarly dampened. As seen in Figure 12b, if we do not employ any of the CT EMF AVG algorithms, the results appear to be the same with UCT HLL scheme. While investigating the other schemes like, UCT GFORCE[62], ARITHMETIC[63], CT FLUX, UCT HLLD[62], and CT CONTACT[59], we notice that UCT HLLD (shown by the cyan line) and CT CONTACT (represented by the orange line) exhibit the least amount of dissipation [See Figure 12b]. The results from GMHD3D and PLUTO4.4 still differ significantly from one another. This discrepancy may be owing to the fact that PLUTO4.4 has a higher numerical viscosity.

To eliminate the impact of numerical viscosity and double-check the resolution effect, we have increased the grid resolution in PLUTO4.4 from 1282128^2 to 204822048^2. Since UCT_HLLD [See Figure 13a] and CT_CONTACT [See Figure 13b] are the least dissipative with respect to others [See Figure 12b] at grid resolution 1282128^2, we focus our attention only on these two schemes for our higher resolution analysis. We can see that the PLUTO4.4 data at grid resolution 5122512^2 (represented by magenta line) agrees with the GMHD3D data at grid resolution 1282128^2 (shown by blue dotted line) by comparing Figure 13a & Figure 13b.

Two-panel plots of shifted kinetic energy for 2D Cats Eye Flow at multiple grid resolutions

Figure 13. The shifted kinetic energy for 2D Cats Eye Flow from GMHD3D at grid resolution: 1282128^2 and PLUTO4.4 at grid resolution: 1282128^2, 2562256^2, 5122512^2, 102421024^2 and 204822048^2 using (a) UCT_HLLD scheme (b) CT_CONTACT scheme. PLUTO4.4 data at grid resolution 5122512^2 (magenta line) agrees with the GMHD3D data at grid resolution 1282128^2 (blue dotted line). Simulation Details: Time stepping dt=10−4dt = 10^{-4}

Moreover, we compare the kinetic [See Figure 14a & Figure 14c] and magnetic [See Figure 14b & Figure 14d] energy contours from GMHD3D and PLUTO4.4 data. It has been established that the contours for both codes are identical.

Figure 14. 2D Cats Eye Flow kinetic energy contour and magnetic energy contour from GMHD3D code (a & b) and PLUTO4.4 code (c & d). Simulation Details: Reynolds number Re=Rm=105Re = Rm = 10^{5}, Grid resolution N=5122N = 512^{2}, Time stepping dt=10−4dt = 10^{-4}, initial fluid velocity u0=1.0u_{0} = 1.0, Alfven Mach number MA=1.0M_{A} = 1.0.

Test 5 [Magnetohydrodynamics]: Coherent nonlinear oscillations using 3-dimensional astrophysical Flows

In this subsection, we discuss the dynamics of some well-known three-dimensional astrophysical flows for example: Taylor-Green flow (See Appendix A for details), Archontis flow (See Appendix A for details), Cats Eye flow (See Appendix A for details) & Arnold–Beltrami–Childress Flow.

3D Arnold–Beltrami–Childress [ABC] Flow

Lastly, we look at the most well-known flow in astrophysics, which is called the 3D Arnold-Beltrami-Childress flow, or 3D ABC flow in short. The flow is divergence-free, and it is widely acknowledged in the astrophysical research area for its complicated nature and numerous symmetries. The velocity profile for 3D ABC flow is given by,

ux=u0[Asin⁡(k0z)+Ccos⁡(k0y)]uy=u0[Bsin⁡(k0x)+Acos⁡(k0z)]uz=u0[Csin⁡(k0y)+Bcos⁡(k0x)](11)\begin{aligned} u_x &= u_0[A\sin(k_0z)+C\cos(k_0y)]\\ u_y &= u_0[B\sin(k_0x)+A\cos(k_0z)]\\ u_z &= u_0[C\sin(k_0y)+B\cos(k_0x)] \tag*{(11)} \end{aligned}

with A=B=C=1.0A=B=C=1.0 and k0=1.0k_0=1.0. The remaining parameters are identical to those used in previous numerical experiments. The consistent and periodic exchange of energy between kinetic and magnetic modes is shown in Figure 15a in the form of coherent non-linear oscillation. The oscillation periods are measured to be T=30.171T = 30.171 for both codes.

Figure 15. (a) The shifted kinetic and magnetic energies for 3D Arnold-Beltrami-Childress [ABC] flow from GMHD3D and PLUTO4.4 code. The visualization of velocity iso-surface (Iso-V) for 3D Arnold-Beltrami-Childress [ABC] flow at any arbitrary time from (b) GMHD3D code and (c) PLUTO4.4 code. Simulation Details: Reynolds number Re=Rm=1000R_e = R_m = 1000, Grid resolution N=1283N = 128^3, Time stepping dt=10−4dt = 10^{-4}, initial fluid velocity u0=1.0u_0 = 1.0, Alfven Mach number MA=1.0M_A = 1.0.

We also visualize the velocity iso-surface (Iso-V) using data from both codes and confirm that the two isosurfaces (Iso-V) are identical [See Figure 15b & Figure 15c]. Using 3D ABC flow from both codes, we further study some parameter scanning.

We begin by investigating the impact of Alfven speed on coherent non-linear oscillations. As shown in Figure 16a & Figure 16b, the period of oscillation of energy (kinetic and magnetic) linearly increases with increasing of Alfven Mach number (MAM_A) from both codes when the initial wave number (k0k_0) remains constant at 1.0 and the Alfven Mach number (MAM_A) is varied over the range 0.1, 0.2, 0.3, 0.4, 0.5, 1.0, 1.5. According to recently published works, these results are quite predictable and consistent[64].

Shifted magnetic energy for 3D Arnold-Beltrami-Childress flow

Figure 16. Shifted magnetic energy for 3D Arnold–Beltrami–Childress [ABC] flow with Alfven mach number (a) Ma=0.1,0.2,0.3,0.4M_a = 0.1, 0.2, 0.3, 0.4, (b) MA=0.5,1.0,1.5M_A = 0.5, 1.0, 1.5 (c) growth of magnetic energies with Alfven Mach number (MAM_A) = 10.0, 100.0, 1000.0 from GMHD3D code and PLUTO4.4 code at grid resolution 64364^3. It appears that the results from both codes are identical. Simulation Details: Time stepping dt=10−4dt = 10^{-4}.

Moreover, we see that the oscillation completely disappears as the Alfven Mach number (MAM_A) is increased to very high limits such as MA=10,100,1000M_A = 10, 100, 1000. In addition, a notable saturation of magnetic energy is observed, followed by a growth [See Figure 16c]. The concept of “dynamo action” is used to describe this process of magnetic energy growth.

In the following, we present findings from both codes for a wide range of the initial wave number (k0k_0).

We fix the Alfven Mach number (MAM_A) at 1.0 and vary k0k_0 value in the range k0=1,2,4,8,16k_0 = 1, 2, 4, 8, 16 to analyze the impact of the initial mode number on coherent nonlinear oscillation. It can be seen in Figure 17a & Figure 17b that when the initial wave number (k0k_0) increases, the frequency of oscillation increases, i.e. the time period of oscillation reduces. For lower wave numbers, such as k0=1.0,2.0,4.0k_0 = 1.0, 2.0, 4.0, the energy oscillation is reproduced identically by both codes [See Figure 17a]; however, for higher wave numbers, such as k0=8.0,16.0k_0 = 8.0, 16.0, the oscillation of energy is heavily damped for the PLUTO4.4 code [See Figure 17b].

Shifted kinetic energy for 3D Arnold–Beltrami–Childress flow with initial wave numbers $k_0=1.0, 2.0, 4.0$ and $k_0=8.0, 16.0$

Figure 17. Shifted kinetic energy for 3D Arnold–Beltrami–Childress [ABC] Flow with initial wave number (a) k0=1.0,2.0,4.0k_0 = 1.0, 2.0, 4.0 and (b) k0=8.0,16.0k_0 = 8.0, 16.0 from GMHD3D code and PLUTO4.4 code at grid resolution 64364^3. For lower wave numbers, the oscillations are exactly reproduced from both codes but for higher wave numbers the oscillations are heavily damped for PLUTO4.4 code. Simulation Details: Time stepping dt=10−4dt = 10^{-4}.

In order to improve PLUTO4.4 results, we first adopt the same approach as we conducted for the 2D Cats Eye [CE] flow, i.e., we investigate all of the available CT EMF AVG schemes.

From Figure 17a, we can see that when k0=1,2,4k_0 = 1, 2, 4, the data for both codes are exactly the same. For the sake of completeness, we investigate all electric field averaging techniques for the k0=4.0k_0 = 4.0 scenario and find that, with the exception of the UCT HLL[60, 61] and UCT GFORCE[62] schemes, the findings are identical for all schemes [See Figure 18a]. In comparison to other schemes, these two are discovered to have a height diffusive effect.

Shifted kinetic energy using 3D Arnold–Beltrami–Childress flow with initial wave numbers $k_0=8.0$ and $k_0=16.0$

Figure 18. Shifted kinetic energy using 3D Arnold–Beltrami–Childress [ABC] Flow with initial wave number (a) k0=8.0k_0 = 8.0 and (b) k0=16.0k_0 = 16.0 from GMHD3D code and PLUTO4.4 code (with all available CT EMF AVG schemes) at grid resolution 64364^3. Among all the schemes UCT HLL and UCT GFORCE have height diffusivity. Simulation Details: Time stepping dt=10−4dt = 10^{-4}.

Figure 17b shows that the oscillations are significantly dampened for larger wave numbers (k0=8,16k_0 = 8, 16). We now look into the most efficient emf averaging techniques for k0=8.0k_0 = 8.0. From Figure 18b, we can see that, of all the possible schemes, only the ARITHMETIC63 and CT_CONTACT59 ones are the best ones for this present case. It is important to notice that the PLUTO4.4 data does not perfectly correspond with the GMHD3D data at the 64364^3 grid resolution, even if we are utilizing ARITHMETIC63 and CT_CONTACT59 schemes.

To further improve our results, we provide higher resolution runs for PLUTO4.4 with ARITHMETIC and CT_CONTACT averaging schemes. It is easy to observe from Figure 19 that the PLUTO4.4 data with a greater resolution, i.e. 1283128^3, matches the GMHD3D data with a lower resolution, 64364^{3}, exactly [See Fig. 19]. The results are quite encouraging, as it demonstrate the superior accuracy of the spectral solver compared to that of the grid-based solver in a triply periodic domain.

Shifted kinetic energy plots for 3D Arnold–Beltrami–Childress flow with $k_0=8.0$

Figure 19. Shifted kinetic energy using 3D Arnold–Beltrami–Childress [ABC] Flow for k0=8.0k_0 = 8.0 from GMHD3D code and PLUTO4.4 code with (a) ARITHMETIC scheme (b) CT_CONTACT scheme. PLUTO4.4 code needs atleast 1283128^3 grid resolution to reproduce the oscillation that we get from GMHD3D code at 64364^3 grid resolution. Simulation Details: Time stepping dt=10−4dt = 10^{-4}.

Based on our findings in the k0=8k_{0} = 8 case, we investigate all of the electric field averages techniques in the k0=16k_{0} = 16 case as well [See Fig. 20a]. We can see that none of the available algorithms work sufficiently well to make the oscillation we are looking for, which we obtained from GMHD3D at 64364^{3}. All of the emf averaging approaches demonstrate that the oscillations are significantly damped. So, the best way to re-solve this right now is to improve the grid resolution. As in the previously stated scenario, we increase the grid resolution while keeping the ARITHMETIC[63] and CT_CONTACT[59] averaging schemes in place. In conjunction with the ARITHMETIC averages scheme, it is evident from Figure 20b that the high resolution simulation

Shifted kinetic energy plots for 3D Arnold–Beltrami–Childress flow with $k_0=16.0$

Figure 20. (a) Shifted kinetic energy using 3D Arnold–Beltrami–Childress [ABC] Flow for k0=16.0k_0 = 16.0 from GMHD3D code and PLUTO4.4 code (with all available CT-EMF_AVG schemes). High resolution study of 3D Arnold–Beltrami–Childress Flow for k0=16.0k_0 = 16.0 from GMHD3D code [at grid resolution 64364^3, 1283128^3, 2563256^3 & 5123512^3] and PLUTO4.4 code [at grid resolution 64364^3, 1283128^3, 2563256^3 & 5123512^3] with (b) ARITHMETIC scheme (c) CT_CONTACT scheme. PLUTO4.4 code requires atleast 2563256^3 grid resolution to reproduce the oscillation that we get from GMHD3D code at 64364^3 grid resolution. Simulation Details: Time stepping dt=10−4dt = 10^{-4}.

at grid resolution 1283128^{3} is unable to reproduce the oscillation in its entirety. It is likely to be reproduced with a 2563256^{3} grid resolution. The same issue is seen when using the CT_CONTACT scheme [See Fig. 20c].

These findings are also quite promising. Our analysis shows that PLUTO4.4 requires 2563256^{3} grid resolution, but GMHD3D only need 64364^{3} grid resolution, in order to re-solve the highest initial wave number (k0=16.0k_{0} = 16.0). This finding once again demonstrates the superior accuracy of a spectral solver over a grid-based solver.

So far, we have discussed about several of the well-known flows without a driving mechanism. The impact of an external driver on these flows are also investigated both in two and three dimensions (See Appendix B for details).

Test 6 [Magnetohydrodynamics]: Recurrence Dynamics in 3D MHD plasma

Coherent nonlinear oscillations of kinetic and magnetic energy in the form of Alfven waves have been demonstrated in 3D single fluid magnetohydrodynamic plasmas, as detailed above. A periodic reconstruction of the initial fluid flow and magnetic variables mediated by coherent non-linear oscillations is predicted when the energy alternates between kinetic and magnetic forms. This phenomenon is called recurrence. It is recently discovered that astrophysical plasmas exhibit two distinct types of flow[65]. Unlike regular 3D Arnold-Beltrami-Childress [ABC] flow, the initial velocity and magnetic field surface of these flows cannot be reconstructed, hence they are characterized as non-recurring[65]. Another type of flow is the 3D Taylor-Green [TG] flow, which shows full recurrence by reconstructing the structure of the isosurfaces of kinetic and magnetic energy[65]. These flows are called Recurring flows.

Here, we use the recently developed code GMHD3D and the open-source code PLUTO4.4 to investigate the recurrent phenomenon.

Non-Recurring 3D ABC flow

For the recurrence study, we first focus on the divergence-free 3D Arnold-Beltrami-Childress [ABC] flow. In this case, the profile of the flow is given by

ux=u0[Asin⁡(k0z)+Ccos⁡(k0y)]uy=u0[Bsin⁡(k0x)+Acos⁡(k0z)]uz=u0[Csin⁡(k0y)+Bcos⁡(k0x)](12)\begin{aligned} u_x &= u_0[A\sin(k_0z) + C\cos(k_0y)] \\ u_y &= u_0[B\sin(k_0x) + A\cos(k_0z)] \\ u_z &= u_0[C\sin(k_0y) + B\cos(k_0x)] \tag*{(12)} \end{aligned}

where A=B=C=1A = B = C = 1 and k0=1k_0 = 1. For the 3D ABC flow, which we have previously investigated, oscillations of the kinetic and magnetic energy in the form of coherent non-linear oscillation have been observed. We now plot the velocity and magnetic field iso-surfaces from the GMHD3D code and the PLUTO4.4 code, respectively, as illustrated in Figure 21 & Figure 22. As depicted in Figure 21, the values of the velocity isosurfaces are 0.1 (Red), 0.09 (Green), 0.08 (Blue), 0.05 (Cyan), and 0.03 (Yellow).

Figure 21. Non-Recurring 3D Arnold–Beltrami–Childress [ABC] Flow Velocity iso-surface from GMHD3D code [Top two rows (a–h)] and PLUTO4.4 code [Bottom two rows (i–p)]. Values of iso-surface: 0.1 (Red), 0.09 (Green), 0.08 (Blue), 0.05 (Cyan) and 0.03 (Yellow). Simulation Details: Reynolds number Re=Rm=1000R_e = R_m = 1000, Grid resolution N=1283N = 128^3, Time stepping dt=10−4dt = 10^{-4}, initial fluid velocity u0=1.0u_0 = 1.0, Alfven Mach number MA=1.0M_A = 1.0.

Non-Recurring 3D ABC Flow Magnetic Field iso-surface from GMHD3D and PLUTO4.4 codes at times 0.0, 6.2, 12.5, 25.0, 31.2, 37.5, 43.7, and 56.2

Figure 22. Non-Recurring 3D ABC Flow Magnetic Field iso-surface from from GMHD3D code [Top two rows (a–h)] and PLUTO4.4 code [Bottom two rows (i–p)]. Values of iso-surface: 0.02 (Red), 0.185 (Green), 0.166 (Blue), 0.133 (Cyan) and 0.1 (Yellow). Simulation Details: Reynolds number Re=Rm=1000R_e = R_m = 1000, Grid resolution N=1283N = 128^3, Time stepping dt=10−4dt = 10^{-4}, initial fluid velocity u0=1.0u_0 = 1.0, Alfven Mach number MA=1.0M_A = 1.0.

Also, the values for the magnetic field iso-surface are as follows: 0.02 (Red), 0.185 (Green), 0.166 (Blue), 0.133 (Cyan), and 0.10 (Yellow) [See Figure 22]. From Figure 21 & Figure 22 it can be seen that neither the velocity isosurface nor the magnetic field isosurface are reconstructed back. This is a clear sign of non-recurrence for both the velocity and the magnetic field[65]. The same thing has been observed using both GMHD3D code and PLUTO4.4 code.

Recurring 3D Taylor-Green flow

Finally, we consider the Taylor-Green [TG] flow in three dimensions as the initial velocity profile. The flow profile looks like,

ux=u0[Acos⁡(k0x)sin⁡(k0y)cos⁡(k0z)]uy=−u0[Asin⁡(k0x)cos⁡(k0y)cos⁡(k0z)]uy=0(13)\begin{aligned} u_x &= u_0[A\cos(k_0x)\sin(k_0y)\cos(k_0z)] \\ u_y &= -u_0[A\sin(k_0x)\cos(k_0y)\cos(k_0z)] \\ u_y &= 0 \tag*{(13)} \end{aligned}

with A=1A = 1 and k0=1k_0 = 1. Iso-surfaces of velocity and magnetic field for this flow are shown in Figure 23 & Figure 24.

Recurring 3D Taylor-Green Flow Velocity iso-surface from GMHD3D and PLUTO4.4

Figure 23. Recurring 3D Taylor-Green Flow Velocity iso-surface from GMHD3D code [Top two rows (a–h)] and PLUTO4.4 code [Bottom two rows (i–p)]. Values of iso-surface: 0.001 (red), 0.01 (Green), 0.03 (Blue), 0.04 (Cyan), 0.05 (Yellow). Simulation Details: Reynolds number Re=Rm=1000R_e = R_m = 1000, Grid resolution N=1283N = 128^3, Time stepping dt=10−4dt = 10^{-4}, initial fluid velocity u0=1.0u_0 = 1.0, Alfven Mach number MA=1.0M_A = 1.0.

Recurring 3D Taylor-Green Flow Magnetic Field iso-surfaces from GMHHD3D and PLUTO4.4 at times 0.0, 6.2, 12.5, 25.0, 31.2, 37.5, 43.7, and 56.2

Figure 24. Recurring 3D Taylor-Green Flow Magnetic Field iso-surface from GMHHD3D code [Top two rows (a–h)] and PLUTO4.4 code [Bottom two rows (i–p)]. Values of iso-surface: 0.13 (Green), 0.15 (Yellow), 0.16 (Blue), 0.2 (Red). Simulation Details: Reynolds number Re=Rm=1000Re = Rm = 1000, Grid resolution N=1283N = 128^3, Time stepping dt=10−4dt = 10^{-4}, initial fluid velocity u0=1.0u_0 = 1.0, Alfven Mach number MA=1.0M_A = 1.0.

Figure 23 shows an iso-velocity surface with values of 0.001 (red), 0.01 (green), 0.02 (blue), 0.04 (cyan), and 0.05 (yellow).

The values of the magnetic field iso-surface are also 0.13 (Red), 0.15 (Green), 0.16 (Blue), 0.18 (Cyan), and 0.20 (Yellow) [See Figure 24].

Both the GMHD3D and PLUTO4.4 codes show that the velocity isosurface and the magnetic field isosurface continue to recur back. This 3D Taylor-Green [TG] flow is recognized as a recurrent flow since both the isosurfaces (velocity and magnetic field) are recurring[65].

A plausible explanation for Recurrence

To gain a thorough knowledge of the recurrence and non-recurrence phenomena, we use Thyagaraja’s[66] mathematical description in terms of Rayleigh Quotient (Q(t)). Recently, Mukherjee et al.[65] employed a modified form of the Rayleigh Quotient for MHD systems as,

Q(t)=∫V[(∇⃗×u⃗)2+12(∇⃗×B⃗)2] dV∫V[∣u⃗∣2+12∣B⃗∣2] dV(14)Q(t) = \frac{\int_V \left[(\vec{\nabla} \times\vec{u})^2 + \frac{1}{2}(\vec{\nabla} \times\vec{B})^2\right]\,dV}{\int_V \left[|\vec{u}|^2 + \frac{1}{2}|\vec{B}|^2\right]\,dV} \tag*{(14)}

Physically, Q(t)Q(t) is a measure of the number of active degrees of freedom possible in the system. It is already known that for typical hydrodynamic flows, the Rayleigh Quotient [Q(t)] is found to be bound in nature to demonstrate a recurrence phenomenon[66]. Recent study has shown that the Rayleigh Quotient [Q(t)] is bounded with time for Taylor-Green [TG] flow, but is unbounded for 3D Arnold-Beltrami-Childress [ABC] flow[65] in the presence of homogeneous ambient magnetic field.

Similar features, such as the time-dependent unbounded Rayleigh Quotient for 3D Arnold-Beltrami-Childress (ABC) flow and the time-dependent bounded

Rayleigh Quotient for 3D Taylor-Green (TG) flow, are also observed in this present work from both codes (GMHD3D and PLUTO4.4) [See Fig. 25].

Summary and Conclusion

In this study, we have examined the performance of two MHD codes, GMHD3D[44] (which is developed in-house at the Institute for Plasma Research, India) and PLUTO4.4[25] (which is freely available) on a number of different physics problems. For solving coupled partial differential equations, GMHHD3D employs a pseudo-spectral technique, while PLUTO4.4 employs a finite volume method.

The points to take home are:

Rayleigh Quotient calculation from GMHD3D code and PLUTO4.4 code

Figure 25. Rayleigh Quotient calculation from GMHD3D code and PLUTO4.4 code. Simulation Details: Grid resolution N=1283N = 128^3, Time stepping dt=10−4dt = 10^{-4}.

  • Using the GMHD3D and PLUTO4.4 codes, we have investigated the hydrodynamic 2D Kelvin-Helmholtz (KH) instability for two broken jets traveling in opposing directions at the compressible limit. We obtain the identical growth rates for the KH instability at grid resolution 5122512^2 from both codes at different sonic Mach numbers and different mode numbers, which is the same as what Keppens et al.[55] has reported.

  • Taylor-Green vortex evolution in 3D is investigated in the pure hydrodynamic limit utilizing both codes. The time evolution of mean square vorticity is estimated from both codes, and for a given value of Reynolds number, the PLUTO4.4 code requires a grid resolution of at least 2563256^3 to replicate the GMHD3D code’s 64364^3 grid resolution data. This numerical observation indicates that the spectral solver is more superior than a grid-based solver.

  • The period doubling bifurcation of the TG vortex is reproduced using both codes as part of this investigation. The results of our numerical observation are found to agree with those of some prior studies[57].

  • We explore the problem of coherent nonlinear oscillation using some well-known two-dimensional flows in the presence of a homogeneous and ambient magnetic field, and find that both codes produce almost identical results. The spatial distributions of kinetic and magnetic energy produced by both codes are similar.

  • In 2D-Orszag-Tang flow, the findings from both codes are identical, but in 2D-Cat’s Eye [CE] flow, the predicted oscillations of energy are found to be significantly dampened when utilizing the PLUTO4.4 solver. One possible explanation is that the PLUTO4.4 solver has a greater numerical viscosity.

  • To get rid of the observed damping, we use different electric field averaging techniques that are available in PLUTO4.4. It has been observed that the PLUTO4.4 solver is most efficient with the UCT_HLLD and CT_CONTACT EMF averaging techniques. However, with these two techniques present, obtaining the desired oscillations in PLUTO4.4 requires at least 5122512^2 grid resolution, whereas GMHD2D resolves that at 1282128^2 grid resolution. This essentially demonstrates the superiority of the spectral solver over the grid based solver.

  • For a variety of well-known 3D astrophysical processes, the GMHD3D and PLUTO4.4 codes produce identical oscillations of kinetic and magnetic energy in the form of Alfven waves.

  • We also look closely at the different parameter regimes for a certain flow, namely the 3D Arnold–Beltrami–Childress [ABC] flow. The findings from both codes are found to be identical for a range of Alfven speeds, from sub to super-Alfvenic. When the wave number is larger (k0=8,16k_0 = 8,16), however, the spectral solver (GMHD3D) is observed to be more superior than the grid based solver (PLUTO4.4), as the latter requires a grid resolution of 1283128^3&2563256^3 in order to recreate the results obtained by GMHD3D at a grid resolution of 64364^3.

  • It is found that both codes yield the same outcomes for externally driven flows in both 2-dimensions and 3-dimensions.

  • We finally reproduce the Recurrence phenomenon[65] in 3-dimensional MHD plasma using both the codes. The analytical description of the same seems to support it perfectly from both the solver.

The finite volume version of the PLUTO4.4 code maintains global second-order accuracy in time, whereas GMHD3D solver maintains up to 6’th order accuracy in space and 4’th order accuracy in time. Hence, the pseudo-spectral technique offers superior accuracy compared to finite volume and finite difference schemes. On the other hand, the PLUTO4.4 can handle non-periodic boundary conditions, Hall-MHD dynamics, can capture plasma flow with shock wave and so on. Also, the current version of the GMHD3D solver does not have plasma transport terms such as thermal diffusion (for example, Braginskii fluid equations[67]) as GMHD3D solves single fluid equations. Though GMHD3D can solve both dynamics energy equation and equation of state, in the current study, we have invoked equation of state (Eq. (3) as closure). A cost metric study for both the solvers (GMHD3D & PLUTO4.4) have also been carried out on GPUs & CPUs. It has been identified that as the number of grid points increases, the computational cost for both CPUs and GPUs also increases. It has been observed that the GPU solver is more cost-effective than the CPU solver for a substantial amount of computational load.

To conclude, we have attempted a systematic comparison of two MHD codes in both the hydrodynamic and magnetohydrodynamic limit. Our numerical analysis shows that while both algorithms produce comparable answers in most circumstances, the spectral solver surpasses the grid-based solver in periodic domain for a subset of physics-related challenges. We also believe this work highlights the advantages of a spectral solver over a grid-based solver. To the best of our knowledge, this is the first work ever attempted which makes a thorough comparision of a pseudo-spectral code with the freely available grid based MHD solver PLUTO4.4. We intend to extend this comparative study in the near future to include advanced grid-based solvers such as FLASH, HYDRA, MIRANDA, PENCIL, etc.

ACKNOWLEDGMENTS

The simulations and visualizations presented here are performed on GPU nodes and visualization nodes of the ANTYA cluster at the Institute for Plasma Research (IPR), India. We would like to express our gratitude to the anonymous referees for their insightful feedback, which significantly enhanced the manuscript’s quality. We are grateful to Dr. Dipanjan Mukherjee at IUCAA and Dr. Bhargav Vaidya at IITI for their insightful discussions regarding the PLUTO code. One of the author S.B is thankful to Dr. Rupak Mukherjee at Central University of Sikkim (CUS), Gangtok, Sikkim, India for providing an initial version of GMHD3D code. S.B. is thankful to N. Vydyanathan, Bengaluru, and B. K. Sharma at NVIDIA, Bengaluru, India, for extending their help with basic GPU methods. S.B. is grateful to Dr. Soumen De Karmakar at IPR for many helpful discussions regarding GPUs, and the HPC support team of IPR for extending their help related to the ANTYA cluster.

DATA AVAILABILITY

The data underlying this article will be shared on reasonable request to the corresponding author.

Conflict of Interest

The authors have no conflicts to disclose.

References

Coherent Nonlinear oscillations

3D Taylor-Green [TG] Flow

One of the most well-known flows in astrophysical plasma is the Taylor-Green (TG) flow. This is a divergence free flow. The flow profile looks like,

ux=u0[Acos⁡(k0x)sin⁡(k0y)cos⁡(k0z)]uy=−u0[Asin⁡(k0x)cos⁡(k0y)cos⁡(k0z)]uz=0(15)\begin{aligned} u_x &= u_0[A\cos(k_0x)\sin(k_0y)\cos(k_0z)] \\ u_y &= -u_0[A\sin(k_0x)\cos(k_0y)\cos(k_0z)] \\ u_z &= 0 \tag*{(15)} \end{aligned}

where A=1.0A = 1.0 and k0=1k_0 = 1 (the mode number). We also use the values MA=1.0M_A = 1.0 for the Alfven Mach number, u0=0.1u_0 = 0.1 for the initial fluid speed, and Ms=0.1M_s = 0.1 for the sonic Mach number. In the presence of a uniform and ambient beginning magnetic field, we observe an oscillation in the kinetic and magnetic energies due to the constant conversion and exchange of energy between the two modes [See Figure 26a]. From Fig. 26a it is observed that the outputs of the GMHD3D and PLUTO4.4 codes are identical.

Figure 26. (a) The shifted kinetic and magnetic energies for 3D Taylor-Green Flow (TG) from GMHD3D and PLUTO4.4 code. The visualization of velocity iso-surface (Iso-V) for 3D Taylor-Green (TG) Flow at any arbitrary time from (b) GMHD3D code and (c) PLUTO4.4 code. Simulation Details: Reynolds number Re=Rm=1000R_e = R_m = 1000, Grid resolution N=1283N = 128^3, Time stepping dt=10−4dt = 10^{-4}, initial fluid velocity u0=1.0u_0 = 1.0, Alfven Mach number MA=1.0M_A = 1.0.

We also visualize the velocity iso-surface (Iso-V surface) at any arbitrary time using GMHD3D data and PLUTO4.4 data, and discover that the iso-surfaces are indistinguishable from one another [See Figure 26b & Figure 26c].

3D Archontis [No cosine] Flow

The 3D Archontis [No cosine] flow is another interesting flow in the astrophysical plasmas. For this flow, we consider a three-dimensional velocity profile of the from,

ux=u0[Asin⁡(k0z)]uy=u0[Bsin⁡(k0x)]uy=u0[Csin⁡(k0y)](16)\begin{aligned} u_x &= u_0[A\sin(k_0z)] \\ u_y &= u_0[B\sin(k_0x)] \\ u_y &= u_0[C\sin(k_0y)] \tag*{(16)} \end{aligned}

where AA, BB, and CC are constants with unity value, and k0=1k_0 = 1 is the mode number. All other parameters remain the same as in the preceding cases.

Like in the earlier cases, we can see the periodic exchange of energy between the kinetic and magnetic regimes [See Figure 27a]. Furthermore, we find that the periods of oscillation in both codes are identical, coming in at T=30.680T = 30.680 [See Figure 27a].

Figure 27. (a) The shifted kinetic and magnetic energies for 3D Archontis [No cosine] flow from GMHD3D and PLUTO4.4 code. The visualization of velocity iso-surface (Iso-V) for 3D Archontis [No cosine] flow at any arbitrary time from (b) GMHD3D code and (c) PLUTO4.4 code. Simulation Details: Reynolds number Re=Rm=1000R_e = R_m = 1000, Grid resolution N=1283N = 128^3, Time stepping dt=10−4dt = 10^{-4}, initial fluid velocity u0=1.0u_0 = 1.0, Alfven Mach number MA=1.0M_A = 1.0.

Figures 27b & 27c illustrate the velocity iso-surface (Iso-V) of 3D Archontis flow as calculated using GMHD3D and PLUTO4.4 data, respectively. The iso-surface (Iso-V) representation reveals that the results from both codes are identical.

3D Cats Eye [CE] Flow

Cats Eye [CE] flow is another well-known flow profile that we choose for the sake of completeness. The equation describing the flow profile of 3D Cats Eye [CE] flow is as follows,

ux=u0[Bsin⁡(k0y)]uy=u0[Asin⁡(k0x)]uy=u0[Acos⁡(k0x)−Bcos⁡(k0y)](17)\begin{aligned} u_x &= u_0[B\sin(k_0y)]\\ u_y &= u_0[A\sin(k_0x)]\\ u_y &= u_0[A\cos(k_0x)-B\cos(k_0y)] \tag*{(17)} \end{aligned}

where A=35A=\sqrt{\frac{3}{5}} and B=2AB=2A are real constants. The rest of the parameters are kept the same throughout the simulation, just like in the previous case. From Figure 28a it is identified that for Cats Eye [CE] flow, the GMHD3D code and the PLUTO4.4 code reveal a periodic energy exchange between the kinetic and magnetic modes.

Figure 28. (a) The shifted kinetic and magnetic energies for 3D Cats Eye [CE] flow from GMHD3D and PLUTO4.4 code. The visualization of velocity iso-surface (Iso-V) for 3D Cats Eye [CE] flow at any arbitrary time from (b) GMHD3D code and (c) PLUTO4.4 code. Simulation Details: Reynolds number Re=Rm=1000R_e=R_m=1000, Grid resolution N=1283N=128^3, Time stepping dt=10−4dt=10^{-4}, initial fluid velocity u0=1.0u_0=1.0, Alfven Mach number MA=1.0M_A=1.0.

For additional cross-checking, we visualize the velocity iso-surface (Iso-V) using data extracted from both codes and determine that both iso-surfaces (Iso-V) are identical [See Figure 28b & Figure 28c].

Coherent Nonlinear oscillations for driven flows

Proofs

Test B1 [Magnetohydrodynamics]: Dynamics of a externally driven 2D flow

Until now, we have discussed the flow dynamics without any driving force in two dimensions. Here in this subsection, we discuss about externally forced flows in 2-dimensions. Here, we start with a 2D Orszag-Tang [OT] flow profile that looks like this:

ux=−u0[Asin⁡(k0y)]uy=u0[Asin⁡(k0x)](18)\begin{aligned} u_x &= -u_0[A\sin(k_0y)]\\ u_y &= u_0[A\sin(k_0x)] \tag*{(18)} \end{aligned}

Moreover, we force the flow profile to be equal to itself, i.e

f⃗=α[−sin⁡(k0y)sin⁡(k0x)](19)\vec{f}=\alpha \begin{bmatrix} -\sin(k_0y)\\ \sin(k_0x) \end{bmatrix} \tag*{(19)}

The forcing amplitude is α=0.1\alpha=0.1, and the drive mode number is k0=1.0k_0=1.0. Similar to the previous unforced Orszag-Tang [OT] scenario, we find that the kinetic and magnetic energy oscillate in the form of a coherent non-linear oscillation, but this time the peak magnitude varies with time from both the code [See Figure 29].

Shifted kinetic and magnetic energies for 2D forced Orszag-Tang Flow

Figure 29. The shifted kinetic and magnetic energies for 2D forced Orszag-Tang Flow from GMHD2D and PLUTO4.4 (with CT-CONTACT scheme) code at grid resolution 1282128^{2}. Coherent non-linear oscillation with a fluctuating peak magnitude with time is observed from both the code. Simulation Details: Time stepping dt=10−4dt = 10^{-4}

Test B2 [Magnetohydrodynamics]: Dynamics of a externally driven 3D flow (Forced ABC Flow)

In the previous section, we have talked about Arnold-Beltrami-Childress [ABC] Flow and how it relates to coherent non-linear oscillation. Now we talk about the case where we force the 3D ABC flow with the 3D ABC flow itself. The profile of forcing can be written as,

f⃗=α[Asin⁡(k0z)+Ccos⁡(k0y)Bsin⁡(k0x)+Acos⁡(k0z)Csin⁡(k0y)+Bcos⁡(k0x)](20)\vec{f} = \alpha \begin{bmatrix} A\sin(k_{0}z) + C\cos(k_{0}y) \\ B\sin(k_{0}x) + A\cos(k_{0}z) \\ C\sin(k_{0}y) + B\cos(k_{0}x) \end{bmatrix} \tag*{(20)}

where α=0.1\alpha= 0.1, A=B=C=1.0A = B = C = 1.0, and k0=1.0k_{0} = 1.0.

Coherent non-linear oscillations of kinetic and magnetic energy are seen in both the GMHD3D and PLUTO4.4 codes, much like in forced 2D Orszag-Tang [OT] flow [See Figure 30a]. The system appears to operate as a forced relaxed system, despite the presence of forcing.

Shifted kinetic and magnetic energies for 3D forced Arnold–Beltrami–Childress FlowLissajous curve for kinetic and magnetic energy

Figure 30. (a) The shifted kinetic and magnetic energies or 3D forced Arnold–Beltrami–Childress [ABC] Flow from GMHD3D code and PLUTO4.4 code. System effectively acts as a forced relaxed system in spite of presence of driver. (b) The Lissajous curve for kinetic and magnetic energy from GMHD3D and PLUTO4.4 code. Simulation Details: Grid resolution N=643N = 64^{3}, Time stepping dt=10−4dt = 10^{-4}.

Both codes show an identical consistency in the dynamics of the Lissajous curve for kinetic and magnetic energy [See Figure 30b].

References

  1. [1]R. H. Kraichnan, The Physics of Fluids 8, 1385 (1965), https://aip.scitation.org/doi/pdf/10.1063/1.1761412, URL https://aip.scitation.org/doi/abs/10.1063/1.1761412.
  2. [2]M. E. Brachet, M. Meneguzzi, H. Politano, and P. L. Sulem, Journal of Fluid Mechanics 194, 333–349 (1988).
  3. [3]S. Biswas and R. Ganesh, Physics of Fluids 34, 065101 (2022), https://doi.org/10.1063/5.0092212, URL https://doi.org/10.1063/5.0092212.
  4. [4]D. Biskamp, Magnetohydrodynamic turbulence (Cambridge University Press, 2003).
  5. [5]A. Beresnyak, Living Reviews in Computational Astrophysics 5, 1 (2019).
  6. [6]A. A. Schekochihin, Journal of Plasma Physics 88, 155880501 (2022).
  7. [7]A. R. Choudhuri et al., The physics of fluids and plasmas: an introduction for astrophysicists (Cambridge University Press, 1998).
  8. [8]R. Tanna, H. Raj, J. Ghosh, R. Kumar, S. Aich, T. Macwan, D. Kumawat, K. Jadeja, K. Patel, M. Kalal, et al., Nuclear Fusion 59, 112006 (2019), URL https://doi.org/10.1088/1741-4326/ab0a9e.
  9. [9]D. Mukherjee, G. Bodo, A. Mignone, P. Rossi, and B. Vaidya, Monthly Notices of the Royal Astronomical Society 499, 681 (2020), ISSN 0035-8711, https://academic.oup.com/mnras/article-pdf/499/1/681/33857350/staa2934.pdf, URL https://doi.org/10.1093/mnras/staa2934.
  10. [10]S. A. Balbus and J. F. Hawley, Rev. Mod. Phys. 70, 1 (1998), URL https://link.aps.org/doi/10.1103/RevModPhys.70.1.DOI
  11. [11]J. Ghosh, R. L. Tanna, P. K. Chattopadhyay, A. Sen, H. Raj, P. Dhyani, S. Dolui, S. V. Kulkarni, K. Mishra, R. Singh, et al., Physics of Plasmas 28, 052501 (2021), https://doi.org/10.1063/5.0037384, URL https://doi.org/10.1063/5.0037384.
  12. [12]E. N. Parker, The Astrophysical Journal 122, 293 (1955).
  13. [13]A. A. Schekochihin, S. C. Cowley, J. L. Maron, and J. C. McWilliams, Phys. Rev. Lett. 92, 054502 (2004), URL https://link.aps.org/doi/10.1103/PhysRevLett.92.054502.DOI
  14. [14]A. B. Iskakov, A. A. Schekochihin, S. C. Cowley, J. C. McWilliams, and M. R. E. Proctor, Phys. Rev. Lett. 98, 208501 (2007), URL https://link.aps.org/doi/10.1103/PhysRevLett.98.208501.DOI
  15. [15]J. Squire and A. Bhattacharjee, Phys. Rev. Lett. 115, 175003 (2015), URL https://link.aps.org/doi/10.1103/PhysRevLett.115.175003.DOI
  16. [16]D. A. St-Onge and M. W. Kunz, The Astrophysical Journal 863, L25 (2018), URL https://doi.org/10.3847/2041-8213/aad638.
  17. [17]D. A. St-Onge, M. W. Kunz, J. Squire, and A. A. Schekochihin, Journal of Plasma Physics 86, 905860503 (2020).
  18. [18]V. Skoutnev, J. Squire, and A. Bhattacharjee, The Astrophysical Journal 906, 61 (2021), URL https://doi.org/10.3847/1538-4357/abc8ee.
  19. [19]S. Biswas and R. Ganesh, Physica Scripta 98, 075607 (2023), URL https://dx.doi.org/10.1088/1402-4896/acdccf.
  20. [20]S. Biswas and R. Ganesh, Physics of Plasmas 30, 112902 (2023), ISSN 1070-664X, https://pubs.aip.org/aip/pop/article-pdf/doi/10.1063/5.0170796/18210251/112902_1.5.0170796.pdf, URL https://doi.org/10.1063/5.0170796.
  21. [21]J. M. Stone and M. L. Norman, The Astrophysical Journal Supplement Series 80, 753 (1992).
  22. [22]K. G. Powell, P. L. Roe, T. J. Linde, T. I. Gombosi, and D. L. De Zeeuw, Journal of Computational Physics 154, 284 (1999), ISSN 0021-9991, URL https://www.sciencedirect.com/science/article/pii/S0021999199629X.
  23. [23]B. Fryxell, K. Olson, P. Ricker, F. X. Timmes, M. Zingale, D. Q. Lamb, P. MacNeice, R. Rosner, J. W. Truran, and H. Tufo, The Astrophysical Journal Supplement Series 131, 273 (2000), URL https://doi.org/10.1086/317361.
  24. [24]A. Dubey, L. B. Reid, and R. Fisher, Physica Scripta T132, 014046 (2008), URL https://doi.org/10.1088/0031-8949/2008/t132/014046.
  25. [25]A. Mignone, G. Bodo, S. Massaglia, T. Matsakos, O. Tesileanu, C. Zanni, and A. Ferrari, The Astrophysical Journal Supplement Series 170, 228 (2007), URL https://doi.org/10.1086/513316.
  26. [26]J. M. Stone, T. A. Gardiner, P. Teuben, J. F. Hawley, and J. B. Simon, The Astrophysical Journal Supplement Series 178, 137 (2008), URL https://doi.org/10.1086/588755.
  27. [27]U. Ziegler, Computer Physics Communications 179, 227 (2008), ISSN 0010-4655, URL https://www.sciencedirect.com/science/article/pii/S0010465508000969.
  28. [28]S. C. Jardin, N. Ferraro, X. Luo, J. Chen, J. Breslau, K. E. Jansen, and M. S. Shephard, Journal of Physics: Conference Series 125, 012044 (2008), URL https://doi.org/10.1088/1742-6596/125/1/012044.
  29. [29]P. D. Mininni, GHOST, http://wp.df.uba.ar/mininni/ghost/.
  30. [30]P. D. Mininni, D. Rosenberg, R. Reddy, and A. Pouquet, Parallel Computing 37, 316 (2011), ISSN 0167-
  31. [31]Gudiksen, B. V., Carlsson, M., Hansteen, V. H., Hayek, W., Leenaarts, J., and Martínez-Sykora, J., A&A 531, A154 (2011), URL https://doi.org/10.1051/0004-6361/201116520.
  32. [32]M. W. Kunz, J. M. Stone, and X.-N. Bai, Journal of Computational Physics 259, 154 (2014), ISSN 0021-9991, URL https://www.sciencedirect.com/science/article/pii/S0021999113007973.
  33. [33]G. L. Bryan, M. L. Norman, B. W. O’Shea, T. Abel, J. H. Wise, M. J. Turk, D. R. Reynolds, D. C. Collins, P. Wang, S. W. Skillman, et al., The Astrophysical Journal Supplement Series 211, 19 (2014), URL https://dx.doi.org/10.1088/0067-0049/211/2/19.
  34. [34]F. D. Lora-Clavijo, A. Cruz-Osorio, and F. S. Guzmán, The Astrophysical Journal Supplement Series 218, 24 (2015), URL https://doi.org/10.1088/0067-0049/218/2/24.
  35. [35]E. L. Shi, G. W. Hammett, T. Stoltzfus-Dueck, and A. Hakim, Journal of Plasma Physics 83, 905830304 (2017).
  36. [36]N. R. Mandell, A. Hakim, G. W. Hammett, and M. Franciscquez, Journal of Plasma Physics 86, 905860109 (2020).
  37. [37]D. Hu, A. Bhattacharjee, and Y.-M. Huang, Physics of Plasmas 25 (2018), ISSN 1070-664X, 062305, https://pubs.aip.org/aip/pop/article-pdf/doi/10.1063/1.5022292/15941745/062305_1_online.pdf, URL https://doi.org/10.1063/1.5022292.
  38. [38]Y. Matsumoto, Y. Asahina, Y. Kudoh, T. Kawashima, J. Matsumoto, H. R. Takahashi, T. Minoshima, S. Zenitani, T. Miyoshi, and R. Matsumoto, Publications of the Astronomical Society of Japan 71 (2019), ISSN 0004-6264, 83, https://academic.oup.com/pasj/article-pdf/71/4/83/29163378/psz064.pdf, URL https://doi.org/10.1093/pasj/psz064.
  39. [39]K. J. Burns, G. M. Vasil, J. S. Oishi, D. Lecoanet, and B. P. Brown, Phys. Rev. Research 2, 023068 (2020), URL https://link.aps.org/doi/10.1103/PhysRevResearch.2.023068.DOI
  40. [40]T. P. C. Collaboration, A. Brandenburg, A. Johansen, P. A. Bourdin, W. Dobler, W. Lyra, M. Rheinhardt, S. Bingert, N. E. L. Haugen, A. Mee, et al., Journal of Open Source Software 6, 2807 (2021), URL https://doi.org/10.21105/joss.02807.
  41. [41]W. Zhang, S. Jardín, Z. Ma, A. Kleiner, and H. Zhang, Computer Physics Communications 269, 108134 (2021), ISSN 0010-4655, URL https://www.sciencedirect.com/science/article/pii/S0010465521002460.
  42. [42]R. Mukherjee, Turbulence, Flows and Magnetic Field Generation in Plasmas using a Magnetohydrodynamic Model [PhD thesis] (Institute for Plasma Research, India, 2019).
  43. [43]R. Mukherjee, R. Ganesh, V. Saini, U. Maurya, N. Vydyanathan, and B. Sharma, in 2018 IEEE 25th International Conference on High Performance Computing Workshops (HiPCW) (2018), pp. 46–55.DOI
  44. [44]S. Biswas and R. Ganesh, GPU Technology Conference 2022, https://www.nvidia.com/en-us/on-demand/session/gtcspring22-s41199/.DOI
  45. [45]ANTYA, Institute for Plasma Research (IPR), India, https://www.ipr.res.in/documents/high-performance_computing__it.html. A. Gholami, J. Hill, D. Malhotra, and G. Biros, http://arxiv.org/abs/1506.07933 (2016).
  46. [47]Paulo-herrera, PyEVTK, https://github.com/paulo-herrera/PyEVTK.
  47. [48]LLNL, VisIt, https://wci.llnl.gov/simulation/computer-codes/visit.
  48. [49]Kitware, Paraview, https://www.paraview.org/.
  49. [50]G. L. Koij, M. A. Botchev, E. M. Frederix, B. J. Geurts, S. Horn, D. Lohse, E. P. van der Poel, O. Shishkina, R. J. Stevens, and R. Verzicco, Computers & Fluids 166, 1 (2018), ISSN 0045-7930, URL https://www.sciencedirect.com/science/article/pii/S0045793018300100.
  50. [51]F. Capuano, N. Beratis, F. Zhang, Y. Peet, K. Squires, and E. Balaras, European Journal of Mechanics - B/Fluids 102, 91 (2023), ISSN 0997-7546, URL https://www.sciencedirect.com/science/article/pii/S0997754623000164.
  51. [52]G. S. Patterson and S. A. Orszag, The Physics of Fluids 14, 2538 (1971), https://aip.scitation.org/doi/pdf/10.1063/1.1693365, URL https://aip.scitation.org/doi/abs/10.1063/1.1693365.
  52. [53]A. Mignone, Journal of Computational Physics 270, 784 (2014), ISSN 0021-9991, URL https://www.sciencedirect.com/science/article/pii/S002199914002538.
  53. [54]A. Mignone, Journal of Computational Physics 225, 1427 (2007), ISSN 0021-9991, URL https://www.sciencedirect.com/science/article/pii/S0021999107000617.
  54. [55]R. KEPPENS, G. TÓTH, R. H. J. WESTERMANN, and J. P. GOEDBLOED, Journal of Plasma Physics 61, 1–19 (1999).
  55. [56]M. E. Brachet, D. I. Meiron, S. A. Orszag, B. G. Nickel, R. H. Morf, and U. Frisch, Journal of Fluid Mechanics 130, 411–452 (1983).
  56. [57]N. Sharma and T. K. Sengupta, Physics of Fluids 31, 035106 (2019), https://doi.org/10.1063/1.5083870, URL https://doi.org/10.1063/1.5083870.
  57. [58]S. A. Orszag and C.-M. Tang, Journal of Fluid Mechanics 90, 129–143 (1979).
  58. [59]T. A. Gardiner and J. M. Stone, Journal of Computational Physics 205, 509 (2005), ISSN 0021-9991, URL https://www.sciencedirect.com/science/article/pii/S0021999104004784.
  59. [60]Del Zanna, L., Bucciantini, N., and Londrillo, P., A&A 400, 397 (2003), URL https://doi.org/10.1051/0004-6361:20021641.
  60. [61]P. Londrillo and L. Del Zanna, Journal of Computational Physics 195, 17 (2004), ISSN 0021-9991, URL https://www.sciencedirect.com/science/article/pii/S0021999103005102.
  61. [62]A. Mignone and L. Del Zanna, Journal of Computational Physics 424, 109748 (2021), ISSN 0021-9991, URL https://www.sciencedirect.com/science/article/pii/S002199120305222.
  62. [63]D. S. Balsara and D. S. Spicer, Journal of Computational Physics 149, 270 (1999), ISSN 0021-9991, URL https://www.sciencedirect.com/science/article/pii/S002199198961538.
  63. [64]R. Mukherjee, R. Ganesh, and A. Sen, Physics of Plasmas 26, 024121 (2019), https://doi.org/10.1063/1.5083001, URL https://doi.org/10.1063/1.5083001.
  64. [65]R. Mukherjee, R. Ganesh, and A. Sen, Physics of Plasmas 26, 022101 (2019), https://doi.org/10.1063/1.5083052, URL https://doi.org/10.1063/1.5083052.
  65. [66]A. Thyagaraja, The Physics of Fluids 22, 2093 (1979), https://aip.scitation.org/doi/pdf/10.1063/1.862519, URL https://aip.scitation.org/doi/abs/10.1063/1.862519.
  66. [67]F. I. Parra, Braginskii fluid equations (2019), URL http://www-thphys.physics.ox.ac.uk/people/FelixParra/CollisionalPlasmaPhysics/notes/lecIV_braginskii.pdf.

Paper details

vlnc:W_5varygqc
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 8

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