betaFeedback

WorksPhysics and Astronomy

PIC-JRho_{ho}m: a pseudo-spectral Particle-In-Cell formulation with arbitrary charge and current densities time dependencies for the modeling of relativistic plasmas

Abstract

This paper introduces a novel formulation of the Particle-In-Cell (PIC) method for the modeling of relativistic plasmas, which leverages the ability of the Pseudo-Spectral Analytical Time-Domain solver (PSATD) to handle arbitrary time dependencies of the charge and current densities during one PIC cycle. The new formulation is applied to a modified set of Maxwell's equations that was proposed earlier in the context of divergence cleaning, and to recently proposed extensions of the PSATD-PIC algorithm. Detailed analysis and testings revealed that, under some condition, the new formulation can expand the range of numerical parameters under which PIC simulations are stable and accurate when modeling relativistic plasmas such as, e.g., plasma-based particle accelerators.

Introduction

Simulations of relativistic plasmas often rely on the electromagnetic particle-in-cell (PIC) method [2, 3] [1], with variations of the method that have been proposed and are chosen based on the application. For the modeling of plasma-based accelerators [4, 5], a variation that has gained in popularity uses the “infinite-order” (in space and time) Pseudo-Spectral Analytical Time-Domain (PSATD) method [6, 7], instead of the (almost universally adopted) second-order (in space and time) Finite-Difference Time-Domain (FDTD) “Yee” method [8], to solve Maxwell’s equations at discrete points in space and time. In contrast to the Yee solver, the PSATD solver offers no numerical dispersion and no Courant condition on the field solve. Extensions of the PSATD PIC method includes the use of finite-order spatial stencils [9, 10], alternating nodal-staggered representations of the field quantities during one PIC loop [11], time-averaging of the fields gathered onto the particles [12], and integration of the equations in a Galilean frame moving at a given velocity [13, 14]. The combination of the Galilean PSATD PIC (also labeled as Galilean PIC for convenience) method with the other extensions has led to stable modeling of plasma accelerators, free of the numerical Cherenkov instability (NCI) [15] when using the Lorentz boosted frame method to speed up simulations [16]. In some cases, however, the method, which relies on the user setting a predefined Galilean velocity, can become inaccurate when it cannot be assumed that the local plasma velocity is close to that predefined velocity. As a possible remedy, this paper introduces and starts exploring a novel formulation of the PIC algorithm where the standard assumption that the current density that is produced by the particles is constant over a time step is relaxed.

The reminder of the paper is organized as follows. The formulation of the novel algorithm is derived first, followed by the presentation of its finite-order stencil, alternating nodal-staggered and time-averaged extensions. The connection between the new algorithm and the Galilean PIC formulation is discussed next. The effectiveness of the new algorithm at mitigating the NCI is then explored theoretically and numerically on a simple uniform plasma case. Finally, the new scheme is tested in simulations of laser-plasma accelerators in a Lorentz boosted frame.

Novel PIC-JRhom algorithm

Presentation of the algorithm

The following modified system of Maxwell’s equations is considered

∂E∂t=c2∇×B−Jε0+c2∇F,(1)\frac{\partial\mathbf{E}}{\partial t}=c^2\nabla\times\mathbf{B}-\frac{\mathbf{J}}{\varepsilon_0}+c^2\nabla F, \tag*{(1)}
∂B∂t=−∇×E,(2)\frac{\partial\mathbf{B}}{\partial t}=-\nabla\times\mathbf{E}, \tag*{(2)}
∂F∂t=∇⋅E−ρε0.(3)\frac{\partial F}{\partial t}=\nabla\cdot\mathbf{E}-\frac{\rho}{\varepsilon_0}. \tag*{(3)}

In addition to the usual Maxwell-Faraday and Ampère-Maxwell equations, the system contains an extra equation for the scalar field FF, which propagates deviations to Gauss’ law. (Note that, in the case where Gauss’ law is verified in the PIC simulation, Eq. (1c) leads to F=0F=0, and Eqs. (1a),(1b) reduce to the standard Maxwell’s equations.) These additional terms were introduced in [17] from the potential formulation in the Lorentz gauge and used as a propagative divergence cleaning procedure, as an alternate to the Langdon-Marder [18] or Marder [19] diffusive ones. This type of divergence cleaning was also proposed independently and analyzed more formally in [20]. A connection to the formulation of Eqs. (1) in potential form, derived more formally than in [17], is instructive and given in Appendix A.

While the abovementioned earlier work [17, 20] considered this formulation in the context of the standard PIC method using FDTD discretization of Eqs. (1), this article focuses on the PSATD [6, 21] [1] discretization of Eqs. (1), where the equations are integrated analytically over one timestep, in Fourier space. The expression of (1) in Fourier space reads

∂E^∂t=ic2k×B^−J^ε0+ic2F^k,(4)\frac{\partial\hat{E}}{\partial t}=ic^{2}\mathbf{k}\times\hat{B}-\frac{\hat{J}}{\varepsilon_{0}}+ic^{2}\hat{F}\mathbf{k}, \tag*{(4)}
∂B^∂t=−ik×E^,(5)\frac{\partial\hat{B}}{\partial t}=-i\mathbf{k}\times\hat{E}, \tag*{(5)}
∂F^∂t=ik⋅E^−ρ^ε0,(6)\frac{\partial\hat{F}}{\partial t}=i\mathbf{k}\cdot\hat{E}-\frac{\hat{\rho}}{\varepsilon_{0}}, \tag*{(6)}

where f^\hat{f} denotes the Fourier transform of function ff. The analytical integration of Eqs (2) in time requires an assumption on the time dependency of the current and charge densities J^\hat{J} and ρ^\hat{\rho} over the integration interval, i.e., over a timestep that goes from t=nΔtt=n\Delta t to t=(n+1)Δtt=(n+1)\Delta t. In the standard PSATD algorithm [6, 7], J^\hat{J} is assumed to constant in time, and ρ^\hat{\rho} is assumed to be linear in time, within a given timestep Δt\Delta t.

This paper considers more general time dependencies for J^\hat{J} and ρ^\hat{\rho} within one timestep, which is divided into mm subintervals of equal size δt=Δt/m\delta t=\Delta t/m. During these subintervals, J^\hat{J} and ρ^\hat{\rho} are considered to be either piecewise constant, piecewise linear, or piecewise quadratic in time. This is illustrated in Figure 1. In the rest of this paper, the notation “PIC-JRho_m” is used, where J and Rho (J,Rho ∈{C\in\{C (constant), L (linear), Q (quadratic)}\}) indicate the (piecewise) time dependency of the current density J^\hat{J} and charge density ρ^\hat{\rho}, respectively, and mm is the number of subintervals. For example, “PIC-LL2“ refers to the novel PIC algorithm with linear time dependency of both J^\hat{J} and ρ\rho and 2 subintervals. Note that, in this notation, “PIC-CL1“ refers to the standard PSATD PIC algorithm [21], where J^\hat{J} is constant and ρ^\hat{\rho} is linear in time over one time step.

Figure 1. Diagrams illustrating various time dependencies of the current density JJ and charge density ρ\rho for constant/linear (CL), both constant (CC), linear (LL) and quadratic (QQ) dependencies with mm subintervals: (first column) m=1m = 1, (second) m=2m = 2 and (third) m=4m = 4. CL1 corresponds to the standard PSATD PIC method. The triangle and circle glyphs represent the times at which the macroparticles deposit ρ\rho and JJ on the grid, respectively. The dashed and solid lines represent the assumed time dependency of ρ\rho and JJ within one timestep, when integrating the Maxwell equations analytically.

More specifically:

  • When ρ^(t)\hat{\rho}(t) is assumed to be piecewise constant: macroparticles deposit their charge density in the middle of each time subinterval, i.e., at tn+(ℓ+1/2)/m≡nΔt+(ℓ+1/2)δtt_{n+(\ell+1/2)/m}\equiv n\Delta t+(\ell+1/2)\delta t with ℓ∈[0,m−1]\ell\in[0,m-1], and ρ^\hat{\rho} is then assumed to be constant in each subinterval:

ρ^(t)=ρn+(ℓ+1/2)/m,t∈[nΔt+ℓδt,nΔt+(ℓ+1)δt].\hat{\rho}(t)=\rho^{n+(\ell+1/2)/m},\quad t\in[n\Delta t+\ell\delta t,n\Delta t+(\ell+1)\delta t].
  • When ρ^(t)\hat{\rho}(t) is assumed to be piecewise linear: macroparticles deposit their charge density at the edge of each time subinterval, i.e., at tn+ℓ/m≡nΔt+ℓδtt_{n+\ell/m}\equiv n\Delta t+\ell\delta t, with ℓ∈[0,m]\ell\in[0,m],

and ρ^\hat{\rho} is then assumed to be linear in each subinterval:

ρ^(t)=ρ^n+(ℓ+1)/m−ρ^n+ℓ/mδt(t−tn+(ℓ+1/2)/m)+ρ^n+(ℓ+1)/m+ρ^n+ℓ/m2,\hat{\rho}(t)=\frac{\hat{\rho}^{n+(\ell+1)/m}-\hat{\rho}^{n+\ell/m}}{\delta t}\left(t-t_{n+(\ell+1/2)/m}\right)+\frac{\hat{\rho}^{n+(\ell+1)/m}+\hat{\rho}^{n+\ell/m}}{2},
t∈[nΔt+ℓδt,nΔt+(ℓ+1)δt].t\in\left[n\Delta t+\ell\delta t,n\Delta t+(\ell+1)\delta t\right].
  • When ρ^(t)\hat{\rho}(t) is assumed to be piecewise quadratic: macroparticles deposit their charge density at the middle and edge of each time subinterval, i.e., at tn+(ℓ+1/2)/mt_{n+(\ell+1/2)/m} with ℓ∈[0,m−1]\ell\in[0,m-1] and tn+ℓ/mt_{n+\ell/m} with ℓ∈[0,m]\ell\in[0,m]. ρ^\hat{\rho} is then assumed to be quadratic in each subinterval:

ρ^(t)=2(ρ^n+(ℓ+1)/m−2ρ^n+(ℓ+1/2)/m+ρ^n+ℓ/m)δt2(t−tn+(ℓ+1/2)/m)2+ρ^n+(ℓ+1)/m−ρ^n+ℓ/mδt(t−tn+(ℓ+1/2)/m)+ρ^n+(ℓ+1/2)/m,\begin{aligned} \hat{\rho}(t)=\frac{2\left(\hat{\rho}^{n+(\ell+1)/m}-2\hat{\rho}^{n+(\ell+1/2)/m}+\hat{\rho}^{n+\ell/m}\right)}{\delta t^{2}}\left(t-t_{n+(\ell+1/2)/m}\right)^{2} \\ &+\frac{\hat{\rho}^{n+(\ell+1)/m}-\hat{\rho}^{n+\ell/m}}{\delta t}\left(t-t_{n+(\ell+1/2)/m}\right)+\hat{\rho}^{n+(\ell+1/2)/m}, \end{aligned}
t∈[nΔt+ℓδt,nΔt+(ℓ+1)δt].t\in\left[n\Delta t+\ell\delta t,n\Delta t+(\ell+1)\delta t\right].

with similar definitions for J^\hat{\mathbf{J}}, when J^(t)\hat{\mathbf{J}}(t) is assumed to be piecewise constant, piecewise linear, or piecewise quadratic respectively.

Overall, the time dependency of J^\hat{\mathbf{J}} and ρ^\hat{\rho} can thus be expressed, for t∈[nΔt+ℓδt,nΔt+(ℓ+1)δt]t\in[n\Delta t+\ell\delta t,n\Delta t+(\ell+1)\delta t], with ℓ∈[0,m−1]\ell\in[0,m-1], as:

J^(t)=2aJτδt2(t−tn+(ℓ+1/2)/m)2+bJτδt(t−tn+(ℓ+1/2)/m)+cJτ,(7)\hat{\mathbf{J}}(t)=\frac{2\mathbf{a}_{J}^{\tau}}{\delta t^{2}}\left(t-t_{n+(\ell+1/2)/m}\right)^{2}+\frac{\mathbf{b}_{J}^{\tau}}{\delta t}\left(t-t_{n+(\ell+1/2)/m}\right)+\mathbf{c}_{J}^{\tau}, \tag*{(7)}
ρ^(t)=2aρτδt2(t−tn+(ℓ+1/2)/m)2+bρτδt(t−tn+(ℓ+1/2)/m)+cρτ,(8)\hat{\rho}(t)=\frac{2a_{\rho}^{\tau}}{\delta t^{2}}\left(t-t_{n+(\ell+1/2)/m}\right)^{2}+\frac{b_{\rho}^{\tau}}{\delta t}\left(t-t_{n+(\ell+1/2)/m}\right)+c_{\rho}^{\tau}, \tag*{(8)}

where the coefficients of the polynomials are given in Table I.

It is important to note that the particles’ momenta are not updated during one time step, i.e., the proposed scheme does not involve subcycling of the macroparticles motion. As in standard PSATD PIC, macroparticles move in straight line from their known position at tn=nΔtt_n=n\Delta t to time tt, using their known momentum at tn+1/2t_{n+1/2}:

x(t)=xn+pn+1/2m1+(pn+1/2/mc)2(t−tn)\mathbf{x}(t)=\mathbf{x}^{n}+\frac{\mathbf{p}^{n+1/2}}{m\sqrt{1+\left(\mathbf{p}^{n+1/2}/mc\right)^{2}}}(t-t_n)

where xn\mathbf{x}^{n} and pn+1/2\mathbf{p}^{n+1/2} follow the standard leap-frog time stepping that is commonly used in PIC simulations. Thus, here, even though the charge and current density may be deposited several times per timestep Δt\Delta t, the macroparticles’ momentum p\mathbf{p} is only updated once per

PCconstant (τ=0\tau= 0)linear (τ=1\tau= 1)quadratic (τ=2\tau= 2)
aJτa_{\mathbf J}^{\tau}0000J^n+(ℓ+1)/m−2J^n+(ℓ+1/2)/m+J^n+ℓ/m\widehat{\mathbf J}^{n+(\ell+1)/m} - 2\widehat{\mathbf J}^{n+(\ell+1/2)/m} + \widehat{\mathbf J}^{n+\ell/m}
bJτb_{\mathbf J}^{\tau}00J^n+(ℓ+1)/m−J^n+ℓ/m\widehat{\mathbf J}^{n+(\ell+1)/m} - \widehat{\mathbf J}^{n+\ell/m}J^n+(ℓ+1)/m−J^n+ℓ/m\widehat{\mathbf J}^{n+(\ell+1)/m} - \widehat{\mathbf J}^{n+\ell/m}
cJτc_{\mathbf J}^{\tau}J^n+(ℓ+1/2)/m\widehat{\mathbf J}^{n+(\ell+1/2)/m}(J^n+(ℓ+1)/m+J^n+ℓ/m)/2(\widehat{\mathbf J}^{n+(\ell+1)/m} + \widehat{\mathbf J}^{n+\ell/m})/2J^n+(ℓ+1/2)/m\widehat{\mathbf J}^{n+(\ell+1/2)/m}
aρ^τa_{\widehat{\rho}}^{\tau}0000ρ^n+(ℓ+1)/m−2ρ^n+(ℓ+1/2)/m+ρ^n+ℓ/m\widehat{\rho}^{n+(\ell+1)/m} - 2\widehat{\rho}^{n+(\ell+1/2)/m} + \widehat{\rho}^{n+\ell/m}
bρ^τb_{\widehat{\rho}}^{\tau}00ρ^n+(ℓ+1)/m−ρ^n+ℓ/m\widehat{\rho}^{n+(\ell+1)/m} - \widehat{\rho}^{n+\ell/m}ρ^n+(ℓ+1)/m−ρ^n+ℓ/m\widehat{\rho}^{n+(\ell+1)/m} - \widehat{\rho}^{n+\ell/m}
cρ^τc_{\widehat{\rho}}^{\tau}ρ^n+(ℓ+1/2)/m\widehat{\rho}^{n+(\ell+1/2)/m}(ρ^n+(ℓ+1)/m+ρ^n+ℓ/m)/2(\widehat{\rho}^{n+(\ell+1)/m} + \widehat{\rho}^{n+\ell/m})/2ρ^n+(ℓ+1/2)/m\widehat{\rho}^{n+(\ell+1/2)/m}

Table I. Polynomial coefficients (PC), based on the time dependency of the current and charge densities J^\widehat{\mathbf J} and ρ^\widehat{\rho} over ℓ\ell-th time subinterval [nΔt+ℓδt,nΔt+(ℓ+1)δt][n\Delta t + \ell\delta t, n\Delta t + (\ell+ 1)\delta t].

timestep, and therefore the fields E\mathbf E and B\mathbf B are gathered onto macroparticles to update p\mathbf p only once per timestep also.

Using the piecewise definition of ρ^\widehat{\rho} and J^\widehat{\mathbf J} given in Eqs. (3), Eqs. (2) can be integrated analytically over one timestep Δt\Delta t, i.e., from t=nΔtt = n\Delta t to t=(n+1)Δtt = (n + 1)\Delta t. In practice, this is done by sequentially integrating these equations over each subinterval ℓ∈[0,m−1]\ell\in[0, m - 1]:

E^n+(ℓ+1)/m=CE^n+ℓ/m+ic2Sckk×B^n+ℓ/m+ic2SckF^n+ℓ/mk+1ε0ck(Y3aJ+Y2bJ−ScJ)+ic2ε0ck2(Y1aρ−Y5bρ−Y4cρ)k,(9)\widehat{\mathbf E}^{n+(\ell+1)/m} = C\widehat{\mathbf E}^{n+\ell/m} + i c^2 \frac{S}{ck}\mathbf k \times\widehat{\mathbf B}^{n+\ell/m} + i c^2 \frac{S}{ck}\widehat{F}^{n+\ell/m}\mathbf k + \frac{1}{\varepsilon_0 ck}(Y_3a_{\mathbf J} + Y_2b_{\mathbf J} - Sc_{\mathbf J}) + \frac{i c^2}{\varepsilon_0 c k^2}(Y_1a_{\rho} - Y_5b_{\rho} - Y_4c_{\rho})\mathbf k, \tag*{(9)}
B^n+(ℓ+1)/m=CB^n+ℓ/m−iSckk×E^n+ℓ/m−iε0c2k2k×(Y1aJ−Y5bJ−Y4cJ),(10)\widehat{\mathbf B}^{n+(\ell+1)/m} = C\widehat{\mathbf B}^{n+\ell/m} - i\frac{S}{ck}\mathbf k \times\widehat{\mathbf E}^{n+\ell/m} - \frac{i}{\varepsilon_0 c^2 k^2}\mathbf k \times(Y_1a_{\mathbf J} - Y_5b_{\mathbf J} - Y_4c_{\mathbf J}), \tag*{(10)}
F^n+(ℓ+1)/m=CF^n+ℓ/m+iSckk⋅E^n+ℓ/m+iε0c2k2k⋅(Y1aJ−Y5bJ−Y4cJ)+1ε0ck(Y3aρ+Y2bρ−Scρ)(11)\widehat{F}^{n+(\ell+1)/m} = C\widehat{F}^{n+\ell/m} + i\frac{S}{ck}\mathbf k \cdot\widehat{\mathbf E}^{n+\ell/m} + \frac{i}{\varepsilon_0 c^2 k^2}\mathbf k \cdot(Y_1a_{\mathbf J} - Y_5b_{\mathbf J} - Y_4c_{\mathbf J}) + \frac{1}{\varepsilon_0 ck}(Y_3a_{\rho} + Y_2b_{\rho} - Sc_{\rho}) \tag*{(11)}

where

C=cos⁡(ckδt),S=sin⁡(ckδt),Y1=(1−C)(8−c2k2δt2)−4Sckδt2c2k2δt2,Y2=2(C−1)+Sckδt2ckδt,Y3=S(8−c2k2δt2)−4ckδt(1+C)2c2k2δt2,Y4=(1−C),Y5=(1+C)ckδt−2S2ckδt.(12)\begin{aligned} C &= \cos(ck\delta t), \quad S = \sin(ck\delta t),\\ Y_1 &= \frac{(1-C)(8-c^2k^2\delta t^2)-4Sc k\delta t}{2c^2k^2\delta t^2},\\ Y_2 &= \frac{2(C-1)+Sc k\delta t}{2ck\delta t},\\ Y_3 &= \frac{S(8-c^2k^2\delta t^2)-4ck\delta t(1+C)}{2c^2k^2\delta t^2},\\ Y_4 &= (1-C), \quad Y_5 = \frac{(1+C)ck\delta t-2S}{2ck\delta t}. \tag*{(12)} \end{aligned}

The steps of the novel PIC-JRhom_{\mathrm{hom}} cycle with sub-timestepping are summarized in the diagram shown in Figure 2.

Diagram of the PIC-JRhom algorithm, showing particle pushing, density deposition, Maxwell-field updates, substeps, and field gathering

Figure 2. Diagram of the PIC-JRhom_{\mathrm{hom}} algorithm.

Extensions

As shown in Refs. [9, 10, 11, 12], the PSATD PIC algorithm can be extended to (a) arbitrary-order spatial stencils, (b) a scheme that alternates between nodal and staggered representations of the field components on the simulation grid, and (c) a scheme that averages the fields to be gathered over one timestep. Such extensions are presented in the next sections for the PSATD PIC-JRρm\mathrm{JR}_{\rho m} algorithm.

Extension to finite-order stencils

When using domain decomposition to run PSATD PIC methods on parallel computers, it is advantageous to alter the wave vector in the Fourier representation of the equations so as to emulate a finite-difference approximation of the spatial derivatives at a finite order pp, since this enhances the locality of the field solvers and thus reduces the required number of guard cells around each subdomain [9, 10]. The modified [kup][k_u^p] at order pp along the direction u=x,y,zu=x,y,z are then given by

[kup]nodal=∑j=1p/2[αjp]nodalsin⁡(kujΔu)jΔu,u=x,y,z,(13)[k_u^p]_{\mathrm{nodal}}=\sum_{j=1}^{p/2}[\alpha_j^p]_{\mathrm{nodal}}\frac{\sin(k_u j\Delta u)}{j\Delta u},\qquad u=x,y,z, \tag*{(13)}
[kup]staggered=∑j=1p/2[αjp]staggeredsin⁡(ku(j−1/2)Δu)(j−1/2)Δu,u=x,y,z,(14)[k_u^p]_{\mathrm{staggered}}=\sum_{j=1}^{p/2}[\alpha_j^p]_{\mathrm{staggered}}\frac{\sin(k_u(j-1/2)\Delta u)}{(j-1/2)\Delta u},\qquad u=x,y,z, \tag*{(14)}

for a nodal and staggered representation, respectively, with the following Fornberg coefficients [22]:

[αjp]nodal=(−1)j+12[(p/2)!]2(p/2−j)!(p/2+j)!,(15)[\alpha_j^p]_{\mathrm{nodal}}=(-1)^{j+1}\frac{2[(p/2)!]^2}{(p/2-j)!(p/2+j)!}, \tag*{(15)}
[αjp]staggered=(−1)j+1[p!2p(p/2)!]24(2j−1)(p/2−j)!(p/2+j−1)!.(16)[\alpha_j^p]_{\mathrm{staggered}}=(-1)^{j+1}\left[\frac{p!}{2^p(p/2)!}\right]^2\frac{4}{(2j-1)(p/2-j)!(p/2+j-1)!}. \tag*{(16)}

These modified wave numbers can be readily used with the PIC-JRρm\mathrm{JR}_{\rho m} algorithm to limit the number of guard cells and enable efficient parallel simulations, just as with other flavors of PSATD PIC algorithms [9, 10].

Extension to alternating nodal-staggered grids

Just like the standard and averaged formulations of PSATD PIC, the PIC-JRhom_{\mathrm{ho}m} algorithm can readily adopt the “hybrid nodal-staggered” scheme presented in [11] where the field alternate between nodal and staggered representations on the simulation grid. More precisely, the Maxwell solve and guard cell exchanges are performed on a staggered “Yee” grid while the charge/current depositions and fields gather are performed with field quantities on a separate nodal grid. This “hybrid” alternating nodal-staggered extension allows to retain the advantages of low numerical dispersion and compact stencils of the integration of Maxwell’s equations on a staggered grid with the stability associated with the interpolation of fields onto the particles from a nodal grid [11] (esp. for NCI-prone boosted-frame simulations). The application of the “hybrid” alternating nodal-staggered scheme to PIC-JRhom_{\mathrm{ho}m} leads to the steps shown in Figure 3.

Diagram of the alternating nodal-staggered PIC-JRho_m algorithm

Figure 3. Diagram of the alternating nodal-staggered PIC-JRhom_m algorithm.

Extension to the time-averaged PSATD PIC algorithm

In Refs. [12], an extension to PSATD PIC, named time-averaged PSATD PIC (also labeled as averaged PIC for convenience), is presented that enables stable boosted-frame simulations even when the time step is larger than the Courant condition along a given axis, e.g., cΔt=Δz>Δxc\Delta t = \Delta z > \Delta x. With the time-averaged algorithm, the field quantities that are gathered onto the particles are given by time averages of the fields on the grid obtained by analytically integrating the E^\widehat{\mathbf{E}} and B^\widehat{\mathbf{B}} fields from t=nΔtt = n\Delta t to t=(n+2)Δtt = (n + 2)\Delta t. The time-averaged PIC-JRhom_{\mathrm{ho}m} algorithm consists of the steps shown in Figure 4, where the analytical average of E^\widehat{\mathbf{E}} and B^\widehat{\mathbf{B}} at time t=(n+1)Δtt = (n + 1)\Delta t are,

Diagram of the time-averaged PIC-JRhom algorithm

Figure 4. Diagram of the time-averaged PIC-JRhom algorithm.

⟨E^n+1⟩=12Δt∑ℓ=02m−1[SCkE^n+ℓ/m+ic2Y4c2k2k×B^n+ℓ/m+iY42ckδtF^n+ℓ/m+1ε0c2k2(Y1aJτ−Y5bJτ−Y4cJτ)−ic2k×(Y6aρτ+Y7bρτ+Y8cρτ)],(17)\left\langle\widehat{\mathbf{E}}^{n+1} \right\rangle= \frac{1}{2\Delta t}\sum_{\ell=0}^{2m-1}\left[ \frac{S}{Ck}\widehat{\mathbf{E}}^{n+\ell/m} +\frac{ic^{2}Y_{4}}{c^{2}k^{2}}\mathbf{k}\times\widehat{\mathbf{B}}^{n+\ell/m} +\frac{iY_{4}}{2ck\delta t}\widehat{\mathbf{F}}^{n+\ell/m} +\frac{1}{\varepsilon_{0}c^{2}k^{2}}\left(Y_{1}a_{J}^{\tau}-Y_{5}b_{J}^{\tau}-Y_{4}c_{J}^{\tau}\right) -ic^{2}\mathbf{k}\times\left(Y_{6}a_{\rho}^{\tau}+Y_{7}b_{\rho}^{\tau}+Y_{8}c_{\rho}^{\tau}\right) \right], \tag*{(17)}
⟨B^n+1⟩=12Δt∑ℓ=02m−1[SCkB^n+ℓ/m−iY4c2k2k×E^n+ℓ/m+ik×(Y6aJτ+Y7bJτ+Y8cJτ)].(18)\left\langle\widehat{\mathbf{B}}^{n+1} \right\rangle= \frac{1}{2\Delta t}\sum_{\ell=0}^{2m-1}\left[ \frac{S}{Ck}\widehat{\mathbf{B}}^{n+\ell/m} -\frac{iY_{4}}{c^{2}k^{2}}\mathbf{k}\times\widehat{\mathbf{E}}^{n+\ell/m} +i\mathbf{k}\times\left(Y_{6}a_{J}^{\tau}+Y_{7}b_{J}^{\tau}+Y_{8}c_{J}^{\tau}\right) \right]. \tag*{(18)}

For a detailed derivation see Appendix § 1.0.1.

Relation to the Galilean PSATD PIC algorithm

This section examines the relationship between the Galilean PIC algorithm, the standard PSATD PIC algorithm and the new PIC-JRhom_m algorithm. To this end, it is instructive to “deconstruct” the Galilean PIC algorithm by separating it in two independent steps: (i) a shift of the quantities to recenter them on a grid moving at vgal\mathbf{v}_{\mathrm{gal}}, (ii) the integration of the PSATD equations assuming that the current source is constant along the flow moving at the Galilean velocity vgal\mathbf{v}_{\mathrm{gal}}.

The standard Galilean PIC scheme [13, 14] can then be written highlighting terms that

arise from step (i) in Eqs. (9a)–(9b) and those from step (ii) in Eqs. (10a)–(10d) and Eq. (25):

B^n+1=Cθ2B^n−iSωk×(θ2E^n)+iX1k×θJ^n+12,(19)\widehat{\mathbf{B}}^{n+1}=C\theta^2\widehat{\mathbf{B}}^n-i\frac{S}{\omega}\mathbf{k}\times\left(\theta^2\widehat{\mathbf{E}}^n\right)+iX_1\mathbf{k}\times\theta\widehat{\mathbf{J}}^{n+\frac{1}{2}}, \tag*{(19)}
E^n+1=Cθ2E^n+ic2Sωk×(θ2B^n)+X4θJ^n+12+i(X3θ2ρ^n−X2ρ^n+1)k,(20)\widehat{\mathbf{E}}^{n+1}=C\theta^2\widehat{\mathbf{E}}^n+ic^2\frac{S}{\omega}\mathbf{k}\times\left(\theta^2\widehat{\mathbf{B}}^n\right)+X_4\theta\widehat{\mathbf{J}}^{n+\frac{1}{2}}+i\left(X_3\theta^2\widehat{\rho}^n-X_2\widehat{\rho}^{n+1}\right)\mathbf{k}, \tag*{(20)}

where the coefficients X1X_1, X2X_2, X3X_3 and X4X_4 are defined as

X1:=1ε0(ω2−Ω2)(θ∗−θC+iΩθSω),(21)X_1:=\frac{1}{\varepsilon_0(\omega^2-\Omega^2)}\left(\theta^*-\theta C+i\Omega\theta\frac{S}{\omega}\right), \tag*{(21)}
X2:=c2θ∗−θ(θ∗χ1ε0ω2−θ1−Cε0ω2),(22)X_2:=\frac{c^2}{\theta^*-\theta}\left(\theta^*\frac{\chi_1}{\varepsilon_0\omega^2}-\theta\frac{1-C}{\varepsilon_0\omega^2}\right), \tag*{(22)}
X3:=c2θ∗−θ(θ∗χ1ε0ω2−θ∗1−Cε0ω2),(23)X_3:=\frac{c^2}{\theta^*-\theta}\left(\theta^*\frac{\chi_1}{\varepsilon_0\omega^2}-\theta^*\frac{1-C}{\varepsilon_0\omega^2}\right), \tag*{(23)}
X4:=iΩX1−θε0Sω,(24)X_4:=i\Omega X_1-\frac{\theta}{\varepsilon_0}\frac{S}{\omega}, \tag*{(24)}

with

χ1:=ω2ω2−Ω2(θ∗−θC+iΩθSω),(25)\chi_1:=\frac{\omega^2}{\omega^2-\Omega^2}\left(\theta^*-\theta C+i\Omega\theta\frac{S}{\omega}\right), \tag*{(25)}

where Ω:=vgal⋅k\Omega:=\mathbf{v}_{\mathrm{gal}}\cdot\mathbf{k}, ω:=ck\omega:=c k, C:=cos⁡(ωΔt)C:=\cos(\omega\Delta t), S:=sin⁡(ωΔt)S:=\sin(\omega\Delta t), θ:=eiΩΔt/2\theta:=e^{i\Omega\Delta t/2} and θ∗:=e−iΩΔt/2\theta^*:=e^{-i\Omega\Delta t/2}.

When setting vgal=0\mathbf{v}_{\mathrm{gal}}=0, the system (9a)–(11) converges to the standard PSATD algorithm, as expected.

Step (i), which corresponds to the multiplication of some of the terms by θ\theta or θ2\theta^2, in red in (9a)–(9b), is the easiest to interpret: noting that a multiplication by θ:=eiΩΔt/2\theta:=e^{i\Omega\Delta t/2} in Fourier space corresponds to shifting the terms spatially by the distance vgalΔt/2\mathbf{v}_{\mathrm{gal}}\Delta t/2 in real space, the terms known at time n+1/2n+1/2 are multiplied by θ\theta, hence shifted by vgalΔt/2\mathbf{v}_{\mathrm{gal}}\Delta t/2 while the terms known at time nn are multiplied by θ2\theta^2, hence shifted by vgalΔt\mathbf{v}_{\mathrm{gal}}\Delta t. These are exactly the shifts that are needed to bring the corresponding quantities to their new grid location after one time step, when assuming a Galilean frame of reference moving at vgal\mathbf{v}_{\mathrm{gal}}.

Understanding the terms associated with step (ii) requires a more detailed comparisons between how the standard and the Galilean PIC equations are obtained. The standard PSATD algorithm is derived assuming that the current density (source term) is constant over one time step on a fixed grid. The Galilean algorithm makes the same assumption but in a Galilean frame, i.e., that the current density (source term) is constant over one time step on a Galilean grid. Following this comparison, it flows logically that step (ii) ought to correspond to an integration of the PSATD equations on a fixed grid assuming that the currents are constant along a segment of length vgalΔtv_{\mathrm{gal}}\Delta t. Indeed, it was verified that integrating the PSATD equations based on these assumptions leads to the system (9a)-(11) with the terms highlighted in red replaced by 1 in (9a)-(9b).

From this, it follows that the new algorithm PIC-JRhom_{\mathrm{hom}} is related to step (ii) of the Galilean PIC algorithm in the following way. While step (ii) of Galilean PIC provides a more accurate analytical integration of the PSATD equations over one time step for a flow that moves uniformly at vgalv_{\mathrm{gal}}, the PIC-JRhom_{\mathrm{hom}}, with its arbitrary time-dependence of JJ and ρ\rho and its subintervals, provides a more accurate analytical integration of the PSATD equations over one time step for an arbitrary local flow of particles. The new PIC-JRhom_{\mathrm{hom}} algorithm can thus be viewed as a possible generalization of step (ii) of the Galilean PIC algorithm. Indeed, the numerical tests discussed below show that, like the Galilean PIC algorithm, PIC-JRhom_{\mathrm{hom}} can lead to simulations that are very stable with regard to the numerical Cherenkov instability, and that it can also remain accurate in cases where the Galilean assumption is becoming less appropriate.

Numerical tests

This section presents various physics applications to test the novel PIC-JRhom_{\mathrm{hom}} algorithm. All simulations and results have been performed and obtained with the open-source electromagnetic PIC code WarpX [23, 24, 25]. The current implementation provides the flexibility to:

  • choose an arbitrary polynomial time dependency of JJ and ρ\rho among the following combinations:

    • JJ and ρ\rho constant in time (CCm);

    • JJ constant in time and ρ\rho linear in time (CLm);

    • JJ and ρ\rho linear in time (LLm);

    • JJ and ρ\rho quadratic in time (QQm);

  • choose the number of subintervals mm within one time step;

  • turn on/off the divergence cleaning term, that is, solve Maxwell’s equations (1) with or without the scalar field FF;

  • turn on/off the time averaging of the EE and BB fields gathered on the macro-particles, as in (8).

To assess the stability of the novel PIC-JRhom method theoretically, the analytical dispersion equation was derived (see appendix). This allows to predict the growth rates of the numerical Cherenkov instability in the case of a uniform drifting plasma. Moreover, a variety of WarpX simulation tests were run to further investigate the method’s stability and accuracy. These tests include: 2D simulations of a uniform plasma drifting with a relativistic velocity v0\mathbf{v}_0 (with/without divergence cleaning, with/without subintervals, and with small/large time steps) and 3D simulations of laser wakefield acceleration (LWFA).

Stability of a uniform plasma drifting at relativistic velocity

This section presents WarpX simulations of a uniform electron-proton plasma with density n0=ϵ0mec2γ0/e2n_0 = \epsilon_0 m_e c^2 \gamma_0/e^2 (where ϵ0\epsilon_0 is the permittivity of free space, cc is the speed of light in free space, and ee and mem_e are respectively the electron charge and mass), drifting along zz with a relativistic velocity v0=(0,0,v0)\mathbf{v}_0 = (0, 0, v_0), with v0=c1−1/γ02v_0 = c\sqrt{1 - 1/\gamma_0^2} and Lorentz factor γ0=130\gamma_0 = 130, through a two-dimensional computational domain with xmin⁡=zmin⁡=−6.45 μmx_{\min} = z_{\min} = -6.45\,\mu\mathrm{m} and xmax⁡=zmax⁡=6.45 μmx_{\max} = z_{\max} = 6.45\,\mu\mathrm{m}, periodic boundary conditions and 600×200600 \times200 grid cells along xx and zz, respectively. The simulations were performed with 4 particles per cell, per species, 1 pass of bilinear filter in the transverse direction xx and 4 passes in the longitudinal direction zz (the direction along which the plasma is drifting). Four cases were considered:

  1. PIC-JRhom with cΔt=Δx=Δzc\Delta t = \Delta x = \Delta z without divergence cleaning (Figure 5);

Total electromagnetic energy versus $\omega_{p,r}t$ for PIC-CCm, PIC-LLm, PIC-QQm, PIC-CLm, and PIC-LQm

Figure 5. WarpX simulations of a uniform plasma with a time step at the Courant condition limit cΔt=Δx=Δzc\Delta t = \Delta x = \Delta z and a stencil at infinite order, without divergence cleaning. The total electromagnetic (EM) energy of a uniform plasma drifting at relativistic velocity v0v_0 along the zz-axis is plotted versus the time of the simulation with (a) the same time-dependencies for J\mathbf{J} and ρ\rho and (b) different time-dependencies for J\mathbf{J} and ρ\rho, for various combinations of time-dependency and number of timestep subintervals. Here, ωp,r=ωp/γ0\omega_{p,r} = \omega_p/\sqrt{\gamma_0} is the relativistic plasma frequency, where time ωp,rt=4000\omega_{p,r}t = 4000 corresponds to roughly 6.7×1046.7 \times10^4 time steps.

  1. PIC-JRhom with cΔt=Δx=Δzc\Delta t = \Delta x = \Delta z with divergence cleaning (Figure 6 and Figure 7);

Image not extracted

Figure 6. WarpX simulations of a uniform plasma with a time step at the Courant limit cΔt=Δx=Δzc\Delta t = \Delta x = \Delta z and a stencil at infinite order, with divergence cleaning. The total electromagnetic (EM) energy of a uniform plasma drifting at relativistic velocity v0v_0 along the zz-axis is plotted versus the time of the simulation with (a) same time-dependencies for J^\hat{J} and ρ^\hat{\rho} and (b) different time-dependencies for J^\hat{J} and ρ^\hat{\rho}, for various combinations of time-dependency and number of timestep subintervals. The energy history from a simulation using the Galilean PIC algorithm [13] is also plotted for comparison in plot (a).

Figure 7. NCI growth rates, with small time step cΔt=Δxc\Delta t = \Delta x. Normalized NCI growth rates Im⁡(ω)/ωp,r\operatorname{Im}(\omega)/\omega_{p,r} in spectral space (kx,kz)(k_x,k_z), calculated from (a) the analytical dispersion equation and (b) WarpX simulations (b) using four different solvers: Galilean PIC with matched velocity (vgal=v0)(v_{\mathrm{gal}} = v_0), standard PIC-CL1, novel PIC-CC1 and PIC-LL1. All numerical and physical parameters are the same as in Figure 6: divergence cleaning is used in all cases except for Galilean PIC. The simulation time step is cΔt=Δx=Δzc\Delta t = \Delta x = \Delta z and the transverse and longitudinal cell sizes are Δx=6.45×10−2kp,r−1\Delta x = 6.45 \times10^{-2} k_{p,r}^{-1}, where

kp,r2=n0e2ϵ0mec2γ0.k_{p,r}^{2} = \frac{n_0 e^2}{\epsilon_0 m_e c^2 \gamma_0}.

for J^\hat{J} and ρ^\hat{\rho} (CL and LQ).

  1. averaged PIC-JRhom with cΔt=6Δx=Δzc\Delta t = 6\Delta x = \Delta z with divergence cleaning (Figure 8);

Total electromagnetic energy of a uniform plasma for PIC-CLm, PIC-CCm, PIC-LLm, and PIC-QQm simulations

Figure 8. WarpX simulation of a uniform plasma with cΔt=Δz=6Δxc\Delta t = \Delta z = 6\Delta x. Total electromagnetic (EM) energy of a uniform plasma drifting at relativistic velocity v0\mathbf{v}_0 along the zz-axis. Simulations were performed with time steps of cΔt=Δz=6Δxc\Delta t = \Delta z = 6\Delta x and divergence cleaning, for (a) PIC-CLmm, (b) PIC-CCmm, (c) PIC-LLmm and (d) PIC-QQmm, with m=1,2,4m = 1, 2, 4. The results from a simulation using the average Galilean PIC solver is also plotted for comparison.

  1. PIC-JRhom with cΔt=Δx=Δzc\Delta t = \Delta x = \Delta z and averaged PIC-JRhom with cΔt=6Δx=Δzc\Delta t = 6\Delta x = \Delta z, with divergence cleaning and finite order stencils (Figure 9 and 10);

Total electromagnetic energy of a uniform plasma for standard and averaged PIC algorithms at spectral orders $p=8$, $16$, and $32$

Figure 9. WarpX simulation of a uniform plasma at finite spectral order pp. Total electromagnetic (EM) energy of a uniform plasma drifting at relativistic velocity v0\mathbf{v}_0 along the zz-axis. Simulations were performed with (a) the standard PICp\mathrm{PIC}^{p}-JRhom\mathrm{JR}_{\mathrm{hom}} algorithm with cΔt=Δz=Δxc\Delta t=\Delta z=\Delta x and (b) the averaged PICp\mathrm{PIC}^{p}-JRhom\mathrm{JR}_{\mathrm{hom}} algorithm with cΔt=Δz=6Δxc\Delta t=\Delta z=6\Delta x, using linear time dependency for J^\hat{\mathbf{J}} and ρ^\hat{\rho} in all cases and varying the spectral order p=8p=8, 1616, 3232.

and are discussed below in detail.

PIC-JRhom with cΔt=Δx=Δzc\Delta t = \Delta x = \Delta z without divergence cleaning

Figure 5 shows the total electromagnetic energy as a function of ωp,rt=ωpt/γ0=2n0/(meϵ0)\omega_{p,r}t = \omega_pt/\sqrt{\gamma_0} = \sqrt{2n_0/(m_e\epsilon_0)} obtained from WarpX simulations using PIC-JRhom with CCm, LLm, QQm, CLm and LQm, for m=1,2,5,10m = 1, 2, 5, 10, without divergence cleaning. In this case, increasing the order of the polynomial dependency (from C, L to Q), or timestep subintervals (m>1m > 1), helps delaying the onset of the instability and lowering the growth rate. When using the same time dependency for J\mathbf{J} and ρ\rho (CC, LL and QQ), for a given number of depositions per step, it is more advantageous to increase the order of the polynomial than to increase the subintervals number mm. Conversely, when using a different time dependency for J\mathbf{J} and ρ\rho (CL, LQ), it is more advantageous to increase the number of subintervals mm than to increase the order of the polynomial.

Matching the time dependency of J^\hat{J} and ρ^\hat{\rho} (as in CC, LL, QQ) is also increasing stability, with PIC-LL5 and PIC-QQ2 being more stable than PIC-LQ10.

PIC-JRhom_m with cΔt=Δx=Δzc\Delta t = \Delta x = \Delta z with divergence cleaning

Figure 6 shows the total electromagnetic energy as a function of ωp,rt\omega_{p,rt} obtained from WarpX simulations using PIC-JRhom_m with CCmm, LLmm, QQmm, CLmm and LQmm for m=1,2,5,10m = 1, 2, 5, 10, with divergence cleaning. The energy history from a simulation using the Galilean PIC algorithm [13] is also plotted for comparison.

In contrast to the previous case, when divergence cleaning is used, having the same time dependency for J^\hat{J} and ρ^\hat{\rho} leads to an extraordinary level of stability that is comparable to the one of the Galilean PSATD method. Conversely, turning on the divergence cleaning degrades significantly the stability when using different time dependencies The remarkable stability reported in Figure 6 when matching the time-dependencies is confirmed with a theoretical NCI analysis. Figure 7 shows the NCI growth rates, Im⁡(ω)/ωp,r\operatorname{Im}(\omega)/\omega_{p,r}, obtained from theoretical calculations and WarpX simulations for the Galilean PIC, the standard PSATD PIC (CL1), PIC-CC1 and PIC-LL1, with an excellent agreement between theory and simulations.

A detailed derivation of the two-dimensional dispersion equation for the PIC ⁣− ⁣JRhoM\mathrm{PIC\!-\!JR}_{\mathrm{hoM}} scheme, for time dependencies of J^\hat{J} and ρ^\hat{\rho} up to quadratic, is presented in Appendix § 3.7, clarifying the origin of the remarkable stability that is observed with PIC-CC1, PIC-LL1 and PIC-QQ1. As explained in the appendix, it can be shown that under some conditions that include having the same time dependency for J^\hat{J} and ρ^\hat{\rho}, key terms cancel out in the analysis matrix, leading to stable real solutions of the determinant.

Averaged PIC-JRho_{\mathrm{ho}}m with cΔt=6Δx=Δzc\Delta t = 6\Delta x = \Delta z with divergence cleaning

In this test, the transverse cell size is intentionally set to a much smaller value than the longitudinal cell size, as typical in plasma accelerator simulations in a Lorentz boosted frame of reference [16, 26] with a high Lorentz factor γ0\gamma_0 [12], while keeping the time step at the CFL limit of the longitudinal cell size: cΔt=Δz=6Δxc\Delta t = \Delta z = 6\Delta x. The results from Figure 8 show that this case is more challenging for all schemes, and even the averaged Galilean PIC scheme is not stable beyond 1000 plasma periods. Increasing the order of the polynomial and the number of subintervals mm both help delaying the onset and lowering the growth rate of the instability, slowly for CLmm but quite effectively for CCmm, LLmm and QQmm, with increasing the number of subintervals mm being the most effective strategy for a given number of depositions per time step.

PICp^{p}-JRho_{\mathrm{ho}}m with cΔt=Δx=Δzc\Delta t = \Delta x = \Delta z and averaged PICp^{p}-JRho_{\mathrm{ho}}m with cΔt=6Δx=Δzc\Delta t = 6\Delta x = \Delta z, with divergence cleaning and finite order stencils

This test shows numerical (Figure 9) and theoretical (Figure 10) evidence that using a stencil at finite-order pp with PICp^{p}-LLmm leads to a degradation of the stability that increases as the order pp decreases. This is because the NCI resonant modes, caused by temporal and spatial aliasing, depends on the stencil order:

Figure 10. NCI growth rates. Normalized NCI growth rates Im⁡(ω)/ωp,r\operatorname{Im}(\omega)/\omega_{p,r} in spectral space (kx,kz)(k_x,k_z), calculated from the analytical dispersion equation of the PICp_p-LL1 algorithm with different stencil order p=8,16,32p=8,16,32. Solid grey lines correspond to (mt,mz)=(0,0)(m_t,m_z)=(0,0) mode, which blue-shifts as the stencil order increases. All numerical and physical parameters are the same as the ones used for the results reported in Figure 9.

[kx,resp]=([kzp]v0c+mz2πΔzv0c−2πmtcΔt)2−[kzp]2(26)[k_{x,\mathrm{res}}^{p}] = \sqrt{\left([k_{z}^{p}]\frac{v_{0}}{c} + m_{z}\frac{2\pi}{\Delta z}\frac{v_{0}}{c} - \frac{2\pi m_{t}}{c\Delta t}\right)^{2} - [k_{z}^{p}]^{2}} \tag*{(26)}

for any mz,mt∈Zm_z,m_t \in\mathbb{Z}, where mzm_z is the spatial alias index and mtm_t is the temporal alias index [27]. As the stencil order gets lower, such resonant modes relocate to lower wavenumbers where the resonance is stronger, as can be seen on Figure 10 that shows the theoretical NCI growth rate at different spectral orders, p=8,16,32p = 8, 16, 32. A nonzero growth rate is observed solely along the NCI resonant mode that is caused by aliasing between the temporal mt=0m_t = 0 and spatial mz=0m_z = 0 modes. The results from Figure 9-Figure 10 indicate that the choice of stencil order will depend on the total duration of the simulations (as measured in plasma periods) for a given application.

Laser-plasma acceleration

This section demonstrates the extension of the stability properties observed in the uniform plasma cases to realistic 3D simulations of laser wakefield acceleration (LWFA) [4]. In these runs, a Gaussian laser pulse with amplitude a0=1.7a_0=1.7, duration τ=73.3 fs\tau=73.3\ \mathrm{fs} and waist w0=50 μmw_0=50\ \mu\mathrm{m} is injected at the entrance of a parabolic plasma channel with a background density n0=1018 cm−3n_0=10^{18}\ \mathrm{cm}^{-3} on axis. The simulations are run in a Lorentz boosted frame of reference [16] with γ0=60\gamma_0=60 using the novel PIC16\mathrm{PIC}^{16}-JRhom\mathrm{JR}_{\mathrm{hom}} scheme (stencil order p=16p=16 in all directions) with a hybrid alternating nodal-staggered grids (using field and current centering of order 16 in all directions) [11]. Similarly to the uniform plasma case, a bilinear filter was applied to the current and charge densities at each time step, with 4 passes in the zz direction and 1 pass in the other directions. The simulations were run on the Oak Ridge Leadership Computing Facility (OLCF) supercomputer Summit using 24 nodes (144 GPUs), with domain decomposition along both xx and zz, using 24 guard cells in each direction. The longitudinal resolution (in the boosted frame) was set to Δz=(1+β0)γ0λlab/24=4.08 μm\Delta z=(1+\beta_0)\gamma_0\lambda_{\mathrm{lab}}/24=4.08\ \mu\mathrm{m}, where β0=1−1/γ02\beta_0=\sqrt{1-1/\gamma_0^2} and λlab=0.8 μm\lambda_{\mathrm{lab}}=0.8\ \mu\mathrm{m} is the driving laser wavelength in the laboratory frame, while the transverse resolution was Δx=0.68 μm\Delta x=0.68\ \mu\mathrm{m}, so that Δz=6Δx\Delta z=6\Delta x. Simulations were also performed with the standard and averaged Galilean PIC16\mathrm{PIC}^{16} algorithm [13, 14] for reference.

Figure 11 displays snapshots of the longitudinal electric field EzE_z from simulations running the Galilean PIC16\mathrm{PIC}^{16} and the PIC16-JRho\mathrm{PIC}^{16}\text{-}\mathrm{JR}_{ho} algorithms at time t=28.05t=28.05 ps (which corresponds to ωp,rt=84.3\omega_{p,r}t=84.3) with different simulation time steps: (a) cΔt=Δx=Δz/6c\Delta t=\Delta x=\Delta z/6, (b) cΔt=3Δx=Δz/2c\Delta t=3\Delta x=\Delta z/2 and (c) cΔt=6Δx=Δzc\Delta t=6\Delta x=\Delta z. Figure 12 shows the corresponding lineouts at x=0x=0 for a selection of runs. Table II compares the performance of the various runs in each case.

Figure 11. WarpX simulation of LWFA with various time steps. Snapshots of the longitudinal electric field Ez(x,z)E_z(x,z) slice at time t=28.05t = 28.05 ps from the 3D simulation of two consecutive laser-driven plasma accelerator stages using the Galilean PIC16^{16} and PIC16^{16}-JRhoM_{\mathrm{hoM}} (with JRhoM_{\mathrm{hoM}} = CC1, CL1, LL1, CC2, LL2 or QQ1) algorithms with time step (a) cΔt=Δx=Δz/6c\Delta t = \Delta x = \Delta z/6, (b) cΔt=3Δx=Δz/2c\Delta t = 3\Delta x = \Delta z/2 and (c) cΔt=6Δx=Δzc\Delta t = 6\Delta x = \Delta z. The laser (not shown) that drives the wake propagates from left to right.

Figure 12. WarpX simulation of LWFA. EzE_z-field lineouts at x=0x = 0 for selected cases of the results reported in Figure 11.

PSATD PIC solvercΔt/Δxc\Delta t/\Delta xAveraged in timeStabilityAverage time per step [s]Total time at t=1.33t = 1.33 ps [s]
Galilean PIC1nostable0.465279.27
PIC-CL11nounstable0.44261.67
PIC-CC11nostable0.39231.15
Galilean PIC3yesstable, but inaccurate0.54109.6
PIC-CC13yesstable0.7144.83
PIC-LL13yesstable0.9182.9
Galilean PIC6yesstable, but inaccurate0.57462.68
PIC-CC16yesunstable0.8388.39
PIC-CC26yesstable1.24130.64
PIC-LL16yesunstable1.05111.545
PIC-LL26yesstable1.61165.4
PIC-QQ16yesstable1.47152.23

Table II. Performance comparison of runtimes for 3D LWFA simulations shown in Figure 11 using different spectral PIC solvers, run on the Summit supercomputer without I/O, using 24 nodes (144 GPUs), with domain decomposition in xx and zz and with 24 guard cells in each direction. Average time per step is from running up to time t=1.33t = 1.33 ps which corresponds to the first 600 time steps with cΔt=Δxc\Delta t = \Delta x, or first 300 time steps with cΔt=3Δxc\Delta t = 3\Delta x, or the first 100 time steps cΔt=6Δxc\Delta t = 6\Delta x.

When using the “small” time step cΔt=Δx=Δz/6c\Delta t=\Delta x=\Delta z/6, the PIC16-CC1\mathrm{PIC}^{16}\text{-}\mathrm{CC1} algorithm is as effective as the standard Galilean PIC16\mathrm{PIC}^{16} algorithm for mitigating the NCI instability (which is emerging at the end of the second stage in the simulations using PIC16-CL1\mathrm{PIC}^{16}\text{-}\mathrm{CL1}), with around 20% speedup. For larger time steps cΔt=3Δx=Δz/2c\Delta t=3\Delta x=\Delta z/2 and cΔt=6Δx=Δzc\Delta t=6\Delta x=\Delta z, although the averaged Galilean PIC16\mathrm{PIC}^{16} method is stable, it does not produce accurate physics results, leading to a very diminished amplitude of the electric field in the second stage. Instead, the novel averaged PIC16-JRhom\mathrm{PIC}^{16}\text{-}\mathrm{JR}_{ho}m method is stable and produces accurate results provided that the numbers of deposition and the number of timestep subintervals are high enough. For cΔt=3Δx=Δz/2c\Delta t=3\Delta x=\Delta z/2, both PIC16-CC1\mathrm{PIC}^{16}\text{-}\mathrm{CC1} and PIC16-LL1\mathrm{PIC}^{16}\text{-}\mathrm{LL1} are stable and accurate, with respective speedups of approximately 1.9x and 1.5x as compared to the Galilean reference case with small time steps. For cΔt=6Δx=Δzc\Delta t=6\Delta x=\Delta z, both PIC16-CC1\mathrm{PIC}^{16}\text{-}\mathrm{CC1} and PIC16-LL1\mathrm{PIC}^{16}\text{-}\mathrm{LL1} are unstable, while PIC16-CC2\mathrm{PIC}^{16}\text{-}\mathrm{CC2}, PIC16-LL2\mathrm{PIC}^{16}\text{-}\mathrm{LL2} and PIC16-QQ1\mathrm{PIC}^{16}\text{-}\mathrm{QQ1} are stable and accurate, with respective speedups of approximately 2.1x, 1.7x and 1.8x as compared to the

Galilean reference case with small time steps.

These results show that the PICp^{p}-JRhom method is effective, efficient and versatile for controlling the numerical Cherenkov instability in plasma accelerator simulations, both in cases for which other methods (e.g., averaged Galilean PIC) apply as well, and in other cases that happen to be more challenging for the other methods.

Conclusion

A novel formulation of pseudo-spectral analytical time-domain Particle-In-Cell algorithm is proposed and analyzed. The formulation includes an additional term of “hyperbolic divergence cleaning” and a relaxation of the standard assumption of constant time dependency of the current density over one time step. Extensions of the algorithm to finite-order stencils, alternating nodal-staggered grids and time-averaging over a time step were also presented.

Tests and analyses revealed that assuming the same time dependency for the evolution of the charge and current densities over one time step leads to excellent stability with regard to the numerical Cherenkov instability. Detailed analysis of the dispersion relation of the new algorithm (see Appendix § 3.7) provides a hint that explains the stability.

The new algorithm is found to be effective, efficient and versatile for controlling the numerical Cherenkov instability in plasma accelerator simulations, both in cases for which other methods (e.g., Galilean PIC) apply and, more importantly, in other cases that happen to be more challenging for the other methods. A possible extension of the algorithm for this particular application could be to incorporate the Galilean PIC algorithm in each subinterval, which should provide enhanced stability while preserving the versatility of the new scheme.

While the application of the new algorithm to the modeling of plasma acceleration as proven successful, the application to other domains must be explored with care. For example, initial testings of the application of the method to the modeling of relativistic plasma shocks [28] has led to the observation of unphysical effects that have been tentatively attributed to unphysical coupling between the unphysical longitudinal electric field waves associated with divergence cleaning (from the term FF in Eq. 1c) and the plasmas. Further studies are needed to fully understand the underlying mechanisms and propose possible remedies.

Connection between the modified system of Maxwell's equations and a potential formulation

It is instructive to derive the modified system of Maxwell's equations (1) in its potential form, starting with

∂E∂t=c2∇×B−Jε0+c2∇F,(27)\frac{\partial\mathbf{E}}{\partial t}=c^{2}\nabla\times\mathbf{B}-\frac{\mathbf{J}}{\varepsilon_{0}}+c^{2}\nabla F, \tag*{(27)}
∂B∂t=−∇×E,(28)\frac{\partial\mathbf{B}}{\partial t}=-\nabla\times\mathbf{E}, \tag*{(28)}
∂F∂t=∇⋅E−ρε0,(29)\frac{\partial F}{\partial t}=\nabla\cdot\mathbf{E}-\frac{\rho}{\varepsilon_{0}}, \tag*{(29)}
∇⋅B=0.(30)\nabla\cdot\mathbf{B}=0. \tag*{(30)}

Eq. (A.1d) implies that B\mathbf{B} can be derived from a potential B=∇×A\mathbf{B}=\nabla\times\mathbf{A} which, when inserted in Eq. (A.1b), gives ∇×(E+∂A∂t)=0\nabla\times\left(\mathbf{E}+\frac{\partial\mathbf{A}}{\partial t}\right)=0. This means that E+∂A∂t\mathbf{E}+\frac{\partial\mathbf{A}}{\partial t} can be written as the gradient of a potential Φ\Phi, giving

E=−∇Φ−∂A∂t.(31)\mathbf{E}=-\nabla\Phi-\frac{\partial\mathbf{A}}{\partial t}. \tag*{(31)}

Plugging (A.2) in (A.1a) and (A.1c) leads to

∇2Φ+∂∂t(∇⋅A)=−ρε0−∂F∂t,(32)\nabla^{2}\Phi+\frac{\partial}{\partial t}(\nabla\cdot\mathbf{A})=-\frac{\rho}{\varepsilon_{0}}-\frac{\partial F}{\partial t}, \tag*{(32)}
∇2A−∂2Ac2∂t2−∇(∇⋅A+1c2∂Φ∂t)=−μ0J+∇F,(33)\nabla^{2}\mathbf{A}-\frac{\partial^{2}\mathbf{A}}{c^{2}\partial t^{2}}-\nabla\left(\nabla\cdot\mathbf{A}+\frac{1}{c^{2}}\frac{\partial\Phi}{\partial t}\right)=-\mu_{0}\mathbf{J}+\nabla F, \tag*{(33)}

which, choosing Φ\Phi and A\mathbf{A} that verify the Lorentz gauge ∇A+1c2∂Φ∂t=0\nabla\mathbf{A}+\frac{1}{c^{2}}\frac{\partial\Phi}{\partial t}=0, gives

∇2Φ−∂2Φc2∂t2=−ρε0−∂F∂t,(34)\nabla^{2}\Phi-\frac{\partial^{2}\Phi}{c^{2}\partial t^{2}}=-\frac{\rho}{\varepsilon_{0}}-\frac{\partial F}{\partial t}, \tag*{(34)}
∇2A−∂2Ac2∂t2=−μ0J+∇F(35)\nabla^2 A-\frac{\partial^2 A}{c^2\partial t^2}=-\mu_0 J+\nabla F \tag*{(35)}
∇2F−∂2Fc2∂t2=μ0(∇⋅J+∂ρ∂t)(36)\nabla^2 F-\frac{\partial^2 F}{c^2\partial t^2}=\mu_0\left(\nabla\cdot J+\frac{\partial\rho}{\partial t}\right) \tag*{(36)}

A gauge transformation

A′=A−∇Λ(37)A'=A-\nabla\Lambda \tag*{(37)}
ϕ′=ϕ+1c2∂Λ∂t(38)\phi'=\phi+\frac{1}{c^2}\frac{\partial\Lambda}{\partial t} \tag*{(38)}

with

F=(−∇2+1c2∂2∂t2)Λ=−(∇⋅A′+1c2∂Φ′∂t)(39)F=\left(-\nabla^2+\frac{1}{c^2}\frac{\partial^2}{\partial t^2}\right)\Lambda=-\left(\nabla\cdot A'+\frac{1}{c^2}\frac{\partial\Phi'}{\partial t}\right) \tag*{(39)}

then leads to

∇2Φ′−∂2Φ′c2∂t2=−ρε0(40)\nabla^2\Phi'-\frac{\partial^2\Phi'}{c^2\partial t^2}=-\frac{\rho}{\varepsilon_0} \tag*{(40)}
∇2A′−∂2A′c2∂t2=−μ0J(41)\nabla^2 A'-\frac{\partial^2 A'}{c^2\partial t^2}=-\mu_0J \tag*{(41)}

This is consistent with the derivation given in [17], where Eqs. (A.1a)–(A.1c) were derived from Maxwell’s equations in the Lorentz gauge form (i.e., the form of (A.7a)–(A.7b)) with the assumption that J=J0+δJJ=J_0+\delta J where J0J_0 is the portion of JJ that verifies the continuity equation ∂ρ∂t+∇J0=0\frac{\partial\rho}{\partial t}+\nabla J_0=0, and defining F=−∇δAF=-\nabla\delta A such that A′=A+δAA'=A+\delta A with ∇A+1c2∂Φ′∂t=0\nabla A+\frac{1}{c^2}\frac{\partial\Phi'}{\partial t}=0.

In addition to showing that the term FF can arise from considerations other than a “divergence cleaning” term, this derivation also highlights how FF relates more directly to the continuity equation via Eq. (A.4c) and gauges via Eq. (A.6).

Derivation of the PIC-JRhom equations

We first rewrite Eqs. (2) in an equivalent second-order differential form,

∂2E^∂t2+c2k2E^=−1ε0(∂J^∂t+ic2kρ^)(42)\frac{\partial^2\widehat{E}}{\partial t^2}+c^2k^2\widehat{E}=-\frac{1}{\varepsilon_0}\left(\frac{\partial\widehat{J}}{\partial t}+ic^2\mathbf{k}\widehat{\rho}\right) \tag*{(42)}
∂2B^∂t2+c2k2B^=1ε0ik×J^(43)\frac{\partial^2\widehat{B}}{\partial t^2}+c^2k^2\widehat{B}=\frac{1}{\varepsilon_0}i\mathbf{k}\times\widehat{J} \tag*{(43)}
∂2F^∂t2+c2k2F^=−1ε0(∂ρ^∂t+k⋅J^)(44)\frac{\partial^2\widehat{F}}{\partial t^2}+c^2k^2\widehat{F}=-\frac{1}{\varepsilon_0}\left(\frac{\partial\widehat{\rho}}{\partial t}+\mathbf{k}\cdot\widehat{J}\right) \tag*{(44)}
f^\widehat{f}∂tf^(tn+ℓ/m)\partial_t\widehat{f}(t_n+\ell/m)C3C_3C4C_4C5C_5
E^\widehat{E}ic2k×Bn+ℓ/m−Jn+ℓ/mε0+ic2F^n+ℓ/mkic^2\mathbf{k}\times\mathbf{B}^{n+\ell/m}-\frac{\mathbf{J}^{n+\ell/m}}{\varepsilon_0}+ic^2\widehat{F}^{n+\ell/m}\mathbf{k}−i2aτc2kε0δt2-i\frac{2a_\tau c^2\mathbf{k}}{\varepsilon_0\delta t^2}−4aτ+ibτc2δtkε0δt2-\frac{4a_\tau+ib_\tau c^2\delta t\mathbf{k}}{\varepsilon_0\delta t^2}4iaτc2k−bτω2δt−ic2cρω2δt2kε0δt2\frac{4ia_\tau c^2\mathbf{k}-b_\tau\omega^2\delta t-ic^2c_\rho\omega^2\delta t^2\mathbf{k}}{\varepsilon_0\delta t^2}
B^\widehat{B}−ik×En+ℓ/m-i\mathbf{k}\times\mathbf{E}^{n+\ell/m}i2aτc2k×ε0δt2i\frac{2a_\tau c^2\mathbf{k}\times}{\varepsilon_0\delta t^2}ibτk×ε0δti\frac{b_\tau\mathbf{k}\times}{\varepsilon_0\delta t}−4iaτk×+icJω2δt2k×ε0δt2ω2\frac{-4ia_\tau\mathbf{k}\times+ic_J\omega^2\delta t^2\mathbf{k}\times}{\varepsilon_0\delta t^2\omega^2}
F^\widehat{F}ik⋅E^n+ℓ/m−ρ^n+ℓ/mε0i\mathbf{k}\cdot\widehat{E}^{n+\ell/m}-\frac{\widehat{\rho}^{n+\ell/m}}{\varepsilon_0}−i2aτkε0δt2-i\frac{2a_\tau\mathbf{k}}{\varepsilon_0\delta t^2}−i4aτk+ibJδtkε0δt2-i\frac{4a_\tau\mathbf{k}+ib_J\delta t\mathbf{k}}{\varepsilon_0\delta t^2}4iaτk−bρδtω2−icJω2δt2kε0δt2ω2\frac{4ia_\tau\mathbf{k}-b_\rho\delta t\omega^2-ic_J\omega^2\delta t^2\mathbf{k}}{\varepsilon_0\delta t^2\omega^2}

Table III. Integration coefficients over ℓ\ell-th time subinterval [nΔt+ℓδt,nΔt+(ℓ+1)δt][n\Delta t+\ell\delta t,n\Delta t+(\ell+1)\delta t].

and then sequentially integrate them analytically over each subinterval [nΔt+ℓδt,nΔt+(ℓ+1)δt][n\Delta t+\ell\delta t,n\Delta t+(\ell+1)\delta t], ℓ∈[0,m−1]\ell\in[0,m-1] with δt=Δt/m\delta t=\Delta t/m, assuming that the current and charge densities are piecewise functions of time, given by Eqs. (3a)–(3b). Each of those equations can be expressed in the following generalized form with a right part as time polynomial up to order two:

(∂2∂t2+c2k2)f^=∑j=02a0jtj(45)\left(\frac{\partial^2}{\partial t^2}+c^2k^2\right)\widehat{f}=\sum_{j=0}^{2}a_{0j}t^j \tag*{(45)}

where {a0j}j=02\{a_{0j}\}_{j=0}^{2} are known coefficients for any given f^=E^,B^,F^\widehat{f}=\widehat{E},\widehat{B},\widehat{F}. The general solution of such a second-order PDE equation with constant coefficients is:

f^(t)=C1cos⁡(ω(t−tn+ℓ/m))+C2sin⁡(ω(t−tn+ℓ/m))+1ω2(C3(t−tn+(ℓ+1/2)/m)2+C4(t−tn+(ℓ+1/2)/m)+C5)(46)\widehat{f}(t)=C_1\cos\left(\omega(t-t_{n+\ell/m})\right)+C_2\sin\left(\omega(t-t_{n+\ell/m})\right)+\frac{1}{\omega^2}\left(C_3(t-t_{n+(\ell+1/2)/m})^2+C_4(t-t_{n+(\ell+1/2)/m})+C_5\right) \tag*{(46)}

where {Ck}k=15\{C_k\}_{k=1}^{5} are integration coefficients to be determined. The coefficients CkC_k with indexes k=3,4,5k=3,4,5 for any given f^=E^,B^,F^\widehat{f}=\widehat{E},\widehat{B},\widehat{F} can be determined by solving a system of linear equations, obtained from substitution of Eq. (B.3) into the corresponding Eq. (B.1) and calculated at time steps tn+ℓ/mt_{n+\ell/m}, tn+(ℓ+1/2)/mt_{n+(\ell+1/2)/m} and tn+(ℓ+1)/mt_{n+(\ell+1)/m}. While the remaining coefficients C1C_1 and C2C_2 can be determined from the initial conditions f^(t)∣tn+ℓ/m\widehat{f}(t)|_{t_{n+\ell/m}} and ∂tf^(t)∣tn+ℓ/m\partial_t\widehat{f}(t)|_{t_{n+\ell/m}}, respectively,

C1=f(tn+ℓ/m)−(C3(δt/2)2+C4(δt/2)+C5)/ω2,C_1=f(t_n+\ell/m)-\left(C_3(\delta t/2)^2+C_4(\delta t/2)+C_5\right)/\omega^2,
C2=∂tf(tn+ℓ/m)−(2C3(δt/2)+C4)/ω2.(47)C_2=\partial_t f(t_n+\ell/m)-\left(2C_3(\delta t/2)+C_4\right)/\omega^2. \tag*{(47)}

The expression of the field components f^(tn+(ℓ+1)/m)\hat{f}(t_{n+(\ell+1)/m}) at the next time subinterval are then given by:

f^(tn+(ℓ+1)/m)=C1cos⁡(ωδt)+C2sin⁡(ωδt)+1ω2(C3(δt/2)2+C4(δt/2)+C5).(48)\hat{f}(t_{n+(\ell+1)/m}) = C_1 \cos(\omega\delta t) + C_2 \sin(\omega\delta t) + \frac{1}{\omega^2}\left(C_3(\delta t/2)^2 + C_4(\delta t/2) + C_5\right). \tag*{(48)}

Derivation of the averaged PIC-JRhom_{\mathrm{hom}} equations

The notation ⟨f^(t)⟩n+1\left\langle\hat{f}(t)\right\rangle^{n+1} is introduced to refer to the average of any given function f^(t)\hat{f}(t) over the time interval [nΔt,(n+1)Δt][n\Delta t,(n+1)\Delta t] as,

⟨f^⟩n+1=12Δt∫tntn+2Δtf^(t′) dt′, where f^=E^,B^.(49)\left\langle\hat{f}\right\rangle^{n+1} = \frac{1}{2\Delta t}\int_{t_n}^{t_n+2\Delta t}\hat{f}(t')\,\mathrm{d}t',\ \text{where }\hat{f}=\hat{E},\hat{B}. \tag*{(49)}

For any given number of subintervals mm, the integral in Eq. (C.1) can be split into a sum over 2m2m integrals over [tn+ℓδt,tn+(ℓ+1)δt][t_n+\ell\delta t,t_n+(\ell+1)\delta t], ℓ=0,…,2m−1\ell=0,\ldots,2m-1 as,

⟨f^⟩n+1=12Δt∑ℓ=02m−1∫tn+ℓδttn+(ℓ+1)Δtf^(t′) dt′, where f^=E^,B^.(50)\left\langle\hat{f}\right\rangle^{n+1} = \frac{1}{2\Delta t}\sum_{\ell=0}^{2m-1}\int_{t_n+\ell\delta t}^{t_n+(\ell+1)\Delta t}\hat{f}(t')\,\mathrm{d}t',\ \text{where }\hat{f}=\hat{E},\hat{B}. \tag*{(50)}

The averaged ⟨E^⟩\left\langle\hat{E}\right\rangle and ⟨B^⟩\left\langle\hat{B}\right\rangle fields are obtained through sequential integration of Eq. (B.3) over each subinterval [tn+ℓδt,tn+(ℓ+1)δt][t_n+\ell\delta t,t_n+(\ell+1)\delta t], ℓ=0,…,2m−1\ell=0,\ldots,2m-1 and then susbstituted into Eq. (C.2),

∫tn+ℓδttn+(ℓ+1)δtE^(t′)dt′=SckE^n+ℓ/m+ic2Y4c2k2k×B^n+ℓ/m+iY42ckδtF^n+ℓ/m+1ε0c2k2(Y1aJ−Y5bJ−Y4cJ)−ic2k×(Y6aρ+Y7bρ+Y8cρ),(51)\begin{aligned} \int_{t_n+\ell\delta t}^{t_n+(\ell+1)\delta t}\hat{E}(t')\mathrm{d}t'={}&\frac{S}{ck}\hat{E}^{n+\ell/m}+\frac{i c^2Y_4}{c^2k^2}\mathbf{k}\times\hat{B}^{n+\ell/m}+\frac{iY_4}{2ck\delta t}\hat{F}^{n+\ell/m}\\ &+\frac{1}{\varepsilon_0c^2k^2}(Y_1\mathbf{a}_J-Y_5\mathbf{b}_J-Y_4\mathbf{c}_J)-i c^2\mathbf{k}\times(Y_6\mathbf{a}_\rho+Y_7\mathbf{b}_\rho+Y_8\mathbf{c}_\rho), \tag*{(51)} \end{aligned}
∫tn+ℓδttn+(ℓ+1)δtB^(t′)dt′=SckB^n+ℓ/m−iY4c2k2k×E^n+ℓ/m+ik×(Y6aJ+Y7bJ+Y8cJ),(52)\int_{t_n+\ell\delta t}^{t_n+(\ell+1)\delta t}\hat{B}(t')\mathrm{d}t'=\frac{S}{ck}\hat{B}^{n+\ell/m}-\frac{iY_4}{c^2k^2}\mathbf{k}\times\hat{E}^{n+\ell/m}+i\mathbf{k}\times(Y_6\mathbf{a}_J+Y_7\mathbf{b}_J+Y_8\mathbf{c}_J), \tag*{(52)}

with

Y6=16ε0c5k5δt2((ckδt)2−3δ(ckδt)2S−12ckδt(1+C)+24S),(53)Y_6=\frac{1}{6\varepsilon_0c^5k^5\delta t^2}\left((ck\delta t)^2-3\delta(ck\delta t)^2S-12ck\delta t(1+C)+24S\right), \tag*{(53)}
Y7=12ε0c4k4δt(ckδtS+2(C−1)),(54)Y_7=\frac{1}{2\varepsilon_0c^4k^4\delta t}\left(ck\delta tS+2(C-1)\right), \tag*{(54)}
Y8=δtε0c2k2(1−Sckδt).(55)Y_8=\frac{\delta t}{\varepsilon_0c^2k^2}\left(1-\frac{S}{ck\delta t}\right). \tag*{(55)}

Dispersion relation for the PIC-JRhom_{\mathrm{hom}} algorithm

The 2D dispersion relation for Eqs. (4) is derived to analyze the algorithm’s stability with respect to the numerical Cherenkov instability (NCI), for a uniform plasma flowing through a periodic grid along the zz-axis with a velocity v0=(0,0,v0)\mathbf{v}_0=(0,0,v_0), where v0=c(1−1/γ02)1/2v_0=c(1-1/\gamma_0^2)^{1/2}.

Following the analysis from [12, 13], we consider the discretized perturbed Vlasov equation, expressed in Fourier space:

δf^n+1/2(km,p)eikm⋅vΔt/2−f^n−1/2(km,p)e−ikm⋅vΔt/2+qΔtS^(km)[E^n(k)+v×B^n(k)]⋅∂f0∂p=0,(56)\delta\hat{f}^{n+1/2}(\mathbf{k}_m,\mathbf{p})e^{i\mathbf{k}_m\cdot\mathbf{v}\Delta t/2} -\hat{f}^{n-1/2}(\mathbf{k}_m,\mathbf{p})e^{-i\mathbf{k}_m\cdot\mathbf{v}\Delta t/2} +q\Delta t\hat{S}(\mathbf{k}_m)\left[\hat{\mathbf{E}}^n(\mathbf{k})+\mathbf{v}\times\hat{\mathbf{B}}^n(\mathbf{k})\right]\cdot\frac{\partial f_0}{\partial\mathbf{p}}=0, \tag*{(56)}

where f0=n0δ(p−mγ0v0)f_0=n_0\delta(\mathbf{p}-m\gamma_0\mathbf{v}_0) is the distribution function of the uniform plasma in a state of equilibrium, and δf\delta f is a perturbation to f0f_0. The discretized formulas for the deposited current and charge in Fourier space at any time ℓΔt\ell\Delta t, ℓ∈[n,n+1]\ell\in[n,n+1] centered around δf^n+1/2\delta\hat{f}^{n+1/2}, are given by

J^ℓ(k)=∑mS(km)∫dpqvδf^n+1/2(km,p)e−ikm⋅v(ℓ−(n+1/2))Δt,(57)\hat{\mathbf{J}}^\ell(\mathbf{k})=\sum_m S(\mathbf{k}_m)\int d\mathbf{p}q\mathbf{v}\delta\hat{f}^{n+1/2}(\mathbf{k}_m,\mathbf{p})e^{-i\mathbf{k}_m\cdot\mathbf{v}(\ell-(n+1/2))\Delta t}, \tag*{(57)}
ρ^ℓ(k)=∑mS(km)∫dpqδf^n+1/2(km,p)e−ikm⋅v(ℓ−(n+1/2))Δt.(58)\hat{\rho}^\ell(\mathbf{k})=\sum_m S(\mathbf{k}_m)\int d\mathbf{p}q\delta\hat{f}^{n+1/2}(\mathbf{k}_m,\mathbf{p})e^{-i\mathbf{k}_m\cdot\mathbf{v}(\ell-(n+1/2))\Delta t}. \tag*{(58)}

Then, assuming the same e−iωte^{-i\omega t} time evolution for E^\hat{\mathbf{E}}, B^\hat{\mathbf{B}}, F^\hat{\mathbf{F}}, J^\hat{\mathbf{J}}, ρ^\hat{\rho} and δf^\delta\hat{f} with the following anzatz,

E^n(k)=E^(k)e−iωnΔt,(59)\hat{\mathbf{E}}^n(\mathbf{k})=\hat{\mathbf{E}}(\mathbf{k})e^{-i\omega n\Delta t}, \tag*{(59)}
δf^n+1/2(km,p)=δf^(km,p)e−iω(n+1/2)Δt,(60)\delta\hat{f}^{n+1/2}(\mathbf{k}_m,\mathbf{p})=\delta\hat{f}(\mathbf{k}_m,\mathbf{p})e^{-i\omega(n+1/2)\Delta t}, \tag*{(60)}
J^n(k)=J^(k)e−iωnΔt,(61)\hat{\mathbf{J}}^n(\mathbf{k})=\hat{\mathbf{J}}(\mathbf{k})e^{-i\omega n\Delta t}, \tag*{(61)}
ρ^n(k)=ρ^(k)e−iωnΔt.(62)\hat{\rho}^n(\mathbf{k})=\hat{\rho}(\mathbf{k})e^{-i\omega n\Delta t}. \tag*{(62)}

equation (D.1) yields

δf^(km,p)=−iqΔt2S^(km)E^(k)+v×B^(k)sin⁡((ω−km⋅v)Δt/2).(63)\delta\hat{f}(\mathbf{k}_m,\mathbf{p})=-i\frac{q\Delta t}{2}\hat{S}(\mathbf{k}_m)\frac{\hat{\mathbf{E}}(\mathbf{k})+\mathbf{v}\times\hat{\mathbf{B}}(\mathbf{k})}{\sin((\omega-\mathbf{k}_m\cdot\mathbf{v})\Delta t/2)}. \tag*{(63)}

Substituting the Vlasov equation (D.1) into (D.2)–(D.3) gives the following expressions for the deposited current J^(k)\hat{\mathbf{J}}(\mathbf{k}) and the charge ρ^(k)\hat{\rho}(\mathbf{k}):

J^=ickε0T(ξ^0+(ξ⋅Q^)vc),(64)\hat{\mathbf{J}}=i\frac{ck\varepsilon_0}{T}\left(\hat{\boldsymbol{\xi}}_0+(\boldsymbol{\xi}\cdot\hat{\mathbf{Q}})\frac{\mathbf{v}}{c}\right), \tag*{(64)}
ρ^=ikε0T(ξ⋅Q^).(65)\hat{\rho}=\frac{ik\varepsilon_0}{T}(\boldsymbol{\xi}\cdot\hat{\mathbf{Q}}). \tag*{(65)}
Q^(k)=E^(k)+v×B^(k)−(v⋅E^(k))v/c2(66)\widehat{\mathbf{Q}}(\mathbf{k})=\widehat{\mathbf{E}}(\mathbf{k})+\mathbf{v}\times\widehat{\mathbf{B}}(\mathbf{k})-(\mathbf{v}\cdot\widehat{\mathbf{E}}(\mathbf{k}))\mathbf{v}/c^2 \tag*{(66)}
ξ0=T^ωp2γ0ck∑m=−∞+∞S^2(km)⋅12Δtsω′(67)\xi_0=\frac{\widehat{T}\omega_p^2}{\gamma_0 c k}\sum_{m=-\infty}^{+\infty}\widehat{S}^2(\mathbf{k}_m)\cdot\frac{1}{\frac{2}{\Delta t}s'_\omega} \tag*{(67)}
ξ=T^ωp2γ0k∑m=−∞+∞S^2(km)⋅kmcω′(2Δtsω′)2(68)\boldsymbol{\xi}=\frac{\widehat{T}\omega_p^2}{\gamma_0 k}\sum_{m=-\infty}^{+\infty}\widehat{S}^2(\mathbf{k}_m)\cdot\frac{\mathbf{k}_m c'_\omega}{\left(\frac{2}{\Delta t}s'_\omega\right)^2} \tag*{(68)}

where T^=∏i[1−sin⁡(kiΔi/2)]\widehat{T}=\prod_i[1-\sin(k_i\Delta_i/2)] is one pass of a binomial smoothing operator, and ωp=(n0q2/meε0)1/2\omega_p=(n_0q^2/m_e\varepsilon_0)^{1/2} is the plasma frequency, and S^(km)\widehat{S}(\mathbf{k}_m) is the particle shape factor. Still following [12, 13], Eqs. (4) are then rewritten in the time-symmetrical form

(E^n+(ℓ+1)/m−E^n+ℓ/m)=iS(1+C)ckc2k×(B^n+(ℓ+1)/m+B^n+ℓ/m)+iS(1+C)ckc2k(F^n+(ℓ+1)/m+F^n+ℓ/m)+1ε0ω(Y9aj−2S(1+C)−1cj)−ic2ε0c2k2Y10kbρ(69)\left(\widehat{\mathbf{E}}^{n+(\ell+1)/m}-\widehat{\mathbf{E}}^{n+\ell/m}\right) =i\frac{S}{(1+C)ck}c^2\mathbf{k}\times\left(\widehat{\mathbf{B}}^{n+(\ell+1)/m}+\widehat{\mathbf{B}}^{n+\ell/m}\right) +i\frac{S}{(1+C)ck}c^2\mathbf{k}\left(\widehat{\mathbf{F}}^{n+(\ell+1)/m}+\widehat{\mathbf{F}}^{n+\ell/m}\right) +\frac{1}{\varepsilon_0\omega}\left(Y_9\mathbf{a}_j-2S(1+C)^{-1}\mathbf{c}_j\right)-\frac{ic^2}{\varepsilon_0c^2k^2}Y_{10}\mathbf{k}b_\rho \tag*{(69)}
(B^n+(ℓ+1)/m−B^n+ℓ/m)=−S(1+C)ckik×(E^n+(ℓ+1)/m+E^n+ℓ/m)+ik×bjε0c2k2Y10(70)\left(\widehat{\mathbf{B}}^{n+(\ell+1)/m}-\widehat{\mathbf{B}}^{n+\ell/m}\right) =-\frac{S}{(1+C)ck}i\mathbf{k}\times\left(\widehat{\mathbf{E}}^{n+(\ell+1)/m}+\widehat{\mathbf{E}}^{n+\ell/m}\right)+\frac{i\mathbf{k}\times\mathbf{b}_j}{\varepsilon_0c^2k^2}Y_{10} \tag*{(70)}
(F^n+(ℓ+1)/m−F^n+ℓ/m)=S(1+C)ckik(E^n+(ℓ+1)/m+E^n+ℓ/m)−1ε0c2k2ikbjY10+1ε0ck(Y9aρ−2S(1+C)−1cρ)(71)\left(\widehat{\mathbf{F}}^{n+(\ell+1)/m}-\widehat{\mathbf{F}}^{n+\ell/m}\right) =\frac{S}{(1+C)ck}i\mathbf{k}\left(\widehat{\mathbf{E}}^{n+(\ell+1)/m}+\widehat{\mathbf{E}}^{n+\ell/m}\right)-\frac{1}{\varepsilon_0c^2k^2}i\mathbf{k}b_jY_{10}+\frac{1}{\varepsilon_0ck}\left(Y_9\mathbf{a}_\rho-2S(1+C)^{-1}\mathbf{c}_\rho\right) \tag*{(71)}

Substitution of equations (D.4) in Eqs.(D.11a)-(D.11c) gives

sωE^=−tckcωk×cB^−cωtckkcF^+i(Y9a~ωτ/2−tckc~ωτ)J^+(Y10b~ωτ/2)kρ^(72)s_\omega\widehat{\mathbf{E}}=-t_{ck}c_\omega\mathbf{k}\times c\widehat{\mathbf{B}}-c_\omega t_{ck}k c\widehat{\mathbf{F}}+i\left(Y_9\widetilde{\mathbf{a}}_\omega^\tau/2-t_{ck}\widetilde{\mathbf{c}}_\omega^\tau\right)\widehat{\mathbf{J}}+\left(Y_{10}\widetilde{\mathbf{b}}_\omega^\tau/2\right)\mathbf{k}\widehat{\rho} \tag*{(72)}
sωB^=tckcωk×E^−(Y10b~ωτ/2)k×J^(73)s_\omega\widehat{\mathbf{B}}=t_{ck}c_\omega\mathbf{k}\times\widehat{\mathbf{E}}-\left(Y_{10}\widetilde{\mathbf{b}}_\omega^\tau/2\right)\mathbf{k}\times\widehat{\mathbf{J}} \tag*{(73)}
sωcF^=−tckcωk⋅E^+k⋅J^(Y10b~ωτ/2)+i(Y9a~ωτ/2−tckc~ω)ρ^(74)s_\omega c\widehat{\mathbf{F}}=-t_{ck}c_\omega\mathbf{k}\cdot\widehat{\mathbf{E}}+\mathbf{k}\cdot\widehat{\mathbf{J}}\left(Y_{10}\widetilde{\mathbf{b}}_\omega^\tau/2\right)+i\left(Y_9\widetilde{\mathbf{a}}_\omega^\tau/2-t_{ck}\widetilde{\mathbf{c}}_\omega\right)\widehat{\rho} \tag*{(74)}

Projecting Eqs. (D.6) and (D.12a) along xx and zz and Eqs. (D.12b) along yy gives the following 2D dispersion equation in matrix form:

MUT=0(75)M U^T=0 \tag*{(75)}
M=[−sω0cωk^ztck−cωk^xtckiTχτJ0−iTk^xψτρ0−sω−cωk^xtck−cωk^ztck0iTχτJ−iTk^zψτρcωk^ztck−cωk^xtck−sω0iTk^zψτJ−iTk^xψτJ0−cωk^xtck−cωk^ztck0−sω−iTk^xψτJ−iTk^zψτJiTχτρiTξ00−iTξ0β00−100iTξxβ0iT(1−β02)(ξ0+ξzβ0)−iTξxβ0200−10iTξxiTξz(1−β02)−iTξxβ0000−1](76)M = \begin{bmatrix} -s_{\omega} & 0 & c_{\omega}\hat{k}_{z}t_{ck} & -c_{\omega}\hat{k}_{x}t_{ck} & iT\chi_{\tau_J} & 0 & -iT\hat{k}_{x}\psi_{\tau_{\rho}} \\ 0 & -s_{\omega} & -c_{\omega}\hat{k}_{x}t_{ck} & -c_{\omega}\hat{k}_{z}t_{ck} & 0 & iT\chi_{\tau_J} & -iT\hat{k}_{z}\psi_{\tau_{\rho}} \\ c_{\omega}\hat{k}_{z}t_{ck} & -c_{\omega}\hat{k}_{x}t_{ck} & -s_{\omega} & 0 & iT\hat{k}_{z}\psi_{\tau_J} & -iT\hat{k}_{x}\psi_{\tau_J} & 0 \\ -c_{\omega}\hat{k}_{x}t_{ck} & -c_{\omega}\hat{k}_{z}t_{ck} & 0 & -s_{\omega} & -iT\hat{k}_{x}\psi_{\tau_J} & -iT\hat{k}_{z}\psi_{\tau_J} & iT\chi_{\tau_{\rho}} \\ \frac{i}{T}\xi_{0} & 0 & -\frac{i}{T}\xi_{0}\beta_{0} & 0 & -1 & 0 & 0 \\ \frac{i}{T}\xi_{x}\beta_{0} & \frac{i}{T}(1-\beta_{0}^{2})(\xi_{0}+\xi_{z}\beta_{0}) & -\frac{i}{T}\xi_{x}\beta_{0}^{2} & 0 & 0 & -1 & 0 \\ \frac{i}{T}\xi_{x} & \frac{i}{T}\xi_{z}(1-\beta_{0}^{2}) & -\frac{i}{T}\xi_{x}\beta_{0} & 0 & 0 & 0 & -1 \end{bmatrix} \tag*{(76)}

where U=(E^x,E^z,cB^y,cF^,J^x/(ckε0),J^z/(ckε0),ρ^/(kε0))U = \left(\hat{E}_{x},\hat{E}_{z},c\hat{B}_{y},c\hat{F},\hat{J}_{x}/(ck\varepsilon_{0}),\hat{J}_{z}/(ck\varepsilon_{0}),\hat{\rho}/(k\varepsilon_{0})\right) and k^=k/k\hat{k} = \mathbf{k}/k is the normalized wave vector. The matrix coefficients in MM that depend on the time dependency of the current and charge densities J^\hat{\mathbf{J}} and ρ^\hat{\rho} are summarized in Table IV. For example, the upper index τJ/ρ\tau_{J/\rho} in the coefficients ψτJ/ρ\psi_{\tau_{J/\rho}} and χτJ/ρ\chi_{\tau_{J/\rho}} indicates the time dependency of J^\hat{\mathbf{J}} and ρ^\hat{\rho} and can be constant (C), linear (L) or quadratic (Q).

CoefficientsTime dependency of J^\hat{\mathbf{J}} or ρ^\hat{\rho}
constant (τ=0\tau= 0)linear (τ=1\tau= 1)quadratic (τ=2\tau= 2)
a~ωτ\tilde{a}_{\omega}^{\tau}00(cω−1)(c_{\omega}-1)
b~ωτ\tilde{b}_{\omega}^{\tau}0sωs_{\omega}sωs_{\omega}
c~ωτ\tilde{c}_{\omega}^{\tau}1cωc_{\omega}1
χτ\chi_{\tau}−tck-t_{ck}−cωtck-c_{\omega}t_{ck}Y9(cω−1)−tckY_{9}(c_{\omega}-1)-t_{ck}
ψτ\psi_{\tau}0−isωY10-is_{\omega}Y_{10}−isωY10-is_{\omega}Y_{10}

Table IV. Matrix coefficients of the dispersion equation (D.13a), based on the time dependency of the current and charge densities J^\hat{\mathbf{J}} and ρ^\hat{\rho} over one time subinterval, δt=Δt/m\delta t = \Delta t/m.

The other coefficients are given by:

cω=cos⁡(ωδt/2),(77)c_{\omega} = \cos(\omega\delta t/2), \tag*{(77)}
sω=sin⁡(ωδt/2),(78)s_{\omega} = \sin(\omega\delta t/2), \tag*{(78)}
tω=sω/cω,(79)t_\omega= s_\omega/c_\omega, \tag*{(79)}
cω′=cos⁡((ω−km⋅v)Δt/2),(80)c'_\omega= \cos\left((\omega-\boldsymbol{k}_m\cdot\boldsymbol{v})\Delta t/2\right), \tag*{(80)}
sω′=sin⁡((ω−km⋅v)Δt/2),(81)s'_\omega= \sin\left((\omega-\boldsymbol{k}_m\cdot\boldsymbol{v})\Delta t/2\right), \tag*{(81)}
km=k+2πm/Δr,m∈Z,(82)\boldsymbol{k}_m=\boldsymbol{k}+2\pi\boldsymbol{m}/\Delta r,\quad\boldsymbol{m}\in\mathbb{Z}, \tag*{(82)}
tck=tan⁡(ckδt/2),(83)t_{ck}=\tan(ck\delta t/2), \tag*{(83)}
Y9=tck(8−c2k2δt2)−4ckδt(1+C)(ckδt)2,(84)Y_9=\frac{t_{ck}(8-c^2k^2\delta t^2)-4ck\delta t}{(1+C)(ck\delta t)^2}, \tag*{(84)}
Y10=(1−2tckckδt),(85)Y_{10}=\left(1-\frac{2t_{ck}}{ck\delta t}\right), \tag*{(85)}
χτ=Y9a~ω′−tckc~ω′,(86)\chi_\tau=Y_9\tilde{a}'_\omega-t_{ck}\tilde{c}'_\omega, \tag*{(86)}
ψτ=Y10b~ω′.(87)\psi_\tau=Y_{10}\tilde{b}'_\omega. \tag*{(87)}

The dispersion relation is given by computing the determinant of (M)(M) using the Sarrus rule.

Interestingly, when the charge and current densities have the same temporal dependency, e.g., with CC, LL or QQ, the determinant simplifies to the straightforward expression

det⁡(M)=α1α2,(88)\det(M)=\alpha_1\alpha_2, \tag*{(88)}

where

α1=T^3[ξ0(β0k^z(χτcωtck+ψτsω)−(χτsω+ψτcωtck))+(cω2tck2−sω2)],(89)\alpha_1=\hat{T}^3\left[\xi_0\left(\beta_0\hat{k}_z(\chi_\tau c_\omega t_{ck}+\psi_\tau s_\omega)-(\chi_\tau s_\omega+\psi_\tau c_\omega t_{ck})\right)+(c_\omega^2t_{ck}^2-s_\omega^2)\right], \tag*{(89)}
α2=(cω2tck2−sω2)+(1−β02)[(ξxk^x+ξzk^z)(χτcωtck+ψτsω)+ψτcωtck(ξ0+ξzβ0)+χτ(ξzcωβ0+ξ0sω)].(90)\alpha_2=(c_\omega^2t_{ck}^2-s_\omega^2)+(1-\beta_0^2)\left[(\xi_x\hat{k}_x+\xi_z\hat{k}_z)(\chi_\tau c_\omega t_{ck}+\psi_\tau s_\omega)+\psi_\tau c_\omega t_{ck}(\xi_0+\xi_z\beta_0)+\chi_\tau(\xi_zc_\omega\beta_0+\xi_0s_\omega)\right]. \tag*{(90)}

Here, such simplification is possible due to the presence of similar terms of opposite sign that cancel each other when the charge and current densities have the same time dependency. For example, terms like (ψτJ)2kxkzcω2−ψτJψτρkxkzcω2=0(\psi_{\tau_J})^2k_xk_zc_\omega^2-\psi_{\tau_J}\psi_{\tau_\rho}k_xk_zc_\omega^2=0, since ψτJ=ψτρ=ψτ\psi_{\tau_J}=\psi_{\tau_\rho}=\psi_\tau (τJ=τρ=τ\tau_J=\tau_\rho=\tau). Moreover, at the asymptotic limit, assuming that (i) δω=ω−kmv0\delta\omega=\omega-\boldsymbol{k}_m v_0 is small and (ii) considering an ultra-relativistic regime, e.g., β0=v0/c=1\beta_0=v_0/c=1, the determinant equation reduces to:

ξ0(k^z(χτckmv0tck+ψτskmv0)−(χτskmv0+ψτckmv0tck))+(ckmv02tck2−skmv02)=0(91)\xi_0\left(\hat{k}_z(\chi_\tau c_{k_mv_0}t_{ck}+\psi_\tau s_{k_mv_0})-(\chi_\tau s_{k_mv_0}+\psi_\tau c_{k_mv_0}t_{ck})\right)+(c_{k_mv_0}^2t_{ck}^2-s_{k_mv_0}^2)=0 \tag*{(91)}

where ckmv0=cos⁡(kmv0δt/2)c_{k_mv_0}=\cos(k_mv_0\delta t/2), skmv0=sin⁡(kmv0δt/2)s_{k_mv_0}=\sin(k_mv_0\delta t/2), and ξ0τ\xi_0^\tau is proportional to 1/δω1/\delta\omega and reads

ξ0τ=T^ωp2S2(km)γ0ck1δω+T^ωp2γ0ck∑j=−∞, m≠j+∞S2(kj)1Δt2skjv0′=αmδω+βm.(92)\xi_0^\tau=\frac{\widehat{T}\omega_p^2 S^2(\mathbf{k}_m)}{\gamma_0 c k}\frac{1}{\delta\omega}+\frac{\widehat{T}\omega_p^2}{\gamma_0 c k}\sum_{\substack{j=-\infty,\ m\ne j}}^{+\infty}S^2(\mathbf{k}_j)\frac{1}{\frac{\Delta t}{2}s'_{k_jv_0}}=\frac{\alpha_m}{\delta\omega}+\beta_m. \tag*{(92)}

Finally, we obtain a first order equation for δω\delta\omega with real coefficients,

δω=−αm(k^z(χτckmv0tck+ψτskmv0)−(χτskmv0+ψτckmv0tck))βm(k^z(χτckmv0tck+ψτskmv0)−(χτskmv0+ψτckmv0tck))+(ckmv02tck2−skmv02).(93)\delta\omega=-\frac{\alpha_m\left(\widehat{k}_z(\chi_\tau c_{kmv_0}t_{ck}+\psi_\tau s_{kmv_0})-(\chi_\tau s_{kmv_0}+\psi_\tau c_{kmv_0}t_{ck})\right)}{\beta_m\left(\widehat{k}_z(\chi_\tau c_{kmv_0}t_{ck}+\psi_\tau s_{kmv_0})-(\chi_\tau s_{kmv_0}+\psi_\tau c_{kmv_0}t_{ck})\right)+(c_{kmv_0}^2t_{ck}^2-s_{kmv_0}^2)}. \tag*{(93)}

It follows that, under the assumptions (i)-(ii), the determinant has only real coefficients, δω\delta\omega is real, and the algorithm is stable.

Acknowledgments

This research used the open-source particle-in-cell code WarpX (https://github.com/ECP-WarpX/WarpX). We acknowledge all WarpX contributors. This research was supported by the Exascale Computing Project (17-SC-20-SC), a collaborative effort of the U.S. Department of Energy Office of Science and the National Nuclear Security Administration. This research was performed in part under the auspices of the U.S. Department of Energy by Lawrence Berkeley National Laboratory under Contract DE-AC02-05CH11231. This research used resources of the Oak Ridge Leadership Computing Facility, which is a DOE Office of Science User Facility supported under Contract DE-AC05-00OR22725. The data that support the findings of this study are available from the corresponding author upon reasonable request.

Olga Shapoval derived and implemented the algorithms in their final forms, performed the numerical analyses and numerical tests. Edoardo Zoni contributed to the derivation, implementation and testing of the algorithms. Rémi Lehe wrote the initial implementation of the algorithm in WarpX and discussed the results. Maxence Thévenet performed numerical tests of an early prototype implemented in the code Warp. Jean-Luc Vay proposed the concept and implemented an early prototype in the code Warp.

References

References

  1. [2]J. M. Dawson, Particle Simulation Of Plasmas, Reviews Of Modern Physics 55, 403 (1983).DOI
  2. [3]C. K. Birdsall and A. B. Langdon, Plasma Physics Via Computer Simulation (Adam-Hilger, 1991) pp. Xxvi+479 Pp.
  3. [4]T. Tajima and J. Dawson, Laser Electron-Accelerator, Physical Review Letters 43, 267 (1979).DOI
  4. [5]P. Chen, J. M. Dawson, R. W. Huff, and T. Katsouleas, Acceleration of Electrons by the Interaction of a Bunched Electron Beam with a Plasma, Physical Review Letters 54, 693 (1985).
  5. [6]I. Haber, R. Lee, H. Klein, and J. Boris, Advances in electromagnetic simulation techniques, in Proc. Sixth Conf. Num. Sim. Plasmas, Berkeley, CA (1973) pp. 46–48.
  6. [7]J.-L. Vay, I. Haber, and B. B. Godfrey, A domain decomposition method for pseudo-spectral electromagnetic simulations of plasmas, Journal of Computational Physics 243, 260 (2013).DOI
  7. [8]K. Yee, Numerical Solution Of Initial Boundary Value Problems Involving Maxwells Equations In Isotropic Media, Ieee Transactions On Antennas And Propagation Ap14, 302 (1966).DOI
  8. [9]H. Vincenti and J.-L. Vay, Detailed analysis of the effects of stencil spatial variations with arbitrary high-order finite-difference Maxwell solver, Computer Physics Communications 200, 147 (2016).DOI
  9. [10]H. Vincenti and J.-L. Vay, Ultrahigh-order Maxwell solver with extreme scalability for electromagnetic PIC simulations of plasmas, Computer Physics Communications 10.1016/j.cpc.2018.03.018 (2018).DOI
  10. [11]E. Zoni, R. Lehe, O. Shapoval, D. Belkin, N. Zaïm, L. Fedeli, H. Vincenti, and J. L. Vay, A hybrid nodal-staggered pseudo-spectral electromagnetic particle-in-cell method with finite-order centering, Computer Physics Communications 279, 108457 (2022).DOI
  11. [12]O. Shapoval, R. Lehe, M. Thévenet, E. Zoni, Y. Zhao, and J.-L. Vay, Overcoming timestep limitations in boosted-frame particle-in-cell simulations of plasma-based acceleration, Physical Review E 104, 055311 (2021).DOI
  12. [13]R. Lehe, M. Kirchen, B. B. Godfrey, A. R. Maier, and J.-L. Vay, Elimination of numerical Cherenkov instability in flowing-plasma particle-in-cell simulations by using Galilean coordinates, Physical Review E 94, 053305 (2016).DOI
  13. [14]M. Kirchen, R. Lehe, B. B. Godfrey, I. Dornmair, S. Jalas, K. Peters, J.-L. Vay, and A. R. Maier, Stable discrete representation of relativistically drifting plasmas, Physics of Plasmas 23, 100704 (2016).DOI
  14. [15]B. Godfrey, Numerical Cherenkov Instabilities In Electromagnetic Particle Codes, Journal of Computational Physics 15, 504 (1974).DOI
  15. [16]J.-L. Vay, Noninvariance Of Space- And Time-Scale Ranges Under A Lorentz Transformation And The Implications For The Study Of Relativistic Interactions, Physical Review Letters 98, 1 (2007).
  16. [17]J.-L. Vay and C. Deutsch, A three-dimensional electromagnetic particle-in-cell code to simulate heavy ion beam propagation in the reaction chamber, Fusion Engineering and Design 32–33, 10.1016/S0920-3796(96)00502-9 (1996).DOI
  17. [18]A. B. Langdon, On Enforcing Gauss Law In Electromagnetic Particle-In-Cell Codes, Computer Physics Communications 70, 447 (1992).DOI
  18. [19]B. Marder, A Method For Incorporating Gauss Law Into Electromagnetic Pic Codes, Journal of Computational Physics 68, 48 (1987).DOI
  19. [20]C.-D. Munz, P. Omnes, R. Schneider, E. Sonnendrücker, and U. Voß, Divergence Correction Techniques for Maxwell Solvers Based on a Hyperbolic Model, Journal of Computational Physics 161, 484 (2000).DOI
  20. [21]J.-L. Vay, I. Haber, and B. Godfrey, A domain decomposition method for pseudo-spectral electromagnetic simulations of plasmas, Journal of Computational Physics 243, 260 (2013).DOI
  21. [22]B. Fornberg, High-order finite differences and the pseudospectral method on staggered grids, SIAM Journal on Numerical Analysis 27, 904 (1990).
  22. [23]J.-L. Vay, A. Almgren, J. Bell, L. Ge, D. Grote, M. Hogan, O. Kononenko, R. Lehe, A. Myers, C. Ng, J. Park, R. Ryne, O. Shapoval, M. Thévenet, and W. Zhang, Warp-X: A new exascale computing platform for beam–plasma simulations, Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment 909, 476 (2018), 3rd European Advanced Accelerator Concepts workshop (EAAC2017).DOI
  23. [24]J.-L. Vay, A. Almgren, L. Amorim, J. Bell, L. Ge, K. Gott, D. Grote, M. Hogan, A. Huebl, R. Jambunathan, R. Lehe, A. Myers, C. Ng, J. Park, M. Rowan, O. Shapoval, M. Thévenet, W. Zhang, Y. Zhao, and E. Zoni, Toward the modeling of chains of plasma accelerator stages with WarpX, Journal of Physics: Conference Series 1596, 012059 (2020).DOI
  24. [25]J. L. Vay, A. Huebl, A. Almgren, L. D. Amorim, J. Bell, L. Fedeli, L. Ge, K. Gott, D. P. Grote, M. Hogan, R. Jambunathan, R. Lehe, A. Myers, C. Ng, M. Rowan, O. Shapoval, M. Thévenet, H. Vincenti, E. Yang, N. Zaïm, W. Zhang, Y. Zhao, and E. Zoni, Modeling of a chain of three plasma accelerator stages with the WarpX electromagnetic PIC code on GPUs, Physics of Plasmas 28, 23105 (2021).
  25. [26]J. L. Vay, C. G. R. Geddes, E. Esarey, C. B. Schroeder, W. P. Leemans, E. Cormier-Michel, and D. P. Grote, Modeling Of 10 Gev-1 Tev Laser-Plasma Accelerators Using Lorentz Boosted Simulations, Physics Of Plasmas 18, 10.1063/1.3663841 (2011).
  26. [27]M. Kirchen, R. Lehe, S. Jalas, O. Shapoval, J. L. Vay, and A. R. Maier, Scalable spectral solver in Galilean coordinates for eliminating the numerical Cherenkov instability in particle-in-cell simulations of streaming plasmas, Physical Review E 102, 13202 (2020).DOI
  27. [28]A. Spitkovsky, Particle Acceleration in Relativistic Collisionless Shocks: Fermi Process at Last?, The Astrophysical Journal 682, L5 (2008).DOI

Paper details

vlnc:W_mcpke7ek
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

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

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents