You are on page 1of 6

IET Science, Measurement & Technology

Research Article

Multiclass multistep discontinuous Galerkin ISSN 1751-8822


Received on 26th February 2018
Revised 10th August 2018
discretisation for multiscale electromagnetic Accepted on 17th September 2018
E-First on 31st October 2018
wave propagation simulations doi: 10.1049/iet-smt.2018.5009
www.ietdl.org

Fidel Souza1 , Renato C. Mesquita1, Elson J. Silva1


1Graduate Program in Electrical Engineering, Universidade Federal de Minas Gerais, Av. Antônio Carlos 6627, 31270-901, Belo Horizonte, MG,

Brazil
E-mail: fidelsouza00@gmail.com

Abstract: The classical numerical methods for the simulation of wave propagation phenomena in multiscale systems can
demand an unnecessary computational cost. This study proposes a multiclass strategy to be applied to the linear multistep time
integration methods in the formalism of the discontinuous Galerkin (DG) space discretisation. The multiclass scheme is applied
specifically to linear multistep strong stability preserving method (SSPMS). The potential of this strategy is demonstrated by
numerical tests applied to two electromagnetic problems. The results show that the proposed schemes promote a significant
improvement compared with the standard SSPMS and the fourth-order Runge–Kutta. Furthermore, in order to obtain a real
speed up, this study presents a class parameter that produces a previous knowledge to determine the number of classes.

1 Introduction explicit methods with local time stepping (LTS). The hybrid
methods try to combine the advantages of explicit and implicit
In science and engineering, multiscale modelling imposes techniques. However, the classical scheme suffers from loss of
computational challenges to solve problems using numerical accuracy and stability [2]. The LTS strategies are most commonly
methods, mainly if the method is based on a mesh or grid used due to its simplicity. Many LTS methods have been proposed
discretisation of the domain. In time-domain methods, mesh size to improve the performance of the DGTD method in terms of
discrepancy can become a prohibitive issue since time step must be computational cost. These methods allow marching with solutions
adjusted to the smallest spatial discretisation step. Hence several of elements of different sizes, in different time steps, maintaining
strategies to select time step have been studied in order to obtain an the stability of the scheme. Important aspects of LTS schemes are
accurate solution with a reduced simulation time. Usually, when a the time integration method and the way in which the numerical
time-dependent system is simulated by numerical methods, it flux is imposed, which strongly affect the accuracy and
evolves in a discretised space. The choice of an appropriate spatial computational efficiency of the time integration.
discretisation is an important step to develop stable, accurate and A lot of the early LTS schemes were based on a leap-frog (LF)
efficient numerical solvers. Also, the discrete structure must scheme to solve problems modelled by Maxwell's curl equations.
preserve all significant information of the continuous model The LF scheme is a central difference method, whose major
equations. advantages are the simplicity and the symplectic feature. However,
The discontinuous Galerkin time-domain (DGTD) method, has the LTS version is not necessarily symplectic. A rigorous study on
several features that make it suitable for modelling time-domain the stability and dispersion of LTS-LF schemes applied to the
multiscale problems. As the finite element method (FEM), the second-order wave equation is presented by Diaz and Grote [3].
discontinuous Galerkin uses unstructured meshes to represent the Applying eigenvalue analysis they show that LTS introduces
computational domain. This provides a better adaptation to numerical dispersion and can produce instabilities if the global
contours at different scales, improving accuracy, when compared time step is not slightly reduced compared with a classical scheme.
with methods that use structured grids. The DG is able to use high- A method called recursive LF is presented in 2008 by Montseny et
order polynomials functions. This is an important feature in wave al. [4]. This method is based on the second-order LF (LF2).
propagation simulations when dealing with numerical dispersion Nowadays, the most used methods with DG are the explicit
and dissipation. In DG, the element solutions are discontinuous and Runge–Kutta (RK) family [5, 6]. These methods present a high
the global continuity must be imposed through numerical flux order of approximation and large time steps. We also find papers
making it easy to consider each element separately. addressing LTS strategies applied to RK family methods. In 2014,
The DGTD method supports time integration based on implicit Angulo et al. [7] proposed an LTS method named causal path LTS.
or explicit techniques. Implicit time integration methods are It was applied in two integration methods, the LF2 and the fourth-
unconditionally stable, i.e. the numerical stability does not depend order low-storage RK. These schemes were constructed to deal
on the time step. However, implicit methods solve a linear with the Maxwell curl equations. The results were compared with
equations system at each time step. It can take a long central Montseny LTS-LF2 and showed a reduction in the numerical
processing unit (CPU) time. Explicit methods are simple and dissipation and dispersion. The Causal-Path Local Time-Stepping
straightforward. In these schemes, matrix inversion is not (CPLTS) take additional calculations of the electromagnetic fields
necessary. Therefore, these methods are much more to satisfy the numerical flux and to maintain the stability. In 2013,
computationally efficient. However, explicit methods are Seny et al. [8] present a RK multirate time-stepping scheme
conditionally stable. Therefore, the time step is subject to a applied to geophysical flows. Following, in 2014, this same author
Courant-Friedrichs-Lewy (CFL) condition to guarantee the presented a parallel implementation of the method [9]. The author
stability. presented two strategies. The first strategy is conservative but only
In multiscale problems, the time integration can become a second-order accurate. The second one is not conservative and
limiting factor, since both implicit and explicit methods present third-order accurate. Recently, in 2016, an LTS scheme named
severe limitations. To overcome these limitations, researchers multirate method was proposed by Kameni et al. [10]. This scheme
basically adopt two approaches: apply hybrid methods [1] or use is based on two-stage second-order RK and was applied to

IET Sci. Meas. Technol., 2019, Vol. 13 Iss. 1, pp. 97-102 97


© The Institution of Engineering and Technology 2018

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DE MINAS GERAIS. Downloaded on April 07,2020 at 16:54:23 UTC from IEEE Xplore. Restrictions apply.
Maxwell equations. RK schemes are based on a multistage ε(r)∂t E(r, t) = ∇ × H(r, t) (1)
strategy. In one stage methods, it is necessary to calculate the slope
of the function once in each time step. To improve the accuracy μ(r)∂t H(r, t) = − ∇ × E(r, t) (2)
order of the approximation in the multistage methods, slopes are
calculated at intermediary instants. The accuracy order of the The hyperbolic equations (1) and (2) relate the electric, E, and
approximation is proportional to the number of the substages. magnetic, H, fields. The electrical permittivity, ε, and the magnetic
Therefore, the high-order RK methods can need a lot of function permeability, μ, can depend on the position vector r. Furthermore,
evaluations. This can bring a larger computational cost compared the time is represented in (1) and (2) by t. The discontinuous
with one stage methods. Galerkin discretisation must be applied in the Maxwell curl
An alternative form to get high-order approximations is to use equations in the conservation law form [18]
one stage multistep methods. In this family of methods, one
function evaluation is enough to get the solution. However, these Q∂tq(r, t) + ∇ F(q) = 0 (3)
schemes use past solutions to increase the accuracy order. Hence,
these methods require more storage than the RK. We also found Q is the material matrix
papers addressing LTS technics based on multistep methods. The
first scheme was presented by Gear and Wells in 1984 and was ε(r) 0
denominated multirate linear multistep methods [11]. In this paper, Q=
they presented error and convergence analysis; furthermore, a 0 μ(r)
rigorous stability analysis of the multirate scheme. After that, other
LTS multistep schemes have been developed. For instance, q represents the state vector
Schomann et al. [12] present a scheme based on Taylor expansion
in 2010. The results compare the scheme proposed to the low- E(r, t)
q(r, t) =
storage four-order four stages RK LSRK(4,4) and demonstrate H(r, t)
computational efficiency gain when the LTS approach is used.
More recently, in 2013, Grote and Mitkova proposed a space and and the flux term is given by F(q)
time discretisation scheme with an arbitrary order, applied to the
damped wave equation [13]. The time discretisation was done Fx
using the Adams–Bashforth method with LTS. The scheme is e^i × E(r, t)
F = Fy , Fi = .
based on an FEM discretisation; therefore, the resulting time −e^i × H(r, t)
marching scheme is fully explicit only if the mass matrix is Fz
diagonal, introducing an additional complication.
A subset of the ordinary differential equations (ODE) solvers is where e^i are the Cartesian unit vectors and i = x, y, z. After the
named strong stability preserving (SSP). There are two types of discontinuous Galerkin space discretisation, we get to the semi-
SSP methods: based on RK methods and based on linear multistep discrete form. Grouping the matrix operators resulting from the DG
methods. The SSP method class was developed for non-linear discretisation and assembling all local systems, we can represent
hyperbolic problems. However, these methods are equally effective the semi-discrete form by
when applied to linear operators [14]. The SSP methods are
suitable for applying to electromagnetic propagation. Owing to the ∂tq = ℋ(q, t) . (4)
arbitrary order of approximation, it is possible to maintain the time
discretisation order consistent with the polynomial discretisation There are several texts describing details of the DG discretisation;
order, preserving the DG high order of convergence [15]. The therefore, we will not do it in this paper. For the one interested in
consistency of the time discretisation and spatial discretisation the DG general features, an efficient nodal version is described in
orders can reduce the dispersion and dissipation errors, providing a [19]. In the space discretisation of the test problems presented in
higher accuracy without a significant increase in the computational Section 4, we consider only admissible meshes.
cost [16].
The local time step approach has already been introduced for
almost all types of ODE integrators. However, our goal in this
3 Linear multistep SSP method
paper is to demonstrate the effectiveness of the high-order linear The method proposed here is based on the linear multistep SSP
multistep SSP methods when applied to electromagnetic wave methods. There is a great difference compared with strong-
propagation simulations. To improve the computational efficiency stability-preserving Runge-Kutta (SSPRK) since these methods do
of the scheme, we create an LTS strategy, which can produce a not have intermediary stages in the calculation of the solution. This
speed up compared with the standard version, in multiscale reduces the number of mathematical operations and brings the
problems. We called this strategy, multiclass method (mC-SSPMS). possibility of computational gain. The high order is achieved by
This strategy is general and can be applied to any linear multistep using past solutions, leading to an increase in storage requirements.
method. Furthermore, we create a class-based parameter to guide Thus, in this paper, we propose a trade-off between storage and
the classification process. This parameter is used to determine how computational cost.
many classes should be used to achieve good performance. To The time integration methods are ODE solvers. A priori, we are
validate the proposal, we applied the multiclass strategy to the interested in first-order methods to solve the problem
third-order SSPMS(6,3) and simulate two electromagnetic wave
propagation problems. Moreover, we compare the CPU time of the q′(t) = ℋ(q, t); q(t0) = q0 . (5)
multiclass method to the standard SSPMS(6,3) and the
conventional fourth-order RK. A simple first-order accuracy method to solve (5) is the Euler
forward method. The Euler forward is explicit and can be derived,
2 Discontinuous Galerkin space discretisation among other forms, using Taylor series expansion. Considering a
time interval Δt we have
The DG space discretisation is based on the local formulation and
discontinuous elements. Generally, the continuity of the global q(t0 + Δt) = q(t0) + Δtq′(t0) + O(Δt2) . (6)
solution and the boundary conditions are imposed by the numerical
flux. The method, originally proposed by Reed and Hill in 1973
By (5) we know that q′(t) = ℋ(q, t). Therefore, applying the
[17], appears in different versions depending on how certain
calculus fundamental theorem
features are supported.
In this paper, we are concerned in the discontinuous Galerkin
discretisation of Maxwell's curl equations

98 IET Sci. Meas. Technol., 2019, Vol. 13 Iss. 1, pp. 97-102


© The Institution of Engineering and Technology 2018

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DE MINAS GERAIS. Downloaded on April 07,2020 at 16:54:23 UTC from IEEE Xplore. Restrictions apply.
t0 + Δt
q(t0 + Δt) − q(t0) = ∫ t0
ℋ(q, t)dt . (7)

If Δt is small enough, we can approximate the integral on the right-


hand side of (7) by Δtℋ(q0, t0). In this way, the Euler forward
method, for the interval t0 → t0 + Δt, is given by
Fig. 1  Time evolution classes
q(t0 + Δt) = q0 + Δtℋ(q0, t0) . (8)
Before proceeding further, there is an important issue that
This can be generalised for any interval. Changing the notation and should be considered: the numerical flux must be satisfied with the
rewriting (8) for a generic interval tn → tn + 1, where tn + 1 = tn + Δt interfaces of each element. Elements of a class which has
neighbours of another class, with a larger time step, will need nodal
qn + 1 = qn + Δtℋ(qn, tn) . (9) values at instants which are not calculated in the time marching. To
satisfy the numerical flux at these interfaces, we perform
Considering higher terms of the Taylor series we can arrive in the interpolations of the solution of the edge nodes of the class element
high-order linear multistep schemes, named Adams–Bashforth with the larger time step, in the required instants. Additional
[20]. In multistep methods, in addition to the current solution, calculations will not be required to satisfy the numerical flux. We
previous solutions are used to increase the accuracy order. In the can use several types of interpolations. The choice should be made
RK methods with precision order greater than four, the number of based on accuracy and computational cost. For example, a linear
stages can grow significantly, increasing the computational cost. interpolation may be suitable when the discrepancy between the
Unlike RK, in multistep methods, the growing of the order of adjacent class time steps is not very large. However, when the time
accuracy does not increase significantly the number of step jump from one class to another is very large, we can use
mathematical operations. Therefore, the order of the time higher-order polynomial interpolations to improve accuracy. It is
integration can be adjusted by the order of the spatial discretisation important to highlight that with the use of the interpolation it is not
without increasing the computational cost. possible to guarantee that the multiclass scheme will achieve the
The Euler forward is a total variation diminishing scheme. SSP property.
Therefore, it has the SSP property [21]. This means that for a given The time marching, in each class, can be performed in parallel.
norm ∥ ⋅ ∥ we have: ∥ qn + Δtℋ(qn, tn) ∥ ≤ ∥ qn ∥ for However, in this work, we developed an efficient sequential
0 ≤ Δt ≤ ΔtFE. ΔtFE is the required time step in the Euler forward implementation which can produce a significant speed up in large
systems with long simulation time and multiscale meshes. In this
method. The SSP property ensures the monotonicity of the time
way, we consider that the solution can be represented as
integration. The reference time step for the DG can be calculated
by [22] m m

2 rD
qn + 1 = ∑ A jqn +1 = ∑ qnj +1 . (13)
j=1 j=1
ΔtFE = r min . (10)
3 G |c|
In (13), A j are sparse arrays with elements equal to unity in some
In (10), c is a module of the wave speed, rD is a vector containing main diagonal entries. These entries refer to nodes contained in the
the ratio of the area and the semi-perimeter of each element and rG class j. The solution of the nodes of the class j is represented as
is the lowest distance of grid points distributed in the qnj + 1. In the class j = 1, when the time goes from tn to tn + 1, the
unidimensional standard element, [−1 1], with the same order of calculation is done in one step. In the j > 1 classes, it is necessary
the spatial approximation. Linear SSP multistep methods can be to perform intermediate calculations, or sub-steps, until you get
built taking a linear combination of Euler forward [23] tn + 1. The number of sub-steps will depend on κ j.
s
We can better understand the process by looking at the
following example. We consider that we have three classes with
qn + 1 = ∑ (αiqn +1−i + βiΔtℋ(qn + 1 − i, tn + 1 − i)) . (11)
κ1 = 0, κ2 = 1 and κ3 = 2. Therefore, Δt2 = Δt1 /2 and
i=1
Δt3 = Δt2 /2 = Δt1 /4. Thus, at tn, all nodal values are defined. In
To maintain the consistency, it is required that ∑i αi = 1. The time tn + 1/4, qn3 + 1/4 is updated. At the time tn + 1/2, qn3 + 1/2 and qn2 + 1/2 are
step is an SSP coefficient function: Δt = C(αi, βi)ΔtFE with [14] updated. In tn + 3/4, qn3 + 3/4 is updated. Finally, at the time tn + 1, qn3 + 1,
qn2 + 1 and qn1 + 1 are updated. This process is illustrated in Fig. 1.
βi
C = min , if αi > 0 for all i (12) The number of sub-steps of a class j is: nsj = 2κ j − 1. It should
αi not be overlooked that multistep methods require prior solutions to
calculate the solution in the next step. However, this is necessary
Therefore, in an SSP high-order method: ∥ qn + 1 ∥ ≤ ∥ qn ∥ for only in the first instant of the time marching. After this, the
0 ≤ Δt ≤ CΔtFE. A detailed description of SSP methods solutions can be stored and used in the calculations of future
characteristics can be found in [14]. solutions. In general, a sub-step in a class j with a number ns of
sub-steps can be calculated as
3.1 Multiclass SSPMS method
s
Our strategy is simple, it does not increase the storage requirements qnj + ℓ /2κ j = ∑ j
αiqn + (ℓ − i)/2κ j
and reduces the number of mathematical operations to be i=1 (14)
performed at each time step. In the proposed scheme, the elements j
will be classified by its local time steps. First, we calculate the +βiΔt jℋ j qn + (ℓ − i)/2κ j, tn + (ℓ − i)/2κ j
local time step Δtk of each local element. Then, we group the
elements into m classes according to the Δtk. The least restrictive with ℓ varying between 1 and ns + 1. It is important to highlight
Δtk is taken as a reference, that is, Δtref = max (Δtk). Then, the that ℋ j operates only on elements contained in the class j.
However, to calculate the numerical flux in ℋ j, it is necessary to
time steps in each class j are given by: Δt j = Δtref /2κ j, κ j ∈ ℕ. In
provide information on neighbour elements including interpolated
this way, the synchronisation is easily done. In the first class,
values. The procedure to one step of time marching is
κ1 = 0, in the other classes κ j will depend on the size difference of
demonstrated in the following pseudocode (see Fig. 2).
the elements of the mesh.

IET Sci. Meas. Technol., 2019, Vol. 13 Iss. 1, pp. 97-102 99


© The Institution of Engineering and Technology 2018

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DE MINAS GERAIS. Downloaded on April 07,2020 at 16:54:23 UTC from IEEE Xplore. Restrictions apply.
the largest and the smallest edges is 16.13 and, in this case, the
time step range is (2/3)rG[4.21, 60.44] × 10−9 s (see Fig. 3).
The aim of this test is to compare the computational efficiency
(CPU time) and accuracy of the mC-SSPMS(6,3) with
m = 2, 3, 4, 5 and the standard SSPMS(6,3) and the fourth-order
RK (RK4). We measure the accuracy at each time step by
computing the L2 error as
2
∫Ω Eh2 − E2 dΩ
er = (16)
∫Ω E2 dΩ

where Eh is the DGTD solution and E is the reference solution. We


considered the upwind numerical flux and simple linear
interpolation. The execution time and the error for each scheme are
presented in Table 1. Here, the error is calculated at t = 1.3 μs,
Fig. 2  Procedure to one step of time marching when we stop the simulation.
The first fact to be highlighted is the efficiency gain of the
standard SSPMS(6,3) regarding RK4. The marching execution
time with the multistep method was around 42% lesser than with
the RK4. It occurs due to the additional stages in the RK4. Even
having a time step of about half but having only one stage against
four for RK4, a gain of about 50% was expected. According to the
results in Table 1, in general, the multiclass division showed
computational efficiency improvement. Only the 2C-SSPMS(6,3)
presented inferior performance integration to the standard version.
The 4C-SSPMS(6,3) presented the best performance, being on
average 34% faster than the standard version and 61% faster than
RK4.
As expected, the performance of the multiclass scheme is
affected by the number of classes as well as a number of elements
inside each one. For example, when increasing the number of
classes, the interpolation process at each time step may degrade the
efficiency by introducing computational time overhead. Therefore,
it is necessary to develop a strategy to find the number of classes to
obtain a good performance in terms of computational effort. One
way to solve this problem is to group the elements such that the
time step discrepancy between classes is reduced. To capture this
information we define a class average parameter as
Fig. 3  Triangular multiscale mesh for solving the pulse propagation
problem 1 Δtk j
K j ∑ min (hk j)
rj = . (17)
4 Numerical tests: electromagnetic wave
propagation K j is the number of elements in class j, Δtk j is the local time step of
Now, we solve an electromagnetic wave propagation problem, in k j and min (hk j) is the length of the minor edge of the element k j.
order to verify the efficiency and the accuracy of the order three Thus, we propose the following pre-processing step for a given
and six steps multiclass SSP method, mC-SSPMS(6,3). The mesh supporting until M classes. First, the r j, j = 1, …, m
coefficients which define the standard SSPMS(6,3) are [14] parameters are computed for each class m = 2, …, M. The
generated points ( j, r j) are marked in a plane and the slope, ϕm, of
α1 = 0.850708871672579 β1 = 1.459638436015276 the best first-order fitting is a non-uniformity parameter that
indicates the variation of Δt among classes. Fig. 4 illustrates the
α5 = 0.030664864534383 β5 = 0.052614491749200 application of this process. We can observe a saturation in ϕm
parameter when increasing the number of classes. The saturation
α6 = 0.118626263793039 β6 = 0.203537849338252 indicates the uniformisation of the Δt of classes tends to a limit.
Therefore, the arbitrary increase in the number of classes does not
The SSPMS(6,3) has an optimal SSP coefficient C = 0.52. ensure improvement in CPU time. Thus, a suitable number of
classes will be near the saturation regime. In Test 1, m = 4 is the
4.1 Test 1: Gaussian pulse propagation first value before the saturation regime.
To ensure the stability and accuracy, the time step has to be
In this first test case, a TMz (H x, Hy, Ez) mode is loaded using the
reduced by a factor of Cr (Table 2). Thus, the effective time step
initial condition of the electric field in the domain. The Gaussian
used in each simulation is given by Δt j = CrΔt. This reduction
pulse waveform is
rises when we increase the polynomial order and the number of
2 + y2 /4σ 2 classes. Probably this is due to the increase in interpolations. It
Ez(x, y, t = 0) = e− x (15) seems that the interpolation process affects the stability of the
scheme. Consequently, we have to reduce the time step. A stability
where the pulse's spatial width is σ = 0.1 m. To truncate the analysis of the proposed method is not performed in this work but
computational domain, the Silver–Muller boundary condition we intend to do it in future works.
n × E − Zn × (H × n) = 0 is applied to the boundary of the circle
of radius r = 5, 0 m. We work with a triangular non-uniform mesh
constituting of 2721 vertices and 5312 triangles. The ratio between

100 IET Sci. Meas. Technol., 2019, Vol. 13 Iss. 1, pp. 97-102
© The Institution of Engineering and Technology 2018

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DE MINAS GERAIS. Downloaded on April 07,2020 at 16:54:23 UTC from IEEE Xplore. Restrictions apply.
Table 1 Test 1: efficiency and accuracy of mC-SSPMS(6,3), SSPMS(6,3) and RK4
N 2C-SSPMS(6,3) 3C-SSPMS(6,3) 4C-SSPMS(6,3) 5C-SSPMS(6,3) SSPMS(6,3) RK4
texec er texec er texec er texec er texec er texec er
1 3.643 0.2441 2.715 0.2405 2.313 0.2394 2.478 0.2445 3.970 0.2470 6.803 0.2442
2 11.325 0.0785 7.365 0.0794 5.810 0.0817 6.928 0.0772 10.079 0.0791 16.994 0.0789
3 29.803 0.0433 18.490 0.0437 14.973 0.0443 16.536 0.0422 21.384 0.0438 36.706 0.0437
4 53.082 0.0257 39.366 0.0258 30.658 0.0258 35.490 0.0246 42.696 0.0260 73.069 0.0259
5 98.662 0.0163 68.136 0.0164 55.198 0.0164 60.182 0.0156 75.315 0.0166 129.889 0.0166

Table 2 Cr factor for the Test 1 numerical experiments


N 2C-SSPMS 3C-SSPMS 4C-SSPMS 5C-SSPMS
1 0.88 0.81 0.81 0.73
2 0.71 0.65 0.65 0.52
3 0.63 0.60 0.60 0.50
4 0.63 0.58 0.56 0.46
5 0.61 0.58 0.54 0.44

where d = 2(a − ra) is the width of the waveguide, σ = 5 × 10−14 s


defines the bandwidth of the pulse and f m = 2 × 1014 Hz is the
centre frequency. Furthermore, in the simulations we consider
Fig. 4  Non-uniformity parameters among classes t0 = 2.5 × 10−13 s.
The Silver–Muller first-order absorbing condition is applied on
the sides of the computational domain. We work with a non-
uniform mesh consisting of 17,935 vertices and 35,524 triangles.
The ratio between the largest and smallest edges is 13.5810 and, in
this case, the time step range is (2/3)rG[1.40, 20.15] × 10−17 s. The
simulations have been done using the mC-SSPMS(6,3) with three
classes m = 2, 3, 4 for N = 1, 2, 3, 4, 5. The solution for the
electrical field component, Ez, is presented in Fig. 5b. Furthermore,
Fig. 5c is the quiver plot of the Poynting vector that indicates the
propagation direction. The CPU times are shown in Table 3, the
performances of the standard SSPMS(6,3) and the RK4 are also
reported.
Again, we can note a speed up in the computational time. It
occurs when m = 2, 3. The percentage gain, Gp, is shown in
Table 3. We observe the main efficiency improvement occurred for
m = 3. The 3C-SSPMS was on average 33% faster than the
standard version and 63% compared with the RK4 method. While
N increases, the time step of the mC-SSPMS method tends to
become more restrictive. Table 3 presents the experimental values
for Cr. It is important to emphasise that the excessive increase in
the number of classes can degrade the stability of the scheme. This
behaviour was observed in the numerical experiments and is
illustrated by Cr values.
Observing the mesh presented in Fig. 5a, we note that the
Fig. 5  Test 2: an L-shaped bent photonic waveguide elements have approximately three distinct sizes. Therefore,
(a) Multiscale mesh to solve a photonic crystal waveguide, (b) Ez 3C-SSPMS(6,3) intuitively the 3C-SSPMS seems to be the most appropriate for this
solution for N = 5, (c) Poynting vector distribution problem. Fig. 4 presents the ϕm parameter for Test 2. Once again
we realise that ϕm tends to saturation. The value m = 3 is the first
4.2 Test 2: an L-shaped bent photonic waveguide before the saturation regime. It indicates that when m > 3 there
may be no computational gain. This really happens, as we can see
The second test case that we consider is a two-dimensional (2D)
in Table 3. The ϕm parameter does not determine an optimal
photonic crystal L-shaped bent waveguide. Fig. 5a shows the
triangular mesh representing the computational domain. The number of classes. However, ϕm produces a previous knowledge
photonic crystal base used here is composed of circular dielectric which helps to define the number of classes to achieve a speed up.
pillars, ε1 = 11.4, in air background, ε2 = 1, on a square array with
lattice constant a = 0.57 μm. The radii of pillars are 5 Conclusions
ra = 0.114 μm. The photonic crystal has a band gap for transverse
In this paper, we presented a multiclass linear multistep method
magnetic (TM) polarised field whose frequency range is extended suitable for dealing with multiscale domain discretisations. In our
from 0.35( = ωa/2πc) to 0.42. The optical waveguide is created by strategy, the elements in the computational domain were classified
removing one row of pillars where the light is confined. The according to the time step requirements and the time integration
incident pulse, with frequency band inside the band gap, is taken as was done locally. To validate the approach, we solved 2D
electromagnetic wave problems using the discontinuous Galerkin
πy 2
Ez(x, y, t) = E0 cos cos(2π f mt)e− t − t0 /2σ (18) FEM and the mC-SSPMS(6,3). In addition, the proposed method
d was compared with its standard version and with the fourth-order
RK method.

IET Sci. Meas. Technol., 2019, Vol. 13 Iss. 1, pp. 97-102 101
© The Institution of Engineering and Technology 2018

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DE MINAS GERAIS. Downloaded on April 07,2020 at 16:54:23 UTC from IEEE Xplore. Restrictions apply.
Table 3 Test 2: efficiency of mC-SSPMS(6,3), SSPMS(6,3) and RK4
N 2C-SSPMS(6,3) 3C-SSPMS(6,3) 4C-SSPMS(6,3) SSPMS(6,3) RK4
texec Gp Cr texec Gp Cr texec Gp Cr texec texec
1 445 35 1.00 378 45 1.00 462 33 0.83 690 1410
2 1773 16 0.85 1347 37 0.85 1836 14 0.63 2125 3942
3 4695 10 0.79 3445 35 0.79 4637 12 0.50 5251 8815
4 9693 7 0.75 7821 25 0.75 10,431 −0.4 0.38 10,389 18,300
5 17,512 6 0.73 14,333 23 0.73 20,387 −10 0.36 18,570 33,264

In Test 1, we observed that the SSPMS(6,3) was more [4] Montseny, E., Pernet, S., Ferrières, X., et al.: ‘Dissipative terms and local
time-stepping improvements in a spatial high order discontinuous Galerkin
computationally efficient than the fourth-order RK, it took 42% scheme for the time-domain Maxwell's equations’, J. Comput. Phys., 2008,
less time. The results also proved the multiclass strategy is able to 227, (14), pp. 6795–6820
produce a speed up in the time integration. Moreover, no [5] Runge, C.: ‘Ueber die numerische auflösung von differentialgleichungen’,
significant additional error was introduced by the proposed Math. Ann., 1895, 46, (2), pp. 167–178
[6] Kutta, W.: ‘Beitrag zur näherungweisen integration totaler
strategy. For the solved problem, the 4C-SSPMS(6,3) was most differentialgleichungen’, Z. Math. Phys., 1901, 46, pp. 435–453
efficient, the four classes yielded a maximum computational [7] Angulo, L., Alvarez, J., Teixeira, F.L., et al.: ‘Causal-path local time-stepping
improvement of 42% compared with standard SSPMS(6,3) and in the discontinuous Galerkin method for Maxwell's equations’, J. Comput.
66% compared with the RK4. Phys., 2014, 256, pp. 678–695
[8] Seny, B., Lambrechts, J., Comblen, R., et al.: ‘Multirate time stepping for
In Test 2, the mC-SSPMS(6,3) also presented a speed up accelerating explicit discontinuous Galerkin computations with application to
compared with the standard version and the fourth-order RK. We geophysical flows’, Int. J. Numer. Methods Fluids, 2013, 71, (1), pp. 41–64
could observe a maximum speed up of 45% over the SSPMS(6,3) [9] Seny, B., Lambrechts, J., Toulorge, T., et al.: ‘An efficient parallel
and 73% over the RK4. The 3C-SSPMS showed the best implementation of explicit multirate Runge–Kutta schemes for discontinuous
Galerkin computations’, J. Comput. Phys., 2014, 256, pp. 135–160
computational time gain. Considering the fact of a simulation with [10] Kameni, A., Seny, B., Pichon, L.: ‘Multirate technique for explicit
a time window in the μs order can take several hours, the speed up discontinuous Galerkin computations of time-domain Maxwell equations on
presented by the proposed method was really significant. complex geometries’, IEEE Trans. Magn., 2016, 52, (3), pp. 1–4
A class parameter, based on the ratio between the time steps and [11] Gear, C.W., Wells, D.: ‘Multirate linear multistep methods’, BIT Numer.
Math., 1984, 24, (4), pp. 484–502
the smallest edge of each element, helped in defining the number [12] Schomann, S., Godel, N., Warburton, T., et al.: ‘Local time-stepping
of classes for good performance. We found that this parameter techniques using Taylor expansion for modeling electromagnetic wave
must tend to saturation for a given set of classes. Moreover, we propagation with discontinuous Galerkin-FEM’, IEEE Trans. Magn., 2010,
observed that the performance degraded with a large number of 46, (8), pp. 3504–3507
[13] Grote, M.J., Mitkova, T.: ‘High-order explicit local time-stepping methods for
classes. damped wave equations’, J. Comput. Appl. Math., 2013, 239, pp. 270–289
The numerical experiments showed a negative fact. The CPU [14] Gottlieb, S., Ketcheson, D.I., Shu, C.-W.: ‘Strong stability preserving Runge–
time speed up decreased when the polynomial order increased. It Kutta and multistep time discretizations’ (World Scientific Publishing,
occurred because the effective time step became more restrictive Singapore, 2011)
[15] Chen, M.H., Cockburn, B., Reitich, F.: ‘High-order RKDG methods for
when the polynomial order increased; in some cases, this produced computational electromagnetics’, J. Sci. Comput., 2005, 22, (1–3), pp. 205–
a negative gain. In fact, this was a drawback since the high-order 226
polynomial base is suitable for control of the numerical dissipation [16] Sármány, D., Botchev, M.A., van der Vegt, J.J.W.: ‘Dispersion and dissipation
and dispersion. error in high-order Runge–Kutta discontinuous Galerkin discretisations of the
Maxwell equations’, J. Sci. Comput., 2007, 33, (1), pp. 47–74
[17] Los Alamos Scientific Lab.: ‘Triangular mesh methods for the neutron
6 Acknowledgment transport equation’. Proc. American Nuclear Society, Ann Arbor, Michigan,
USA, 1973, pp. 1–23
The authors gratefully acknowledge FAPEMIG (Research Support [18] Busch, K., König, M., Niegemann, J.: ‘Discontinuous Galerkin methods in
Foundation of Minas Gerais) and CNPq (National Council for nanophotonics’, Laser Photonics Rev., 2011, 5, (6), pp. 773–809
[19] Hesthaven, J.S., Warburton, T.: ‘Nodal discontinuous Galerkin methods:
Scientific and Technological Development) for financial support. algorithms, analysis, and applications’ (Springer, New York, 2007)
[20] Booth, A.D.: ‘Bashforth–Adams method for the numerical solution of
differential equations’, Nature, 1954, 173, (4405), pp. 635–636
7 References [21] Bresten, C., Gottlieb, S., Grant, Z., et al.: ‘Explicit strong stability preserving
[1] Chen, J., Tobon, L.E., Chai, M., et al.: ‘Efficient implicit–explicit time multistep Runge–Kutta methods’, Math. Comput., 2017, 86, (304), pp. 747–
stepping scheme with domain decomposition for multiscale modeling of 769
layered structures’, IEEE Trans. Compon. Packag. Manuf. Technol., 2011, 1, [22] Hesthaven, J.S., Warburton, T.: ‘Nodal high-order methods on unstructured
(9), pp. 1438–1446 grids: I. Time-domain solution of Maxwell's equations’, J. Comput. Phys.,
[2] Zhang, H., Sandu, A., Blaise, S.: ‘High order implicit–explicit general linear 2007, 181, (1), pp. 186–221
methods with optimized stability regions’, SIAM J. Sci. Comput., 2016, 38, [23] Ruuth, S.J., Hundsdorfer, W.: ‘High-order linear multistep methods with
(3), pp. A1430–A1453 general monotonicity and boundedness properties’, J. Comput. Phys., 2005,
[3] Diaz, J., Grote, M.J.: ‘Energy conserving explicit local time stepping for 209, (1), pp. 226–248
second-order wave equations’, SIAM J. Sci. Comput., 2009, 31, (3), pp. 1985–
2014

102 IET Sci. Meas. Technol., 2019, Vol. 13 Iss. 1, pp. 97-102
© The Institution of Engineering and Technology 2018

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DE MINAS GERAIS. Downloaded on April 07,2020 at 16:54:23 UTC from IEEE Xplore. Restrictions apply.

You might also like