Bull. Comput. Appl. Math., Vol.1, No.
1, 2013
ISSN 2244-8659
www.
ompama.
o.usb.ve
Solving the KPI wave equation with a
moving adaptive FEM grid
∗
Granville Sewell
CompAMa Vol.1, No.1, pp.55-71, 2013 - A
epted April 28, 2013
Abstra
t
The Kadomtsev-Petviashvili I (KPI) equation is the di
ult nonlinear wave equa-
tion Uxt + 6Ux2 + 6U Uxx + Uxxxx = 3Uyy . We solve this equation using PDE2D
(www.pde2d.
om) with initial
onditions
onsisting of two lump solitons, whi
h
ollide and reseparate. Sin
e the solution has steep, moving, peaks, an adaptive
nite element grid is used with a grading whi
h moves with the peaks.
1 Introdu
tion
The Kadomtsev-Petviashvili I (KPI) wave equation:
Uxt + 6Ux2 + 6UUxx + Uxxxx = 3Uyy
is used to model waves in thin lms with high surfa
e tension. It has been
extensively studied in the mathemati
al
ommunity sin
e the 1970 paper by
Boris Kadomtsev and Vladimir Petviashvili [1℄. [2℄ and [3℄ report that only
two kinds of numeri
al methods have been used to solve the KPI equation: -
nite dieren
e methods (whi
h these two papers apply) and a pseudo-spe
tral
method developed by [4℄.
All previous su
essful attempts to solve this di
ult wave equation re-
quired development of numeri
al methods espe
ially tailored for the equation,
here we attempt to solve it using a robust, general-purpose nite element pro-
gram developed by the author.
∗
Mathemati
s Dept., University of Texas El Paso, USA, e-mail: sewelutep.edu
55
Solving the KPI wave equation with a moving adaptive FEM grid 56
2 The Finite Element Method Used
PDE2D ([5℄,[6℄,[7℄) is a general-purpose partial dierential equation solver
whi
h solves very general systems of nonlinear, steady-state, time-dependent
and eigenvalue PDEs in 1D intervals, general 2D regions (with
urved bound-
aries), and a wide range of simple 3D regions, with general boundary
ondi-
tions. It uses a
ollo
ation nite element method, with
ubi
elements, for
3D problems, and either a
ollo
ation or Galerkin nite element method
an
be used for 1D and 2D problems. If the Galerkin algorithm is used for 2D
problems, as in this paper, triangular elements of up to 4th degree
an be
used, on a triangulation whi
h is automati
ally re ned and graded, either
adaptively or a
ording to user spe
i
ations.
To use PDE2D, we have to redu
e this fourth order equation to a system
of three rst or se
ond order equations, by introdu
ing the variables V ≡
Ux , W ≡ Uxx :
0 = Ux − V
0 = Uxx − W
Vt = −Wxx + 3Uyy − 6V 2 − 6UW
[3℄ give a two-lump soliton analyti
al solution of the KPI equation, ex-
pressed as Q(x, y, t) = 2[ΦΦxx − Φ2x ]/Φ2 , where Φ(x, y, t) is de ned as the
determinant of a
ertain 4 by 4 matrix. We will use this analyti
al solution
for de ning initial
onditions, and for
omputing errors.
Initial
onditions for the problems solved in this paper are
U(x, y, 0) = Q(x, y, 0)
V (x, y, 0) = Qx (x, y, 0)
W (x, y, 0) = Qxx (x, y, 0)
Two of the problems solved in [3℄ will be solved here:
1. An oblique
ollision" problem, where two solitons of equal size
ollide
o
at a 90 angle and pass through ea
h other.
59 Sewell
easy way to estimate the numeri
al error. At t = 5, the error in I1 was
97%, and the error in I2 was 9.4%. The I1 integral is mu
h more sensitive
than the I2 integral to the smaller values near the boundary, far from the
peaks, so the problem was resolved with U, V, W set to the true solution on
the entire boundary, and the error in I1 de
reased to 3.3%, while the error
in I2 in
reased slightly, to 10.0%. Sin
e U is not always positive, it may
be more reasonable to divide by the integral of |U|, rather than the integral
of U, in
al
ulating the I1 relative error; when this is done, we get a more
respe
table-looking gure of 0.7% for the I1 error. Lu, Tian and Grimshaw
report errors in I1 and I2 of order 0.1% and 1%, respe
tively, with a time
step of ∆t = 0.0001. Although our time step is 125 times larger, we have to
solve a large linear system every time step, while they used an expli
it time
integration. They used a 600 by 600 uniform nite dieren
e grid in spa
e,
whi
h means their problem has 360,000 unknowns.
For the dire
t
ollision
ase, results are shown in Figures 5-8. Again the
moving grid follows the peaks ni
ely, and the PDE2D solution agrees well
with the analyti
solution until about t = 5.0, when the taller, faster, peak
at
hes the smaller one (Figures 6a,9a). After that, the peaks
omputed
by PDE2D separate more slowly than they should: the PDE2D solution
at t = 10 looks mu
h like the true solution at t = 8 (Figures 8a,9b)! [3℄
report similar results on this problem, but they attribute the slow evolution
in time of their nite dieren
e solution to the fa
t that they are using the
approximate initial
ondition 1 rather than Q(x, y, 0) at t = 0. But we are
using Q(x, y, 0) as our initial
ondition, so our slower evolution
annot be
explained similarly. In fa
t, when we used the approximate initial
ondition
1, our solution developed even more slowly.
For the dire
t
ollision problem, a time step of ∆t = 0.025 was used, and
again the grid was updated adaptively every 10 steps. The I2 integral diers
from the true value at t = 10 by about 5.9%. Lu, Tian and Grimshaw report
an error in I2 of only 0.45% at t = 10 for this problem, using a time step of
∆t = 0.0001.
Finally, we re-solved both problems with the same number of elements and
same time step sizes, but this time using a
onstant, uniform triangulation.
The resulting solutions, shown in Figures 10a-b, are very bad, and
learly
illustrate the importan
e of the moving, adaptive grid. The error in the
2
integral of U for the oblique
ollision problem at t = 2.5 is now 500%,
and 4000% at t=5 for the dire
t
ollision problem! Noti
e that the dire
t
ollision solution is not only quite noisy, but the peaks are very far from
Solving the KPI wave equation with a moving adaptive FEM grid 70
where they should be (
ompare Figures 9a and 10b). The fa
t that a uniform
triangulation of 4800
ubi
elements produ
es su
h a poor solution illustrates
how di
ult this nonlinear problem is.
4 Con
lusions
Although our PDE2D results appear to be substantially less a
urate than
those in Lu, Tian and Grimshaw, to judge by the errors in the integrals
2
of U and U , they, and to our knowledge every other su
essful attempt
to solve this notoriously di
ult PDE, used a numeri
al method
arefully
tailored to the KPI equation. We have shown that it is possible to get
reasonable a
ura
y using a general-purpose nite element program, provided
an adaptive, moving grid is used whi
h follows the peaks.
Referen
es
[1℄ B. Kadomtsev and V. Petviashvili. On the stability of solitary waves in
weakly dispersive media. Sov. Phys. Dok, 15:539541, 1970.
[2℄ B. Feng and T. Mitsui. A nite dieren
e method for the Korteweg-de
Vries and the kadomtsev-Petviashvili equations. Journal of Computa-
tional and Applied Mathemati
s, 90:95116, 1998.
[3℄ Z. Lu, E. Tian, and R. Grimshaw. Intera
tion of two lump solitons de-
s
ribed by the kadomtsev-petviashvili i equation. Wave Motion, 40:123
135, 2004.
[4℄ B. Fornberg and G.B. Whitham. A numeri
al and theoreti
al study of
ertain nonlinear wave phenomena. Phil. Trans. Royal So
iety London,
289:373404, 1978.
[5℄ G. Sewell. The Numeri
al Solution of Ordinary and Partial Dierential
Equations, se
ond edition. John Wiley & Sons, 2005.
[6℄ G. Sewell. Solving pdes in non-re
tangular 3d regions using a
ollo
ation
nite element method. Advan
es in Engineering Software, 5:748753,
2010.
71 Sewell
[7℄ G. Sewell. (free) software for general partial dierential equation problems
in non-re
tangular 2d and 3d regions. Bulletin of Computational Applied
Mathemati
s, 1(1):5154, 2013.
[8℄ I. Du and J. Reid. The multifrontal solution of unsymmetri
sets of
linear equations. SIAM Journal of S
ienti
and Statisti
al Computing,
5:633641, 1984.
