25
1 CHAPTER 11 Numerical Schemes for Hyperbolic PDEs ------------------------------------------------------------------------------------------------------- Copyright © 2015 by R.M. Barron. All rights reserved. No part of these notes may be reproduced or distributed in any form or by any means, mechanical or electronic, including but not limited to photocopying, recording, storage or retrieval system, without prior written permission from the author. 0 ; c cu u x t ] ) 1 ( 2 ) 1 ( [ 2 1 ) ( 1 1 i i i i x i x u u u x u u 1. Model Hyperbolic Equation (Linear Case) The space derivative can be approximated by either 1 st order one-sided differencing, or by 2 nd order central differencing (see Ch. 9, Slide 2) Note: β = 0 central β = 1 backward; leads to inherently unstable ODEs if c < 0 β = -1 forward; leads to inherently unstable ODEs if c > 0 Since c > 0, we consider only central and backward differencing. (11.1) (11.2)

Chapter+11+-+Numerical+Schemes+for+Hyperbolic+PDEs+F15

Embed Size (px)

DESCRIPTION

cdd

Citation preview

Page 1: Chapter+11+-+Numerical+Schemes+for+Hyperbolic+PDEs+F15

1

CHAPTER 11

Numerical Schemes for Hyperbolic PDEs

-------------------------------------------------------------------------------------------------------

Copyright © 2015 by R.M. Barron. All rights reserved. No part of these notes may be reproduced or distributed

in any form or by any means, mechanical or electronic, including but not limited to photocopying, recording,

storage or retrieval system, without prior written permission from the author.

0; ccuu xt

])1(2)1([2

1)( 11

iiiixix uuu

xuu

1. Model Hyperbolic Equation (Linear Case)

The space derivative can be approximated by either 1st – order one-sided

differencing, or by 2nd – order central differencing (see Ch. 9, Slide 2)

Note: β = 0 → central

β = 1 → backward; leads to inherently unstable ODEs if c < 0

β = -1 → forward; leads to inherently unstable ODEs if c > 0

Since c > 0, we consider only central and backward differencing.

(11.1)

(11.2)

Page 2: Chapter+11+-+Numerical+Schemes+for+Hyperbolic+PDEs+F15

2

A. Explicit Schemes

(i) Explicit Euler (for time march)

For β = 0 (see FTCS; Ch. 5, Slide 8)

• O(∆t, ∆x2)

• unconditionally unstable

)(2

)(2

11111

1 ni

ni

ni

ni

ni uu

x

tcuuu

• O(∆t, ∆x2)

• conditionally stable; Courant number =

• 1-step method

])1(2)1([2

11

1ni

ni

ni

ni

ni uuu

x

c

t

uu

Lax Method: average in FTCSniu

1

x

tcC

(11.3; HC:6-6)

Page 3: Chapter+11+-+Numerical+Schemes+for+Hyperbolic+PDEs+F15

3

• conditionally stable;

• 1-step method

)( 11 n

ini

ni

ni uu

x

tcuu

For β = 1 (FTBS)

• O(∆t, ∆x)

• σ - λ relationship for Explicit Euler is σ = 1 +λΔt

• from Ch. 9, Slide 4:

M

mi

M

m

x

cm

2sin

2cos1

1

x

tcC

(11.4; HC:6-4)

Page 4: Chapter+11+-+Numerical+Schemes+for+Hyperbolic+PDEs+F15

4

(ii) β = 0, Leapfrog (for time march)

• see Ch. 5, Slides 9,10,14-28

• O(∆t2, ∆x2)

• conditionally stable;

• 2-step method

1

x

tcC

)( 1111 n

ini

ni

ni uu

x

tcuu

(11.5; HC:6-7)

Page 5: Chapter+11+-+Numerical+Schemes+for+Hyperbolic+PDEs+F15

5

(iii) Lax-Wendroff Method

• see Ch. 8, Slide 5 (for 2D)

• O(∆t2, ∆x2)

• conditionally stable;

• 1-step method

1

x

tcC

)2(2

)(2

112

22

111 n

ini

ni

ni

ni

ni

ni uuu

x

tcuu

x

tcuu

(11.6; HC:6-11)

Note: Before applying central space differencing, the Lax-Wendroff

technique leads to the equation

n

i

n

i

ni

ni

x

utc

x

uc

t

uu

2

221

2

artificial dissipation ↑

Page 6: Chapter+11+-+Numerical+Schemes+for+Hyperbolic+PDEs+F15

6

B. Implicit Schemes

(i) Implicit Euler (for time march)

])1(2)1([2

11

111

1

ni

ni

ni

ni

ni uuu

x

c

t

uu

For β = 0 (see BTCS; Ch. 6, Slide 1)

• O(∆t, ∆x2)

• unconditionally stable

• 1-step method, using tridiagonal solver

)(2

11

11

1

n

ini

ni

ni uu

x

tcuu (11.7; HC:6-12)

Page 7: Chapter+11+-+Numerical+Schemes+for+Hyperbolic+PDEs+F15

7

For β = 1 (BTBS)

• O(∆t, ∆x)

• σ - λ relationship for Implicit Euler is σ = (1 – λΔt)-1

• from Ch. 9, Slide 4:

M

mi

M

m

x

cm

2sin

2cos1

• unconditionally stable

• 1-step method, using tridiagonal solver

)( 11

11

n

ini

ni

ni uu

x

tcuu (11.8; HC:6-13)

Page 8: Chapter+11+-+Numerical+Schemes+for+Hyperbolic+PDEs+F15

8

(ii) Crank-Nicolson, with β = 0

• see Ch. 6, Slide 2

• O(∆t2, ∆x2)

• unconditionally stable

• 1-step method, using tridiagonal solver

)(4

111

11

11 n

ini

ni

ni

ni

ni uuuu

x

tcuu

(11.9; HC:6-15)

Page 9: Chapter+11+-+Numerical+Schemes+for+Hyperbolic+PDEs+F15

9

C. Splitting Methods

• for multidimensional problems

• Approximate Factorization – see, eg.,

Lax-Wendroff for 2D (Ch. 8, Slides 5-8)

C-N:

Factored form (Ch. 8, Slides 15,16)

Factored delta form (Ch. 8, Slides 17,18)

ADI (Ch. 8, Slide 19)

ADI (Ch. 7, Slides 12-14,16,17)

Fractional Step (Ch. 7, Slides 15,18)

Page 10: Chapter+11+-+Numerical+Schemes+for+Hyperbolic+PDEs+F15

10

D. Multi-Step Methods

• better suited for nonlinear problems

(i) Richtmyer/Lax-Wendroff (2-step)

Richtmyer:

• apply equations at i

• O(∆t2, ∆x2)

• conditionally stable;

• see equations in H&C: 6-17, 6-18

Step 1: n → n + ½; use Lax method

Step 2: n + ½ → n + 1; use Leapfrog method

2

x

tcC

Page 11: Chapter+11+-+Numerical+Schemes+for+Hyperbolic+PDEs+F15

11

Lax-Wendroff:

• apply equations at i + ½

• O(∆t2, ∆x2)

• conditionally stable;

• see equations in H&C: 6-19, 6-20

1

x

tcC

(ii) MacCormack Method

• see Ch. 5, Slides 3,4

• O(∆t2, ∆x2)

• conditionally stable;

• see equations in H&C: 6-22, 6-23

1

x

tcC

See Hoffmann & Chiang, p. 191-206 for a computational example

and comparison of methods (linear problem).

Page 12: Chapter+11+-+Numerical+Schemes+for+Hyperbolic+PDEs+F15

12

2. Model Hyperbolic Equation (Nonlinear Case)

x

uu

t

u

(11.10)

x

E

t

u

where E = u2/2or (11.11)

• this are referred to as the “inviscid Burgers equation”

• wave propagates with different velocity at each point

• may eventually form a discontinuity - similar to shock wave formation

from a series of compression waves

Page 13: Chapter+11+-+Numerical+Schemes+for+Hyperbolic+PDEs+F15

13

A. Explicit Schemes

(i) Lax Method

Applying Explicit Euler and central space differencing to (11.11):

)(2

111 n

ini

ni

ni EE

x

tuu

Stability is improved by averaging to give niu

)(2

)(2

11111

1 ni

ni

ni

ni

ni EE

x

tuuu

or, in terms of u,

Page 14: Chapter+11+-+Numerical+Schemes+for+Hyperbolic+PDEs+F15

14

• explicit method

• nonlinear terms appear at level n on RHS; no linearization needed

• O(Δt, Δx2)

• conditionally stable: Courant number =

• O(Δt) implies dissipative errors may occur – initial discontinuity

smeared over several grid points

• see H&C, p. 208, Fig. 6-20

1max

x

tuC

))()((4

)(2

1 2

1

2

111

1 n

i

n

i

n

i

n

i

n

i uux

tuuu

(11.12; HC:6-34)

Page 15: Chapter+11+-+Numerical+Schemes+for+Hyperbolic+PDEs+F15

15

(ii) Lax-Wendroff Method (see Chapter 8)

Expand in Taylor series about , and replace time

derivatives using the PDE.

Note that E = E(u) so, for example,

niu1n

iu

x

EA

x

E

u

E

t

u

u

E

t

E

where Jacobian

u

EA

Since E = u2/2, we get A = u, so the above becomes

x

Eu

t

E

Then, one can show that

x

Eu

xx

EA

xt

u2

2

Page 16: Chapter+11+-+Numerical+Schemes+for+Hyperbolic+PDEs+F15

16

Substituting into the Taylor series gives

)(2

21

tOt

x

EA

xx

E

t

uu ni

ni

↑ like an error term in modified PDE

Central space differencing and some averaging leads to

)(2

111 n

ini

ni

ni EE

x

tuu

)])(())([()(4

)(11112

2ni

ni

ni

ni

ni

ni

ni

ni EEuuEEuu

x

t

(11.13; HC:6-39)

- see H&C for derivation, page 210.

Page 17: Chapter+11+-+Numerical+Schemes+for+Hyperbolic+PDEs+F15

17

• explicit method

• nonlinear terms appear at level n on RHS; no linearization needed

• O(Δt2, Δx2)

• conditionally stable: Courant number =

• O(Δt2) implies dispersion errors may occur – oscillations near

the discontinuity

• see H&C, p. 211-212, Figs. 6-22, 6-23

1max

x

tuC

Page 18: Chapter+11+-+Numerical+Schemes+for+Hyperbolic+PDEs+F15

18

(iii) MacCormack Method

• predictor-corrector or multi-level type scheme

• explicit method; no linearization needed

• O(Δt2, Δx2)

• conditionally stable: Courant number =

• well-behaved; no oscillations and very little smearing when

C = 1, due to the splitting which applies a forward, then a

backward difference for the space derivative

1max

x

tuC

)( 1* n

ini

nii EE

x

tuu

)(2

1 *1

**1iii

ni

ni EE

x

tuuu (11.14; HC:6-40,6-41)

• see H&C, p. 212-213, Figs. 6-24, 6-25

Page 19: Chapter+11+-+Numerical+Schemes+for+Hyperbolic+PDEs+F15

19

(iv) FTBS

Applying Explicit Euler and backward space differencing to (11.11):

)( 11 n

ini

ni

ni EE

x

tuu

(11.15; HC:6-49)

• explicit method

• nonlinear terms appear at level n on RHS; no linearization needed

• O(Δt, Δx)

• conditionally stable: Courant number =

• O(Δt) implies dissipative errors may occur – initial discontinuity

smeared over several grid points; but O(Δx) upwinding introduces

artificial viscosity which may suppress these errors

• see H&C, p. 219, Fig. 6-29

1max

x

tuC

Page 20: Chapter+11+-+Numerical+Schemes+for+Hyperbolic+PDEs+F15

20

(i) Beam and Warming Method

• implicit method; use tridiagonal solver

• FDE is nonlinear; linearized by lagging the Jacobian

• O(Δt2, Δx2)

• unconditionally stable

• O(Δt2) implies dispersion error (very large); add 4th order

damping term (like adding artificial viscosity). Since 4th

order, it doesn’t affect the O(Δx2) accuracy.

• advantages of implicit scheme (stable, larger Δt) is lost due

to dispersion errors (large oscillations)

B. Implicit Schemes

Page 21: Chapter+11+-+Numerical+Schemes+for+Hyperbolic+PDEs+F15

21

4

44

111111 )(44

)(2 x

uxuA

x

tuA

x

tEE

x

tu e

ni

ni

ni

ni

ni

ni

ni

1

1111

1144

ni

ni

ni

ni

ni uA

x

tuuA

x

t

(11.16; HC:6-47,6-48)

• for stability, we must choose 0 < εe ≤ 0.125

• the damping term can be approximated by

)464()( 21124

44 n

ini

ni

ni

niee uuuuu

x

ux

• see H&C, p. 216-218, Figs. 6-26, 6-27, 6-28

Page 22: Chapter+11+-+Numerical+Schemes+for+Hyperbolic+PDEs+F15

22

(ii) BTBS

Applying Implicit Euler and backward space differencing to (11.11):

)( 11

11

n

ini

ni

ni EE

x

tuu

• nonlinear terms appear at level n+1 on RHS

• linearization needed – lag the value of u, i.e., write

1211

2

1)(

2

1 nnnn uuuE

Page 23: Chapter+11+-+Numerical+Schemes+for+Hyperbolic+PDEs+F15

23

• implicit method

• use bidiagonal solver (lower bidiagonal matrix – forward sweep)

• O(Δt, Δx)

• unconditionally stable

• see H&C, p. 220, Fig. 6-30

(11.17; HC:6-51)

This scheme can be expressed in the bidiagonal form

ni

ni

ni

ni

ni uuu

x

tuu

x

t

1111

21

2

Page 24: Chapter+11+-+Numerical+Schemes+for+Hyperbolic+PDEs+F15

24

C. RK Schemes

• popular explicit methods for solving ODEs; easy to program

• uses weighted average of several solutions over time interval

Δt – improves accuracy

• RK2 or 2-stage RK method – 2nd order accurate

• RK4 or 4-stage RK method – 4th order accurate; most commonly

used RK method

• Modified RK methods – reduces storage requirements

• better stability than comparable explicit schemes

• requires significantly more computations per time step

• difficult to estimate errors

• see H&C, p. 219-227.

Page 25: Chapter+11+-+Numerical+Schemes+for+Hyperbolic+PDEs+F15

25

D. Other Schemes

• Flux Corrected Transport

• Total Variation Diminishing (TVD) Schemes

- 1st order TVD – monotone schemes

- 2nd order TVD

- often use “flux limiters”: Harten-Yee upwind limiter,

Roe-Sweby upwind limiter, Davis-Yee symmetric

limiter

• Essentially Non-Oscillatory (ENO) Schemes

• Weighted ENO (WENO) Schemes

• Compact Schemes

• see H&C, p. 233-252.