4:09 PM

CH-2-16(MO) 1/31 4:09 PM

Set of ODE’s
2/31

Numerical Methods
 A set of ODEs are solved very similarly
(Ordinary Differential Equations-3&4)
dy
 f1 ( x, y , z ) with y  yo at x  xo
dx
Kannan Iyer dz
[email protected]  f 2 ( x, y , z ) with z  zo at x  xo
dx
 Modified Euler’s method

k11  f1 ( xn , yn , z n ) k 21  f1 xn  h, y n 1 , z n 1 
k12  f 2 ( xn , yn , z n ) k 22  f 2 xn  h, y n 1 , z n 1 
Department of Mechanical Engineering
Indian Institute of Technology Jammu y n 1  y n  hk11  y n 1  y n  h / 2k11  k 21 
z n 1  z n  hk12  z n 1  z n  h / 2k12  k 22 

4:09 PM 3/31 4:09 PM 4/31

Higher Order equations
This equation is a stiff
Runge-Kutta Fourth Order Method
𝑑 𝑦 𝑑𝑦 equation with Solution
+ 3.1 + 0.3𝑦 = 0 𝑘 = 𝑓 (𝑥 , 𝑦 , 𝑧 )
𝑑𝑥 𝑑𝑥 y=e-3x+e-0.1x
𝑑𝑦 𝑘 = 𝑓 (𝑥 , 𝑦 , 𝑧 )
𝑤𝑖𝑡ℎ 𝑦 𝑥 = 0 = 2, (𝑥 = 0) = −3.1
𝑑𝑥 𝑘 = 𝑓 𝑥 + 0.5ℎ, 𝑦 + ℎ(0.5𝑘 ), 𝑧 + ℎ(0.5𝑘 )
 We can split the above equation as 𝑘 = 𝑓 𝑥 + 0.5ℎ, 𝑦 + ℎ(0.5𝑘 ), 𝑧 + ℎ(0.5𝑘 )
𝑑𝑦 𝑘 = 𝑓 𝑥 + 0.5ℎ, 𝑦 + ℎ(0.5𝑘 ), 𝑧 + ℎ(0.5𝑘 )
=𝑧 𝑤𝑖𝑡ℎ 𝑦(𝑥 = 0) = 2, 𝑘 = 𝑓 𝑥 + 0.5ℎ, 𝑦 + ℎ(0.5𝑘 ), 𝑧 + ℎ(0.5𝑘 )
𝑑𝑥
𝑘 = 𝑓 𝑥 + ℎ, 𝑦 + ℎ𝑘 , 𝑧 + ℎ𝑘
𝑑𝑧
= −3.1𝑧 − 0.3𝑦 𝑤𝑖𝑡ℎ 𝑧(𝑥 = 0) = −3.1 𝑘 = 𝑓 𝑥 + ℎ, 𝑦 + ℎ𝑘 , 𝑧 + ℎ𝑘
𝑑𝑥
𝑦 = 𝑦 + ℎ/6 𝑘 + 2𝑘 + 2𝑘 + 𝑘
𝑧 = 𝑧 + ℎ/6 𝑘 + 2𝑘 + 2𝑘 + 𝑘

1
4:09 PM 5/31 4:09 PM 6/31

Sample Problem Adaptive method

 Analytical Solution 2.00
for the given equation 2.00 A step size of dx = 0.05 is would require
1000 steps upto x = 50.
y  e 0.1 x  e 3 x 1.60
Adaptive method with a tolerance of 10-5
1.60
for relative error needs only 75 steps
1.20

y
1.20
A step sixe of dx = 0.05 is y 0.80
required for accurate sol.
0.80
0.40

0.40 0.00
0.00 2.00 4.00 6.00
x 0.00 20.00
x 40.00 60.00

4:09 PM 7/31 4:09 PM 8/31

Shooting method
Boundary Value Problem
The equation is split into a system of three first
ff   2 f   0 ; f ( 0 )  0 , f ( 0 )  0 , f (  )  1. order equations

 This equation is the classical Blasius Equation 𝑑𝑓

=𝑓, 𝑤𝑖𝑡ℎ 𝑓(0) = 0
 It does not have an analytical solution 𝑑𝑥
 Numerical solution obtained suggests x =10 can be
considered infinite
𝑑𝑓
 The solution can be found by the IVP approach =𝑓, 𝑤𝑖𝑡ℎ 𝑓 (0) = 0
iteratively 𝑑𝑥
 For this f’’(0) is first assumed and adjusted till f’(10)
obtained numerically is 0
 This approach is called shooting method
𝑑𝑓 −𝑓𝑓
= 𝑤𝑖𝑡ℎ 𝑓 0 = 0 (𝑎𝑠𝑠𝑢𝑚𝑒𝑑)
𝑑𝑥 2

2
4:09 PM 9/31 4:09 PM 10/31

Comments on Shooting Method Direct Solutions of BVP

• Shooting methods need iterative solutions • Finite difference methods can be used to obtain
solutions that will satisfy boundary conditions
• This may create convergence problems automatically
but usually it can be circumvented by • For a non-linear system the equations have to
judicial under relaxation be linearized, as otherwise solutions become
messy
• The advantage is that we can easily get 4th
• Step size sensitivity studies have to be
order solutions performed before accepting the solutions as
• Non-linearity does not require any special satisfactory
treatment

4:09 PM 11/31 4:09 PM 12/31

Finite difference principles Finite differences-I

• In this method, the derivatives are replaced by • The finite differences for derivatives can be
finite differences obtained very easily by Newton interpolating
• The domain is discretised into finite number of polynomials derived earlier
regions (say N) • The same can also be obtained by Taylor series
• A system of linear equations is formed for the N • Since Taylor series derivation is easy for first
unknown values of the functions and second derivatives upto second order it is
• Several approaches with varying accuracy are illustrated first.
possible
• Popular approaches restrict the order of method
upto second order 1 N+1
i-1 i i+1

3
4:09 PM 13/31 4:09 PM 14/31

Finite differences -II Finite differences-III

f ( x  h )  f ( x )  hf ( x )  h2! f ( x )  h3! f ( x )  O( h 4 )
2 3

• To get consistent accuracies near boundaries, often

f ( x  h )  f ( x )  hf ( x )  h2
f ( x )  h3
f ( x )  O( h )
4 we need to get forward and backward differences
2! 3!
• This is easily obtained by using Newton’s forward
 f ( x  h )  f ( x  h )  2 hf ( x )  2 h3! f ( x )  O( h 4 )
3
interpolating polynomial
• A system of linear equations is formed for the N
 f ( x  h )  f ( x  h )  2 f ( x )  h 2 f ( x )  O( h 4 ) unknown values of the functions
f ( x  h ) f ( x  h ) • Consider four points in the neighbourhood that are
 f ( x )   O( h 2 ) a distance h from each other
2h
sh h
f ( x  h ) f ( x  h ) 2 f ( x )
 f ( x )  2
 O( h 2 )
h x
X0 X1 X2 X3
The above two relations are called the centered approximations

4:09 PM 15/31 4:09 PM 16/31

Finite differences-IV Finite differences -V

• With four points we can fit a polynomial of third  Third Order Polynomial
order which will be fourth order accurate s( s  1) 2
P3 ( x0  sh)  f (0)  sf (0)   f (0)
• When the first derivative is taken, then this 2
approximation will drop to third order accuracy s ( s  1)( s  2) 3
• The same will become second order accurate,   f (0)  O ( h 4 )
6
when second derivative is expressed
• First we shall derive second order accurate s2  s
formulas by dropping one of the term and compare
 f 0  s( f1  f 0 )  ( f 2  2 f1  f 0 )
2
the results with the previously obtained ones.
s 3  3s 2  2 s
 ( f 3  3 f 2  3 f1  f 0 )  O( h 4 )
6

4
4:09 PM 17/31 4:09 PM 18/31

Finite differences -VI Finite differences -VII

 The first derivative
• To get derivatives at x0 the value of s will be 0 and
to get the same at x1, x2 and x3, the values of s will
 2s  1 be 1, 2 and 3 respectively
P3( x0  sh)  ( f1  f 0 )  ( f 2  2 f1  f 0 )
 2 • Thus, we can get backward, forward and centered
3s 2  6s  2 1 differences from a single expression just by
 ( f 3  3 f 2  3 f1  f 0 )  O( h 4 )  changing the value of s.
6 h • First, let us get relations for first derivatives that are
 The second derivative second order accurate at these points
 The first derivative One sided Difference at x0
P3 ( x0  sh)  ( f 2  2 f1  f 0 )  can be expressed as
6s  6  1  2s  1 1
( f 3  3 f 2  3 f1  f 0 )  O( h 4 )  2 P2( x0  sh)  ( f1  f 0 )  ( f 2  2 f1  f 0 )  O(h 2 )
6 h  2 h

4:09 PM 19/31 4:09 PM 20/31

Finite differences -IX
Finite differences -VIII  Putting s = 2, we get
 Putting s = 0, we get
 3 1
 1 1 P2( x2 )  ( f1  f 0 )  ( f 2  2 f1  f 0 )   O(h 2 )
P2( x0 )  ( f1  f 0 )  ( f 2  2 f1  f 0 )   O( h 2 )  2 h
 2 h (3 f 2  4 f1  f 0 )
( f 2  4 f1  3 f 0 )   O( h 2 ) ItBackward
has become
  O(h 2 ) Forward Difference 2h Difference
2h  We can get third order accurate one sided differences
 Putting s = 1, we get by using 3 terms and putting s = 0 and 3
 2s  1
 1 1 P3( x0  sh )  ( f1  f 0 )  ( f 2  2 f1  f 0 )
P2( x1 )  ( f1  f 0 )  ( f 2  2 f1  f 0 )  O( h 2 )  2
 2 h
3s 2  6 s  2 1
(f  f )  ( f 3  3 f 2  3 f1  f 0 )  O ( h 4 ) 
 2 0  O (h 2 ) It has become
6
2h centered Difference h
Forward Difference

5
4:09 PM 21/31 4:09 PM 22/31
Finite differences -X Simple Application-I
 1
P3( x0 )  ( f1  f 0 )  ( f 2  2 f1  f 0 )
 2
2 1 d2y dy
 ( f 3  3 f 2  3 f1  f 0 )   O ( h 3 ) 2
 a  by  0
6 h dx dx
2 f  9 f 2  18 f1  11 f 0 )
 3  O(h 3 ) with y ( x  0)  y0 , y ( x  L)  y L
6h Forward Difference
 5
P3( x3 )  ( f1  f 0 )  ( f 2  2 f1  f 0 )
 2
11 1 1 i-1 i i+1 n
 ( f 3  3 f 2  3 f1  f 0 )  O(h 3 )
6 h
11 f 3  18 f 2  9 f1  2 f 0 )
  O (h3 ) Backward Difference
6h

4:09 PM 23/31 4:09 PM 24/31

Simple Application-II Simple Application-III

d2y y( i  1 )  2 y ( i )  y ( i  1 ) Multiplying by h2 and collecting coefficients we get
2

dx i h2
dy y( i  1 )  y( i  1 ) y( i  1 )( 1  0.5 ah )  y( i )( bh 2  2 )  y( i  1 )( 1  0.5 ah )  0
 1 0 0 0   y1   y0 
dx i 2h
a a22 a23 0  y   b 
 21   2    2 
The finite difference equation for node I is 0 a32 a33 a34   yn 1  bn 1 
 
0 0 0 1   yn   y L 
y( i  1 )  2 y( i )  y( i  1 ) y( i  1 )  y( i  1 )
2
a  by( i )  0
h 2h We can solve for y’s using TDMA

6
4:09 PM 25/31 4:09 PM 26/31
Simple Application-IV Simple Application-V
 Writing FDE at point i
 Treatment of Neumann Boundary Condition
d2y dy y( i  1 )  2 y( i )  y( i  1 ) y( i  1 )  y( i  1 )
a  by( i )  0
2
 a  by  0 h 2
2h
1

dx dx  Boundary Condition at point i

dy y (i  1)  y (i  1)
with y ( x  0)  y0 , ( x  L)  yL  O(h 2 )  yL
dx 2h
 METHOD-1 Extended Domain Method  y (i  1)  y (i  1)  2hyL  O(h 3 ) 2

 Substituting Eq. (2) in Eq. (1), we get

y (i  1)  2hy L  O( h3 )  2 y (i )  y (i  1)
 There is
h2 degeneration
i-3 i-2 i-1 i i+1

y (i  1)  2hy L  O (h )  y (i  1)
3 of accuracy
X=L
a  by(i )  0
2h

4:09 PM 27/31 4:09 PM 28/31

Simple Application-V Simple Application-VI
 However, the solution can be obtained as for the Dirichlet
Boundary Condition as the matrix is tri-diagonal The above can be rearranged as
 The loss of accuracy near the boundary condition may not 6hyL  11yi  18 yi 1  9 yi  2  2 yi 3
be acceptable
 This can be overcome by using higher order formulation This formulation will break the tri-diagonal structure
at the boundary
1 0 0 0 0 0   y1   y0 
 METHOD-2 Higher Order Boundary Method a
 21 a22 a23 0 0 0   y2   b2 
We have shown that a third order accurate derivative can    
be expressed at the boundary as 0 a32 a33 a34 0 0   y3   b3 
     
yL 
11 yi  18 yi 1  9 yi  2  2 yi 3 )
 O(h 3 ) 0 0 a43 a44 a( i 2 )(i 1) 0   y4   b4 
6h 0 0 0 a54 a( i 1)(i 1) a(i 1)i   yi 1   bi 1 
    
 0 0 a( i  3 ) i a( i  2 ) i a(i 1) i aii   yi  6hyL 
i-3 i-2 i-1 i i+1

X=L

7
4:09 PM 29/31 4:09 PM 30/31
Simple Application-VII Treatment of Non-Linearity-I
 A tri-diagonal matrix will be obtained by performing two  Consider a non-linear Equation
Gauss operations
 First by performing Gauss Operation between i-2 and i y   2 y 2 y   0
rows, a(i-3),i can be reduced to 0  When a finite difference equation is written for a node, it
 Then by performing a Gauss operation between i-1 and I will lead to a non-linear equation due to the presence of
rows, we can reduce a(i-2),i to 0 higher order powers
 Thus, tri-diagonal structure is restored and can be solved  In such cases to get a linear form of the equation, we need
by TDMA to resort to iterations
 The procedure is to assume a y distribution
 Linearise and solve for y
 Iterate until convergence is reached
 The underlying principles used in linearisation are
discussed in next slide

4:09 PM 31/31
Treatment of Non-Linearity-II
 The term y 2 y  is linearised as

 y   y 
2 k k 1

 Thus, while solving, y2 value is always known and becomes

a coefficient in the matrix
 Frequently, the methods tend to diverge
 To facilitate convergence, under-relaxation is employed

 y k 1    y k 1  (1   ) y k
 α is assumed to have a value between 0 and 1
 The above suppresses wild variations of y introduced by
the iteration method
 Severe non-linearity may force the value of α close to zero
 Convergence criterion is similar to what we normally do by
controlling the normalized values between the iterations