hydroinformatics module 4: numerical methods i · pdf filei.popescu:numerical methods...
TRANSCRIPT
I.Popescu:Numerical methods 1/16/2012
1
Hydroinformatics yModule 4: Numerical Methods I
Lecture 1 and 2: Models, ODE, Explicit and implicit methods
I.Popescu
Remarks:1. Schedule2. Prerequisites
• Basic mathematics, • Fluid dynamics and fluid mechanics,• Mathematical representation of fluid flow equations
3 Marking
1/16/2012 I.Popescu: Numerical Methods 2
3.Marking• Assignments (15%)• Written examination (30%)
I.Popescu:Numerical methods 1/16/2012
2
Course
Main aimintroduce numerical techniques which can be used on computersused on computersoverview of
WHAT it can be doneHOW it can be done
At the end of lectures you will:apply standard methods to simple flow problems and
1/16/2012 I.Popescu: Numerical Methods 3
apply standard methods to simple flow problems and determine the optimal values for time steps and grid sizeassess simulation results produced by different methodsevaluate the relevance of their useidentify and analyse the reason for failure of particular methods and, possibly, provide solutions
1-Why Numerical Methods?
I.Popescu:Numerical methods 1/16/2012
3
Computational Fluid Dynamics
Computational Fluid Dynamics (CFD) is the science of predicting fluid flow, heat transfer mass transfer chemical reactionstransfer, mass transfer, chemical reactions, and related phenomena by solving the mathematical equations which govern these processes using a numerical process (that is, on a computer).
1/16/2012 I.Popescu: Numerical Methods 5
Solution of Equations Phenomena in nature -described in terms of properties that prevail at each point of space and time separately
Formulating PDE’s governing the phenomena
SOLUTIONS
HOW ??
1/16/2012 I.Popescu: Numerical Methods 6
I.Popescu:Numerical methods 1/16/2012
4
Solution of Equations Phenomena in nature -described in terms of properties that prevail at each point of space and time separately
Formulating PDE’s governing the phenomena
Analytical
SOLUTIONS
Too complex equations?
1/16/2012 I.Popescu: Numerical Methods 7
Separationof
variables
Integral Solution(Green)
Integraltransforms(Fourier
& Laplace)
Solution of Equations Phenomena in nature -described in terms of properties that prevail at each point of space and time separately
Formulating PDE’s governing the phenomena
Analytical
SOLUTIONS
1/16/2012 I.Popescu: Numerical Methods 8
Separationof
variables
Integral Solution(Green)
Integraltransforms(Fourier
& Laplace)
I.Popescu:Numerical methods 1/16/2012
5
What is a model?
Is this a model?
1/16/2012 I.Popescu: Numerical Methods 9
What is a model?
Is this a model?
1/16/2012 I.Popescu: Numerical Methods 10
I.Popescu:Numerical methods 1/16/2012
6
What is a model?
Is this a model?
– Naomi Campbel
1/16/2012 I.Popescu: Numerical Methods 11
What is a model?
Is this a model?
1/16/2012 I.Popescu: Numerical Methods 12
I.Popescu:Numerical methods 1/16/2012
7
What is a model?
Is this a model?
1/16/2012 I.Popescu: Numerical Methods 13
What is a model?
Is this a model?
1/16/2012 I.Popescu: Numerical Methods 14
I.Popescu:Numerical methods 1/16/2012
8
What is a model?
1/16/2012 I.Popescu: Numerical Methods 15
What is a model?A simplification of reality, but with essential features -- an example:
RealityReality ModelModel
1/16/2012 I.Popescu: Numerical Methods 16
I.Popescu:Numerical methods 1/16/2012
9
What is a model?
Different formsEngineering problem solving
– A model is a simplified, schematic representations of the real world, a representations that retains enough aspects of the original system to make it useful to the modeller.
In fluid dynamics– physical models (scale models);
1/16/2012 I.Popescu: Numerical Methods 17
– physical models (scale models);– mathematical models
– conceptual– simulation models
Error and uncertainty in modellingDefinitions of model:
"A system which will convert a given input (geometry, boundary conditions, force, etc.) into an output (flow
t l l t ) t b d i i ilrates, levels, pressures, etc.) to be used in civil engineering design and operation" (Novak)"A simplified representation of the natural system" (Resfgaard)
It is not reasonable to expect that a "simplified
1/16/2012 I.Popescu: Numerical Methods 18
representation" convert input into exact output. Errors are inevitable. The actual values of the output variables are uncertain.
I.Popescu:Numerical methods 1/16/2012
10
What’s the difference between a modeling program, and a model?
Modelling program = softwareSoftware with an array of equations that simulate river flowsimulate river flow
Model = software + dataA modelling program + data specific to a certain river
1/16/2012 I.Popescu: Numerical Methods 19
What is a model?
Mathematical models
Q
1/16/2012 I.Popescu: Numerical Methods 20
H
I.Popescu:Numerical methods 1/16/2012
11
Solution of Equations Phenomena in nature -described in terms of properties that prevail at each point of space and time separately
Formulating PDE’s governing the phenomena
Analytical Numericalt h i
SOLUTIONS
1/16/2012 I.Popescu: Numerical Methods 21
techniquesSeparation
of variables
Integral Solution(Green)
Integraltransforms(Fourier
& Laplace)
FDM FVM FEMBEM
Ph i l
DifferentialEquations
AnalyticalMethods
SimpleDifferentialEquations
Why numerical methods?
Physicalphenomenon
Equations+
BoundaryConditions
ApproximateMethods
FDMor
FVMFEM,
BEM
1/16/2012 I.Popescu: Numerical Methods 22
BEMGeneral differential equations are too complicated to be solved analyticallyFDM, FEM, FVM, BEM are approximate numerical approaches for solving differential equations
I.Popescu:Numerical methods 1/16/2012
12
Why Numerical Methods?
Analytical solutions– derived only for few engineering problems – limited number of closed form solutions– useful to provide insight to the behaviour of an
engineering system
Numerical solutions
1/16/2012 I.Popescu: Numerical Methods 23
– fill the gap between theory and experiment
Why Numerical Methods?
Easy input - preprocessor.Solves many types of problemsy yp pModular design
– fluids, dynamics, heat, etc.
Can run on PC’s now.Relatively low cost, even sometimes
1/16/2012 I.Popescu: Numerical Methods 24
cheaper
I.Popescu:Numerical methods 1/16/2012
13
Why Numerical Methods?
Models are no longer restricted to closed solution problemsRepresent accurately realistically complex system such as:
– non-linear systems– time dependent state variables
Sometimes physical experiments are
1/16/2012 I.Popescu: Numerical Methods 25
Sometimes physical experiments are impossible to be simulated
Simple Mathematical Models
( )var var , ,Dependent iable f independent iable parameters forcing function=
Dependent variableh h ll fl h b h– A characteristic that usually reflects the behavior or
state of the system
Independent variable– Are usually dimensions, such as time and space
Parameters
1/16/2012 I.Popescu: Numerical Methods 26
– Are reflective of system’s properties or compositions
Forcing functions– External influences acting on the system
I.Popescu:Numerical methods 1/16/2012
14
Basic principle of Numerical Methods
DISCRETISATIONAn exact ‘continuous’ solution is represented by an approximate ‘discrete’ solution at a number of points.
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
1/16/2012 I.Popescu: Numerical Methods 27
0 1 2 3 4 5 6 7-1
-0.8
0 1 2 3 4 5 6 7-1
-0.8
Exact solution Discretised approximation
Basic principle of Numerical Methods
DISCRETISATION of space and time
Continuous in time
Discrete in timeApproximation
of the derivative
t
( ) ( )
1/16/2012 I.Popescu: Numerical Methods 28
t4t3t2t1
Δt
t ( ) ( )2 1 2 1
2 1 2 1
U t U tU U Ut t t t t
−Δ −= =
Δ − −
I.Popescu:Numerical methods 1/16/2012
15
Example - simple mathematical model
Newton Second LawFF ma or am
= =m
c
g
( ) ( ) ( ) ( )1 1
new value old value slope x ste
i i i i icv t v t g v t t tm
+ +
= +
⎡ ⎤= + − −⎢ ⎥⎣ ⎦↑ ↑ ↑ ↑
p size
1/16/2012 I.Popescu: Numerical Methods 29
/( ) 1 ct mgmv t ec
−⎡ ⎤= −⎣ ⎦Analytical solution -
Example-simple mathematical model
Newton Second Law- Falling objectM=68.1 kg
c=12 5 kg/sc=12.5 kg/s
t step= 2 secTime (s) Computed velocity (m/s)
Numerically Analytically0.0 0.000 0.0002.0 19.600 16.4044.0 32.005 27.769
1/16/2012 I.Popescu: Numerical Methods 30
6.0 39.856 35.6418.0 44.824 41.09510.0 47.969 44.87312.0 49.959 47.490
: 53.390 53.390
I.Popescu:Numerical methods 1/16/2012
16
Example - simple mathematical model
Newton Second Law- Falling object60
0
10
20
30
40
50
Velo
city
M=68.1 kg
c=12.5 kg/s
t step= 2 sec
1/16/2012 I.Popescu: Numerical Methods 31
00 10 20 30 40 50
Time
numerically analitically
2- Differential equations
- Quick review -
I.Popescu:Numerical methods 1/16/2012
17
Introduction and First Definitions1
Differential equations are equations that contain derivatives
In an algebraic equation, the unknown is a number. InIn an algebraic equation, the unknown is a number. In a differential equation, the unknown is a function.Most often the unknown function we seek, expresses the behavior of the state variable or variables as a function of time. This function it is called state equation
The order of the differential equation is the order
1/16/2012 Review of Mathematics 33
The order of the differential equation is the order of the derivative of the unknown function involved in the equation.
Introduction and First Definitions2
A linear differential equation of order n is a differential written in the following form:
)()()(...)()( 011
1
1 xfyxadxdyxa
dxydxa
dxydxa n
n
nn
n
n =++++ −
−
−
where an(x) is not the zero function.
Note: Some books may use the notation y’, ’’ ’’’ (4) f th d i ti
1/16/2012 Review of Mathematics 34
y’’, y’’’, y(4), …for the derivatives.
I.Popescu:Numerical methods 1/16/2012
18
Linear Differential EquationsA linear equation obliges the unknown function y(x) to have some restrictions. Accepted operation on y are:Accepted operation on y are:
Differentiating yMultiplying y and its derivatives by a function of the variable x.Adding what you obtained in
1/16/2012 Review of Mathematics 35
Adding what you obtained in previous step and let it be equal to a function of x.
Solving differential equations (DE)A major field of study
Existence: Does a differential equation have a solution?Uniqueness: Does a differential equation have more than one solution? If yes, how can we find a solution which satisfies particular y , pconditions?
Differential equations require integration to be solvedIf the values of the unknown function y(x) and its derivatives at some point are known the solution to DE is called an initial value problem (in short IVP). If no initial conditions are given, the description of all solutions to the DE is called the general solution
Some kinds of D E ’s can be solved exactly; many not
1/16/2012 Review of Mathematics 36
Some kinds of D.E. s can be solved exactly; many not for those that cannot, there are various ways to approximate the solution
I.Popescu:Numerical methods 1/16/2012
19
d
Single VariableWe know:
Differential Equations
)( yFdtdy
=
We want to know:)(tfy =
1/16/2012 Review of Mathematics 37
Several VariableWe know: We want to know:
Differential Equations
),...,,(1 qxyFdtdy
=
),...,,(2 qxyFdtdx
=
M M
)(2 tfx =
)(1 tfy =
1/16/2012 Review of Mathematics 38
),...,,( qxyFdtdq
n=
M M)(3 tfq =
I.Popescu:Numerical methods 1/16/2012
20
Methods for solving DEs
Solving a differential equation can be done in three major ways:
l i ll- analytically, - qualitatively, - numerically.
1/16/2012 Review of Mathematics 39
Partial derivatives
When a function changes with more than one variable at once – for example, both time and space – we use partial derivativesspace we use partial derivativesNotation:
= means the rate of change in y with tty∂∂
1/16/2012 Review of Mathematics 40
= means the rate of change in y with xxy∂∂
I.Popescu:Numerical methods 1/16/2012
21
Total time derivative
The total time derivative is used to describe the change in a variable in elation to a mo ing (as opposed torelation to a moving (as opposed to
stationary) objectExample:
xyu
ty
DtDy
∂∂
+∂∂
=
1/16/2012 Review of Mathematics 41
u is velocity
xtDt ∂∂
Time & space derivatives: reviewNotation:
= ordinary derivative: y changes only with tdy ordinary derivative: y changes only with t
= partial derivative: change in y with t at a fixed point
= total time derivative: change in y as oneDy
ty∂∂dt
1/16/2012 Review of Mathematics 42
total time derivative: change in y as one moves about (e.g., in a fluid)Dt
I.Popescu:Numerical methods 1/16/2012
22
Classification of DE ( Clasess)dimension of unknown:
ordinary differential equation (ODE) – unknown is a function of one variable, e.g. y(t)partial differential equation (PDE) – unknown is a function of multiple variables, e.g. u(t,x,y)p , g ( , ,y)
number of equations:single differential equation, e.g. y’=ysystem of differential equations (coupled), e.g. y1’=y2, y2’=-g
ordernth order DE has nth derivative, and no higher, e.g. y’’=-g
linear & nonlinear:linear differential equation: all terms linear in unknown and its
1/16/2012 Review of Mathematics 43
qderivativese.g x’’+ax’+bx+c=0 , x’=t2x – linear;x’’=1/x – nonlinear
Classification of PDEs2 2 2
2 2
u u u u ua b c d e fu gx x y y x y
∂ ∂ ∂ ∂ ∂+ + = + + +
∂ ∂ ∂ ∂ ∂ ∂2 4b acΔ = −
elliptic PDE - b2- 4ac < 0 – smooth solutions
hyperbolic PDE - b2- 4ac > 0 – solutions with discontinuities
4b acΔ
1/16/2012 I.Popescu: Numerical Methods 44
parabolic PDE - b2- 4ac =0
I.Popescu:Numerical methods 1/16/2012
23
General features of PDEs
Elliptic PDEs – describe processes that have already reached
steady state– time-independent (e.g. steady-state aquifer flow or
heat diffusion).
Hyperbolic PDEs – describe time-dependent, conservative physical
processes, such as convection, that are not evolving towards steady state (e.g. free surface flow, pipe
1/16/2012 I.Popescu: Numerical Methods 45
y ( g , p pflow).
Parabolic PDEs – describe time-dependent, dissipative physical
processes, such as diffusion, that are evolving towards steady state. (e.g. transient aquifer flow).
What do we need to solve DE?
Bou
ndar
y va
lues
Bou
ndar
y va
lues
Problem domain
1/16/2012 I.Popescu: Numerical Methods 46
a b
BB
Initial values
I.Popescu:Numerical methods 1/16/2012
24
Boundary and initial conditions
Initial value problems (IVP)derivative with respect to time (either dU/dt, or d2U/dt2 etc ) in the equation it is necessaryor d2U/dt2, etc.) in the equation, it is necessary to know the initial value of U in order to compute its value in the future, within the model domain. initial condition of U must be given. this problems are also called initial value
1/16/2012 I.Popescu: Numerical Methods 47
this problems are also called initial value problems
Boundary and initial conditions
Boundary value problems (BVP)– Initial conditions are not always sufficient to ensure
the existence of a unique solutions. Conditions onthe existence of a unique solutions. Conditions on the boundary of the domain of the problem must be given as well. These conditions are called boundary conditions.
C CComes from the boundary
time t time t + Δ t
1/16/2012 I.Popescu: Numerical Methods 48
x x
I.Popescu:Numerical methods 1/16/2012
25
Reminder - Basic formulas
What is a derivative?– Rate of change of a function
slope of the tangent line to the graph of a function– slope of the tangent line to the graph of a function– best linear approximation to a function
f
Real derivative
( ) ( )( ) limx o
f x x f xf xxΔ →
+ Δ −′ =Δ
1/16/2012 I.Popescu: Numerical Methods 49
Approx.derivatives
xx2x2x1 ( ) ( )( ) f x x f xf xx
+Δ −′ ≈Δ
Δx is small
Reminder - Basic formulas
Taylor series expantion2 3 4 5
( ) ( ) ( ) ( ) ( ) ( ) ( )2! 3! 4! 5!x x x xf x x f x xf x f x f x f x f xΔ Δ Δ Δ′ ′′ ′′′ ′′′′ ′′′′′±Δ = ±Δ + ± + ± +K( ) ( ) ( ) ( ) ( ) ( ) ( )2! 3! 4! 5!
f f f f f f f
2 3( ) ( ) 1 ( ) ( ) ( ) ( ) ( )2! 3!
f x x f x x xxf x f x f x f x O xx x
⎡ ⎤+ Δ − Δ Δ′ ′′ ′′′ ′= Δ + + + = + Δ⎢ ⎥Δ Δ ⎣ ⎦K
32( ) ( ) 1 2 ( ) 2 ( ) ( ) ( )f x x f x x xxf x f x f x O x
⎡ ⎤+ Δ − − Δ Δ′ ′′′ ′Δ + + + Δ⎢ ⎥
first order accurate in xΔ
1/16/2012 I.Popescu: Numerical Methods 50
2 ( ) 2 ( ) ( ) ( )2 2 3!
xf x f x f x O xx x
= Δ + + = + Δ⎢ ⎥Δ Δ ⎣ ⎦K
second order accurate in xΔ
I.Popescu:Numerical methods 1/16/2012
26
Reminder - Basic formulas
( ) ( )f x f x x− − Δf(x )
fo rw a rd d i f fe r e n c eb a c k w a rd d i f fe r e n c e
c e n tra l d i f fe r e n c e
( ) ( )f x x f xx
+ Δ −Δ
( ) ( )2
f x x f x xx
+ Δ − −ΔΔ
xΔ
1/16/2012 I.Popescu: Numerical Methods 51
xx xx x xi-1 i i+ 1= x
2 xΔ
Why study Numerical Methods ?NO numerical method is completely trouble free in all situations!
– How should one choose/use an algorithm to get trouble free and accurate answers?trouble free and accurate answers?
No numerical method is error free!,– What level of error/accuracy do one take in the way
the problem is solved?
No numerical method is optimal for all types/forms of an equation!
1/16/2012 I.Popescu: Numerical Methods 52
yp qIMPORTANT : In order to solve a physical problem numerically, you must understand the behavior of the numerical methods used as well as the physics of the problem
I.Popescu:Numerical methods 1/16/2012
27
Typical difficulties with NM
Instability
1/16/2012 I.Popescu: Numerical Methods 53
Typical difficulties with NM
Inconsistency
1/16/2012 I.Popescu: Numerical Methods 54
I.Popescu:Numerical methods 1/16/2012
28
Typical difficulties with NM
Inaccurate
1/16/2012 I.Popescu: Numerical Methods 55
Errors
Errors due to computers: – Round-off error
Exponent underflow– Exponent underflow – Exponent overflow
Errors due to numerical methods – Truncation error – Propagated error
M f
1/16/2012 I.Popescu: Numerical Methods 56
Measures for errors– absolute error– relative error
I.Popescu:Numerical methods 1/16/2012
29
Errors
1/16/2012 I.Popescu: Numerical Methods 57
Hydroinformaticianas a developer and user must understand how a numerical method performs for his/her given problemhis/her given problemunderstanding how various numerical algorithmsare derived,how the physics effects the numericshow the numerics effects the physics
h t th / t bilit ti
1/16/2012 I.Popescu: Numerical Methods 58
what are the accuracy/stability propertieswhat are the cost of a method for given level ofaccuracy (this varies substantially from method tomethod )
I.Popescu:Numerical methods 1/16/2012
30
Properties of Numerical Methods
Convergence– numerical scheme solution is convergent if it comes
closer and closer to the analytical solution of the realcloser and closer to the analytical solution of the realODE/PDE when the time step decreases;
LaxTheorem:2conditions need it for convergence– Consistency; and– Stability
1/16/2012 I.Popescu: Numerical Methods 59
Properties of Numerical Methods
ConsistencyA scheme is consistent if it gives a correct
i ti f th ODE/PDE thapproximation of the ODE/PDE as thetime/space step is decreasedverified using Taylor Series expansion
1/16/2012 I.Popescu: Numerical Methods 60
I.Popescu:Numerical methods 1/16/2012
31
Properties of Numerical Methods
StabilityA scheme is stable if any initially finite
t b ti i b d d tiperturbation remains bounded as time grows
Stability verificationMatrix methodFourier methodDomains of dependence
1/16/2012 I.Popescu: Numerical Methods 61
Domains of dependence
Notations
Un = U (tn)Uj = U (xj)Uj
n = U (tn, xj)t
n + 1
Δ t
Δ x
1/16/2012 I.Popescu: Numerical Methods 62
x
n
j -1 j j+ 1
I.Popescu:Numerical methods 1/16/2012
32
4-Finite difference methods for ODE
The general Initial Value Problem1
Given
),( tufdu=
and initial condition given
If you know values at time n, find the
),(fdt
1/16/2012 I.Popescu: Numerical Methods 64
values at time n+1
I.Popescu:Numerical methods 1/16/2012
33
The general IVP2
All initial value problems are solved by Integrating forward in tThere are two main types of integration yp gprocedure
– One-step methods– Multi-step methods
u u
1/16/2012 I.Popescu: Numerical Methods 65
t0 ti ti+1
t tt0
General classification
Numerical scheme– a particular discretization of a differential equation
Schemes:Schemes: – Explicit (Euler method) - the value of the variable at
time level n+1 can be computed directly (or explicitly) from the value at time level n.
– Implicit (Improved Euler) - involves values of the variable at time level n+1, there is an implicit
1/16/2012 I.Popescu: Numerical Methods 66
relationship between the derivative and the variable which has to be computed.
– Mixed (Crank-Nicholson ) - schemes in which the explicit and implicit schemes are combined.
I.Popescu:Numerical methods 1/16/2012
34
Explicit schemes/Euler/One Step Method
u
Forward difference approximation
= ),( tufdtdu
u
Δt∑−
=
+
Δ+=
⋅Δ+≈
1
0
1
),(
),(
N
n
nninitialfinal
nnnn
tuftuu
tuftuu
dt
1/16/2012 I.Popescu: Numerical Methods 67
This is the same as saying:new value = old value + (slope) x (step size)t n t n+1
These schemes are also called one step methods or “marching in time”, because new value is calculated as one step forward from the old value.
Implicit schemesBackward difference approximation
U ( , )dU f U t=U
Δt( )1 1 1,n n n nU f U t t U+ + +− Δ =
( , )f U tdt
Iteration need it, ussualy
1/16/2012 I.Popescu: Numerical Methods 68
t n t n+1
I.Popescu:Numerical methods 1/16/2012
35
Mixed schemesAveraging implicit and explicit schemes
U
Use this “average” slopeto predict Un+1
1/16/2012 I.Popescu: Numerical Methods 69
( ) ( )1 11
, ,
2
n n nnn n
f U t f U tU U t
+ ++
+= + Δ
{t n t n+1
Numerical schemes for ODE - Example
Ground water reservoir
Q
h
dh hdt
α= −
1/16/2012 I.Popescu: Numerical Methods 70
toh h e α−=
-exact solution
I.Popescu:Numerical methods 1/16/2012
36
Multi step methods: Midpoint MethodsA better estimate of the function U would come from evaluating the derivative at the midpoint of the Δt interval.
U
Δuu n+1/2=?
The problem: we know t at the midpoint but we don’t know the u(t) at the midpoint (yet). The solution is to use Euler’s method to estimate Δu and then re-estimate Δu using the derivatives evaluated halfway t n t n+1/2 t n+1
Δt
u
t
1/16/2012 I.Popescu: Numerical Methods 71
along the line segment encompassing the original Δu.This method is called the second order Runge Kutta method, or the midpoint method.
Mixed schemes - Midpoint Methods
tkakauu nn Δ++=+ )( 22111
Where for two point schemes
U(analytical)
U(Midpoint)
U(t)
),(1nn tufk =
),( 11112 tpttkqufk nn ΔΔ= ++
;21
211
,aa
tscoefficienweightedaa−=−
Where for two point schemesU(Euler)
tn ttn+1
1/16/2012 I.Popescu: Numerical Methods 72
21
12 =pa
21
112 =qa
;21 1 aa
I.Popescu:Numerical methods 1/16/2012
37
ExerciseApply the Euler and Mid-point methods to the following problem (with known solutions!):
fdu )(
utufutufttufttuf
fortuf
dtdu
−====
=
),(.4),(.3),(.2),(.1
),(
2
1/16/2012 I.Popescu: Numerical Methods 73
In each case, compare the errors in the two schemes after taking the same number (10, say) of identical timeteps (start with ). Is the mid-point method noticeably superior to Euler?
Mixed schemes - Multi Step Methods
The midpoint method can be extended by considering other intermediate estimates. The most frequently used variation is the fourth-order q yRunge-Kutta method which considers one estimate at the initial point, two estimates at the midpoint, and one estimate at a trial endpoint.This is popular for two reasons
It is easy to program
1/16/2012 I.Popescu: Numerical Methods 74
It is stable
HoweverIt requires more computing time
I.Popescu:Numerical methods 1/16/2012
38
Runge-Kutta Method - 44thth--orderorder
f2uu
f1
f3
f4
f
u
1/16/2012 I.Popescu: Numerical Methods 75
ti ti + h/2 ti + h
( )4321 2261 fffff +++=
t
Rungge Kutta - 4 point scheme
tkakakakauu nn Δ++++=+ )( 443322111
WhereWhere
)()2/,2/()2/,2/(
),(
),(),(
23
12
1
kfktttkufktttkufk
tufk
tuatderivativethetuf
ΔΔΔ+Δ+=Δ+Δ+=
=
−U(analytical)
U(t)
2
3
4
1/16/2012 I.Popescu: Numerical Methods 76
),( 34 tttkufk Δ+Δ+=tn
ttn+1
12
I.Popescu:Numerical methods 1/16/2012
39
Runge-Kutta 4th order - solution
(tn + k1Δt/2, un+ k1Δt/2)
Δt/2
slope = k1
t
u
)(
),(),(
tfk
tuatderivativethetuf −(tn ,un )
(k1Δt/2)
1/16/2012 I.Popescu: Numerical Methods 77
),()2/,2/()2/,2/(
),(
34
23
12
1
tttkufktttkufktttkufk
tufk
Δ+Δ+=Δ+Δ+=Δ+Δ+=
=
point to
Runge-Kutta versus Euler-example
2007.06 uudtdu
−=
Solve the equation below using both Euler and Runge Kutta methods
point to estimate
Problem: estimate the slope to
calculate Δu
Δuu
1/16/2012 I.Popescu: Numerical Methods 78
Δ t = 0.5
I.Popescu:Numerical methods 1/16/2012
40
u0 = 10 analytical u
Euler solution
u
k1 = du/dt at u0
k1 = 6*10-.007*(10)2
Δu = k1*Δt
uest= u0 + Δu
Δu
estimated u
1/16/2012 I.Popescu: Numerical Methods 79
Δ t = 0.5
Step 1: Evaluate slope at current value of state variable.
u0 = 10
k1 = du/dt at u0
k1 = 6*10-.007*(10)2
k1 = 59.3k1=slope 1
u
1/16/2012 I.Popescu: Numerical Methods 80
I.Popescu:Numerical methods 1/16/2012
41
Step 2: Calculate u1at t +Δt/2 using k1.
Evaluate slope at u1.
u1 = u0 + k1* Δt /2
u = 24 82u1 = 24.82
k2 = du/dt at u1
k2 = 6*24.8-.007*(24.8)2
k2 = 144.63k2=slope 2
u1
u
1/16/2012 I.Popescu: Numerical Methods 81
Δ t = 0.5/2
1
Step 3: Calculate u2 at t +Δt/2 using k2.
Evaluate slope at u2.
u2 = u0 + k2* Δt /2
u2 = 46.2 k3 = slope 3
k3 = du/dt at u2
k3 = 6*46.2-.007*(46.2)2
k3 = 263.0
u2
u
1/16/2012 I.Popescu: Numerical Methods 82
Δ t = 0.5/2
I.Popescu:Numerical methods 1/16/2012
42
Step 4: Calculate u3 at t +Δt using k3.
Evaluate slope at u3.
u3 = u0 + k3* Δt
u3 =141.0 k4 = slope 4u3
k4 = du/dt at u2
k4 = 6*141.0-.007*(141.0)2
k4 = 706.9
u2
u
1/16/2012 I.Popescu: Numerical Methods 83
Δ t = 0.5
Step 5: Calculate weighted slope.
Use weighted slope to estimate u at t +Δt
1 2 3 41 ( 2 2 )6
k k k k+ + +weighted slope =
)22(14321
1 kkkktuu nn +++Δ+=+
true value
estimated valueweighted slope
)22(6 4321 kkkktuu +++Δ+
u
1/16/2012 I.Popescu: Numerical Methods 84
Δ t = 0.5
I.Popescu:Numerical methods 1/16/2012
43
4th order Runge-Kutta offers substantial improvement over Eulers.Both techniques provide estimates, not “true” values.The accuracy of the estimate depends on the size of the step used in the algorithm.
Conclusions
Runge-Kutta
Analytical
Euleru
1/16/2012 I.Popescu: Numerical Methods 85
Numerical schemes for ODE - Example
Ground water reservoir
Q
h
dh hdt
α= −
1/16/2012 I.Popescu: Numerical Methods 86
toh h e α−=
-exact solution
I.Popescu:Numerical methods 1/16/2012
44
Numerical schemes for ODE - ExampleSolution methods :1. Explicit(Euler)
The right hand side is determined at current time (hn)
1 hh nn+
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
nh1+nh
)(th
)1(
1
1
thh
ht
hh
nn
nnn
Δ−=
−=Δ−
+
+
α
α
1/16/2012 I.Popescu: Numerical Methods 87
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 50 t
)1(1
th
hn
n
Δ−=+
α
The right hand side is determined at future time (hn+1)
Numerical schemes for ODE - ExampleSolution methods :2. Implicit
11
++
−=Δ− n
nn
ht
hh α
11un+0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
nh1+nh
)(th
1/16/2012 I.Popescu: Numerical Methods 88
)1(
1tu
un Δ+=
α0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
0 t
I.Popescu:Numerical methods 1/16/2012
45
Numerical schemes for ODE - Example
toh h e α−=
h
dh hdt
α= −1 (1 )n nh t hα+ = − Δ
Q
1/16/2012 I.Popescu: Numerical Methods 89
1
1
nn hh
tα+ =
+ Δ
The BIG question
≈≠=1/16/2012 I.Popescu: Numerical Methods 90
I.Popescu:Numerical methods 1/16/2012
46
The BIG question
If equations are approximation then itshould:
behave in the same way as the function to which it• behave in the same way as the function to which itis an approximation, and
• become a better approximation as the time-step isreduced.
1/16/2012 I.Popescu: Numerical Methods 91
0.9
1Comparison of different solution methods
exactexplicit
Explicit vs Implicit
1nh +
)exp()exp
)1(exp 1
t-αtαn(
t)n(-αh
hn
n
Δ=Δ−Δ+
=+
Exact solution
0.3
0.4
0.5
0.6
0.7
0.8
pimplicit nh
1/16/2012 I.Popescu: Numerical Methods 92
0 0.5 1 1.5 2 2.5 30
0.1
0.2
Exact solution lies between explicit
and implicit solutions
tΔα
I.Popescu:Numerical methods 1/16/2012
47
0.9
1Comparison of different solution methods
exact
1
n
n
hh +
Three different solution methods -comparison
0.3
0.4
0.5
0.6
0.7
0.8
exactexplicitimplicitcombined (Crank-Nicolson)
Crank-Nicolson is closest to
exact solution
1/16/2012 I.Popescu: Numerical Methods 93
0 0.5 1 1.5 2 2.5 30
0.1
0.2
tΔα
Stability, Consistency, Convergence
Lax Theorem:
Consistency + stability ⇔ convergence
This is why you can trust your models !This is why you can trust your models !
Consistency : The scheme approximates the ODE more correctlyas the discretisation is refined.
Stability: Any initial perturbation remains bounded.Convergence: The numerical solution becomes closer to the real
one as the discretisation is refined
1/16/2012 I.Popescu: Numerical Methods 94
I.Popescu:Numerical methods 1/16/2012
48
The BIG question
If equations are approximation then itshould:
behave in the same way as the function to which it• behave in the same way as the function to which itis an approximation, and
• become a better approximation as the time-step isreduced.
1/16/2012 I.Popescu: Numerical Methods 95
Stability, Consistency, Convergence
Convergence - Explicit method
tn ehh α−→ 0ehh →
ththhthh
)1()1()1(
2012
01
Δ−=Δ−=Δ−=
ααα
K
1/16/2012 I.Popescu: Numerical Methods 96
non thh )1( Δ−= α
on
n hnth ⎥⎦
⎤⎢⎣
⎡⎟⎠⎞
⎜⎝⎛−+=
α1
I.Popescu:Numerical methods 1/16/2012
49
Stability, Consistency, Convergence
Convergence - Explicit method (cont)
⎥⎤
⎢⎡
⎟⎞
⎜⎛
⎟⎞
⎜⎛ ⋅⎟
⎞⎜⎛⎟⎞
⎜⎛ ⋅ 2
210 ntnt αα
⎥⎥⎦
⎤
⎢⎢⎣
⎡+⎟⎟⎠
⎞⎜⎜⎝
⎛⎟⎠⎞
⎜⎝⎛−+⎟⎟
⎠
⎞⎜⎜⎝
⎛⎟⎠⎞
⎜⎝⎛−+= −− K
11
111 210 n
ntn
nthh nnnn αα
where
!)!(!
rrnn
rn
−=⎟
⎠⎞⎜
⎝⎛
r⎞⎛
1/16/2012 I.Popescu: Numerical Methods 97
!)!( rrn⎠⎝
rnrnnnrrn
n→−−=
−))...(1(
!)!(! !r
nrn r
=⎟⎠⎞⎜
⎝⎛
Stability, Consistency, Convergence Convergence - Explicit method (cont)
nn ntnthh αα⎥⎤
⎢⎡
+⎟⎞
⎜⎛ ⋅
+⎟⎞
⎜⎛ ⋅
+→22
0 1
Back to the equation
tn ehtth
nnhh
ααα −→⎥⎥⎦
⎤
⎢⎢⎣
⎡+⎟
⎠⎞
⎜⎝⎛ ⋅−+⎟
⎠⎞
⎜⎝⎛ ⋅−+→
⎥⎥⎦⎢
⎢⎣
+⎟⎠
⎜⎝−+⎟
⎠⎜⎝−+→
02
0
!21
111
!2!11
K
K
1/16/2012 I.Popescu: Numerical Methods 98
⎥⎦⎢⎣ ⎠⎝⎠⎝ !211
Taylor series for te ⋅−α
I.Popescu:Numerical methods 1/16/2012
50
Stability, Consistency, Convergence
1
1nh +
≤
Stability condition
111
≤≤+nh
Stability - Explicit schemes
1nh≤
αα 20
111 =Δ⇒⎨⎧ ≤Δ−
⇒≤Δ−≤− tt
t
11 ≤≤− nh
1/16/2012 I.Popescu: Numerical Methods 99
ααα
2111 maxΔ⇒
⎩⎨ ≤Δ
⇒≤Δ≤ tt
t
Stability, Consistency, Convergence
Example
1
2210
max
=
==Δ⇒=
hand
tTakeα
α
) d
dt
dt=0.125
a)0<dt<1
1/16/2012 I.Popescu: Numerical Methods 100
dt=0.5
dt=0.25
dt 0 5
I.Popescu:Numerical methods 1/16/2012
51
b)1<=dt<2
Stability, Consistency, Convergence
dt=1.25
d 1
1/16/2012 I.Popescu: Numerical Methods 101
=>Stable But Oscillatory Approximations
dt=1.5dt=1.5
c)2<dtStability, Consistency, Convergence
dt=4.5
dt=3
1/16/2012 I.Popescu: Numerical Methods 102
=>Unstable (i.e. Bad) Approximations
dt=2.5
I.Popescu:Numerical methods 1/16/2012
52
Stability, Consistency, Convergence
Stability - Explicit schemes
H
0 1 1 1 or 0 t tαα
≤ − Δ ≤ ≤ Δ ≤
Δt
1/16/2012 I.Popescu: Numerical Methods 103
t n t n+1
Stability, Consistency, Convergence
Stability - Explicit schemes
H
ΔtΔt0 1/α 2/α
oscillatory unstablestable
1/16/2012 I.Popescu: Numerical Methods 104
t n t n+1
I.Popescu:Numerical methods 1/16/2012
53
Stability, Consistency, Convergence
Consistency - explicit schemes
( )2
2dh t d hh O tΔ+ Δ( )2
22h O t
dt dtα+ = − − Δ
Truncation error - first order accurate
1/16/2012 I.Popescu: Numerical Methods 105
Stability, Consistency, Convergence
Implicit schemes– convergence - FOR YOU TO TEST
stability unconditionally stable– stability - unconditionally stable
Mixed schemes– unconditionally stable
1/16/2012 I.Popescu: Numerical Methods 106
I.Popescu:Numerical methods 1/16/2012
54
Amplification factor
U n+1/U n is called the numerical amplification factor
1/16/2012 I.Popescu: Numerical Methods 107
Amplitude portrait
Method EulerImproved Euler
Region of instability
1
0
1
0 1 2 3 4AN
Region of instability
1/16/2012 I.Popescu: Numerical Methods 108
-2
-1
a Δt
Region of instability
I.Popescu:Numerical methods 1/16/2012
55
4-Finite difference methods for ODE
Quick summary
What you should remember
ODEs can be solved numerically by replacing the d/dt operator by a difference of the considered function at two time levels.function at two time levels.When the derivative is estimated using the values of the function at time step level n, the scheme is said to be explicit. When the derivative is estimated using the values at time n+1, the scheme is said to be implicit.
1/16/2012 I.Popescu: Numerical Methods 110
pThe discretized ODE is consistent to the real one if it comes closer to the real ODE as dt (or dx) comes closer to zero.
I.Popescu:Numerical methods 1/16/2012
56
What you should rememberThe difference between the discretised ODE and the real one is called the truncation error. It is obtained by substituting Taylor series expansions into the discretisation.The solution is said to be stable if initial perturbations in the numerical solution remain bounded.
1/16/2012 I.Popescu: Numerical Methods 111
The stability of the solution is investigated by calculating the value of the amplification factor. If |AN| is smaller than unity, the solution is stable.
What you should remember
In general, explicit schemes are not stable for all values of ∆t (or ∆x). I li i h l blImplicit schemes are always stable.The numerical solution is said to converge to the real one if it comes closer to it as ∆t (or ∆x) comes closer to 0.Consistency and stability of a scheme are
1/16/2012 I.Popescu: Numerical Methods 112
Consistency and stability of a scheme are necessary and sufficient conditions to the convergence of the solution