Courant and all that. Consistency, Convergence Stability Numerical Dispersion Computational grids and numerical anisotropy

Transcription

1 Consistency, Convergence Stability Numerical Dispersion Computational grids and numerical anisotropy c dx ε 1 he goal of this lecture is to understand how to find suitable parameters (i.e., grid spacing dx and time increment ) for a numerical simulation and knowing what can go wrong. 1

2 A simple example: Newtonian Cooling Numerical solution to first order ordinary differential equation d = f (, t) We can not simply integrate this equation. We have to solve it numerically! First we need to discretise time: t = t 0 + and for emperature = ( t )

3 A finite-difference approximation Let us try a forward difference: d t= t 1 = + + O( )... which leads to the following explicit scheme : + f (, t + 1 ) his allows us to calculate the emperature as a function of time and the forcing inhomogeneity f(,t). Note that there will be an error O() which will accumulate over time.

4 Coffee? Let s try to apply this to the Newtonian cooling problem: Air Cappucino How does the temperature of the liquid evolve as a function of time and temperature difference to the air?

5 Newtonian Cooling he rate of cooling (d/) will depend on the temperature difference ( cap - air ) and some constant (thermal conductivity). his is called Newtonian Cooling. With = cap - air being the temperature difference and τ the time scale of cooling then f(,t)=-/ τ and the differential equation describing the system is d = /τ with initial condition = i at t=0 and τ>0.

6 Analytical solution his equation has a simple analytical solution: ( t) = i exp( t / τ ) How good is our finite-difference appoximation? For what choices of will we obtain a stable solution? Our FD approximation is: + 1 = + 1 τ = = (1 (1 ) τ ) τ

7 FD Algorithm + 1 = (1 ) τ 1. Does this equation approximation converge for -> 0? 2. Does it behave like the analytical solution? With the initial condition = 0 at t=0: 2 1 = 0(1 ) τ = 1( 1 ) = 0 (1 )(1 τ τ = 0(1 ) τ leading to : ) τ

8 Understanding the numerical approximation ) 0(1 τ = Let us use =t / where t is the total time up to time step : t + = τ 1 0 his can be expanded using the binomial theorem = t t τ τ

9 ! )! (! r r r =... where we are interested in the case that -> 0 which is equivalent to -> r r r + = 1) 2)...( 1)( ( )! (! as a result r! r r

10 substituted into the series for we obtain: 2 2 t t ! τ 2! τ which leads to 2 t 1 t τ 2! τ... which is the aylor expansion for = 0 exp( t / τ )

11 So we conclude: For the Newtonian Cooling problem, the numerical solution converges to the exact solution when the time step gets smaller. How does the numerical solution behave? = 0 exp( t / τ ) he analytical solution decays monotonically! + 1 = (1 ) τ What are the conditions so that +1 <?

12 Convergence Convergence means that if we decrease time and/or space increment in our numerical algorithm we get closer to the true solution.

13 Stability + 1 = (1 ) τ +1 < requires 0 1 < 1 τ or 0 < τ he numerical solution decays only montonically for a limited range of values for! Again we seem to have a conditional stability.

14 + 1 = (1 ) τ if τ < < 2τ then ( 1 ) < 0 τ the solution oscillates but converges as 1-/τ <1 if > 2τ then / τ > 2 1-/τ<-1 and the solution oscillates and diverges... now let us see how the solution looks like...

15 Matlab solution % Matlab Program - Newtonian Cooling % initialise values nt=10; t0=1. tau=.7; =1. % initial condition =t0; % time extrapolation for i=1:nt, (i+1)=(i)-/tau*(i); end % plotting plot()

16 Example 1 Blue numerical Red - analytical

17 Example 2 Blue numerical Red - analytical

18 Example 3 Blue numerical Red - analytical

19 Example 4 Blue numerical Red - analytical

20 Stability Stability of a numerical algorithm means that the numerical solution does not tend to infinite (or zero) while the true solution is finite. In many case we obtain algoriths with conditional stability. hat means that the solution is stable for values in well-defined intervals.

21 Grid anisotropy In which direction is the solution most accurate and why?

22 Grid anisotropy with ac2d.m FD FD

23 and more a Green s function FD

24 Grids Cubed sphere Hexahedral grid Spectral element implementation etrahedral grid Discontinuous Galerkin method

25 Grid anisotropy

26 Grid anisotropy Grid anisotropy is the directional dependence of the accuracy of your numerical solution if you do not use enough points per wavelength. Grid anisotropy depends on the actual grid you are using. Cubic grids display anisotropy, hexagonal grids in 2D do not. In 3D there are no grid with isotropic properties! Numerical solutions on unstructured grids usually do not have this problem because of averaging effects, but they need more points per wavelength!

27 Real problems Alluvial Basin V S = 300m/s f max = 3Hz λ min = V S /f max = 100m 36 km 30 km Bedrock V S = 3200m/s f max = 3Hz λ min = V S /f max = m

28 Hexahedral Grid Generation

29 Courant ime step v P dx ε ε dx v P Largest velocity Smallest grid size

30 - Summary Any numerical solution has to be checked if it converges to the correct solution (as we have seen there are different options when using FD and not all do converge!) he number of grid points per wavelength is a central concept to all numerical solutions to wave like problems. his desired frequency for a simulation imposes the necessary space increment. he Courant criterion, the smallest grid increment and the largest velocity determine the (global or local) time step of the simulation Examples on the exercise sheet.