MA557/MA578/CS557 Lecture 28 Spring 2003 Prof. Tim Warburton [email protected] 1 Linear Systems of 2D PDEs • A general linear pde system will look like: Find q : [0, t ] d where d dim(q) such that: q q q A B 0 t x y 2 Deriving DG • Ok – by popular demand here’s a brief derivation of the DG scheme for advection. Weak form of the advection equation is, 1 find q : P p T 0, t d such that P p T q q q , A B 0 x t T t Integrate by parts the terms with spatial derivatives: 2 q q q , A B 0 x t T t q , , Anxq Bn y q A , q B , q 0 T t T x T y T 3 q , , A n q B n q A , q B , q x y 0 T t T x T y T 3 modification to approximate q q , , Anx Bn y qˆ t T T A , q B , q 0 x T y T q ˆ , , A n B n q A , q B , q x y 0 T t T x T y T integration by parts of the last two terms 4 T by qˆ qˆ q , q q q q , , A , B , Anx Bn y qˆ q x T y T t T 4 T 0 5 Necessary condition on qhat for stability Scalar case dim(q) 1: q q q ˆ , , a , b , an bn q q x y x T y T t T Let's consider stability (usual approach): T 0 q, q Tk d q q ˆ q , a b q , an bn q q x y dt k 2 x y k k Tk anx bn y q, q Tk d q , q dt k 2 2 k 0 Tk q , anx bn y qˆ q Tk anx bn y q, q Tk d q , q dt k 2 2 k k q , anx bn y qˆ Tk k anx bn y q , q q , anx bn y qˆ 2 k Tk k Tk Tk 0 for stability 5 Tk 0 i.e. q , q 2qˆ 0 where anx bn y 2 k Tk 5 cont We re-express the sum over all elements into a sum over all unique edges. e q , 2 q 2qˆ q , 2 q 2qˆ 0 e e It is sufficient to make sure that each of these terms is negative, i.e.: q , q 2qˆ q , q 2qˆ e 0 having used e Sufficient condition: 6 q , q 2qˆ q , q 2qˆ e e - 0 using - Possible solution: If 0 ("inflow") then qˆ q , qˆ q If 0 ("outflow") then qˆ q , qˆ q 6 Proof of possible solution Sufficient condition: q , q 2qˆ q , q 2qˆ e 0 e Possible solution: If 0 ("inflow") then qˆ q , qˆ q q ,q q , q q , q 2qˆ q , q 2qˆ ... e 2q q ,q q e e 2q e e ... , q 2 q , q 0 qed e e If 0 ("outflow") then qˆ q , qˆ q q ,q q , q q , q 2qˆ q , q 2qˆ ... e 2q e 2q e e ... - q , q q , q 2 q , q 0 qed e e e 7 Upwinding Unveiled Sufficient condition: q , q 2qˆ q , q 2qˆ e 0 e Possible solution: If 0 ("inflow") then qˆ q , qˆ q If 0 ("outflow") then qˆ q , qˆ q This is nothing more than upwinding in disguise 8 Equivalent Conditions Sufficient condition: q , q 2qˆ q , q 2qˆ e 0 e Possible solution: If 0 ("inflow") then qˆ q , qˆ q If 0 ("outflow") then qˆ q , qˆ q Can be written: qˆ 2 q 2 q 1 1 q q q q 2 2 and similarly: 1 1 qˆ q q q q 2 2 1 1 = q 2 q 2 q q using 9 Returning To DG Scheme Scalar case, dim(q) 1: q q q ˆ , , a , b , an bn q q x y x T y T t T becomes: T 0 q q 1 q q q q q q 0 , , a , b , x T y T 2 2 t T T 1 1 using qˆ q q q q 2 2 Finally: q q q q q , , a , b , x T y T 2 t T 1 q q 2 T q q q q , a , b , , x T y T 2 T q q 2 10 T Convergence • Convergence follows as before…however for a given choice of qhat we need to verify consistency to guarantee stability. • In fact for the convergence result to hold we required the truncation error to vanish in the limit of small h and also large p. • For example we required: qˆ - P p q , P p q - - P p q - T 0 as h 0 and/or p • This is easily verified for the choice of qhat we just made. 11 General Choice of qhat • We established the following sufficient conditions on the choice of the function qhat(q+,q-): Sufficient condition: q , q 2qˆ q , q 2qˆ qˆ - P p q , P p q - - P p q - e T 0 e 0 as h 0 and/or p • As long as these are met for a choice of qhat then the scheme will be stable. • However, the choice of qhat may impact accuracy… 12 An Alternative Construction for qhat 1 1 qˆ q q q q where is chosen so 2 2 • Exercise – prove that this is a sufficient condition for stability. i.e. q , q 2qˆ q , q 2qˆ 0 e e • This generates the Lax-Friedrich flux based DG scheme: q q q q q q q , , a , b , , t x y 2 2 T T T T T 13 Lax-Friedrichs Fluxes • Suppose we know the maximum wave speed of the q q q hyperbolic system: A B 0 t x y are • As long as the matrix A, B codiagonalizable for any linear combination. n n • i.e. C A B Has real eigenvalues for any real alpha,beta. • And: max | C , x x for some x n , , | 2 2 1 • then we can use the following DG scheme: 14 Lax-Friedrichs DG Scheme For Linear Systems As before – we can generalize the scalar PDE scheme to a system: Find q : P p T [0, t ] d such that P p T q q q Anx Bn y q , q 0 , t A x B y , 2 T T 2 T where q q q 15 Discrete LF/DG Scheme • We choose to discretize the Pp polynomial space on each triangle using M=(p+1)(p+2)/2 Lagrange interpolatory polynomials. • i.e. P p T hm 1mM • The scheme now reads: Find q m t , m 1,..., M such that: m M h ,h m 1 n m T m M hm dq m m M hm A hn , q m ..... q m B hn , dt x T m1 y T m 1 Anx Bn y I hn , hm q m 0 2 T 16 Tensor Form Find q mcj t , m 1,..., M , c 1,...,dim(q), j=1,...,#tri such that: m M h ,h m 1 n d dim( q ) d 1 d=dim(q ) d 1 d 1 dt .... m M h A cd hn , m q mdj .... m1 x T j m M h B cd hn , m q mdj ..... y T j m1 d dim( q ) m T j dq mcj A cd nx B cd n y I m M hn , hm T q mdj 0 m1 2 cd 17 Simplified Tensor Form Find q ncj t , n 1,..., M , c 1,...,dim(q), j=1,...,#tri such that: dq ncj dt .... d dim( q ) d 1 mM x d=dim(q ) mM y A cd Dnmq mdj B cd Dnmq mdj ..... d 1 m1 m1 A cd nxe B cd n ey I m M e S q nm mdj 0 2 d 1 e 1 cd m1 where: d dim( q ) e 3 M nm hn , hm T D x nm k M hm x 1 x hn , , D M D nm nk km x T k 1 D y nm k M hm y y hn , , Dnm M 1 Dkm nk y T k 1 S e nm hn , hm T , S e e nm k M M k 1 1 nk S ekm 18 Geometric Factors • We use the chain rule to compute the Dx and Dy matrices: r r s s D D D x x r r s s y D D D y y x • In the umSCALAR2d scripts the umDr D r umDs Ds r s r s rx , sx , ry , sy x x y y 19 Surface Terms • In the Matlab code first we extract the nodes on the edges of the elements: • fC = umFtoN*C • The resulting matrix fC has dimension (umNfaces*(umP+1))xumNel • This lists all nodes from element 1, edge 1 then element 1, edge 2 then… • In order to multiply, say fC, by Se we use umNtoF*(Fscale.*fC) where: Fscale = sjac./(umNtoF*jac); 20 Example Matlab Code (Upwind DG Advection Scalar Eqn) • This is not strictly a global Lax-Friedrichs scheme since the lambda is chosen locally!!! 21 Example Matlab Code cont 22 Specific Example • We will try out this scheme on the 2D, transverse mode, Maxwell’s equations. 23 Maxwell’s Equations (TM mode) • In the absence of sources, Maxwell’s equations are: Hx t Ez 0 y H y Ez 0 t x Ez H x H y 0 t y x Hx H y x permittivity • Where y 0 permeability H magnetic field E electric field 24 Free Space.. • For simplicity we will assume that the permeability and permittivity are =1 we then obtain: H x Ez t y H y Ez t x Ez H x H y t y x H x H y x y 0 0 0 0 25 Divergence Condition • We next notice that by taking the x derivative of the first equation and the y derivative of the second equation we find that: H x H y t x y 0 • i.e. the fourth equation (divergence condition) is a natural result of the equations – assuming that the initial condition for the magnetic field H is divergence free. 26 The Maxwell’s Equations We Will Use H x Ez 0 t y H y Ez 0 t x Ez H x H y 0 t y x H y H x x, y , t 0 x, y , t 0 0 x y 27 The Maxwell’s Equations As Conservation Rule H x 0 Ez 0 y x t H y Ez 0 0 y x t Ez H y H x 0 y t x H y H x x, y , t 0 0 x, y , t 0 y x 28 TM Maxwell’s Boundary Condition • We will use the following PEC boundary condition – which corresponds to the boundary being a perfectly, electrical conducting material: x H H x H y H y Ez Ez 29 The Maxwell’s Equations In Matrix Form q q q A B 0 t x y Hx 0 0 0 0 0 1 q H y , A 0 0 1 , B 0 0 0 E 0 1 0 1 0 0 z • We are now assuming that the magnetic field is going to be divergence free – so we just neglect this for the moment. 30 Eigenvalues • Then: 0 0 0 0 0 1 0 A 0 0 1 , B 0 0 0 C 0 0 1 0 1 0 0 0 0 0 • C has eigenvalues which satisfy: 0 C I 0 0 0 C I 2 2 0, 2 2 • So the matrices are co-diagonalizable for all real alpha, beta and 1 given 1 2 2 31 Project Time • Due 04/11/03 • Code up a 2D DG solver for an arbitrary hyperbolic system based on the 2D DG advection code provided. • Test it on TM Maxwell’s – Set A and B as described, set lambdatilde = 1 and test on some domain of your choosing. – Use PEC boundary conditions all round. • Test it on a second set of named equations of your own choosing (determine these equations by research) • Compute convergence rates.. For smooth solutions. 32
© Copyright 2026 Paperzz