National Central University Department of Mathematics Numerical ElectroMagnetics & Semiconductor Industrial Applications 12 2D-NUFFT & Applications Ke-Ying Su Ph.D. 2/67 Part II : 2D-NUFFT 3/67 Outline : 2D-NUFFT 1. Introduction 2. 2D-NUFFT algorithm 3. Approach 4. Results and discussions 5. Conclusion 4/67 I. Introduction Numerical methods • Finite difference time domain approach (FDTD) • Spectral domain approach (SDA) • Finite element method (FEM) • Integral equation (IE) • Mode matching technique Widely used technique Method of Moment (MoM) 5/67 Disadvantages 1) Slow convergence of Green’s functions 2) A large number of basis functions Used solutions 1) 1,2) Using the 2D discrete fast Fourier Transform (FFT) 2D-FFT + Using the first few resonant modes’ current distributions Nonuniform meshs for mixed potential integral equation (MPIE) 6/67 New solution SDA + 2D-NUFFT + Nonuniform meshs NUFFT : 1D 2D 0.7 The core idea of the 1D-NUFFT: 0.6 0.5 q 1 k1 ( s j )e i ( v j q / 21) 2k / mN f 0.4 rl( sj) e iks j sj = 0.5207 q=8 0.3 0.2 1 0.1 S j-q/2 S j-1 sj S j S j+1 S j+q/2 0 -0.1 1 2 3 4 5 l 6 7 8 9 The q+1 nonzero coefficients. 7/67 NUFFT : 1D 2D The square 2D-NUFFT Some of these 2D coefficients approach to zero rapidly. The nonsquare 2D-NUFFT The (q+1)2 nonzero coefficients. coefficients remove directly accuracy least square error accuracy 8/67 II. 2D-NUFFT Algorithm Dst Our aim M / 2 1 N / 2 1 G m M / 2 n N / 2 mn e e (3.1) mn e imx e iny Y s-q/2 Y s-1 s Y s+q/2 rpg ( xt , y s )e i ( vt p ) 2m / cM e i ( us g ) 2n / cN p Xt-1 Xt Xt+1 Ys Y s+1 The key step t Xt-q/2 imxt inys (3.2) g 2D-NUFFT Dst q/2 q/2 r pg p q / 2 g q / 2 ( xt , ys )H pg ( s, t ) (3.3) 2D-FFT 9/67 Xt+q/2 Regular Fourier Matrix mn e imx e iny t s [r1, q / 2 ( vt 1) m ( us q / 2 ) n r0, q / 2 ( vt 0) m ( us q / 2 ) n r1, q / 2 ( vt 1) m ( us q / 2 ) n ] [r1,q / 2 ( vt 1) m ( us q / 2 ) n r0,q / 2 ( vt 0 ) m ( us q / 2 ) n r1,q / 2 ( vt 1) m ( us q / 2 ) n ] For a given (xt,ys), for m = -M/2,…,M/2-1 and n = -N/2,…, N/2-1 Ar(xt,ys) = b(xt,ys) Xt-q/2 A: (MN)(q2/2+3q+1) b : (MN)1 r(xt,ys) = [A*A]-1[A*b(xt,ys)] = Fr-1Pr Xt-1 Xt Xt+1 Xt+q/2 Y s-q/2 Y s-1 (3.5) Ys Y s+1 Y s+q/2 where Fr is the regular Fourier matrix with size (q2/2+3q+1)2 Fr & Pr : closed forms 10/67 Solution Extract Fr and Pr from Ff and Pf where Ff is the regular Fourier matrix with size (q+1)2 1) Define a vector product as -q/2 -q/2 [a1, a2, …, am] [b1, b2, …, bn] = (3.6) [a1b1, a2b1, …, amb1, …, a1bn, a2bn, …, ambn] 0 0 q/2 Let Vp and Vg be the (p+1)th and (g+1)th row of the regular Foruier matrix for 1D problem p and g = 0, 1, …,q. The [g(q+1)+(p+1)]th row of Ff equals VpVg. 11/67 q/2 Let = ei2/cM and = ei2/cN, (3.7) ( N / 2 )( g ) ( N / 2 )( g ) ( N / 2 )( 1) ( N / 2 )( 1) Vg [ , ..., , N, g 1 1 1 ( N / 2 )(1) ( N / 2 )(1) ( N / 2 )( q g ) ( N / 2 )( q g ) , ..., ] 1 q g 1 1 (3.8) 12/67 2) Choose mn = cos(m/cM)cos(n/cN). The [g(q+1) +(p+1)]th element of Pf Pf ( xt , y s ) g ( q 1)( p 1) sin ( 2 p d q 2{xt cM 2 }) 2c d 1,1 i ( 2{ xt cM 2 } q 2 p d ) cM 1 e sin ( 2 g d q 2 { y cN 2 }) s 2 c d 1,1 i ( 2{ y s cN 2 } q 2 g d ) cN 1 e (3.9) where {x} = x - [x]. 13/67 3) Fill Fr and Pr from Ff and Pf For square grid points: 1 = [1, 2, …, (q+1)2] q2+3q/2 q/2 For octagonal grid points: q 2+q+1 q/2+1 q/2+2 q2+3q/2+1 q2+3q/2+2 = [q/2, q/2+1, q/2+2, 3q/2, …, q2+3q/2, q2+3q/2+1, q2+3q/2+2] q+1 2 (q+1) Fr(i, j) = Ff( (i), (j)) Pr(i) = Pf( (i)) 14/67 4 31 40 49 76 6 33 42 51 78 The relation between Fr and Ff, and Pr and Pf 5 32 41 50 77 Example: Let q = 8 9 36 45 54 81 Index = [1, 2, 3, …, 79, 80, 81] Index = [4, 5, 6, 12, 13, 14, 15, 16, …, 76, 77, 78] 1 2 3 4 5 6 7 8 81 44 55 66 31 31 32 32 33 33 99 36 45 45 54 54 36 40 40 41 41 42 42 49 49 50 50 51 51 73 76 77 78 Fr Ff = 81 1 2 3 4 5 6 7 8 28 28 37 37 46 46 1 2 3 4 5 6 7 8 11 1 2 3 4 5 6 7 8 81 81 81 81 15/67 2D-NUFFT 1~3) Ff, Pf Fr, Pr rr = Fr-1 Pr 4) Xt-q/2 H pg Xt+q/2 Y s-q/2 (3.5) Y s-1 Ys Y s+1 Y s+q/2 2D-FFT: M / 2 1 Xt-1 Xt Xt+1 N / 2 1 1 i ( vt p ) 2m / cM i ( us g ) 2n / cN G e mn mne m M / 2 n N / 2 (3.4) If M = N = 210 and c = 2, then a 2D-FFT with size cMcN uses 3.02 seconds (CPU:1.6GHz). 5) Dst rpg ( xt , ys )H pg ( s, t ) p (3.3) g 16/67 III. Approach z The Green’s functions c b substrate thickness t box dimension abc kxm = m/a, kyn = n/b t 0 y a x ~ Gxx ( x, x' , y, y ' ) Gxx cos( k xm x' ) sin (k yn y ' ) cos( k xm x ) sin( k yn y ) m n m n ~ Gxy ( x, x' , y, y ' ) Gxy sin( k xm x' ) cos(k yn y ' ) cos( k xm x ) sin( k yn y ) m n ~ G yy ( x, x' , y, y ' ) G yy sin( k xm x' ) cos(k yn y ' ) sin( k xm x ) cos( k yn y ) ~ where Gxx is the spectral domain Green’s function (3.10) 17/67 Spectral domain Green’s functions z c b t 0 y a x 18/67 Solution procedure Asymmetric rooftop functions and the nonuniform meshs J(x, y) = axJx(x, y) + byJy(x, y) (3.11) source J s(x,y) J x(x,y) Jy(x,y) J y(x,y) J L(x,y) load terminal y J L(x,y) x load terminal 19/67 Asymmetric rooftop function J x,y) y y y y x x x x x Jx = Jxx(x, x)Jxy(y, y) (3.12) cos( k xm x ) cos( k xm ( x x 1 )) cos( k xm x ) cos( k xm ( x x 2 )) ~ J xx 2 2 x 1k xm x 2 k xm (3.14a) cos( k yn ( y y / 2)) cos( k yn ( y y / 2)) ~ J xy k yn k yn (3.14b) 20/67 Galerkin’s procedure Final MoM matrix ~ ~ ~ Z xx (d , e) G xx (m, n ) J xd (m, n )J xe (m, n ) m (3.15) n Trigonometric identities ~ M / 21 N / 21 Guv ( m, n ) sin( k xm ( xd xe )) cos( k yn ( y d y e )) g h m M / 2 n N / 2 k xm k yn (3.16) ~ M / 21 N / 21 G uv ( m, n ) cos( k xm ( xd xe )) sin( k yn ( y d y e )) g h m M / 2 n N / 2 k xm k yn 21/67 Procedure for evaluating the MoM matrix Nonuniform mesh scheme ~ G xx 3 k xm k yn2 ~ G xx k xm4 k yn2 ~ G yx 2 k xm k yn3 ~ G yx 3 k xm k yn3 ~ G yy 2 k xm k yn4 2D-NUFFT 2D-NUFFT 2D-NUFFT MoM 2D-NUFFT 2D-NUFFT 22/67 III. Numerical Results a L4 Hairpin resonator b GPOF g2 w1 w2 L3 w2 y w2 x c b GPOF g1 L2 z L1 y z y t 0 c a x L5 t L5 er er = 10.2, L1 = 0.7, L2 = 1.01, L3 = 2.74, L4 = 8, L5 = 6, w1 = 1, w2 = 1.19, g1 = 0.2 and g2 = 0.8. All dimensions are in mm. 23/67 Table 3.1 Comparison of CPU Time and L2 error of One Call of the 2D-NUFFT in Analysis of a Hairpin Resonator Comparison of Analyses of The Hairpin Resonator with Uniform and Nonuniform Grids Table 3.2.1 24/67 -q/2 -q/2 The measured and calculated S parameters of the hairpin resonator. 0 10 0 |S 11| and | S 21| (dB) 0 q/2 |S11| Measurement -20 -40 -60 -80 2D-NUFFT, q = 4 2D-NUFFT, q = 6 2D-NUFFT, q = 8 |S21| 2 2.2 2.4 2.6 Frequency (GHz) 2.8 3 25/67 q/2 Normalized magnitudes of the current distribution on the hairpin resonator at 2.473 GHz. (a) |Jx(x,y)| (b) |Jy(x,y)| 26/67 Normalized magnitudes of the current distribution on the hairpin resonator at 2.397 GHz. (c) |Jx(x,y)| (d) |Jy(x,y)| 27/67 Interdigital capacitor y d x L4 L2 e s gd L3 L 4 er = 10.2, L1 = 8, L2 = 1.6, L3 = 0.8, L4 = 1.2, L5 = 7.9, d = 0.4, e = 0.4, g = 0.2 and s = 0.2. The thickness of substrate is 1.27. All dimensions are in mm. L1 L1 L5 Comparison of Analyses of The Ingerdigital capacitor with Uniform and Nonuniform Grids Table 3.2.2 28/67 The measured and calculated S parameters of the interdigital capacitor. |S 11| and | S 21| (dB) 0 |S11| -4 -8 |S21| -12 2D-NUFFT Measurement -16 -20 1 2 3 4 5 6 7 8 9 10 11 12 Frequency (GHz) 29/67 Normalized magnitudes of the current distribution on the interdigital capacitor at 5 GHz. (a) |Jx(x,y)| (b) |Jy(x,y)| 30/67 Wideband filter a L1 = 8, L2 = 0.56, L3 = 0.576, L4 = 0.69, L5 = 0.3605, L6 = 0.125, L7 = 0.125, L8 = 0.125, L9 = 5.19, L10 = 4.88, L11 = 0.38, L12 = 2.06, L13 = 1.9, L14 = 7.75, L15 = 11.3, t = 0.635, er = 10.8. All dimensions are in mm. L12 L1 L6 b y L5 L11 L10 L8 L7 x L4 L13 L3 L2 L1 L9 z y of Analyses of LThe Comparison Wideband filter with 14 L15 L15 c Uniform and Nonuniform Grids t er Table 3.2.3 31/67 The measured and calculated S parameters of the wideband filter. 0 |S 11| and | S 21| (dB) -10 -20 |S21| |S11| -30 -40 2D-NUFFT Measurement [37] -50 -60 2 4 10 12 6 8 Frequency (GHz) 14 16 32/67 Normalized magnitudes of the current distribution at 6 GHz (a) |Jx(x,y)| (b) |Jy(x,y)| 33/67 VI. Conclusion • A 2D-NUFFT algorithm with octagonal interpolated coefficients are used to enhance the Computation. • The octagonal 2D-NUFFT uses less CPU time than the square 2D-NUFFT. • The L2 error of the octagonal 2D-NUFFT is the same as that of square 2D-NUFFT. • The scattering parameters of the hairpin resonator, an interdigital capacitor and a wideband filter are calculated and validated by measurements. 34/67 THE END Thank You for your Participation ! 35/67