IEEE TRANSACTIONS ON ELECTROMAGNETIC COMPATIBILITY, VOL. 53, NO. 4, NOVEMBER 2011 959 Multipoint Full-Wave Model Order Reduction for Delayed PEEC Models With Large Delays Francesco Ferranti, Member, IEEE, Michel S. Nakhla, Fellow, IEEE, Giulio Antonini, Senior Member, IEEE, Tom Dhaene, Senior Member, IEEE, Luc Knockaert, Senior Member, IEEE, and Albert E. Ruehli, Life Fellow, IEEE Abstract—The increase of operating frequencies requires 3-D electromagnetic (EM) methods, such as the partial element equivalent circuit (PEEC) method, for the analysis and design of highspeed circuits. Very large systems of equations are often produced by 3-D EM methods and model order reduction (MOR) techniques are used to reduce such a high complexity. When signal waveform rise times decrease and the corresponding frequency content increases, or the geometric dimensions become electrically large, time delays must be included in the modeling. A PEEC formulation, which include delay elements called τ PEEC method, becomes necessary and leads to systems of neutral delayed differential equations (NDDE). The reduction of large NDDE is still a very challenging research topic, especially for electrically large structures, where delays among coupled elements cannot be neglected or easily approximated by rational basis functions. We propose a novel model order technique for τ PEEC models that is able to accurately reduce electrically large systems with large delays. It is based on an adaptive multipoint expansion and MOR of equivalent first-order systems. The neutral delayed differential formulation is preserved in the reduced model. Pertinent numerical examples based on τ PEEC models validate the proposed MOR approach. Index Terms—Delayed partial element equivalent circuit (PEEC) method, model order reduction (MOR), neutral delayed differential equations (NDDE). I. INTRODUCTION HE increasing demand for performance of ICs pushes operation to higher signal bandwidths and accurate modeling of previously neglected second-order effects, such as crosstalk, T Manuscript received December 1, 2010; revised March 30, 2011; accepted April 30, 2011. Date of publication June 27, 2011; date of current version November 18, 2011. This work was supported by the Research Foundation Flanders (FWO) and by the Italian Ministry of University (MIUR) under a Program for the Development of Research of National Interest (PRIN) under Grant 2006095890. F. Ferranti, T. Dhaene, and L. Knockaert are with the Department of Information Technology, Internet Based Communication Networks and Services (IBCN), Ghent University-IBBT, Gaston Crommenlaan 8 Bus 201, B-9050 Gent, Belgium (e-mail: francesco.ferranti@ugent.be; tom.dhaene@ugent.be; luc.knockaert@ugent.be). M. S. Nakhla is with the Department of Electronics, Carleton University, Ottawa, ON K1S 5B6, Canada (e-mail: msn@doe.carleton.ca). G. Antonini is with the UAq EMC Laboratory, Dipartimento di Ingegneria Elettrica e dell’Informazione, Università degli Studi dell’Aquila, 67100 L’Aquila, Italy (e-mail: giulio.antonini@univaq.it). A. E. Ruehli is with the IBM T. J. Watson Research Center, Yorktown Heights, NY 10598 USA, and also with the Missouri University of Science and Technology, Rolla, MO 65409 USA (e-mail: albert.ruehli@gmail.com). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TEMC.2011.2154335 reflection, delay, and coupling, becomes increasingly important during circuit and system simulations [1], [2]. Electromagnetic (EM) 3-D methods [3]–[5] have become increasingly indispensable analysis and design tools for a variety of complex high-speed systems. Large systems of equations are usually generated by the use of these methods, and model order reduction (MOR) techniques are crucial to reduce the complexity of EM models and the computational cost of the simulations, while retaining the important physical features of the original system [6]–[9]. The development of reduced-order models of EM systems has become a topic of intense research over the last years, with applications to vias, high-speed packages, interconnects, and on-chip passive components [1], [10]–[12]. Among all EM methods, the partial element equivalent circuit (PEEC) method [4] has been found particularly useful for modeling printed circuit boards, interconnects, and power systems. The PEEC method uses a circuit interpretation of the electric field integral equation (EFIE) [13], thus, being especially suitable to problems involving both EM fields and circuits [2], [4], [14]. Nonlinear circuit devices, such as drivers and receivers, are usually connected to PEEC equivalent circuits using a time-domain circuit simulator (e.g. SPICE [15]). However, inclusion of the PEEC model directly into a circuit simulator may be computationally intractable for complex structures because the number of circuit elements can be in the tens of thousands. The quasistatic PEEC formulation [4], which approximates the full-wave PEEC approach [16], yields an equivalent RLC circuit by neglecting the time delays between the elements in the full-wave PEEC formulation. Systems of ordinary differential equations (ODE) with constant coefficients in the time domain and complex algebraic systems of equations with frequency-independent matrices in the frequency domain are obtained. Standard MOR techniques for ODE systems can be used to reduce the size of a quasi-static PEEC model [9], [17], [18]. The quasi-static PEEC formulation is reasonable for problems with electrically small geometries. When signal waveform rise times decrease and the corresponding frequency content increases, or the geometric dimensions become electrically large, time delays must be taken into account and included in the modeling. A PEEC formulation, which includes delay elements called τ PEEC method [19], becomes necessary and leads to systems of neutral delayed differential equations (NDDE) [16] with constant coefficients and constant delay times in the time domain and to complex algebraic systems of equations with frequencydependent matrices in the frequency domain. Simply using quasi-static PEEC models can result in significant errors and artifacts in the modeling [20]. While several successful MOR 0018-9375/$26.00 © 2011 IEEE 960 IEEE TRANSACTIONS ON ELECTROMAGNETIC COMPATIBILITY, VOL. 53, NO. 4, NOVEMBER 2011 methods for large ODE systems have been proposed over the years, the reduction of large NDDE systems is still a very challenging research topic since standard MOR techniques for ODE systems cannot be directly applied to NDDE systems. Especially, the reduction of electrically large structures, where delays among coupled elements cannot be neglected or easily approximated by rational basis functions needs to be investigated and addressed. Some techniques for the reduction of NDDE systems have been proposed over the years [21]–[26]. In [23], an equivalent first-order system is computed by means of a Taylor expansion, and then MOR Krylov subspace methods [8], [9] are applied. The NDDE formulation is not preserved in the reduction process. In [26], an equivalent first-order system is computed by means of a single-point Taylor expansion [23], and a corresponding orthogonal projection matrix is computed by means of a block Arnoldi algorithm [9]. Then, an orthogonal projection matrix for the original NDDE system is extracted and a reduced NDDE system is obtained. This paper presents a novel MOR method for τ PEEC models that is able to accurately reduce electrically large structures, where delays among coupled elements cannot be neglected or easily approximated by rational basis functions. It is based on an adaptive multipoint expansion and MOR of equivalent firstorder systems [26]. An adaptive algorithm to select the expansion points is presented. The NDDE formulation is preserved in the reduced model. It should be noted that the proposed MOR technique is not bound to the τ PEEC method since it can be applied to NDDE systems obtained by other methods. This paper is organized as follows. Section II describes the modified nodal analysis (MNA) equations of the τ PEEC method. Section III describes the proposed MOR method for NDDE systems. Finally, some pertinent numerical examples based on the τ PEEC method validate the proposed technique in Section IV. II. DELAYED PEEC FORMULATION The PEEC method [4] stems from the integral equation form of Maxwell’s equations. The main difference of the PEEC method with other integralequation-based techniques such as the method of moments [3] resides in the fact that it provides a circuit interpretation of the EFIE [13] in terms of partial elements, namely, resistances, partial inductances, and coefficients of potential. Thus, the resulting equivalent circuit can be studied by means of SPICE-like circuit solvers [15] in both time and frequency domain. Over the years, the PEEC method has been extended to nonorthogonal geometries [19] and modified to incorporate ideal and lossy dielectrics [27], [28]. Following the standard approach [4], volumes and surfaces are discretized into elementary regions, hexahedra and patches, respectively [19]. The current and charge densities are expanded into a series of basis functions, which are usually assumed as pulse basis functions. The choice of pulse basis functions implies to assume constant current and charge densities over the elementary volume (inductive) and surface (capacitive) cells, respectively. Conductors are modeled by their ohmic resistance, and dielectrics by the excess capacitance [27]. Hence, magnetic and electric field coupling are modeled by partial inductances and coefficients of potential, respectively. The magnetic field coupling between two inductive volume cells α and β is modeled by the partial inductance 1 μ 1 duα duβ (1) Lpα β = 4π aα aβ u α u β Rα β where Rα β is the distance between any two points in volumes uα and uβ , with aα and aβ their cross sections. The electric field coupling between two capacitive surface cells γ and δ is modeled by the coefficient of potential 1 1 1 dSγ dSδ (2) Pγ δ = 4πε Sγ Sδ S γ S δ Rγ δ where Rγ δ is the distance between any two points on surfaces γ and δ, while Sγ and Sδ denote the area of their respective surfaces. Due to the finite value of the speed of light, partial inductances and coefficients of potentials relate causes and effects delayed in time diβ (t − τα β ) dt (t) = Pγ δ qδ (t − τγ δ ) vL ,α (t) = Lpα β (3) vP ,γ (4) where τα β = Rα β /c0 and τγ δ = Rγ δ /c0 are the center-tocenter distances between the corresponding basis-function domains, and c0 denotes the free-space speed of light. Hence, partial inductance and coefficient of potential matrices act as a delay operator for time derivatives of currents and charges, respectively, di(t − τ L ) p (τ L ) di(t) =L dt dt (τ C ) q(t) v(t) = Pq(t − τ C ) = P vL (t) = Lp (5a) (5b) where τ L and τ C denote the center-to-center delay matrices for the magnetic and electric field couplings, respectively. Generalized Kirchoff’s laws, for conductors, can be rewritten as follows: dq(t) − AT i(t) + ie (t) = 0 dt p (τ L ) di(t) − Ri(t) = 0 − Av(t) − L dt (6a) (6b) where A is the connectivity matrix, v(t) denotes the node potentials to infinity, and i(t) and ie (t) represent the currents flowing in volume cells and the external currents, respectively. Equation (6b) has to be modified when dielectrics are considered since the resistance R is substituted by the excess capacitance, which relates the polarization charge and the corresponding voltage drop as vd (t) = C−1 d qd (t) [27]. Hence, for FERRANTI et al.: MULTIPOINT FULL-WAVE MODEL ORDER REDUCTION FOR DELAYED PEEC MODELS WITH LARGE DELAYS into account, read as follows: ⎡ 0n n ,n i 0n n ,n d In n ,n n ⎢ p (τ L ) 0n ,n ⎢ 0n i ,n n L i d ⎢ ⎢ 0 0 C ⎣ n d ,n n n d ,n i d 0n p ,n n 0n p ,n i ⎤ 0n n ,n p ⎡ q(t) 0n p ,n p is (t) x(t) C ⎡ dielectric elementary cells, (6) become dq(t) − AT i(t) + ie (t) = 0 dt p (τ L ) di(t) − vd (t) = 0 − Av(t) − L dt dvd (t) . i(t) = Cd dt T ie (t) = K is (t). + (7a) 0n n +n i +n d ,n p −In p ,n p 0n n ,n d KT Φ 0n i ,n p 0n d ,n d 0n d ,n p 0n p ,n d 0n p ,n p ⎤ ⎡ ⎤ q(t) ⎥ ⎢ ⎥ ⎢ i(t) ⎥ ⎥ ⎥·⎢ ⎥ ⎣ v (t) ⎥ ⎦ d ⎦ is (t) x(t) G · [ vp (t) ] (9) u(t) B (7b) (7c) A selection matrix K is introduced to define the port voltages by selecting node potentials. The same matrix is used to obtain the external currents ie (t) by the currents is (t), which are of opposite sign with respect to the port currents ip (t) vp (t) = Kv(t) −AT 0n n ,n n ⎢ (τ C ) R ⎢ AP = −⎢ ⎢ 0 −ΦT ⎣ n d ,n n (τ C ) 0n ,n −KP p i Fig. 1. Illustration of τ PEEC circuit electrical quantities for a conductor elementary cell. ⎤ ⎥ ⎢ i(t) ⎥ 0n i ,n p ⎥ ⎥ ⎥ d ⎢ ⎢ ⎥ ⎥ 0n d ,n p ⎦ dt ⎣ vd (t) ⎦ 0n p ,n d 961 (8a) (8b) An example of τ PEEC circuit for a conductor elementary cell is illustrated, in the Laplace domain, in Fig. 1, where the currentcontrolled voltage sources sLp,ij Ij and the charge-controlled current sources sQi model the magnetic and electric field couplings, respectively. A. Descriptor Representation of τ PEEC Circuits We assume that the system under analysis consists of conductors and dielectrics. Let the current and charge densities be defined in volumes and surface of conductors and dielectrics, respectively. The Galerkin approach is applied to convert the continuous EM problem described by the EFIE to a discrete problem in terms of electrical circuit quantities. Let us denote with nn the number of nodes and ni the number of branches where currents flow. Among the latter, we denote with nc and nd the number of branches of conductors and dielectrics, respectively. Furthermore, let us assume to be interested in generating an admittance representation having np output currents ip (t) under voltage excitation vp (t). Since dielectrics require the excess capacitance to model the polarization charge [27], additional nd unknowns are needed in addition to currents. Hence, if the MNA approach [29] is used, the global number of unknowns is nu = ni + nd + nn + np . In a matrix form, (6)–(8), taking (5) where In p ,n p is the identity matrix of dimensions equal to the number of ports. Matrix Φ is 0n c ,n d Φ= . (10) In d ,n d (τ C ) is applied to the first equation in If the delay operator P (9), the system can be recast as follows: ⎡ ⎤ ⎡ ⎤ (τ C ) 0n ,n P 0n n ,n d 0n n ,n p q(t) n i ⎢ ⎥ ⎢ p (τ L ) 0n ,n ⎢ 0n i ,n n L i(t) ⎥ 0n i ,n p ⎥ ⎥ i d ⎢ ⎥ d ⎢ ⎢ ⎥ ⎢ ⎥ 0n ,n Cd 0n ,n ⎦ dt ⎣ vd (t) ⎦ ⎣ 0n ,n d n d 0n p ,n n i 0n p ,n i d 0n p ,n d p 0n p ,n p C ⎡ 0n n ,n n ⎢ ⎢ AP (τ C ) = −⎢ ⎢ ⎣ 0n d ,n n (τ C ) −KP ⎡ q(t) ⎤ (τ C ) AT −P 0n n ,n d R Φ −Φ T 0n p ,n i 0n d ,n d is (t) x(t) (τ C ) KT ⎤ P ⎥ 0n i ,n p ⎥ ⎥· ⎥ 0n d ,n p ⎦ 0n p ,n d 0n p ,n p G ⎢ i(t) ⎥ 0n +n +n ,n ⎢ ⎥ n i p d ·⎢ · [ vp (t) ] . ⎥+ ⎣ vd (t) ⎦ −In p ,n p u(t) is (t) B (11) x(t) In a more compact form, (11) can be rewritten as follows: C (τ ) dx(t) = −G (τ ) x(t) + Bu(t) dt ip (t) = LT x(t) (12a) (12b) where x(t) = [q(t) i(t) vd (t) is (t)]T ∈ n u ×1 and τ ∈ n τ ×1 contains all delays τ L , τ C . Since this is an np -port formulation, whereby the only sources are the voltage sources at 962 IEEE TRANSACTIONS ON ELECTROMAGNETIC COMPATIBILITY, VOL. 53, NO. 4, NOVEMBER 2011 the np -port nodes, B = L, where B ∈ n u ×n p . Each delayed entry of matrices C (τ ) and G (τ ) act as a delay operator for the corresponding entry of vector x(t). Hence, (12) can be rewritten in the Laplace domain as follows: sC(s)X(s) = −G(s)X(s) + BVp (s) Ip (s) = BT X(s) C(s) = C0 + nτ (13) (14) Ck e−sτ k (15) Gk e−sτ k . (16) k =1 G(s) = G0 + nτ k =1 The solution of (12) can be carried out using direct or iterative solvers. Among the direct methods, the LU decomposition has been widely used [30]. If the transient analysis is carried out for nt time steps, the complexity of the direct solution scales as O(nt n3u ), where nu is the total number of unknowns. Hence, for a large number of unknowns nu , equations (12) cannot be solved directly and iterative solvers are needed. Assuming that an average number niter of iterations per time step is required for the iterative solver to converge to a specified residual, and n2u is the computational cost for the matrix–vector products involved, the iterative solution for all time steps nt scales as O(niter nt n2u ) [31]. It is evident that the MOR becomes fundamental to provide a significant acceleration in time-domain analyses of large EM delayed systems. III. MOR ALGORITHM In [23], an equivalent first-order system is computed by means of a Taylor expansion, and then it is reduced by means of MOR Krylov subspace methods. The NDDE formulation is not preserved. In [26], an equivalent first-order system is computed by means of a single-point Taylor expansion [23], and a corresponding orthogonal projection matrix is computed by means of a block Arnoldi algorithm [9]. Then, an orthogonal projection matrix for the original NDDE system is extracted and a reduced NDDE system is obtained. The NDDE formulation is preserved in the reduction process. The proposed novel MOR algorithm is based on an adaptive multipoint expansion and MOR of equivalent first-order systems. As in [26], the NDDE formulation is preserved. The multipoint expansion feature allows to reduce electrically large structures with large delays (2πfreqm ax τm ax > 10) [25] that cannot be neglected or easily approximated by a single-point expansion and rational functions. The equivalent first-order system obtained after the singlepoint Taylor expansion of exponential terms has an order equal to qnu , where q is the order of the Taylor expansion and nu the order of the original NDDE system [26]. Since exponential terms with large delays need many terms in the Taylor expansion to be accurately approximated, the reduction of equivalent firstorder systems becomes computationally expensive and sometimes unfeasible. The multipoint expansion [32], [33] addresses this issue and is able to accurately reduce NDDE systems with large delays since a small expansion Taylor order can be used for each expansion point and the accuracy of the reduced model is increased by adding new expansion points. An adaptive algorithm is proposed to choose the expansion points, assuming that the order of the Taylor expansion is fixed for each expansion point. A. Single-Point Expansion Reduction Performing the MOR on the NDDE system (13) and (14) requires the computation of an orthogonal basis for the first nn pr block moments of X(s), i.e., its first nn pr scaled derivatives w.r.t. s, where nr represents the order of the reduced system. Denoting the orthogonal basis as Q ∈ n u ×n r , the reducedorder system is given by sCr (s)χ(s) = −Gr (s)χ(s) + Br Vp (s) Ip (s) = BTr (17) χ(s) Cr (s) = Cr,0 + (18) nτ Cr,k e−sτ k (19) Gr,k e−sτ k (20) k =1 Gr (s) = Gr,0 + nτ k =1 where the following congruence transformations are used: Cr,i = QT Ci Q, i = 0, . . . , nτ (21a) Gr,i = QT Gi Q, i = 0, . . . , nτ (21b) T Br = Q B (21c) Lr = QT L (21d) and χ(t) is a vector containing the state variables in the reduced domain. The reduced-order system (17) and (18) can be efficiently incorporated into SPICE-like simulators for efficient and accurate transient analysis. The construction of an orthogonal basis Q represents a key issue in obtaining a compact and accurate reduced-order model over a wide frequency range. The Arnoldi algorithm [9] is adopted in this paper to carry out the computation of Q since it is numerically reliable and robust. The numerical robustness of the Arnoldi algorithm stems from its capability to obtain the orthogonal basis for the transfer function moments without computing these moments explicitly. Unfortunately, adapting the Arnoldi algorithm for NDDE systems is not a straightforward task, and therefore, the original NDDE system has to be transformed in a suitable form to apply standard Arnoldi-based reduction. To address this issue, an equivalent first-order system is computed by expanding the exponential factors e−sτ k in a Taylor series form and using a companion form [26]. Using the Arnoldi algorithm on the first-order equivalent system, a corresponding orthogonal projection matrix is computed. Then, an orthogonal projection matrix Q for the original NDDE system is extracted and a reduced NDDE system is obtained [26]. This single-point expansion-based MOR algorithm for NDDE systems was proposed assuming s = 0 as an expansion point. If another expansion point s = s1 , s1 = 0, is selected, setting s = s1 + σ, where s1 is a frequency shift and σ is the new Laplace variable, the NDDE system (13), (14) FERRANTI et al.: MULTIPOINT FULL-WAVE MODEL ORDER REDUCTION FOR DELAYED PEEC MODELS WITH LARGE DELAYS 963 becomes σ C(σ)X(σ) = −G(σ)X(σ) + BVp (σ) T Ip (σ) = B X(σ) 0 + C(σ) =C nτ (22) (23) k e−σ τ k C (24) k e−σ τ k G (25) k =1 0 + G(σ) =G nτ k =1 Fig. 2. Pseudocode for computing the orthogonal basis Q in the multipoint expansion case. where 0 = C0 C (26a) 0 = G0 + s1 C0 G (26b) k = Ck e−s 1 τ k , C k = 1, . . . , nτ k = (Gk + s1 Ck )e−s 1 τ k , G k = 1, . . . , nτ (26c) (26d) and the algorithm described in [26] can be applied. Although both real and imaginary shiftings have been investigated in literature [33], we focus on the minimization of the frequencyresponse error between reduced and full-order NDDE systems, and therefore, imaginary expansion points are considered. This single-point MOR approach [26] is able to preserve the NDDE formulation, but it may be not able to reduce NDDE systems with large delays since the reduction of equivalent first-order systems becomes computationally expensive and sometimes unfeasible. B. Multipoint Expansion Reduction The equivalent first-order system obtained after the singlepoint Taylor expansion of exponential terms has an order equal to qnu , where q is the order of the Taylor expansion and nu the order of the original NDDE system [26]. The reduction of equivalent first-order systems becomes computationally expensive and sometimes not feasible when large delays are involved since exponential terms with large delays need many terms in the Taylor expansion to be accurately approximated. The multipoint expansion [32] addresses this issue and is able to accurately reduce NDDE systems with large delays since a small expansion Taylor order can be used for each expansion point and the accuracy of the reduced model is increased by adding new expansion points. Assuming that the order of the Taylor expansion is fixed for each expansion point, an adaptive algorithm is used to choose the expansion points. As in [26], the NDDE formulation is preserved in the reduced model. At each expansion point, the MOR algorithm described in [26] is applied and the corresponding projection matrix Qi , i = 1, . . . , np oints is computed, where np oints denotes the number of expansion points. The final projection matrix Q is based on the orthogonalization of the stack column collection of all single expansion point projection matrices. The computation of Q in the case of multipoint expansion is explained in the flowchart in Fig. 2. The MOR algorithm described in [26] is called one point DM OR in Fig. 2, where nr represents the reduced order for each expansion point. Once Q is computed, it is applied to the original NDDE system (13), (14) and a reduced NDDE system (17), (18) is obtained. C. Algorithm for the Selection of Expansion Points Assuming that the order of the Taylor expansion q is fixed for each expansion point, an algorithm is needed to choose the location of the expansion points. An adaptive and iterative algorithm is proposed to determine the expansion points. It is based on an iterative comparison between reduced and original model. It starts from two expansion points sm in = jωm in and sm ax = jωm ax located at the minimum and maximum frequency of interest, respectively, and the corresponding reduced model is compared with the original model in the midpoint between sm in and s, , therefore sm id,1 = (sm in + sm ax )/2. If an error threshold is satisfied, then the algorithm stops, otherwise this midpoint is considered as a new expansion point and a new reduced model is computed, which contains information from all the expansion points sm in , sm id,1 , and sm ax , and it is compared with the original model in two new points, namely, the midpoints of the intervals [sm in , sm id,1 ] and [sm id,1 , sm ax ]. If no comparison point exceeds the error threshold, then the algorithm stops, otherwise the new expansion points for the next iteration are chosen as the comparison points that exceed the error threshold. If some comparison points satisfy the error threshold, the corresponding intervals defined by expansion points are considered to be accurately described and are not checked for accuracy in the next iterations. The comparison points at each iteration are the midpoints of all intervals defined by expansion points, except those ones considered accurate in the previous iterations. Concerning the error criterion, let us define the weighted rms error as follows: Err = (n p ) 2 K s i=1 k =1 |wY i (sk )(Yr,i (sk ) − Yi (sk ))|2 (27) (np )2 Ks with wY i (s) = |(Yi (s))−1 |. (28) In the comparison step at each iteration, the error is computed on each comparison point separately; therefore, Ks = 1 and the accuracy threshold is chosen equal to 0.05. 964 Fig. 3. IEEE TRANSACTIONS ON ELECTROMAGNETIC COMPATIBILITY, VOL. 53, NO. 4, NOVEMBER 2011 Cross section of the multiconductor system. TABLE I MOR PARAMETERS Fig. 4. Magnitude of Y 1 1 . Fig. 5. Phase of Y 1 1 . Fig. 6. Magnitude of Y 1 6 . IV. NUMERICAL RESULTS Two numerical examples based on τ PEEC systems validate the proposed MOR approach for NDDE systems. The error defined in (27) is used to assess the accuracy of reduced models over a sampled bandwidth of interest. The proposed MOR algorithm was implemented in MATLAB R2009A [34] and all experiments were carried out on Windows platform equipped with Intel Core2 Extreme CPU Q9300 2.53 GHz and 8GB RAM. A. Multiconductor System A multiconductor system composed of six conductors with a length = 7 cm, a width W = 1 mm, a thickness t = 0.25 mm, and a horizontal Sx = 3 mm and a vertical Sy = 2 mm spacing has been modeled in this example. Fig. 3 shows its cross section. The six ports of the system are defined between a conductor and the corresponding one above. The system is analyzed on the frequency range [0.0001, 20] GHz. The order of the original τ PEEC model is equal to nu = 3366, the number of delays is equal to 664, and the largest delay gives 2πfreqm ax τm ax = 29.5. Table I shows some parameters of the proposed MOR method using the adaptive algorithm for the selection of the expansion points. Figs. 4–7 compare the magnitude and phase of the original and reduced NDDE model of Y11 (s) and Y16 (s). As clearly seen, the presented MOR method can reduce a large NDDE system with large delay terms, while preserving the behavior of the original system. The proposed adaptive algorithm for the selection of the expansion points is able to converge, producing a compact and accurate reduced-order model. B. Three-Port Microstrip Power-Divider Circuit A three-port microstrip power-divider circuit [35] has been modeled in this example. The structure is shown in Fig. 8. The dimensions of the circuit are [20, 20, 0.5] mm in the [x, y, z] directions and the width of the microstrips is equal to 0.8 mm. The relative dielectric constant is equal to r = 2.2. The system is analyzed on the frequency range [0.0001, 20] GHz. The order of the original τ PEEC model is equal to nu = 4149, the number of delays is equal to 692, and the largest delay gives 2πfreqm ax τm ax = 12. Some parameters of the proposed MOR method are shown in Table II. FERRANTI et al.: MULTIPOINT FULL-WAVE MODEL ORDER REDUCTION FOR DELAYED PEEC MODELS WITH LARGE DELAYS Fig. 7. Phase of Y 1 6 . Fig. 8. Structure of the three-port microstrip power-divider circuit. Fig. 9. Fig. 10. Phase of Y 1 1 . Fig. 11. Magnitude of Y 1 3 . Fig. 12. Phase of Y 1 3 . 965 Magnitude of Y 1 1 . TABLE II MOR PARAMETERS V. CONCLUSION Figs. 9–12 compare the magnitude and phase of the original and reduced NDDE model of Y11 (s) and Y13 (s). As in the previous example, the proposed MOR method is able to accurately reduce an NDDE system with large delay terms using an adaptive algorithm for the selection of the expansion points. We have presented a new MOR technique for large NDDE systems with large delays, which is applicable to τ PEEC models. It is able to accurately reduce electrically large structures, where delays among coupled elements cannot be neglected or easily approximated by rational basis functions. It is based on an adaptive multipoint expansion and MOR of equivalent firstorder systems. An adaptive algorithm to select the expansion points is discussed. The NDDE formulation is preserved in the 966 IEEE TRANSACTIONS ON ELECTROMAGNETIC COMPATIBILITY, VOL. 53, NO. 4, NOVEMBER 2011 reduced model. Numerical examples based on τ PEEC models have validated the proposed MOR approach, showing that it is able to accurately reduce large NDDE systems with large delays. REFERENCES [1] R. Achar and M. Nakhla, “Simulation of high-speed interconnects,” Proc. IEEE, vol. 89, no. 5, pp. 693–728, May 2001. [2] A. E. Ruehli and A. C. Cangellaris, “Progress in the methodologies for the electrical modeling of interconnects and electronic packages,” Proc. IEEE, vol. 89, no. 5, pp. 740–771, May 2001. [3] R. F. Harrington, Field Computation by Moment Methods. New York: Macmillan, 1968. [4] A. E. Ruehli, “Equivalent circuit models for three dimensional multiconductor systems,” IEEE Trans. Microw. Theory Tech., vol. 22, no. 3, pp. 216–221, Mar. 1974. [5] J. M. Jin, The Finite Element Method in Electromagnetics, 2nd ed. New York: Wiley, 2002. [6] L. T. Pillage and R. A. Rohrer, “Asymptotic waveform evaluation for timing analysis,” IEEE Trans. Comput.-Aided Design Integr. Circuits Syst., vol. 9, no. 4, pp. 352–366, Apr. 1990. [7] K. Gallivan, E. Grimme, and P. Van Dooren, “Asymptotic waveform evaluation via a Lanczos method,” Appl. Math., vol. 7, no. 5, pp. 75–80, Sep. 1994. [8] P. Feldmann and R. W. Freund, “Efficient linear circuit analysis by Padé approximation via the Lanczos process,” IEEE Trans. Comput.-Aided Design Integr. Circuits Syst., vol. 14, no. 5, pp. 639–649, May 1995. [9] A. Odabasioglu, M. Celik, and L. T. Pileggi, “PRIMA: Passive reducedorder interconnect macromodeling algorithm,” IEEE Trans. Comput.Aided Design Integr. Circuits Syst., vol. 17, no. 8, pp. 645–654, Aug. 1998. [10] A. Dounavis, E. Gad, R. Achar, and M. S. Nakhla, “Passive model reduction of multiport distributed interconnects,” IEEE Trans. Microw. Theory Tech., vol. 48, no. 12, pp. 2325–2334, Dec. 2000. [11] B. Denecker, F. Olyslager, L. Knockaert, and D. De Zutter, “Generation of FDTD subcell equations by means of reduced order modeling,” IEEE Trans. Antennas Propag., vol. 51, no. 8, pp. 1806–1817, Aug. 2003. [12] N. A. Marques, M. Kamon, L. M. Silveira, and J. K. White, “Generating compact, guaranteed passive reduced-order models of 3-D RLC interconnects,” IEEE Trans. Adv. Packag., vol. 27, no. 4, pp. 569–580, Nov. 2004. [13] C. A. Balanis, Advanced Engineering Electromagnetics. New York: Wiley, 1989. [14] W. Pinello, A. C. Cangellaris, and A. Ruehli, “Hybrid electromagnetic modeling of noise interactions in packaged electronics based on the partialelement equivalent-circuit formulation,” IEEE Trans. Microw. Theory Tech., vol. 45, no. 10, pp. 1889–1896, Oct. 1997. [15] L. W. Nagel, “SPICE: A computer program to simulate semiconductor circuits,” University of California, Berkeley, Electr. Res. Lab. Report ERL M520, May 1975. [16] P. J. Restle, A. Ruehli, S. G. Walker, and G. Papadopoulos, “Full-wave PEEC time-domain for the modeling of on-chip interconnects,” IEEE Trans. Computer-Aided Design, vol. 20, no. 7, pp. 877–887, Jul. 2001. [17] R. D. Slone, W. T. Smith, and Z. Bai, “Using partial element equivalent circuit full wave analysis and Padé via Lanczos to numerically simulate EMC problems,” in Proc. IEEE Int. Symp. Electromagn. Compat., Austin, TX, Aug. 1997, pp. 608–613. [18] L. Knockaert and D. De Zutter, “Laguerre-SVD reduced-order modeling,” IEEE Trans. Microw. Theory Tech., vol. 48, no. 9, pp. 1469–1475, Sep. 2000. [19] A. E. Ruehli, G. Antonini, J. Esch, J. Ekman, A. Mayo, and A. Orlandi, “Non-orthogonal PEEC formulation for time and frequency domain EM and circuit modeling,” IEEE Trans. Electromagn. Compat., vol. 45, no. 2, pp. 167–176, May 2003. [20] G. Antonini and J. Ekman, “On characterizing artifacts observed in PEEC based modeling,” presented at the IEEE Int. Symp. Electromagn. Compat., Santa Clara, CA, Aug. 2004. [21] H. Heeb, A. E. Ruehli, J. E. Bracken, and R. A. Rohrer, “Three dimensional circuit oriented electromagnetic modeling for VLSI interconnects,” in Proc. IEEE Int. Conf., Comput. Design: VLSI Comput. Processors, Oct. 1992, pp. 218–221. [22] E. Chiprout, H. Heeb, M. Nakhla, and A. Ruehli, “Simulating 3-D retarded interconnect models using complex frequency hopping (CFH),” in IEEE/ACM Int. Conf. Computer-Aided Des. Tech. Dig., Nov. 1993, pp. 66–72. [23] J. R. Phillips, E. Chiprout, and D. D. Ling, “Efficient full-wave electromagnetic analysis via model-order reduction of fast integral transforms,” in Proc. 33rd Annu. Des. Autom. Conf., New York, NY, 1996, pp. 377–382. [24] J. Cullum, A. Ruehli, and T. Zhang, “A method for reduced-order modeling and simulation of large interconnect circuits and its application to PEEC models with retardation,” IEEE Trans. Circuits Syst. II, vol. 47, no. 4, pp. 261–373, Apr. 2000. [25] T. Klemas, L. Daniel, and J. White, “Segregation by primary phase factors: A full-wave algorithm for model order reduction,” in Proc. 42nd Des. Autom. Conf., Jun., 2005, pp. 943–946. [26] W. Tseng, C. Chen, E. Gad, M. Nakhla, and R. Achar, “Passive order reduction for RLC circuits with delay elements,” IEEE Trans. Adv. Packag., vol. 30, no. 4, pp. 830–840, Nov. 2007. [27] A. E. Ruehli and H. Heeb, “Circuit models for three-dimensional geometries including dielectrics,” IEEE Trans. Microw. Theory Tech., vol. 40, no. 7, pp. 1507–1516, Jul. 1992. [28] G. Antonini, A. Ruehli, and C. Yang, “PEEC modeling of dispersive and lossy dielectrics,” IEEE Trans. Adv. Packag., vol. 31, no. 4, pp. 768–782, Nov. 2008. [29] C. Ho, A. Ruehli, and P. Brennan, “The modified nodal approach to network analysis,” IEEE Trans. Circuits Syst., vol. 22, no. 6, pp. 504– 509, Jun. 1975. [30] E. Anderson, Z. Bai, C. Bischof, S. Blackford, J. Demmel, J. Dongarra, J. Du Croz, A. Greenbaum, S. Hammarling, A. McKenney, and D. Sorensen, LAPACK Users’ Guide, 3rd ed. Philadelphia, PA: SIAM, 1999. [31] W. C. Chew, J.-M. Jin, E. Michielssen, and J. Song, Fast and Efficient Algorithms in Computational Electromagnetics, A. House, Ed., Norwood, MA: Artech House, 2001. [32] I. Elfadel and D. Ling, “A block rational Arnoldi algorithm for multipoint passive model-order reduction of multiport RLC networks,” in Proc. IEEE/ACM Int. Conf. Computer-Aided Des., Nov. 1997, pp. 66–71. [33] E. J. Grimme, “Krylov projection methods for model reduction,” Ph.D. thesis, Univ. Illinois Urbana-Champaign, Champaign, Illinois, 1997. [34] MATLAB User’s Guide, The Mathworks, Inc., Natick, 2009. [35] R. Wang and J.-M. Jin, “A flexible time-stepping scheme for hybrid field-circuit simulation based on the extended time-domain finite element method,” IEEE Trans. Adv. Packag., vol. 33, no. 4, pp. 769–776, Nov. 2010. Francesco Ferranti (M’10) received the B.S. degree (summa cum laude) in electronic engineering from the Università degli Studi di Palermo, Palermo, Italy, in 2005, and the M.S. degree (summa cum laude and honors) in electronic engineering from the Università degli Studi dell’Aquila, L’Aquila, Italy, in 2007, and the Ph.D. degree in electrical engineering from the University of Ghent, Ghent, Belgium, in 2011. He is currently a Post-Doctoral Research Fellow with the Department of Information Technology (INTEC), Ghent University, Ghent, Belgium. His research interests include parametric macromodeling, parameterized model order reduction, electromagnetic compatibility numerical modeling, and system identification. Michel S. Nakhla (S’73–M’75–SM’88–F’98) received the Ph.D. degree in electrical engineering from the University of Waterloo, Waterloo, ON, Canada, in 1975. From 1976–1988, he was a Senior Manager of the computer-aided engineering group at Bell-Northern Research, Ottawa, ON. In 1988, he joined Carleton University, Ottawa, Canada, as a Professor and the holder of the Computer-Aided Engineering Senior Industrial Chair established by Bell-Northern Research and the Natural Sciences and Engineering Research Council of Canada, where he currently a Chancellor’s Professor of Electrical Engineering and the founder of the high-speed CAD research group. His research interests include modeling and simulation of high-speed circuits and interconnects, nonlinear circuits, parallel processing, multidisciplinary optimization, and neural networks. Dr. Nakhla is on various international committees, including the standing committee of the IEEE International Signal Propagation on Interconnects Workshop, the technical program committee of the IEEE International Microwave Symposium, the technical program committee of the IEEE Conference on FERRANTI et al.: MULTIPOINT FULL-WAVE MODEL ORDER REDUCTION FOR DELAYED PEEC MODELS WITH LARGE DELAYS Electrical Performance of Electronic Packaging, and the CAD committee of the IEEE Microwave Theory and Techniques Society (MTT-1). He is an Associate Editor of the IEEE TRANSACTIONS ON ADVANCED PACKAGING. He was as Associate Editor of the IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS. He has also been a member of many Canadian and international government-sponsored research grants selection panels He is a technical consultant for several industrial organizations and is the principal investigator for several major sponsored research projects. Giulio Antonini (M’94–SM’05) received the Laurea degree (summa cum laude) in electrical engineering from the Università degli Studi dell’Aquila, L’Aquila, Italy, in 1994, and the Ph.D. degree in electrical engineering from the University of Rome “La Sapienza,” Rome, Italy, in 1998. Since 1998, he has been with the UAq Electromagnetic Compatibility (EMC) Laboratory, Department of Electrical Engineering, University of L’Aquila, where he is currently an Associate Professor. He is the author or coauthor of more than 170 technical papers and 2 book chapters, and holds one European patent. His research interests include EMC analysis, numerical modeling, and signal integrity for high-speed digital systems. Dr. Antonini has given keynote lectures and chaired several special sessions at international conferences. He has been the recipient of the IEEE Transactions on Electromagnetic Compatibility Best Paper Award in 1997, the Computer Simulation Technology University Publication Award in 2004, the IBM Shared University Research Award in 2004, 2005, and 2006, The Institution of Engineering and Technology—Science, Measurement & Technology Best Paper Award in 2008, and a Technical Achievement Award from the IEEE EMC Society “for innovative contributions to computational electromagnetic on the Partial Element Equivalent Circuit technique for EMC applications” in 2006. Tom Dhaene (SM’05) was born in Deinze, Belgium, on June 25, 1966. He received the Ph.D. degree in electrotechnical engineering from the Ghent University, Ghent, Belgium, in 1993. From 1989 to 1993, he was a Research Assistant in the Department of Information Technology, Ghent University, where he was engaged in research on different aspects of full-wave electromagnetic (EM) circuit modeling, transient simulation, and time-domain characterization of high-frequency and high-speed interconnections. In 1993, he joined the EDA company Alphabit (now part of Agilent). He was one of the key developers of the planar EM simulator ADS Momentum. Since September 2000, he has been a Professor in the Department of Mathematics and Computer Science, University of Antwerp, Antwerp, Belgium, and since October 2007, a Full Professor in the Department of Information Technology (INTEC), Ghent University. He is the author or coauthor of more than 150 peer-reviewed papers and abstracts in international conference proceedings, journals, and books, and is the holder of 3 US patents. 967 Luc Knockaert (SM’00) received the M.Sc. degree in physical engineering, the M.Sc. degree in telecommunications engineering, and the Ph.D. degree in electrical engineering from Ghent University, Ghent, Belgium, in 1974, 1977, and 1987, respectively. From 1979 to 1984 and from 1988 to 1995, he was at the Universities of the Democratic Republic of the Congo and Burundi, where he was involved in North–South cooperation and development projects. He is currently with the Interdisciplinary Institute for BroadBand Technology, Ghent University, where he is also a Professor in the Department of Information Technology. He is the author or coauthor of more than 100 international journal and conference publications. His current research interests include the application of linear algebra and adaptive methods in signal estimation, model order reduction, and computational electromagnetics. Dr. Knockaert is a member of the Mathematical Association of America and the Society for Industrial and Applied Mathematics. Albert E. Ruehli (LF’03) received the Ph.D. degree in electrical engineering from the University of Vermont, Burlington, VT, Canada, in 1972, and an honorary doctorate from the Lulea University, Lulea, Sweden, in 2007. He has been a Manager of a very large scale integration (VLSI) design and CAD group, and a member of various projects with IBM, including mathematical analysis, semiconductor circuits, and devices modeling. Since 1972, he has been at IBM T. J. Watson Research Center, Yorktown Heights, NY, where he was a Research Staff Member in the Electromagnetic Analysis Group, and is currently an Emeritus of IBM Research. He is also an Adjunct Professor in the electromagnetic compatibility (EMC) area at the Missouri University of Science and Technology, Rolla. He is the author or coauthor of more than 190 technical papers, and is the editor of two books Circuit Analysis, Simulation and Design (New York: North Holland, 1986 and 1987). Dr. Ruehli has served in numerous capacities for the IEEE. In 1984 and 1985, he was the Technical and General Chairman, respectively, of the ICCD International Conference. He has been a member of the IEEE Administrative Committee for the Circuit and System (CAS) Society. He has been an Associate Editor for the IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN. He has given talks at universities including keynote addresses and tutorials at conferences, and has organized many sessions. He received IBM Research Division or IBM Outstanding Contribution Awards in 1975, 1978, 1982, 1995, and 2000. In 1982, he received the Guillemin-Cauer Prize Award for his study on waveform relaxation, and in 1999, he received a Golden Jubilee Medal, both from the IEEE CAS Society. In 2001, he received a Certificate of Achievement from the IEEE EMC Society for Inductance Concepts and the Partial Element Equivalent Circuit method. He received the 2005 Richard R Stoddart Award, and in 2007, he received the Honorary Life Member Award from the IEEE EMC Society for outstanding technical performance. He is a member of the Society for Industrial and Applied Mathematics.