There were over 25 requests for my longer summary of the "(no subject)" thread about s-parameter simulation in SPICE and about dispersion, so here it is! Happy reading! -matt Matt Herndon, PhD ECAD Group Apple Computer Using s-parameters in SPICE, and comparison of dispersion and loss. ==================================================================== Summary of SI-list thread from December 2002, with subject "(no subject)". by Matthew K. Herndon, PhD ECAD Group Apple Computer, Inc., Cupertino, CA The original question was submitted by Mick Zhou of Agere on Dec 6,2002 [1]: Has "anybody successfully generate broadband (DC-20GHz) SPICE model from S matrix for complex structures using the SPICE generator in ADS?" Responses indicated that the following additional simulators have this capability; there was various discussion of accuracy, speed, ease of use, bugs, etc. PLEASE NOTE THAT THIS LIST IS NOT INTENDED TO BE COMPLETE --- IT CONTAINS ONLY THOSE SIMULATORS WHICH WRERE DISCUSSED IN THE THREAD OR WERE MADE KNOWN TO ME; I DID NOT DO A PRO-ACTIVE SEARCH. Appologies for anything missing: Fullwave SPICE (Ansoft) IConnect (TDA Systems) Nspice (Apache) Broadband Spice (Sigrity) Apsim (Applied Simulation Technology) Spectre (Cadence) ELDO/ADMS (Mentor, formerly Anacad) Micro-Cap (Spectrum Software, requires easy conversion to Y parameters) Some circuit simulators do not support frequency-dependent sources. ADS does not, HSPICE and PSPICE do. [R6] gives a nice example of the process for simulating with s-parmeters in PSPICE. Aside: "Touchstone compatible" is an ASCII format for S-parameter data, based on Agilent EEsof. For more information, type "Touchstone s-parameter format" in Google; or see, for example, [R1]. At this point, we split into two sub-threads: generating SPICE models without frequency-dependent lookup tables, and the relationship between dispersion and loss. Spice models without frequency-dependent sources: ================================================= In theory, one can create accurate lumped-element (RLC) models of measured s-parameters. In practice, one needs to overcome a few difficulties, especially for broadband simulation and for long lines. In particular, it is hard to accurately model complicated frequency dependencies and dispersion [6]. But this approach does work well in some situations; and converges for the same reason that the Fourier transform converges [8] (Note: I would like to know more about this analogy). There are two approaches to simulating s-parameters in SPICE-like simulators (items (1) and (2) here quote and summarize [26]): (1) Convert S-parameters to certain forms, either equivalent circuit representations or certain table lookup format, from which SPICE engines can read and run. (2) Enable a SPICE solver to read S parameters directly. The SPICE solver will then internally do the things in (1), or do convolution directly which can be quite demanding for computer resources for large number of such circuit components. Often the original S-parameters or the circuit model representing them may not be stable, causal, and passive. Also, extrapolation of the S-parameter data to DC is often a problem and separate DC values may be needed. Lumped-element models do have some advantages [17]: frequency-dependent capacitance and inductance are not needed, and we don't have to know anything about the dielectric properties other than what is implicitly contained in the measurements. No assumptions are made about TEM or quasi-TEM propagation. The only assumption is a linear, time-invariant system. In principle, a linear, time-invariant system can certainly be accurately represented by lumped elements to within the accuracy of the measurements themselves, and up to the highest frequency at which the measurements were taken. It takes a lot of lumped elements to (perfectly) match the measured s-parameters. One calculation for a connector came up with 400K elements. So "curve fitting" approximations are used, analogous to (but different from) lossy data compression algorithms like JPEG. For example, a tenth order rational function would require only 40 (?) elements (10 for the numerator, 10 for the denominator, not sure where the other factor of 2 comes from). Questions then arise about how good the approximations are. How high an order of rational function do you need? Clearly it depends on the complexity of the shape of the s-parameters curve. And if you happen to drive the circuit at frequencies where the curve-fit is poor, you can get erroneous results (imagine a single sinusoid at a point where the curves are far apart). One can check the accuracy of the reduced lumped model by running a time-domain simulation, taking the Fourier transform, and comparing to the s-parameters. Of course, you can only do this for the linear parts of the circuit. As [53] says, this "check is only the first step towards a more complex simulation with on-chip and off-chip elements. If we fail the simple check, we cannot trust the results of the more complex simulation." Dispersion and loss =================== These terms are sometimes used interchangeably, which is confusing and should not be done; they are two different concepts. Dispersion means that the velocity of spectral components of a signal depend on frequency. This effect is not the same as loss, either skin effect or dielectric loss. Dispersion means different frequencies travel at different velocities, effectively scattering or "dispersing" the arrival times of the signal wavefront [14]. Dispersion happens when Er varies with frequency (so therefore velocity varies with frequency), as in a prism. You can have dispersion without loss e.g., a rectangular wave guide (RWG). So conceptually the two are not related: dispersion is related to phase velocity, loss is related to energy [16]. However, you cannot have frequency-dependent loss without dispersion. So, for example, the skin effect will cause dispersion. Further, if the dispersion is caused by the dielectric material (the most common situation in SI), there must also be loss; this property follows from the Kramers-Kronig relationship [R7] (aka causality). Kramers-Kronig states that the real and imaginary parts of the dielectric constant are related to each other [15]. The real part determines velocity, the imaginary part the loss (due to dampening of the vibrating dipole moments). So in this sense, you can say that if a MEDIUM is "dispersive" it must also be "lossy". Microwave engineering uses a number of models that satisfy the Kramers-Kronig relation. One example is the Debye model, which was originally developed in the area of studies of the relaxation of crystal lattices. It turns out that the physical mechanism of loss is quite similar to the mechanics and physics of these lattice structures. There are also other models of dielectric permitivity and loss that are also physically consistent. All of them require what we call loss tangent and permitivity (which gives rise to Er) to vary somewhat across frequency. [49] [40] FR4 is non-isotropic and has significant changes in Er across frequency ([R3] says 4.2 to 4.0 over the range 0.5 to 5 GHz). The dependence of Er on frequency is not a simple, intuitive function. Real air exhibits some change in Er with frequency due to humidity; pure (dry?) air does have a flat response and extremely low losses. Likewise some other materials, e.g., Rogers and some PTFE (teflon) and ceramic materials (see [R4] and [R5]). The telegraphers equations (quasi-TEM mode) assume a single, homogeneous, isotropic material. Dispersion will (always?) occur if the medium non-homogeneous, e.g., microstrip. If you neglect the losses (i.e., assume ideal materials), stripline and coax will be non-dispersive. There was some discussion of whether dispersive media could actually SHARPEN an edge. Possible in theory, but wouldn't happen in practice (would it violate Kramers-Kronig?). There are devices which sharpen rising edges (at the expense of lengthening falling edges) by varying propagation velocity with amplitude. These devices construct lumped-approximations to non-linear transmission lines. See, for example, [R2]. In reality, manufacturing tolerances and non-uniformity of materials causes much larger variations in actual transmission line performance than the change in Er across frequency. However, that does not mean that such an effect does not exist or does not provide interesting insights into the workings of E&M phenomena. References =========== [R1] http://www.sigrity.com/download/pub/Touchstone.pdf Touchstone format. [R2] http://www.picosecond.com/objects/AN-13.pdf (edge compressors) [R3] Dielectric and Magnetic Properties of Printed Wiring Boards and Other Substrate Material Baker-Jarvis J., Riddle B, and Janezic M. NIST Technical Note 1512. 1999 [R4] http://www.rogers-corp.com/mwu/techtip9.htm Er vs. frequency. [R5] http://www.dupont.com/kapton/general/spfreq.html Er vs. frequency. [R6] http://wwwinfo.cern.ch/ce/ae/Maxwell/apps/2stripem/2stripem.html Discussion of obtaining s-parameters, using s2spice to create an equivalent circuit, and simulating in PSPICE. [R7] http://www.math.utah.edu/~eyre/research/kk/node1.html Short discussion of Kramers-Kronig. Following are postings to the SI list for this thread. Please note that I tried to provide a citation anytime I quoted more than a sentence from a posting. Appologies if I missed someone. [1] From: "Zhou, Xingling (Mick)" <xlzhou@xxxxxxxxx> Date: Fri Dec 6, 2002 11:42:41 AM US/Pacific [5] From: "Zhou, Xingling (Mick)" <xlzhou@xxxxxxxxx> Date: Fri Dec 6, 2002 6:19:42 PM US/Pacific [6] From: Yu Liu <yu_liu@xxxxxxxxxxxxx> Date: Fri Dec 6, 2002 9:31:32 PM US/Pacific [8] From: Steve Corey <steven.corey@xxxxxxxxxxxxxx> Date: Mon Dec 9, 2002 11:05:10 AM US/Pacific [9] From: "Loyer, Jeff" <jeff.loyer@xxxxxxxxx> Date: Mon Dec 9, 2002 11:30:45 AM US/Pacific [15] From: Steve Corey <steven.corey@xxxxxxxxxxxxxx> Date: Mon Dec 9, 2002 4:44:28 PM US/Pacific [16] From: Xin Wu <lifehappiest@xxxxxxxxxxx> Date: Tue Dec 10, 2002 6:54:00 AM US/Pacific [26] From: Raj Raghuram <raghu@xxxxxxxxxxx> Date: Tue Dec 10, 2002 3:20:06 PM US/Pacific [40] From: Scott McMorrow <scott@xxxxxxxxxxxxx> Date: Wed Dec 11, 2002 10:45:28 AM US/Pacific [49] From: Scott McMorrow <scott@xxxxxxxxxxxxx> Date: Wed Dec 11, 2002 2:43:33 PM US/Pacific Other interesting refences: "Fields and Waves in Communications Electronics" by Ramo, Whinnery, and Van Duzer? ------------------------------------------------------------------ To unsubscribe from si-list: si-list-request@xxxxxxxxxxxxx with 'unsubscribe' in the Subject field or to administer your membership from a web page, go to: //www.freelists.org/webpage/si-list For help: si-list-request@xxxxxxxxxxxxx with 'help' in the Subject field List archives are viewable at: //www.freelists.org/archives/si-list or at our remote archives: http://groups.yahoo.com/group/si-list/messages Old (prior to June 6, 2001) list archives are viewable at: http://www.qsl.net/wb6tpu