Beruflich Dokumente
Kultur Dokumente
Cascade Microtech, Inc., 2430 NW 206 th Avenue, Beaverton, OR, 97006, USA, leonard@cmicro.com
Abstract - This paper examines the extended 12-term error model commonly used in commercial multiport vector network
analyzers, introduces a generalized multiport error model, and applies this error model for the purposes of general N-port comparison
of calibrations. These tools have been implemented in a commercially available calibration and measurement software product [I].
Previous work demonstrated the utility of these tools in the estimation of calibration error associated with ignoring coupling [2] and for
evaluating measurement system repeatability [3]. Equations are presented for bidirectional conversion between an extended 16-term-
like error model and the extended 12-term model as well as for calculation of DUT-specific and worst-case multiport calibration
comparison error bounds.
Index Terms - Vector Network Analyzer, VNA, Calibration, Multiport,Accuracy, Calibration Comparison, Extended 16-Term.
I. INTRODUCTION
Commercial makers of multiport vector network analyzer systems have recently introduced stand-alone systems capable of
dramatically higher maximum frequencies , from a previous high of 20 GHz to the current maximum of 67 GHz [4]. These
systems normally include software features to convert native single-ended measurement data to the common/differential mixed-
mode representation [5] useful for the characterization of balanced circuits and devices. Commercially available systems rely on
calibration to single-ended (non-modal) reference planes using a simple extrapolation of the 12-term error model traditionally
used for two-port systems. The extended 12-term error model is shown in Fig. l(a).
While expedient for makers of multiport analyzers , this extended 12-term error model does not allow for any fixture or other
DUT-dependent coupling . Simplistic isolation terms are available but are only suitable for modeling DUT independent signal
leakage typically found only as leakage paths internal to the VNA (and also normally small).
A 16-term error model for two-port network analyzers has been studied for many years [e.g., 6] and practical methods for
calibration have been developed [e.g., 7]. Historically , these methods have been of limited utility for probing applications since in
many situations the crosstalk related error terms are not independent of the structure being probed. Recently , however, improved
transistor model extraction using a 16-term calibration approach has been demonstrated [8]. Coupled error models have also been
explored for use with multiport analyzers [e.g., 9].
Calibration comparison [10] provides a method for evaluating the measurement differences associated with the differences
between two error models and is a useful tool with many uses such as evaluating system drift and comparing calibration
algorithms and their sensitivity . Using the 16-term error model the calibration comparison process becomes conceptually simple
and the extension to the multiport case becomes a straight-forward exercise, as will be shown .
The following sections of this paper will review the conversion between 12-term and 16-term error models for the two-port
case and extend this basic method to the multiport case with extended 12-term and extended 16-term error model conversions .
a, Q - - - - - - , - - - - - - - - - , - - - - - - - i a,
N+1
b, 0--'~---'------<i=------>.-------i b,
OUT OUT
a2 Errors
b2 r 21
2 N+2
b2 o-- -+- - - ____O=--- ---->.- - - ----i
[8]N'N [8]N'N
[E]2N'2N
aN
E rN1 ELN bN r N1
N 2N
bN C5-- - - - - -----'O=-- ---->.- - - -----j
W 00
Fig. I. Multiport VNA error models : (a) extended 12-term error model (3N 2 terms total), (b) extended 16-term error model
used for advanced calibration (4N terms when excluding crosstalk). In both cases the port one excitat ion is shown. The error
model conversion is obtained using a straight-forward extension of the process described for two-ports by Marks [11]. When
neglecting crosstalk [E] can be replaced by two-port error boxes at each port.
I
I
I I
Forward 2-port error 1_ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - _.!
model EXR
DUT (a) (6 terms) (b)
Fig. 2. (a) Forward and (b) reverse halves of the 12 term error model. The isolation terms EXF and EXR are independent of the
DUT behavior. After removing isolation and accounting for switching terms the error-model is equivalent to this eight-term
model used for advanced cals like TRL.
- '-
-
A
I-
- '- - I--
DDT A DDT
- '- - I--
- B - - I--
00 ~
Fig. 3. Two-port error models after accounting for switching terms. (a) The eight-term error model used for advanced cals like
TRL is equivalent to the 12-term error model after accounting for switching terms and either ignoring or pre-correcting for
isolation errors. The error boxes (A and B) at each port are independent. (b) For the more general 16-term error model after
accounting for switching terms the errors for the two-port calibration are fully contained by one four-port network. This model is
general and allows for any type of coupling (within the context of modes supported by S-parameter corrections).
E\2. Other crosstalk terms are zero:
E21 = EXF'(I-EDRr21) E I2 = EXR'(1-EDFrI 2) (5)-(6)
E I4 = E41 = E 23 = E 32 = E 34 = E43 = 0 (7)
By analyzing an ideal thru case it becomes clear that: [11]-[12]
Ell = EDF E22 = EDR (8)-(9)
E33= ESF E44 = ESR (10)-( 11)
E31'E24 = ETF'(I-EDRr21) E 42'E l3 =E TR'(I-EDFr\2) (12)-(13)
E 3I'E l3 = ERF E24'E42 = ERR (14)-(15)
These last four terms are interdependent and could potentially cause two different answers to the same question. They need to
be consistent so a single estimate is required. [12] suggests the estimate should be obtained from:
K = ( E 42
E 31 )
est = (EE
31
42
)
fwd
(E
E 42
31
)
rev
(16)
with
(~)
E 42 fwd
TF
-- E .(I-EDR
ERR
.r)
21
E 31 )
(E
42 rev
E RF
= ETR (I-EDF .r12 ) '
(17)-(18)
Then to separate the interdependent equations we need further to assume one term since we cannot know all of them. Choosing
E42 = 1 makes it the normalizing term. Then we get:
E31 = K E l3 = ERF/K (19)-(20)
E42 = 1 E24 = ERR (21)-(22)
All of these terms are summarized in the following equation and in Fig. 4.
o 1 b4
~
J(
E DF
E SF
b3
ER F/J( iI'
bl
s. ,
S 12 S 21 DUT
S 22
a2 I
~
a,
E DR E SR
b2 ~
E
X2,N
.(I-E.r):
ill 2,N I
0 E.R~K o
I 2,N
I
I
and
E Ri ETN,i l - EDN rN,i
(39)
ERN ETi,N (I-EDi ['i,N)
provides the estimates equivalent to equation (16) for the two-port case.
The extended 12-term error model can also be obtained from the extended 16-term error model. The terms for the source port
are simply:
E Di = Ei,i, for i = 1...N (40)
E Si= Ejj, fori= 1 N,j=i+N (41)
ERi = Ej,iEij, for i = 1 N,j = i + N. (42)
The terms that account for the behavior between ports are given by:
E T .. =
n, .E./
,fori=l ...N,j=l ...N,itj,k=j+N,I=i+N (43)
I-E.. .r ..
,j I,
I,)
1,1 I,)
E k E k -T..
E =E + for i.j l ... N,itJ,k=i+N (44)
1- E . . r.. '
,1 I, I,) >
u., k,k
1,1 I,)
E
Xi,}
=
E ..
I-E..
1,)
1,1
.. ' for i,j = 1... N, i j.
. r 1,)
* (45)
The extended 12-term error model for convenience can be represented as a single matrix:
E :E : :E
D1 LI 2 L1 N
I ' I I '
E Rl I
I Err 2 I ''E
I Tl N
I ' I I '
E S1 : E .r r 2 : : E Xl N
----~---~'--r--~
E L 2,1: E D 2 ::EL 2,N
I I I
E I E 1 IE
T 2 ,1 I R2 I I T 2' N
E X2 1 : E S2 : : E X2 N
[ Errors extended 12term ] = --~:"i--~--t;--~--;...:- (46)
: I : 11 :
I I I
I I. I
I 11
I I. I
I I I
I 11
----1----1--1----
E L N,1 : E L N,2 : : E DN
I I I
E T NIl E T N 2 I I ERN
, I ' I I
E IE I IE
XN,I! XN,2 !... ! SN
where Qxi is the ratio of the reverse to forward transmission terms for port i (corresponding to Ei,i+N / Ei+N,i) and Qoxi is the
expected value of the ratio (normally unity for the reciprocity assumption with equal and real normalizing impedances).
We can further determine Qx:
(48)
with the defmitions:
1
and Z=-2 . (49)-(50)
at
The ratio U/Ui is found to be given by
a. ET
) I,) for it j. (51)
-;;; = E Ri + E Di (EU,j - EsJ
Applying the complex differential operator after [11] yields a value of z that minimizes the residual f:
LOi QOXi
Z=_i _
(52)
LI
2
Oil
i
with bi given by:
(53)
and
(54)
This yields the solution for the Ui terms:
_+_1_
a] - - -r;' (55)
and
1 E Ri +EDi (Eu,] -EsJ
i for it 1. (56)
a = a] ( %i) = a] E ri,]
The choice of root in (55) must be determined by additional information such as a delay estimate of the error box or minimizing
phase discontinuities.
The revised error model conversion equation replacing (37) is then:
IE /
Ex1,2(I-ED\r;,2) ... Ex1,N(I-ED\.rLN)! R1 ia. 0 ... 0
------------------------------------------------------~--------------------
~ 0 0 : ES1 0 0
o a2 0: 0
I
ES2 0
I
I
I
o o : 0 o
Fig. 5. The preferred method is to directly compare first tier calibrations. The errors of one calibration (denoted by error boxes
with superscript a are corrected by the second calibration (denoted by b superscripts). Mathematically this is performed by
inverting the error box transmission scattering parameters of one cal and multiplying the other cal error box T-parameters. The
residual T-parameter error boxes are denoted by BX and BY. The deviations from the ideal unity matrix error boxes are denoted
by BX and BY.
corresponding to the difference. Of course, when using this two-tier approach, we want to make sure we don't just map the points
where we calibrated so we need to use structures that are distinct from our original calibration standards.
A preferred approach is to directly compare two first-tier calibrations [13], see Fig. 5. In this method the residual error terms
are computed mathematically by correcting the error boxes of one calibration with the errors of the other. This method is
preferred over the two-tier approach since the second-tier calibration is using data that is corrected by the first-tier, skewing the
results. For example, imagine the case with a first-tier SOLT calibration and a second-tier LRRM calibration. Subject to
repeatability the corrected standards measurement from the SOLT is exactly the SOLT calibration definitions. Such an input to
LRRM can only produce an ideal result, not a fair comparison.
A. Two-Port Calibration Comparison Equations
Error bounds on a specific, measured DDT were developed in [10] and [13] and are given by:
11 =. 81 82] =. ~
_I T - I = I1T - I
a 2n 2n (70)
[ 83 84
Since ~T = 1 +~, (68) can be written using the elements of~ :
IMi,JI:5182 i,JI+I 181i,k 1ISak,J 1-181i,i 1ISai,JI +I Isaik 1184k,J 1-ISai,J 1.184 J,J 1+181i,i - 84J,J ISai,j I
k=O k=O (77)
n-I n-I
+ I I I Sai,kl183k,plI Sap ,JI
p=o k=O
and the worst-case passive structure error bound can be determined by setting alllS~jl = 1 yielding:
n-I n-I n-I n-I
worst-case IMi,J I:51 ,JI+ II8Ii,kI-181i,il+II84k,JI-184J,JI+18Ii,i -84J,jI+ II183k,pl
82i (78)
k=O k=O p=O k=O
If a single, frequency-dependent figure of merit is desired, one can simply take the maximum (for all i and j) of the ~Sij traces
at each frequency [10].
VI. CONCLUSION
This paper has provided a detailed look at error models for multiport vector network analyzers. Specific equations were
included for converting between the extended 12-term error model used in many commercial systems and a more general
extended 16-term error model. The extended 16-term error model was used in a simple but more general derivation of an n-port
calibration comparison method. This extension to calibration comparison not only handles any multiport case, but also includes
the effect of the general coupling terms provided by the extended 16-term error model.
The tools and methods described have been implemented in a commercial software application. Applications of these new
tools are very broad and open-ended. Two specific example applications were previously published. This paper provides the
mathematical details supporting these prior publications.
0.45 -
Coupled 0.4 -
Ul
Erro rs
~ 0.35 -
:!::
i 0.3
-g::::l 0.25
o
D?... 0.2
I 8-term
errors
g 0.15 - B
w
===========================_~~:i==
0.1
0.05 -
o +, ------,----...,------,-----------,-------1
o 10 20 30 40 50
Isolation (dB)
Fig. 7. Calibration comparison error bound for a calibration ignoring coupling vs. various levels of simple capacitive coupling
in the actual error box. The error bound magnitude represents a bound on the vector error difference when comparing S-
parameters for any passive DUT measured with and without coupling correction. Points A and B represent the coupling levels
for state-of-the-art GSSG and GSGSG probes respectively (at 40 GHz). Point C indicates typical coupling for two opposing
GSG probes. At this level the error-bound is at a comparable level to typical system repeatability (dashed region) and usually
not significant. This result was previously presented in [2].
.04
.035
.03
1i .025
-"
];
2- .02
~ .015
.01
5e3
10 15 20 25 30 35 40 45 50
Fteouencv IGHz)
(a) (b)
Fig. 8. A deliberate misalignment of probes for purposes of evaluating probe placement sensitivity is shown in (a). The offset is
approximately 15 urn. The sensitivity to small longitudinal offsets in probe position on the load standard for the 4-port SaLT and LRRM-
SOLR hybrid calibration methods is shown in (b). These curves bound the maximum vector error difference on a worst case passive device
measurement associated with the different calibration. SaLT shows more than four times the sensitivity to longitudinal load offset than the
hybrid cal. This result was previously published in [14].
ACKNOWLEDGMENT
The author would like to thank the WinCal development team for their wonderful ability to tum seemingly strange ideas into
useful tools for calibration and measurement.
REFERENCES