T-MTT/27/12//30945

Accuracy Improvements in Microwave Noise

Parameter Measurements

Andrew C. Davidson
Bernard W. Leake
Eric Strid

Reprinted from
IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES
Vol. 37, No. 12, December 1989
 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES. VOL. 37. NO. 12. DECEMBER 1989 1973

Accuracy Improvements in Microwave Noise

Parameter Measurements
ANDREW C. DAVIDSON, BERNARD W. LEAKE, MEMBER, IEEE,
AND ERIC STRID, M E M B E R IEEE

Abstract - Factors contributing to microwave noise parameter measure-

ment accuracy have been examined theoretically and experimentally. It is
shown that for good accuracy the test source impedances need not be
grouped around the impedance that produces minimum noise figure. Sys-
tem calibration and DUT S-parameter accuracy are important to the
derived noise parameter accuracy, and the use of a vector network analyzer
is advantageous A new algorithm has been implemented which avoids
errors caused by different noise source “on” and "off" impedances.

I. I N T R O D UC T I O N Fig. 1. Combined S-parameter and noise parameter measurement

A
system.
LTHOUGH NOISE parameter measurements are
critical for low-noise microwave circuit design and
device characterization, means of gathering accurate noise
parameters have not been generally available. The result is Using Network Vector Information
that measured noise parameters are often doubted [1], FET The system shown in Fig. 1 (Cascade Microtech NPT18)
noise modeling theories remain unverified [2], and progress has been used in the present work. To more accurately
in low-noise device development is generally hampered. In determine the optimal tuning conditions, the input tuner is
this paper we examine various factors which contribute to switched between a set of predetermined impedance points
inaccuracies in noise parameter measurements, and illus- instead of searching for the optimum [4]-[6]. Measurement
trate effective solutions. speed is improved and device oscillation avoided by termi-
nating the output of the DUT in a broad-band low-noise
II. N OISE PARAMETER M EASUREMENT amplifier. Mismatch between the DUT output and the
The classic noise-parameter measurement system uses a amplifier input is calculated from DUT S parameters and
.. the input reflection coefficient of the receiver. To measure
manual or automated tuner on each end of the device
under test (DUT). The tuners are intended to simulate the the DUT S parameters, source tuner impedances. and
input and output matching networks of a low-noise ampli- other mismatches and losses of the system. a vector net-
fier stage so that the noise figure and gain can be measured work analyzer is switched into the DUT ports. Vector
directly. reflection information provides more accuracy than is
If the input tuner is set for minimum indicated noise available from scalar data.
figure, the resulting tuning condition minimizes the com- To characterize the system, the vector network analyzer
bined loss of the tuner and the noise contributions of the is first calibrated at the DUT connection planes. Then, for
DUT and the second stage. Since most tuners exhibit more all test frequencies. source impedances presented to the
loss with increasing reflection coefficient, the typical result DUT are measured for each tuner setting and for the hot
is that the apparent optimum source reflection coefficient noise source. The available gain of the two-port which
magnitude is too low [3]. The second-stage noise contribu- connects the external calibrated hot noise source to the
tion is often significant: in some cases it is possible that DUT is calculated to allow transfer of the excess noise
source tuning for minimum overall noise figure will result ratio (ENR) calibration to the DUT input. The input
in maximizing the DUT gain rather than minimizing its impedance of the second-stage receiver is also measured.
noise figure. Noise parameters of the second stage are calculated from
noise power measurements with a through-connect substi-
Manuscript received April 6, 1989: revised June 7, 1989. tuted for the DUT.
The authors are with Cascade Microtech, Inc.. 14255 S. W. Brigadoon
Court. Beaverton, OR 97005.
The calibrated noise source is the prime standard which
IEEE Log Number 8930945. determines the ultimate accuracy of &in and R,. The

0018-9480/89/1200-1973$01.00 61989 IEEE

 1974 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 37. NO. 12, DECEMBER 1989

TABLE I
NOISE PARAMETERS USED IN SIMULATION

Fmln = 1.5 dB

rapt = .LSLcf

P= 9 ,, +;opt,2 = 2.7
4Rn I r0’0pt - rs l2
F = Fmtn + - l
20 11 +I-opt12(1- [Ts12)

Fig. 2. Two source constellations used in the simulation

other noise parameters are obtained from relative measure-
ments and so are not affected by noise source power
calibration errors.
To characterize a DUT, its S parameters are measured.
Then one measurement is made of noise power with the
hot noise source connected, followed by a number of noise
power measurements with other source impedances at am-
bient temperature. These measurements provide all the
information necessary to calculate the overall system noise
parameters and the DUT noise parameters and associated
gain.

’ III. S OURCE I M P E D A N C E S : WHERE AND How MA N Y-

Measurements for a minimum of four source impedances
are necessary to solve for the four noise parameters F&,
Gopt’ Bopv and R,, but more are advantageous to allow
averaging and help ensure a unique solution. To lend
insight into the problem of noise parameter sensitivity to
measurement errors, a computer simulation of noise figure
measurements for a wide variety of source impedance
configurations has been made. Topics of interest are how
the source impedances should be distributed and how
..,.
many are needed to optimize for accuracy and measure-
ment speed.
The noise parameters of a typical FET (Table I) were 5 Number of Points 2 5
used to calculate noise figure for each of a chosen set of
Fig 4. Noise parameter errors versus number of source points.
source impedance points, and each noise factor was then
assigned a random error of up to 1 percent. The resulting
noise figures were then used to calculate the noise parame- DUT by lengths of transmission line, causing the orienta-
ters using a least-squares routine [4], [5]. The differences tion to change rapidly with frequency.
between the calculated parameters and the original param- Fig. 3 shows the errors in predicted noise parameters as
eters are the errors which were RMS averaged over functions of T_ for extreme orientations of the source
400 runs. points. It is interesting that, in this case, orientation of the
The simulations presented here used source constella- source points has little effect on accuracy in spite of the
tions forming a cross shape on the Smith chart. Fig. 2 fact that the nearest source can be quite far from gamma-
shows two orientations of a nine-point set. Parameters opt. Larger source constellations generally produce better
which describe such a constellation are the maximum accuracy, but improvement is small beyond a magnitude
reflection coefficient of the outer source points (T_), the somewhat less than gamma-opt. Fig. 4 shows that. for a
angular orientation (6,,), and the number of source points. given constellation; increasing the number of intermediate
In all cases. one of the points was positioned at the center points does not help significantly. in contrast to increasing
of the Smith chart, while the remainder were spaced equally the averaging at a fixed point by a factor N, which would
along the lines forming the cross. The angular orientation reduce the error there by a factor l/m.
of the source constellation was varied in the simulation Another simulation was run in which all points in an
because physical tuners will Likely be distanced from the initial constellation were made to converge on gamma-opt
.
DAVIDSON et al.: ACCURACY IMPROVEMENTS JN MICROWAVE NOISE PARAMETER MEASUREMENTS 1975

Fig. 8. Receiver F,,, versus frequency, with and without the assump-
tion that rHoT = COLD.

Two configurations: The smaller corresponds to a scale factor of coefficient cause large errors in noise figure. One possible
0.7. the larger to a scale factor of 0. way to reduce this effect would be to use an orthogonal
fitting routine [6].
A. S-Parameter Accuracy
. .
Network analyzer calibration and probe placement er-
rors [7], [8], which are known to cause S-parameter inaccu-
racies, can seriously affect computed noise parameter ac-
curacy. Modem low-noise devices are poorly matched to
the normal reference impedance, which results in high-Q
circuits.

B. Noise Source ON/OFF Impedance Change

Some procedures used in the derivation of noise parame-
- 0 Scale Factor .9
ters require the measurement of noise figures for a number
Fig. 6. Noise parameter errors versus proximity of source points to of different source impedances. The noise figure is ob-
ropl’
tained from measurements of the noise powers (P,,, P,,,J
at the output of the receiver under test when its input is
connected to a source which may be set to two substan-
ially different known effective temperatures. The noise
figure (F) is calculated from the equation
ENR
F=-
Y-l
where ENR is the source excess noise ratio. and Y=
Pho, /Puold’ Accuracy of the resultant noise figure relies on
the total receiver gain and noise figure remaining constant
0 ’ I
.9
between measurements of P,, and Pcold, and this requires
.3 S o u m r,
that the source impedance not change with effective tem-
Fig. 7. Noise parameter errors far errors in source reflection coeffi- perature of the source [9], [10].
cients.
An improvement to this technique [11] points out that if
the receiver input reflection coefficient is known, only one
by use of a scale factor, as shown in Fig. 5. Results plotted measurement of noise figure is necessary together with
in Fig. 6 show that errors in B become large as the source several “cold’,:’ noise power measurements. This method
constellation shrinks, while the other parameters change was used in the calibration of a noise characterization
only slightly. This is expected, because /3 describes how system (Fig. 1) using a 15 dB ENR source. In Fig. 8, the
rapidly the noise figure varies with r,. two upper traces show the effect of moving the noise
An example of the effect of source positional uncer- source with a small coaxial extender. The ripples are
tainty is shown in Fig. 7, which is similar to Fig. 3 except caused by the difference in noise source “on” and “off
that random source reflection coefficient errors were used reflection coefficients, which was 0.05 magnitude.
instead of noise figure errors. The large errors for large A further improvement (see the Appendix) shows that
rW.X are expected because as a source impedance gets when a number of cold noise power measurements are
closer to the edge of the Smith chart, the gradient of the made for different source impedances, it is not necessary
noise surface increases and small errors in source reflection to measure cold noise power at the source impedance
 1976 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 37, NO. 12, DECEMBER 1989

6 GHz

Fig. 9. Configuration of source points used in measurements at 6 GHz.

5.00 10.00 15.00

(b)
Frequency [GHz]

Fig. 1 0 . F,,, and associated gain of a HEMT from measured data. Fig. 12. A passive two-port verification. (a) Fti,, (O); maximum avail-
able gain (a). ). (b) Gamma-opt: noise (+): available gain (O).

V. VERIFICATION
Measurement of a passive two-port can provide some
assurance that the measurement system is working prop-
erly. The useful property of such a two-port is that F =
l/G,, and the source for minimum noise figure and that
..,. for maximum available gain are the same. This test is not
affected by noise source calibration errors, provided the
same noise source is used to obtain the second stage and
overall noise parameters, because F,/F,, = Gav in this
1
case. The results of a passive verification are show-n in Fig.
Fig. 11. r,, of a HEMT, 2-18 GHz: Measured l , smoothed 0. 12.

produced by the hot source. An implementation of this VI. CONCLUSIONS

technique to a repeat calibration resulted in the lower trace The combination of noise measurement instruments,
in Fig. 8, which exhibits essentially no ripple. together with a vector network analyzer, allows accurate
determination of linear two-port noise parameters. System
IV. ACTIVE D EVICE M EASUREMENT calibration, including S parameters and receiver noise
The ideas described above have been applied to the parameters, removes uncertainties often associated with
measurement of active devices. A typical 25 x 300 pm2 noise characterization measurements.
HEMT was measured over the frequency band 2 to Computer simulations indicate that, in a system using a
18 GHz, using nine source impedances as shown in Fig. 9 number of fixed source impedances, the principal concern
for 6 GHz. Figs. 10 and 11 show Fti, associated gain, and in choosing a constellation is not the number of source
r,_,, using data which were obtained in less than seven points but rather their distribution on the Smith chart. If
seconds for each frequency. Groups of eight and seven the points are well distributed, the noise figure surface will
points from the same measurements were fitted and found be described accurately even though measurements were
to give Fmin typically within 0.05 dB of the values obtained not necessarily made close to the optimum. Given that
with all nine source points. measurement time is proportional to the number of points.
DAVIDSON et al.: ACCURACY IMPROVEMENTS IN MICROWAVE NOISE PARAMETER MEASUREMENTS 1977

there is clear advantage to fewer, well placed source from (A3),

impedances which yield good results independent of the
location of the optimum. k = (PC/k%
A procedure which avoids the need to measure noise Fo
figure (or P, and PO,, for the same source impedance) has
been described, so differences between noise source hot which, from (Al), gives
and cold impedances are irrelevant.
The noise parameter system described, using on-wafer
measurements of modem devices, results in smooth (within
0.1 dB) values of F,, with frequency and has high
k= [( :),-( 31,1&*
throughput. Here g, is the relative gain; ENR is the hot source power
calibration; (PC-g), = kF, is found from the four con-
A P P E N DIX
stants of (A5) and the source admittance Y, produced by
The ratio of the receiver gain ( for a source reflection the hot source; and P,., is a new noise power measurement
coefficient r, to that (GREF) with a perfectly matched with the hot source.
source (r, = 0) is given by Now that the scale factor k has been found, the known
scaled noise parameters kF,, and kR, can be scaled to
Gs 1 - IW &f-&Is Fti and R,. The previously calculated Y, (which gives
G,=li-r,r~12=(~H-~C)REF=gJ F,,) needs no scaling.
where r, is the receiver input reflection coefficient and PH
REFERENCES
and PC are the hot and cold noise powers.
[1] M. Pospieszalski er. al.. “Comments on ‘Design of microwave
The noise figure for any source impedance is given by GaAs MESFETs for broadband. low-noise amplifier,‘” IEEE
Tram. Microwave Theory Tech., vol. MTT-34,p. 194. Jan. 1986.
F _ E N R ENR.(P,), A. Cappy, “Noise modeling and measurement techniques.” IEEE
(A1) [2]
\ s-y-l- (PH--PC)/ Tram. Microwave Theory Tech., vol. 36. pp. l-10, Jan. 1988.
. [3] E. Strid, “Measurement of losses in noise-matching networks.”
IEEE Trans. Microwave Theory and Tech., v o l . MTT-29, pp.
Therefore the ratio of the noise figures (F,, F2) for source 247-252. Mar. 1981.
reflection coefficients I’, and r, is given by [4] R Q. Lane, “The determination of device noise parameters.” Proc.
IEEE. vol. 57. pp. 1461-1462, Aug. 1969.
[5] G. Caruso and M . Sannino, “Computer-aided determination of
microwave two-port noise parameters,” IEEE Trans. Microwave
Theory Tech.. vol., MTT-26, pp. 639-642. Sept. 1978.
[6] M. Mitama and H. Katoh. “An improved computational method
or for noise parameter measurement," IEEE Trans. Microwave Theory
Tech., vol. MTT-27, pp. 612-615, June 1979.
(A3) [7] Cascade Microtech Model 42 Operation Manual, ch. 4. C asc
a de
Microtech. Beaverton. OR.
[8] K. Jones and E. Strid. “Where are my on-wafer reference planes?"
where k is a constant. in IEEE A RFTG Dig., pp. 27-40. Dec. 1987.
The noise figure (F,) depends on the source admittance [9] E. Strid. “Noise measurements for low-noise GaAsFET amplifiers,"
Microwave Syst. News, part I. pp. 62-70, Nov. 1981; part II, Dec.
(Y ) as given b y 1981.
[10] N. J. Kuhn. "Curing a subtle but significant cause of noise figure
error.” Microwave J.. vol. 27. pp. 85-98, June 1984.
F, = F,, + :,Y, - Y,l* (A4) [11] A. Adamian and A. Uhlir, “A novel procedure for receiver noise
I characterization," I E E E Trans. Instrum. Meas., vol. IM-22 p p .
181-182. June 1973.
where F&, R,. and Y0 = Go + jB, are the noise parame-
ters, which are independent of Y,, the source admittance
that produces the noise figure F,. G, is the real part of q.
If F, is scaled by a factor k. then
k.R,
kF, = kF,, + yp.-s - &I*. (A5)
I
So F,, and R, are also scaled by the same factor k while Andrew C. Davidson was born in Montreal,
Canada. in 1963. He received the B.S. and M.Eng.
Y,. the source admittance that produces Fti, is u n - degrees in applied and engineering physics from
changed. Cornell University, Ithaca. NY. in 1985 and 1987.
Four, or more, measured values of ( P,-)i with their respectively. During he year of 1986 he held an
internship at Schlumberger Well Services Hous-
corresponding values of g,, calculated from (r,); and I’,, ton, TX, where he characterized and modeled
provide a set of values for kF, that may be fitted to (AS) to microwave slot antenna radiation in layered
media.
obtain values for YO, kF,,, and kR,. He has been with Cascade Microtech, Beaver-
To find the scale factor. we proceed as follows. For the ton. OR. since 1987, where he has been involved
source reflection coefficient produced by the hot source, in the development of noise characterization systems.
1978 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 37, NO. 12, DECEMBER 1989

Bernard W. Leake (M'57) was born in London, Eric Strid (S'74-M'75) received the B.S.E.E. de-
England, in 1928, and received the B.Sc degree in gree from the Massachusetts Institute of Tech-
physics from London university in 1951. nology, Cambridge, in 1974 and the M.S.E.E.
Between 1957 and 1987 he worked on mi- degree from the University of California at
crowave systems and component design at Berkeley in 1975.
Raytheon Equipment Division in Massachusetts. He first worked on microwave MIC's at Fari-
For some years his interests have included mi- non Transmission Systems, San Carlos, CA. In
crowave computer-aided design and measure- 1979 he joined the GaAs research group at Tek-
ment. He is presently with Cascade Microtech, tronix. (This group evolved into TriQuint Semi-
Beaverton, OR, where he is involved in computer- conductor.) In 1983 he cofounded Cascade Mi-
controlled characterization or low-noise devices. crotech Inc., where he is now president, and
CEO. He has published various papers on power GaAs FET's, noise
measurements, analog and digital GaAs IC's, and high-frequency wafer
probing.