THE design and characterization of novel GaAs

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 47, NO. 2, FEBRUARY 1999 125 Novel MMIC Source-Impedance Tuners for On-Wafer Microwave Noise-Parameter Measurements Caroline E. McIntosh, Member, IEEE, Roger D. Pollard, Fellow, IEEE, and Robert E. Miles, Member, IEEE Abstract Novel monolithic-microwave integrated-circuit source-impedance tuners for use in on-wafer noise-parameter measurement systems are reported, which can be incorporated into a wafer probe tip. These eliminate the effect of cable and probe losses on the magnitude of a reflection coefficient that can be presented to the input of an on-wafer test device, thus enabling higher magnitudes to be synthesized than for conventional tuners, and with the potential of increasing noise-parameter measurement accuracy. Index Terms Impedance tuner, monolithic microwave integrated circuit, noise measurement, noise parameters, on-wafer characterization. I. INTRODUCTION THE design and characterization of novel GaAs monolithic-microwave integrated-circuit (MMIC) source-impedance tuners for use in on-wafer microwave noise-parameter measurements is reported. These tuners can be incorporated directly into a probe tip, eliminating any loss between tuner and test device, and facilitating improved accuracy on-wafer noise-parameter measurements. The noise parameters of a linear two-port are normally determined from a set of noise-figure measurements for different source impedances using a least-squares data-fitting technique [1]. This requires the use of a tuner to vary the impedance and, hence, the reflection coefficient at the input of the test device [2]. Ideally, the tuner should be capable of synthesizing any impedance within the unit circle on the Smith Chart to enable a widely spread constellation of source impedances to be used [3]. Maximum reflection-coefficient magnitudes of 0.9 can be achieved using passive mechanical tuners [4]. However, in on-wafer measurement systems, a conventional source-impedance tuner is separated from the test device by a probe and length of cable, both of which exhibit loss, thus reducing the maximum achievable reflection coefficient at the test device input. With a typical cable and probe loss of about 3 db, this results in the maximum tuner reflection-coefficient magnitude of 0.9 being reduced Manuscript received July 31, 1996; revised September 24, 1998. This work was supported by the U.K. Engineering and Physical Sciences Research Council (EPSRC) under a CASE Studentship and by the Hewlett-Packard Company. C. E. McIntosh is with the Institute of Microwaves and Photonics, School of Electronic and Electrical Engineering, The University of Leeds, Leeds LS2 9JT, U.K. R. D. Pollard and R. E. Miles are with the Institute of Microwaves and Photonics, School of Electronic and Electrical Engineering, The University of Leeds, Leeds LS2 9JT, U.K. Publisher Item Identifier S 0018-9480(99)01158-8. to only 0.45, as seen by the device-under-test. Tuneable wafer probes have been previously described to overcome this effect [5], but these exhibited very limited Smith Chart coverage area and bandwidth. The use of a MMIC tuner, which could be incorporated directly into a probe head, has also been suggested [6], but the results reported show only a limited number of possible impedance points. Three different novel MMIC tuners are described here, each of which offers complete phase coverage of the Smith Chart, synthesizing 50 discrete impedance points. Any one of these tuners could be incorporated into a wafer probe head to minimize losses and potentially increase noise-parameter measurement accuracy. II. MMIC TUNER DESIGNS The tuners designed here achieve phase coverage of the Smith Chart by starting at the short- and open-circuit points and adding a variable offset to transform the impedance seen at the tuner output around the chart. This is accomplished using pseudomorphic high electron-mobility transistor (phemt) switches to add lengths of microstrip transmission line. The switch is designed with a large (400 m) gatewidth to minimize the on -state impedance with a corresponding cost in off -state isolation. A spiral inductor is connected in parallel with each phemt to resonate its off -state capacitance and improve the isolation [7]. A low-pass filter consisting of a single spiral inductor and metal insulator metal (MIM) capacitor is connected at the gate of each phemt to allow dc control voltages to be applied while acting as an RF open circuit. Amplitude coverage of the Smith Chart is achieved by altering the gate bias on each switch so that it acts as a variable attenuator. A design frequency range of 16 20 GHz was chosen for these proof-of-concept tuners in order to minimize the line lengths and, hence, chip area required, while still operating within the range of commercial noise sources. It should be noted however, that the designs are inherently broadband and could be utilized over a much larger frequency range. A block diagram of a typical measurement setup for noiseparameter measurements is shown in Fig. 1. The tuners described here are all two-port circuits. Each tuner can be used in one of two states: its thru state or its variable impedance state. It is anticipated that the MMIC tuners would be used in conjunction with the cold noise-figure measurement technique [8] rather than the -factor technique. The thru state is used to connect a noise source at Port 1 of the tuner directly to a receiver at Port 2 (via an on-wafer thru standard) to facilitate receiver calibration without removal of the tuner. Two noise-figure measurements would be performed with this 0018 9480/99$10.00 1999 IEEE

126 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 47, NO. 2, FEBRUARY 1999 Fig. 4. Simplified schematic diagram of MMIC Tuner B. Fig. 1. Block diagram of typical noise-parameter measurement system. Fig. 5. Photograph of MMIC Tuner B. Fig. 2. Simplified schematic diagram of MMIC Tuner A. Fig. 6. Simplified schematic diagram of MMIC Tuner C. Fig. 3. Photograph of MMIC Tuner A. arrangement one with the noise source delivering its hot temperature and the other cold. The variable impedance state is then used to synthesize varying loads, all at room temperature (i.e., cold ) at the test device input, which would be connected to Port 2 of the tuner during measurement. If the MMIC tuners were bonded into a wafer probe tip, Tuner Port 2 in Fig. 1 would be connected directly to the test device input, eliminating the need for a connecting cable. If an on-wafer noise source is available [9], only the variable impedance state of the tuner need be used. A. Tuner Design A The first tuner consists of a length (1481 m equivalent to a microstrip half-wavelength at 18 GHz) of transmission line with seven phemt switches spaced along it each connected to ground. A similar idea to this has previously been proposed [10] to allow for post-processing on-chip tuning of MMIC s, although a much smaller tuning range was required. It is also similar to the MMIC tuner reported in [6], in which only a few impedance states were demonstrated. The seven switches were initially spaced uniformly over the length of line, although the lengths between each switch varied slightly after simulation and optimization. With all the switches biased off by default, the tuner acts as a thru line. To synthesize varying impedances, the bias on each switch is varied independently with all the others in their off states, thus creating a short-circuit-stub offset by different lengths of transmission line. A simplified schematic diagram of the circuit is shown in Fig. 2 and a photograph of the MMIC, which measures 2880 m by 800 m, is shown in Fig. 3. B. Tuner Design Tuner consists of a 1481- m length of transmission line interrupted in seven places along its length with a switch. Fig. 7. Photograph of MMIC Tuner C. The thru state for this design occurs with each switch in its default on position, and varying impedances are produced by altering the bias on each switch independently with all the others in their on states to produce different lengths of opencircuit stub. A simplified schematic diagram and photograph of the MMIC (3520 m 640 m) are shown in Figs. 4 and 5, respectively. C. Tuner Design The final tuner design is a hybrid of the two previous designs. Eight switches are used to enable each of four lengths of transmission line to be independently switched, either to ground to form a short circuit, or to become an open-circuit stub [11]. Ideally, only half the length of microstrip line is required compared to the previous tuner designs, keeping the final chip to a smaller area of GaAs and making it significantly cheaper to fabricate. This design should provide a greater Smith Chart coverage area and usable bandwidth than the previous versions, as the use of both offset short- and open-circuit stubs enables high values of reflection-coefficient magnitude to be synthesized in all quadrants of the Smith Chart over a greater frequency range. Also, an overall shorter length of transmission line reduces the loss through the circuit. A simplified schematic diagram of this design is shown in Fig. 6. Switches 1 4 are biased on and switches 5 8 off by default creating a thru path between the two ports. As before, varying impedances are produced by independently altering the bias on each switch with the others in their default states. A photograph of the MMIC tuner, which measures 2400 m by 800 m, is shown in Fig. 7.

MCINTOSH et al.: NOVEL MMIC SOURCE-IMPEDANCE TUNERS 127 Insertion loss of tuners in thru state. Simulated. Measured. Return loss of tuners in thru state. Simulated. Measured. Fig. 8. Fig. 9. III. SIMULATED AND MEASURED RESULTS Circuit simulations were performed using Hewlett- Packard s Microwave Design System (MDS), and a vector network analyzer was used in conjunction with a probe station, wafer probes, and dc probe needles to measure the -parameters of the MMIC tuner circuits on-wafer. A. Insertion and Return Loss of Tuners in Thru State Fig. 8 shows the simulated and measured insertion loss, and Fig. 9 the return loss, for each of the tuners in its thru state. The simulated insertion loss of Tuner is a minimum at less than 3 db around the center frequency of 18 GHz and there is a corresponding maximum magnitude of return loss (25 db). The insertion loss increases at frequencies greater and less than the center frequency band, but the increase is sharper below this center band than above (17 db at 10 GHz, 4 db at 26 GHz). This is due to the off -state insertion loss of the phemt switches differing from the maximum at frequencies other than the center frequency. The loss of each switch in its on state is flat over a wide bandwidth. Consequently, the insertion loss of Tuner is also flat at approximately 2 db and the return loss is better than 15 db over the entire band. The insertion loss of Tuner is lower at 10 GHz than Tuner because fewer shunt switches are employed. The loss is not as low as that of Tuner, which includes only series switches. The reason for the sharp maximum in return loss at 12 GHz is not known and it will be seen that this does not appear in the measured data. The measured values of insertion loss are several decibels greater than the simulations. However, the return loss and shape of the insertion-loss curves are consistent with the simulations. Tuner has a constant insertion loss of 3 db and return loss varying between 15 and 20 db. The other two designs exhibit a minimum in insertion loss in the design frequency band at 10 db for Tuner and 6 db for Tuner. This increases outside the center band and is especially high at lower frequencies, as in the simulations, owing to the behavior of the off -state loss of the switches. These two designs also have a gradual maximum in return-loss magnitude in the design frequency range and it should be noted that the sharp return-loss maximum seen at 12 GHz in the simulations is not present in practice. The greater insertion loss exhibited by the measured tuner circuits compared to the simulations suggests that either the on -state loss of the phemt switches is higher than in the process model used for the simulations or the off -state loss of the switch is lower than in the model. In view of the fact that Tuner, which contains no shunt switches, displays less degradation in insertion loss than the other designs, it is likely that the major cause is a lower impedance in the off state of the switches. The tuners are designed to operate in the thru state to enable noise receiver calibration. At first glance, it appears that the loss introduced between noise source and test device by the tuner (as much as 10 db for Tuner ) would be a drawback. In fact, this is not the case. The effect of introducing a 10-dB pad into a noise-measurement system to reduce mismatch has been investigated previously [12]. This pad reduces the excess noise

128 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 47, NO. 2, FEBRUARY 1999 Fig. 10. Reflection coefficients synthesized between 16 20 GHz by varying the gate bias on Switch 1 for Tuner C. Simulated. Measured. Fig. 11. Reflection coefficients synthesized between 16 20 GHz by varying the gate bias on Switch 5 for Tuner C. Simulated. Measured. ratio (ENR) of the noise source from approximately 15 db at its output to approximately 5 db at the test device input, which can be an advantage, as it reduces the noise-source ENR to a value closer to that likely to be measured from a practical device. B. Impedances Synthesized in Variable Impedance State Figs. 10 12 show the simulated and measured impedances achieved for various values of gate bias on a selection of switches in Tuner with all other switches in their default states. For clarity, only results obtained between 16 20 GHz are shown, although the tuners are inherently broad-band. Only results for Tuner are shown as an example, as this tuner is effectively a combination of the other two designs. Figs. 10 12 show the simulated and measured impedances achieved for switches 1, 5, and 6 of Tuner, where switch 1 creates an open and switches 5 and 6 short-circuit stubs. The measured reflection-coefficient magnitudes are lower than simulated due to poorer switch off -state isolation. It can be seen how switching in additional lengths of transmission line moves the phase of the reflection coefficient around the Smith Chart, and, as expected, varying the gate bias alters the magnitude. The maximum achievable reflectioncoefficient magnitude occurs for switches 1 and 5 in Tuner, and decreases with the addition of each subsequent switch because the preceding switches are imperfect. Even though the switches are in the off state, some of the signal will be lost through them to ground. They also cause unwanted reflections, the effect of which is to alter the impedance seen at the ports of the tuner. The maximum reflection-coefficient magnitude for each switch is reduced by a larger amount for the series switches (creating open-circuit stubs) than for the shunt switches (creating short circuits), as the off -state impedance of each switch is lower than simulated. This has a greater effect on an open than on a short circuit because when synthesizing the maximum reflection coefficient for a short circuit, the relevant switch is in its on state, whereas for an open circuit, the relevant switch must be in its off state. Nonetheless, complete phase coverage of the Smith Chart is achieved for each tuner circuit. More phase shift is exhibited in the measured data than in the simulations, and the maximum reflection coefficient

MCINTOSH et al.: NOVEL MMIC SOURCE-IMPEDANCE TUNERS 129 Fig. 13. Measured constellation of source impedances synthesized by Tuner A at 18 GHz. off -state impedance were increased at the expense of slightly higher on -state impedance, which is consistent with the use of a smaller switch size. The use of MESFET s rather than phemt s would also be likely to increase switch off -state impedance. Fig. 12. Reflection coefficients synthesized between 16 20 GHz by varying the gate bias on Switch 6 for Tuner C. Simulated. Measured. measured for each switch is smaller than simulated. This is probably partly due to discrepancies between the phemt equivalent-circuit model used to simulate the designs and fabricated components, and partly due to gate leakage and a lack of drain and source dc grounding. It is likely that some unaccounted for gate leakage occurs in the tuners owing to inadequate simulation of the dc bias circuitry. Obviously, future iterations of these designs would fully account for gate leakage in the simulations and dc grounding of the drain and source of each switch could be applied; this should significantly reduce the discrepancy between measured and modeled performance. Also, as mentioned in Section II, the size of the device used as a switch was chosen to be large in order to minimize it s on -state impedance. The switch equivalentcircuit model used had been developed from measurements on a smaller device, and some discrepancy between the modeled and fabricated components can be attributed to this factor. The use of a smaller switch size in the tuners would eliminate or reduce this factor. It can be deduced from the measurements that the tuner performance could be improved if the switch C. Smith Chart Coverage The full measured constellation of impedances synthesized for each tuner is shown in Figs. 13 15. Although the measured coverage areas are limited, owing to much poorer off - state isolation exhibited in practice than in the simulations by the switches, it can be seen that the use of both offset short and open circuits in Tuner enables high values of reflection coefficient to be synthesized in all quadrants of the Smith Chart. Switch off -state isolation could be greatly increased by using a smaller gatewidth. Nevertheless, high values of reflection-coefficient magnitude can be achieved, the maximum measured values being 0.85 for Tuners and. This value is similar to those obtained at the output of passive mechanical tuners [4] and better than those achieved by many electronic tuners [13], neither of which can be positioned as electrically close to the test device as an MMIC. D. Tuner Noise Output The noise generated by the switches in the tuners was expected to be purely thermal. However, some noise characterization was necessary to determine whether this was truly the case or whether slight deviations from thermal noise occur due to the gate voltage applied. The excess noise produced by one MMIC tuner into a 50- load was measured on-wafer at 17 GHz; a number of measurements were taken with gate bias applied to different switches. The tuner was connected to a precalibrated noise receiver via a wafer probe and length of cable. Most of the cabling used was accounted for in the calibration, but the loss of the probe itself was not, and no deviation from thermal noise was observed.

130 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 47, NO. 2, FEBRUARY 1999 open-circuit stubs, offers the best prospects because it can synthesize high values of reflection coefficient in all quadrants of the Smith Chart and can be fabricated on the smallest area of wafer, thus reducing cost. The coverage areas could be significantly improved with the use of a smaller phemt or perhaps a MESFET switch. Future iterations of the design should fully take account of gate leakage in the simulations and incorporate dc grounding of the drain and source to achieve better agreement between measured and modeled performance. The tuners are inherently broad-band and have larger usable bandwidths than previous attempts to fabricate tunable wafer probes. ACKNOWLEDGMENT The MMIC s were fabricated by the Microwave Technology Division (MWTD), Hewlett-Packard, Santa Rosa, CA. Fig. 14. Measured constellation of source impedances synthesized by Tuner B at 18 GHz. Fig. 15. Measured constellation of source impedances synthesized by Tuner C at 18 GHz. Therefore, any excess noise produced by the phemt s in switch configuration is too small to be measured with standard noise-measurement equipment and will be assumed negligible. REFERENCES [1] R. Q. Lane, The Determination of device noise parameters, Proc. IEEE, vol. 57, pp. 1461 1462, Aug. 1969. [2] 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. [3] A. C. Davidson, B. W. Leake, and E. Strid, Accuracy improvements in microwave noise parameter measurements, IEEE Trans. Microwave Theory Tech., vol. 37, pp. 1973 1977, Dec. 1989. [4] W. E. Pastori and G. R. Simpson, ATS for power and noise characterization using PC-AT based software, Microwave J., pp. 82 92, Jan. 1993. [5] G. Rabjohn and R. Surridge, Tunable microwave wafer probes, in Proc. GaAs IC Symp., Nashville, TN, Nov. 1988, pp. 213 216. [6] M. Dydyk, MMIC reflection coefficient synthesizer, in 39th ARFTG Conf. Dig., Albuquerque, NM, June 1992, pp. 26 41. [7] V. E. Dunn, N. E. Hodges, O. A. Sy, W. Alyassini, M. Feng, and Y. C. Chang, MMIC phase shifters and amplifiers for millimeter-wavelength active arrays, in IEEE MTT-S Int. Microwave Symp. Dig., 1989, pp. 127 130. [8] V. Adamian and A. Uhlir, A novel procedure for receiver noise characterization, IEEE Trans. Instrum. Meas., vol. IM-22, pp. 181 182, June 1973. [9] B. Hughes and P. Tasker, Improve accuracy of on-wafer noiseparameter testing, Microwaves RF, pp. 67 76, Feb. 1991. [10] W. Bischof, Variable impedance tuner for MMIC s, IEEE Microwave Guided Wave Lett., vol. 4, pp. 172 174, June 1994. [11] C. E. Collins, R. D. Pollard, and R. E. Miles, A novel MMIC source impedance tuner for on-wafer microwave noise parameter measurements, in IEEE MMWMC Symp. Dig., June 1996. [12] N. J. Kuhn, Curing a subtle but significant cause of noise figure error, Microwave J., pp. 85 98, June 1984. [13] R. T. Webster, A. J. Slobodnik Jr., and G. A. Roberts, Determination of InP HEMT noise parameters and S-parameters to 60 GHz, IEEE Trans. Microwave Theory Tech., vol. 43, pp. 1216 1225, June 1995. IV. DISCUSSION AND CONCLUSIONS Three novel proof-of-concept MMIC tuner circuits have been reported which can achieve reflection-coefficient magnitudes comparable to those produced at the output of passive mechanical tuners and better than those produced by many electronic tuners. In addition, they can be connected directly to a test device on-wafer, hence, avoiding connector, cable, and probe losses. The tuners can each synthesize up to 50 different discrete impedance points. Any of the three MMIC tuner designs presented here could be further improved. A variation on Tuner, which contains both offset short- and Caroline E. McIntosh (S 94 M 96) was born in Hull, U.K., in 1969. She received the B.Eng. and Ph.D. degrees from The University of Leeds, Leeds, U.K., in 1991 and 1996, respectively. Her Ph.D. research was concerned with improving the speed and accuracy of microwave noise-figure and noiseparameter measurements, with particular reference to on-wafer measurements. From 1995 to 1998, she was a Post-Doctoral Research Fellow at the Institute of Microwaves and Photonics, The University of Leeds, where she was involved with the Terahertz Integrated Technology Initiative (TINTIN) Project, working on the fabrication and characterization of low-cost micromachined waveguide components for use at millimeter-wave and terahertz frequencies. She is currently a Senior MMIC Design Engineer.

MCINTOSH et al.: NOVEL MMIC SOURCE-IMPEDANCE TUNERS 131 Roger D. Pollard (M 77 SM 91 F 97) was born in London, U.K., in 1946. He received the B.Sc. and Ph.D. degrees in electrical and electronic engineering from The University of Leeds, Leeds, U.K. He currently holds the Hewlett-Packard Chair in High Frequency Measurements at the School of Electronic and Electrical Engineering, The University of Leeds, where he has been a faculty member since 1974. He is Deputy Director at the Institute of Microwaves and Photonics, which has over 40 active researchers, a strong graduate program, and has made contributions to microwave passive and active device research. This activity has significant industrial collaboration, as well as a presence in continuing education through its Microwave Summer School. Since 1981, he has been a Consultant to Hewlett-Packard, Santa Rosa, CA. He has published over 100 technical articles and holds three patents. His personal research interests are in microwave network measurements, calibration and error correction, microwave and millimeter-wave circuits, and large-signal and nonlinear characterization. Prof. Pollard is a Chartered Engineer, U.K., and a member of the Institution of Electrical Engineers (IEE), U.K. He was the 1998 President of the IEEE Microwave Theory and Techniques Society, and is currently serving his second term as an elected member of the administrative committee. Robert E. Miles (M 82) was born in Kettering, U.K. He received the B.Sc. and external Ph.D. degrees from Imperial College, London University, London, U.K., in 1964 and 1972, respectively. From 1964 to 1972, he was a Research Scientist with Zenith Radio Research, U.K., where he worked on the surface properties of the IV VI compound semiconductors. After a period as a Teacher, he joined the School of Electronic and Electrical Engineering, The University of Leeds, Leeds, U.K., in 1981, as a Research Engineer, where he worked on problems of ion implantation and device modeling in GaAs. In 1983, he became a Lecturer at the University of Bradford, Bradford, U.K., where he continued his interests in compound semiconductors. In 1985, he then rejoined the Institute of Microwaves and Photonics, Leeds University, as a Lecturer. His current research interests are in III V semiconductors for highfrequency devices, device and process simulation, devices and circuits at very high frequencies, and micromachining.