IEEE MTT V065 I11B 2017 11 Microwave

theory and techniques Fantomasping

Visit to download the full and correct content document:
https://ebookmeta.com/product/ieee-mtt-v065-i11b-2017-11-microwave-theory-and-te
chniques-fantomasping/
More products digital (pdf, epub, mobi) instant
download maybe you interests ...

IEEE MTT V065 I11A 2017 11 Microwave theory and

techniques Ieee Microwave Theory And Tecnique Society

https://ebookmeta.com/product/ieee-
mtt-v065-i11a-2017-11-microwave-theory-and-techniques-ieee-
microwave-theory-and-tecnique-society/

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES

V065 I12A 2017 12 Fantomasping

https://ebookmeta.com/product/ieee-transactions-on-microwave-
theory-and-techniques-v065-i12a-2017-12-fantomasping/

IEEE MTT V065 I05A 2017 05 Fantomasping

https://ebookmeta.com/product/ieee-
mtt-v065-i05a-2017-05-fantomasping/

IEEE MTT V065 I05B 2017 05 Fantomasping

https://ebookmeta.com/product/ieee-
mtt-v065-i05b-2017-05-fantomasping/
IEEE MTT V065 I02 2017 02 2nd Edition Fantomasping

https://ebookmeta.com/product/ieee-mtt-v065-i02-2017-02-2nd-
edition-fantomasping/

IEEE MTT V065 I06 2017 06 6th Edition Fantomasping

https://ebookmeta.com/product/ieee-mtt-v065-i06-2017-06-6th-
edition-fantomasping/

IEEE MTT V065 I03 2017 03 3rd Edition Fantomasping

https://ebookmeta.com/product/ieee-mtt-v065-i03-2017-03-3rd-
edition-fantomasping/

IEEE MTT V065 I01 2017 01 1st Edition Fantomasping

https://ebookmeta.com/product/ieee-mtt-v065-i01-2017-01-1st-
edition-fantomasping/

IEEE MTT V065 I04 2017 04 4th Edition Fantomasping

https://ebookmeta.com/product/ieee-mtt-v065-i04-2017-04-4th-
edition-fantomasping/
NOVEMBER 2017 VOLUME 65 NUMBER 11 IETMAB (ISSN 0018-9480)

PART II OF TWO PARTS

SPECIAL ISSUE ON RF FRONTENDS FOR MOBILE RADIO TERMINAL APPLICATIONS

Guest Editorial ......................................................................................... A. Tag and A. Hagelauer 4529
SPECIAL ISSUE PAPERS
Devices and Modeling
Design Criteria in Sizing Phase-Change RF Switches ...........................................................................
.............................................. G. Slovin, M. Xu, R. Singh, T. E. Schlesinger, J. Paramesh, and J. A. Bain 4531
X-Parameter Characterization of TC SAW Filters With Enhanced Dynamic Range ........................................
....................................................................... W. Akstaller, C. Musolff, R. Weigel, and A. Hagelauer 4541
Prediction of Nonlinear Distortion in Wideband Active Antenna Arrays .....................................................
........... K. Hausmair, S. Gustafsson, C. Sánchez-Pérez, P. N. Landin, U. Gustavsson, T. Eriksson, and C. Fager 4550
Filters and Multiplexers
Quasi-Lumped-Element Filter-Integrated Single-Pole Double-Throw Switch ................................................
.................................................................................... J. Xu, Q.-H. Cai, Z.-Y. Chen, and Y.-Q. Du 4564
Performance Analysis of Reconfigurable Bandpass Filters With Continuously Tunable Center Frequency
and Bandwidth ............... C. Schuster, A. Wiens, F. Schmidt, M. Nickel, M. Schüßler, R. Jakoby, and H. Maune 4572
Frequency-Tunable Tri-Function Filter ................................. T.-H. Lee, B. Lee, S. Nam, Y.-S. Kim, and J. Lee 4584
Reflectionless Adaptive RF Filters: Bandpass, Bandstop, and Cascade Designs .............................................
........................................................................................... D. Psychogiou and R. Gómez-García 4593
Differential Dual-Band Filters With Flexible Frequency Ratio Using Asymmetrical Shunt Branches for Wideband CM
Suppression ........................................................................................ L.-H. Zhou and J.-X. Chen 4606
Twofold Reflectionless Filters of Inverse-Chebyshev Response With Arbitrary Attenuation ...............................
................................................................................... M. Khalaj-Amirhosseini and M.-M. Taskhiri 4616
Adaptive RF Front-Ends Using Electrical-Balance Duplexers and Tuned SAW Resonators ...............................
................................. B. van Liempd, A. Visweswaran, S. Ariumi, S. Hitomi, P. Wambacq, and J. Craninckx 4621
Design Consideration on Converged Rx SAW Duplexer Module for Multiband RF Front End ...........................
.................................................................................. M. Iwaki, T. Tanaka, M. Ueda, and Y. Satoh 4629
Compact High-Isolation LTCC Diplexer Using Common Stub-Loaded Resonator With Controllable Frequencies and
Bandwidths .......................................................................................... J.-X. Xu and X. Y. Zhang 4636
(Contents Continued on Back Cover)
(Contents Continued from Front Cover)
Hybrid and Monolithic RF Integrated Circuits
A CMOS Antiphase Power Amplifier With an MGTR Technique for Mobile Applications ...............................
............................................................................................ J. Park, C. Lee, J. Yoo, and C. Park 4645
Wide-Bandwidth and High-Linearity Envelope-Tracking Front-End Module for LTE-A Carrier Aggregation
Applications ............................................................ W.-T. Tsai, C.-Y. Liou, Z.-A. Peng, and S.-G. Mao 4657
LO Generation With a Phase Interpolator Digital-to-Time Converter .........................................................
.............................................................. P. Preyler, C. Preissl, S. Tertinek, T. Buckel, and A. Springer 4669
A +25-dBm IIP3 1.7–2.1-GHz FDD Receiver Front End With Integrated Hybrid Transformer in 28-nm CMOS .....
..................................................................... I. Fabiano, M. Ramella, D. Manstretta, and R. Castello 4677
A Reconfigurable Dual-Frequency Narrowband CMOS LNA Using Phase-Change RF Switches ........................
.............................................. R. Singh, G. Slovin, M. Xu, T. E. Schlesinger, J. A. Bain, and J. Paramesh 4689
RF Transceiver Systems
A 28-GHz Low-Power Phased-Array Receiver Front-End With 360° RTPS Phase Shift Range ..........................
..................................................................................................... R. Garg and A. S. Natarajan 4703
Active Eight-Path Filter and LNA With Wide Channel Bandwidth and Center Frequency Tunability ...................
................................................................................................ B. Behmanesh and S. M. Atarodi 4715
A Wideband All-Digital CMOS RF Transmitter on HDI Interposers With High Power and Efficiency .................
................ N.-C. Kuo, B. Yang, A. Wang, L. Kong, C. Wu, V. P. Srini, E. Alon, B. Nikolić, and A. M. Niknejad 4724
A 21-dBm I /Q Digital Transmitter Using Stacked Output Stage in 28-nm Bulk CMOS Technology ...................
..................................................................... W. M. Gaber, P. Wambacq, J. Craninckx, and M. Ingels 4744
Built-in-Self-Calibration for I/Q Imbalance in Wideband Millimeter-Wave Gigabit Transmitters .........................
........................................ A. Rezola, J. F. Sevillano, I. Gurutzeaga, D. del Rio, R. Berenguer, and I. Vélez 4758
 IEEE MICROWAVE THEORY AND TECHNIQUES SOCIETY
The Microwave Theory and Techniques Society is an organization, within the framework of the IEEE, of members with principal professional interests in the field of microwave
theory and techniques. All members of the IEEE are eligible for membership in the Society upon payment of the annual Society membership fee of $24.00, plus an annual subscription
fee of $50.00 per year for electronic and print media. For information on joining, write to the IEEE at the address below. Member copies of Transactions/Journals are for personal
use only.
ADMINISTRATIVE COMMITTEE
D. W ILLIAMS , President T. B RAZIL, President Elect J. A. NAVARRO, Secretary A. A BUNJAILEH , Treasurer

A. A BUNJAILEH M. B OZZI K. G HORBANI R. H ENDERSON P. K HANNA S. KOUL M. M ADIHIAN G. P ONCHAK J. E. R AYAS -S ANCHEZ M. S ALAZAR PALMA D. S CHREURS D. W ILLIAMS
S. BARBIN T. B RAZIL R. G UPTA A. JACOB D. K ISSINGER G. LYONS D. PASQUET S. R AMAN S. R EISING A. S ANADA M. S TEER

Honorary Life Members Distinguished Lecturers Past Presidents

J. BARR R. S PARKS 2015–2017 2016–2018 2017–2019 K. W U (2016)
T. I TOH P. S TAECKER T.-W. H UANG A. M ORTAZAWI C. C AMPBELL P. ROBLIN W. Y. A LI -A HMAD T. L EE (2015)
M. JARRAHI H. Z IRATH T. NAGATSUMA N. S HINOHARA N. B ORGES C ARVALHO R. W EIGEL (2014)

MTT-S Chapter Chairs

Albuquerque: E. FARR
Argentina: A. M. H ENZE
Atlanta: K. NAISHADHAM
Austria: A. S PRINGER
Baltimore: I. A HMAD
Bangalore/India: K. V INOY
Beijing: Z. F ENG
Belarus: S. M ALYSHEV
Benelux: G. VANDENBOSCH
Boston: C. G ALBRAITH
Bombay/India: M. V. P ITKE
Brasilia: J. B EZERRA/
M. V INICIUS A LVES N UNES
Buenaventura: C. S EABURY
Buffalo: M. R. G ILLETTE
Bulgaria: K. A SPARUHOVA
Canada, Atlantic: Z. C HEN
Cedar Rapids/Central Iowa: C. G. X IE
Central & South Italy: L. TARRICONE
Central No. Carolina: Z. X IE
Central Texas: J. P RUITT
Centro-Norte Brasil:
M. V. A LVES N UNES
Chengdu: Z. N EI
Chicago: D. E RRICOLO
Cleveland: M. S CARDELLETTI
Columbus: A. O’B RIEN
Connecticut: C. B LAIR
Croatia: D. B ONEFACIC
Czech/Slovakia: J. VOVES
Dallas: R. S ANTHAKUMAR
Dayton: A. T ERZUOLI
Delhi/India: A. BASU
Denver: M. JANEZIC Macau: C. C. P ONG Pikes Peak: K. H U Sweden: A. RYDBERG
Eastern No. Carolina: T. N ICHOLS Madras/India: S. S ALIVAHANAN Poland: W. J. K RZYSZTOFIK Switzerland: M. M ATTES
Egypt: E. H ASHEESH Malaysia: M. K. M. S ALLEH Portugal: J. C ALDINHAS VAZ Syracuse: D. M C P HERSON
Finland: V. V IIKARI Malaysia, Penang: B. L. L IM Princeton/Central Jersey: W. C URTICE Taegu: Y.-H. J EONG
Florida West Coast: J. WANG Melbourne: R. B OTSFORD Queensland: K. B IALKOWSKI Tainan: H.-H. C HEN
Foothills: M. C HERUBIN Mexican Council: Rio de Janeiro: J. R. B ERGMANN Taipei: C. M ENG
France: D. BAJON R. M. RODRIGUEZ -DAGNINO Rochester: M. S IDLEY Thailand:
Germany: G. B OECK Milwaukee: S. G. J OSHI Romania: T. P ETRESCU C. P HONGCHAROENPANICH
Greece: R. M AKRI Monterrey/Mexico: Russia, Moscow: V. A. K ALOSHIN Toronto: G. V. E LEFTHERIADES
Gujarat/India: S. C HAKRABARTY R. M. RODRIGUEZ -DAGNINO Russia, Nizhny-Novgorad: Tucson: H. X IN
Harbin: Q. W U Morocco: M. E SSAAIDI G. L. PAKHOMOV Tunisia: A. G HARSALLAH
Hawaii: K. M IYASHIRO Montreal: K. W U Russia, Novosibirsk: A. YAROSLAVTSEV Turkey: B. S AKA
Hong Kong: H. W ONG Morocco: M. E SSAAIDI Russia, Saratov/Penza: M. D. P ROKHOROV Twin Cities: C. F ULLER
Houston: S. A. L ONG Nagoya: J. BAE Russia, Saint Petersburg: S. P. Z UBKO UK/RI: A. R EZAZADEH
Houston, College Station: G. H. H UFF Nanjing: W. H ONG Russia, Siberia: V. V. S UHOTIN Ukraine, East: N. K. S AKHNENKO
Hungary: L. NAGY Nanjing, Hangzhou: L. S UN Russia, Tomsk: D. Z YKOV Ukraine, Kiev: Y. P ROKOPENKO
Huntsville: H. S CHANTZ New Hampshire: E. H. S CHENK San Diego: J. TWOMEY Ukraine, Rep. of Georgia:
Hyderabad/India: S. R. N OOKALA New Jersey Coast: J. S INSKY Santa Clara Valley/San Francisco: K. TAVZARASHVILI
India: D. B HATNAGER New South Wales: Y. R ANGA N. S HAMS Ukraine, Vinnitsya:
India/Kolkata: S. S ANKARALINGAM New Zealand: A. W ILLIAMSON Seattle: S. E BADI V. M. D UBOVOY
Indonesia: E. T. R AHARDJO North Italy: G. O LIVERI Seoul: C. S EO Ukraine, West: I. I VASENKO
Israel: S. AUSTER North Jersey: A. K. P ODDAR Serbia and Montenegro: B. M ILOVANOVI Ć United Arab Emirates:
Japan: N. S UEMATSU Northern Australia: J. M AZIERSKA Shanghai: J. M AO N. K. M ALLAT
Kansai: T. I SHIZAKI Northern Canada: M. DANESHMAN Singapore: Z. YANG Uttar Pradesh/India: M. J. A KHTAR
Kingston: S. P ODILCHAK Northern Nevada: B. S. R AWAT South Africa: A. LYSKO Vancouver: S. M C C LAIN
Kitchener-Waterloo: R. R. M ANSOUR Norway: M. U BOSTAD South Australia: T. K AUFMANN Venezuela: J. B. P ENA
Lebanon: E. NASSAR Orange County: South Brazil: J. R. B ERGMANN Victoria: K. G HORBANI
Lithuania: B. L EVITAS H. J. DE L OS S ANTOS Southeastern Michigan: T. O ZDEMIR Virginia Mountain: T. A. W INSLOW
Long Island/New York: Oregon: K. M AYS Southern Alberta: E. F EAR Washington DC/Northern Virginia:
S. PADMANABHAN Orlando: K. K ARNATI Spain: J. I. A LONSO T. I VANOV
Los Angeles, Coastal: V. R ADISIC Ottawa: Q. Z ENG Springfield: P. R. S IQUEIRA Western Saudi Arabia: A. S HAMIM
Los Angeles, Metro/San Fernando: Philadelphia: A. S. DARYOUSH Sri Lanka: A. U. A. W. G UNAWARDENA Winnipeg: P. M OJABI
T. C ISCO Phoenix: S. ROCKWELL St. Louis: D. BARBOUR Xian: X. S HI

Editors-In-Chief Associate Editors

L UCA P ERREGRINI
S T ÉPHANE B ILA PATRICK FAY T ZYH -G HUANG M A M ARTIN VOSSIEK
Univ. of Pavia
XLIM Univ. of Notre Dame NTUST Univ. of Erlangen–Nürnberg
Pavia, Italy
Limoges, France Notre Dame, IN, USA Taipei, Taiwan Erlangen, Germany
J OSE C ARLOS P EDRO X UDONG C HEN A NDREA F ERRERO A RUN NATARAJAN J OHN W OOD
Universidade de Aveiro Nat. Univ. of Singapore Keysight Technol. Oregon State Univ. Obsidian Microwave, LLC
Aveiro, Portugal Singapore Santa Rosa, CA, USA Corvallis, OR, USA Raleigh–Durham, NC, USA
Editorial Assistants A LESSANDRA C OSTANZO K AMRAN G HORBANI H ENDRIK ROGIER
L AURA G RIMOLDI Univ. of Bologna RMIT Univ. Univ. of Ghent
Italy Bologna, Italy Melbourne, Vic., Australia Ghent, Belgium
A NA R IBEIRO C HISTIAN DAMM J USEOP L EE C HRISTOPHER S ILVA
Portugal Univ. Ulm Korea Univ. The Aerospace Corporation
Ulm, Germany Seoul, South Korea El Segundo, CA, USA
A. R IDDLE, Editor-in-Chief, IEEE Microwave Magazine J. S TAKE, Editor-in-Chief, IEEE Trans. Terahertz Science and Technology
N. S. BARKER, Editor-in-Chief, IEEE Microwave and Wireless Component Letters R. M IYAMOTO, Web Master

IEEE Officers
K AREN BARTLESON, President S. K. R AMESH, Vice President, Educational Activities
JAMES A. J EFFERIES, President-Elect S AMIR M. E L -G HAZALY, Vice President, Publication Services and Products
W ILLIAM P. WALSH, Secretary M ARY E LLEN R ANDALL, Vice President, Member and Geographic Activities
J OHN W. WALZ, Treasurer F ORREST D. W RIGHT, President, Standards Association
BARRY L. S HOOP, Past President M ARINA RUGGIERI, Vice President, Technical Activities
K AREN S. P EDERSEN, President, IEEE-USA
J ENNIFER T. B ERNHARD, Director, Division IV—Electromagnetics and Radiation
IEEE Executive Staff
D R . E. JAMES P RENDERGAST, Executive Director & Chief Operating Officer
T HOMAS S IEGERT, Business Administration C HERIF A MIRAT, Information Technology
J ULIE E VE C OZIN, Corporate Governance K AREN H AWKINS, Marketing
D ONNA H OURICAN, Corporate Strategy C ECELIA JANKOWSKI, Member and Geographic Activities
JAMIE M OESCH, Educational Activities M ICHAEL F ORSTER, Publications
E ILEEN M. L ACH, General Counsel & Chief Compliance Officer KONSTANTINOS K ARACHALIOS, Standards Association
VACANT, Human Resources M ARY WARD -C ALLAN, Technical Activities
C HRIS B RANTLEY, IEEE-USA
IEEE Periodicals
Transactions/Journals Department
Senior Director, Publishing Operations: F RAN Z APPULLA
Director, Editorial Services: DAWN M ELLEY Director, Production Services: P ETER M. T UOHY
Associate Director, Editorial Services: J EFFREY E. C ICHOCKI Associate Director, Information Conversion and Editorial Support: K EVIN L ISANKIE
Managing Editor: C HRISTOPHER P ERRY Journals Coordinator: C HRISTINA M. R EZES
IEEE T RANSACTIONS ON M ICROWAVE T HEORY AND T ECHNIQUES (ISSN 0018-9480) is published monthly by the Institute of Electrical and Electronics Engineers, Inc. Responsibility for the contents rests upon
the authors and not upon the IEEE, the Society/Council, or its members. IEEE Corporate Office: 3 Park Avenue, 17th Floor, New York, NY 10016-5997. IEEE Operations Center: 445 Hoes Lane, Piscataway,
NJ 08854-4141. NJ Telephone: +1 732 981 0060. Price/Publication Information: Individual copies: IEEE Members $20.00 (first copy only), nonmember $217.00 per copy. (Note: Postage and handling charge
not included.) Member and nonmember subscription prices available upon request. Copyright and Reprint Permissions: Abstracting is permitted with credit to the source. Libraries are permitted to photocopy for
private use of patrons, provided the per-copy fee of $31.00 is paid through the Copyright Clearance Center, 222 Rosewood Drive, Danvers, MA 01923. For all other copying, reprint, or republication permission,
write to Copyrights and Permissions Department, IEEE Publications Administration, 445 Hoes Lane, Piscataway, NJ 08854-4141. Copyright  c 2017 by The Institute of Electrical and Electronics Engineers, Inc.
All rights reserved. Periodicals Postage Paid at New York, NY and at additional mailing offices. Postmaster: Send address changes to IEEE T RANSACTIONS ON M ICROWAVE T HEORY AND T ECHNIQUES, IEEE,
445 Hoes Lane, Piscataway, NJ 08854-4141. GST Registration No. 125634188. CPC Sales Agreement #40013087. Return undeliverable Canada addresses to: Pitney Bowes IMEX, P.O. Box 4332, Stanton Rd.,
Toronto, ON M5W 3J4, Canada. IEEE prohibits discrimination, harassment and bullying. For more information visit http://www.ieee.org/nondiscrimination. Printed in U.S.A.

Digital Object Identifier 10.1109/TMTT.2017.2758479

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 65, NO. 11, NOVEMBER 2017 4529

Guest Editorial

T HE increasing data transmit rates per user (download

speeds of about 1 Gb/s in 4G and 20 Gb/s in 5G), paired
with the increasing number of users (people or devices) in
mobile communications, require the continuous extension of
the wireless network capacity. The demand for additional and
higher usage of the available frequency spectrum is the driving
force for the frequency spectrum adaptation at microwaves and
millimeter waves and the implementation of new 3GPP cel-
lular standards. These standards include multiband operation,
carrier aggregation, and multi-input multi-output air interfaces.
This results in RF modems with higher complexity and chal-
lenging RF requirements on the antennas and RF components
such as switches, filters, and amplifiers. Furthermore, highly
integrated modules are replacing discrete analog component
chains. They enable to handle the complexity of the system
design and provide a smaller footprint, while reducing power
consumption and improving RF performance.
A significant amount of work, both in industry and acad-
emia, is taking place in this interesting research area. There-
fore, we thought that it would be the right time to organize this
TRANSACTIONS’ “Special Issue on RF Frontends for Mobile
Radio Terminal Applications.”
The 22 papers in this Special Issue range across novel
improvements in single RF antenna interface components,
integrated RF modules, and RF receivers and transmitters.
These papers originate from a group of 11 manuscripts spe-
cially submitted to this Special Issue and from a rigorous
selection of 11 other regular issue papers.
The editorial process for this TRANSACTIONS’ Special Issue
was handled by the Guest Editors Dr. Andreas Tag and Dr.
Amelie Hagelauer, who worked with the Editor-in Chief and
the Associate Editor responsible for the regular issues of
this TRANSACTIONS. This policy ensured that all the papers
presented in this Special Issue were evaluated, not only using
the same process as regular issue papers, but also under the
same Editorial Review Board as any other regular issue paper.
Finally, we would like to thank Prof. José Carlos Pedro,
IEEE T RANSACTIONS ON M ICROWAVE T HEORY AND T ECH -
NIQUES Editor-in-Chief, and Prof. Kamran Ghorbani, Asso-
ciate Editor, for their support of this Special Issue. Our special
thanks go to all the reviewers for their effort and valuable time
and, in particular, to all the authors of this TRANSACTIONS’
Special Issue papers, whose curiosity and courage to face
challenging technical problems continue to advance radio
frequency theory, techniques, and applications.
A NDREAS TAG, Guest Editor
BAW R&D
Qorvo Inc.
Apopka, FL 32703 USA
A MELIE H AGELAUER, Guest Editor
Institute for Electronics Engineering
Friedrich-Alexander Universität Erlangen-Nürnberg
91058 Erlangen, Germany

Andreas Tag (S’14–M’15) received the B.S.E.E. and M.S.E.E. degrees from the Technische
Universitat München, Munich, Germany, in 2010 and 2012, respectively, and the Ph.D. degree
in multiphysical modeling of microacoustic RF MEMS components from Friedrich-Alexander
Universität Erlangen-Nürnberg, Erlangen, Germany, in 2016.
In 2015, he joined Qorvo Inc., Apopka, FL, USA, where he is currently involved in the new
generations of RF acoustic wave components.
Dr. Tag is a member of the VDE (German Association for Electrical, Electronic and
Information Technologies). He was nominated for the Best Paper Award of the European
Microwave Conference in 2013 and was the finalist of the Student Paper Competition of the
IEEE International Ultrasonic Symposium in 2014 and 2015, respectively. He is continuously
contributing to the development of the IEEE Microwave Theory and Techniques Society by
organizing workshops, student design competitions, and by acting as a Guest Editor. Examples
are the workshops “RF Acoustics” and “Acoustic Multiplexer for Carrier Aggregation,” as well
as student design competitions “SAW Triplexer” and “BAW Quadplexer Module” for the IEEE MTT-S International Microwave
Symposium in 2015 and 2017, respectively. He has been a Guest Editor for the special issue of IEEE Microwave Magazine
on RF Acoustic for Mobile Communication—Challenges and Solutions. In 2017, he became the Vice Chair of the MTT-2
Microwave Acoustics Technical Committee.

Digital Object Identifier 10.1109/TMTT.2017.2758558

0018-9480 © 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
4530 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 65, NO. 11, NOVEMBER 2017

Amelie Hagelauer (S’08–M’10) was born in Nuremberg, Germany, in 1981. She received the
Dipl.-Ing. degree in mechatronics and Dr.-Ing. degree in electrical engineering from Friedrich-
Alexander-Universität Erlangen-Nürnberg, Erlangen, Germany, in 2007 and 2013, respectively.
In 2007, she joined the Institute for Electronics Engineering, Friedrich-Alexander-Universität
Erlangen-Nürnberg, as a Research Engineer, where she is involved in SAW/BAW components,
RF-MEMS, packaging technologies, and integrated circuits up to 180 GHz. Since 2016, she
has been leading a research group on electronic circuits.
Dr. Hagelauer has been the Chair of MTT-2 Microwave Acoustics since 2015.
IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 65, NO. 11, NOVEMBER 2017 4531

Design Criteria in Sizing Phase-Change

RF Switches
Gregory Slovin , Min Xu, Rahul Singh , Member, IEEE, T. E. Schlesinger, Fellow, IEEE,
Jeyanandh Paramesh, Senior Member, IEEE, and James A. Bain , Senior Member, IEEE

Abstract— This paper presents design criteria for four-terminal TABLE I

phase-change (PC) RF switches as a function of their dimensions, C OMPARISON OF RF S WITCH T ECHNOLOGIES
(heater width, RF gap, and barrier thickness), materials, and
connection topology. Reducing heater width is shown to reduce
switch actuation power, with this reduction ultimately being
limited by the maximum allowed heater current density, as set
by projected reliability. Narrower RF gaps are shown to increase
switch cut-off frequency FCO , because they decrease on-state
resistance more than they increase off-state capacitance. The
balance of this increased FCO versus the estimated reduction
in power handling as the RF gap shrinks is quantified. Barrier
layers of AlN thicker than 200 nm are shown to be sufficient to
electrically decouple the switch from the heater trace. Coupling to
heater pads can also be significant unless pad sizes are reduced as
in a monolithically integrated switch. Finally, it is demonstrated
that division of the switch into multiple parallel segments is a
viable approach for lowering the heater actuation voltage for
CMOS integration. A 12.5% increase in COFF (∼1.5 fF) was associated technologies in which they have been previously
observed and 35% increase in actuation power at temperature deployed can be found in [10].
was needed in a two-segment parallel switch, as compared to a For RF switches, the most promising topology uses a four-
single segment switch. terminal configuration: 1) two actuator terminals connected to
Index Terms— Germanium telluride, GeTe, PCM switch, a resistive heater and 2) two RF terminals that are connected
phase-change (PC) material, PC switch, reconfigurable circuits,
RF switch, tungsten microheaters. by the PC element, as demonstrated in [1]–[4] and [9] and
discussed in detail in the following section. The strength of
I. I NTRODUCTION this approach is the electrical isolation between the electrical

R F SWITCHES using phase change (PC) materials have

been demonstrated in several device topologies [1]–[9].
These devices exploit the fact that PC materials can be
toggled back and forth between high conductivity (ON) and
low conductivity (OFF) states through appropriate high-speed
heating events. An excellent overview of PC materials and the
Manuscript received April 1, 2017; revised June 21, 2017 and
August 29, 2017; accepted September 20, 2017. Date of publication
October 26, 2017; date of current version November 3, 2017. This work
was supported in part by the Defense Advanced Research Projects Agency
through Radio Frequency-Field Programmable Gate Arrays Program under
Grant HR0011-12-C-0095 and in part by the Office of Naval Research
through Defense University Research Instrumentation Program under Equip-
ment Grant N000141310874. (Corresponding author: James A. Bain.)
G. Slovin was with the Department of Electrical and Computer Engineering,
Carnegie Mellon University, Pittsburgh, PA 15213 USA. He is currently with
TowerJazz, Newport Beach, CA USA (e-mail: slovin.greg@gmail.com).
M. Xu was with the Department of Electrical and Computer Engineering,
Carnegie Mellon University, Pittsburgh, PA 15213 USA. He is currently
with the Oracle Corporation, Cupertino, CA 92660 USA (e-mail:
xumin9096@gmail.com).
R. Singh, J. Paramesh, and J. A. Bain are with the Department of
ECE, Carnegie Mellon University, Pittsburgh, PA 15213 USA (e-mail:
rsingh1@andrew.cmu.edu; paramesh@ece.cmu.edu; jbain@cmu.edu).
T. E. Schlesinger is with the Whiting School of Engineering, The Johns
Hopkins University, Baltimore, MD 21218 USA (e-mail: tschles4@jhu.edu).
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/TMTT.2017.2759735
0018-9480 © 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
actuation path (the heater) and the RF signal path, and it
achieves an attractive switch figure of merit (FOM). This FOM
is the cut-off frequency FCO defined as FCO = 1/2π RON C OFF ,
where RON is the ON-state resistance while C OFF is the
OFF -state capacitance. The aforementioned isolation results in
a reduction of C OFF to as low as 10 fF. This excellent electrical
isolation compares favorably to competing technologies, some
of which are summarized in Table I.
Within Table I, only the PC switches and the MEMS tech-
nologies offer the very high FOMs. These two technologies
have contrasting strengths. Electrostatic MEMS devices, typi-
cally have a very low holding power, are volatile and require
relatively large actuation voltages and specialized packaging.
PC switches require large actuation powers (∼1 W) for short
times (100 ns, for actuation energies of 100 nJ), but are
nonvolatile when not being actuated. They operate on CMOS
compatible voltages and can be integrated rather easily into a
CMOS process flow.
An important challenge in implementing these four-terminal
devices is in managing the required temperatures and powers
without compromising this electrical isolation. The use of
a heater external to the PC material adds heated volume,
and because the heater must be hotter than the PC element,
the actuation energy is substantially higher in the four-terminal
4532 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 65, NO. 11, NOVEMBER 2017
Fig. 1. (a) Cartoon cross section of the fabricated switches. A W heater
is built on top of a sapphire substrate and is used to heat the GeTe.
An AlN barrier layer is used to minimize electrical coupling between the
RF signal path and the programming path while still maintaining the thermal
coupling needed to efficiently heat the PC material. (b) SEM of a fabricated
four-terminal RF PC switch. Transformation of GeTe in the RF gap from
the crystalline state to the amorphous state prevents the RF signal from
propagating from one Cu trace to the other.
configuration than if the heating current could flow through the
PC directly [17]. If one seeks to ameliorate this situation by
putting the PC element closer to the heater, electrical isolation
suffers. This paper seeks to quantify this trade-off and identify
optimal design points which balance these considerations.
Other switch figures of merit, like linearity are outside the
scope of this paper, but may also be affected by the sizing
of components considered here. Importantly, very promising
linearity (IIP3 > 30 dBm) has been measured by others on
similar switches to demonstrate the overall viability of this
switch technology [6].
II. D EVICE D ESIGN
A cartoon cross section and SEM top-down view of the
four-terminal RF PC switch can be seen in Fig. 1(a) and (b),
respectively. In this device, the RF signal path and program-
ming path are electrically isolated from each other. The RF
current passes though the Cu traces to the W/Au contacts and
down to the PC material, GeTe. If the PC material is in the
ON -state (low resistance) the RF current flows though the GeTe
and back up though the contact and Cu trace. If the PC material
is in the OFF-state (high resistance), RF current flow is blocked
by the GeTe. The PC material is thermally actuated between
the crystalline ON-state and amorphous OFF-state with the use
of a W heater that runs transversely to the direction of the
RF current under the PC material in the RF gap. To switch
the PC material to the amorphous state, a 100-ns-long voltage
pulse is applied to the heater whose power is enough to raise
the temperature to the GeTe above its melting temperature
723 °C [10]. The pulse is then turned off and the PC element
quickly cools and freezes in the disordered amorphous state.
To switch the PC material to the crystalline state, a 1-μs-long
voltage pulse is applied to the heater to raise the temperature
of the GeTe to its crystallization temperature, approximately
200 °C.
The heater is electricity isolated from the RF signal path
using an AlN barrier layer. This barrier layer is necessary
to minimize the electrical coupling of the heater to the RF
switch. In the OFF-state, the heater can act as an additional
capacitance across the switch and degrade its RF performance.
While other studies have used SiN as the barrier layer [2]–[4],
in this paper, AlN was integrated as the barrier layer due to
its higher thermal conductivity among candidate dielectrics.
The use of AlN allows for the heater to be thermally well-
connected to the PC layer, but electrically isolated from it.
As a result, the barrier layer thickness can be increased to
further reduce the electrical coupling of the heater the RF
switch [1].
In this paper, switches with varying dimensions (as noted
in Fig. 1) were fabricated and tested to achieve an RON
from 1–2  and C OFF < 30 fF. The heater width W H was
varied between 1 and 3 μm, the switch RF path width WRF
between the 10 and 30 μm, and RF gap length L RF between
400 and 900 nm. The thickness of the GeTe layer, tPC , was
varied between 50 and 100 nm while the AlN barrier layer
thickness, t B , was varied between 100 and 200 nm.
A. Device Fabrication Overview
The switches were fabricated using an in-house fabrication
process flow. Devices were fabricated on sapphire substrates
because they are electrically insulating while still having a rea-
sonably high thermal conductivity of 35 W/m·K. A thermally
conductive substrate is important for achieving an adequate
quenching speed of the PC material [18], [19]. As noted
in [19], this can be achieved on a substrate of low thermal
conductivity if a sufficiently thick heat spreading layer of
high thermal conductivity is present immediately under the
switch. Heaters were fabricated using dc magnetron sput-
tering of W (70 nm thick) to deposit films at a substrate
temperature of 850 °C, which produced W with low elec-
trical resistivity (75 n·m), compared to films deposited at
room temperature (typ 230 n·m). W heaters were patterned
using contact photolithography and etched in a fluorine based
parallel plate reactive ion etch (RIE) system. An AlN barrier
layer (100 or 200 nm) was then reactively sputtered over the
W heater.
The PC material, GeTe, was then deposited (50 and 100 nm)
using co-sputtering from a Ge and Te target at a substrate
temperature of 170 °C. The resistivity of the resulting crys-
talline GeTe film was 1.68 μ·m, due to precise control of
composition. The GeTe was patterned and etched using an
Ar+ ion mill. The contacts to the GeTe were made using a lift-
off process, patterned using e-beam lithography. The contacts
consisted of a 10 nm W interface layer followed by 130 nm of
Au. The GeTe and contacts were then passivated with 100 nm
of sputtered SiO2 . Vias were etched down to the Au contacts
in a Fl-based parallel-plate RIE. Next, vias were etched down
to the W heater though the AlN barrier layer using a Cl-based
ICP RIE. The traces and pads were made to the contacts and
heater using a Cu plating process. A 10-nm Ta interface layer
and 100 nm of Cu were sputtered to form the seed layer for
the Cu plating. The traces and pads were patterned and then
Cu plated to thickness of 2 μm. The Cu and Ta seed layer
were etched in an Ar ion mill. The devices were then capped
with another 100 nm of sputtered SiO2 . Openings in the SiO2
over the pads were made with a Fl-based parallel plate RIE.
SLOVIN et al.: DESIGN CRITERIA IN SIZING PC RF SWITCHES
Fig. 2. (a) 1-port and (b) 2-port measurement structures. Heater pulses are
applied using an SG probe across H1 and H2. 1-port RF measurements were
made using an SG probe across S1 and G in (a) and 2-port RF measurements
were made using an SGS probe across S1 and G as port 1 and S2 and G
as port 2 in (b). (c) Heater pulse profile used to turn the switch OFF (red
line) and ON (blue line). (d) Capacitive model (partial) of the switch in the
OFF state.
B. Test Setup
Both 1-port [Fig. 2(a)] and 2-port [Fig. 2(b)] RF test
structures were fabricated to measure the ON-state resistance
(RON ) and OFF-state capacitance (C OFF ) of the various devices.
As fabricated, the devices are in the low resistance ON-state
and were first tested to determine their minimum power to
amorphize (MPA). The dc resistances of the switches were
measured using a Keithley 2400 source meter connected across
S1 and G for the 1-ports devices [Fig. 2(a)] and across
S1 and S2 for the 2-port devices [Fig. 2(b)]. To measure MPA,
100-ns-long voltage pulses of increasing amplitude [Fig. 2(c)]
were applied using an Agilent 8114A pulse generator across
H1 and H2 [Fig. 2(a) and (b)] until the measured dc resistance
of the switch increased above 10 k. Pulses were widely
separated in time (milliseconds or greater) such that the switch
returned to room temperature each time. Moreover, in this
simple protocol, no pulses are applied to return the switch to
the ON state between amorphization pulses.
The applied power (PApp) is calculated using (1) where
VApp is the voltage applied by the pulse generator to a
50- load, R H is the room temperature resistance of the
heater, and R S is the source impedance of the pulse generator
(R S = 50 )
4V 2App R H
PApp = . (1)
(R S +R H )2
It should be noted that the applied power is a convenient
parameter to tabulate, but it differs from the actual time-
varying power absorbed by the heater, because R H is a
function of temperature and time. For clarity in the definition
of PApp and VApp, we note that in the simple case of R H = R S ,
the total power supplied by the signal generator is 2V 2App /R S
and the PApp would be half of this (with the other half being
dissipated in the source). It should also be noted that we
also define VI , the incident voltage, which is equal to the
actual voltage that appears across the heater, since it is not
a matched load. Unlike VApp, VI is time and temperature
dependent and is given by VI = 2VApp R H /(R H + R S ). In all
4533
of the analysis of the voltage waveforms applied to the heater,
electrical transients are ignored as the electrical time constant
of the heater is orders of magnitude smaller than the thermal
time constant.
The minimum power to crystalize (MPC) is determined in
a manner similar to the MPA procedure, by applying, in this
latter case, 1-μs-long voltage pulses of increasing amplitude
[Fig. 2(c)] across the heater until the resistance of the switch
is within 20% of its initial as fabricated RON value.
RF measurements were made using an Agilent 8364A
PNA Network Analyzer. Using the device layout in Fig. 2(a),
1-port S-parameter measurements were made from 45 MHz to
20 GHz across S1 and G. Two-port S-parameter measurements
were made on the device layout in Fig. 2(b) from 45 MHz
to 14 GHz across S1 and S2 while sharing a common
ground (G). All S-parameter measurements were de-embedded
using standard de-embedding techniques [20] to extract the
RON and C OFF of the switches. The switch can be modeled as
a resistance and capacitance, C OFF , in parallel. The total C OFF
can be further broken down into three parallel capacitances:
1) CTT , the capacitance across the RF gap from one trace
to the other; 2) the series combination of two instances of
CTH , the series capacitances between the trace and heater
[Fig. 2(d)]; and 3) series combinations of non de-embedded
fringing capacitance (C F ) to the heater pads. This last term
is discussed in detail in Section III-D on grounded heaters.
All RF-measurements were made with the heater pads left
floating to allow for more accurate measurement of the RON
and C OFF [2] unless otherwise noted.
III. E FFECTS OF D EVICE S IZING AND
C ONNECTION T OPOLOGY
Correctly sizing the switch dimensions is important to
maximize RF performance. This entails trade-offs between
performance and power consumption. This section will discuss
the effects of heater width, RF gap length and AlN barrier layer
thickness on FCO and MPA of the RF switches.
A. Effect of Heater Width
Two-port measurement structures were fabricated to deter-
mine the effect of heater width. Fabricated switches were
20-μm-wide WRF , with heaters widths W H of 1, 2, and 3 μm.
The GeTe and AlN barrier were both 100 nm thick in these
devices. The MPA and MPC were measured for these switches
and are shown in Fig. 3(a). The MPA increases by 0.5 W per
micrometer increase in heater width, while the MPC increases
by 0.1 W per micrometer. This increase in power can be
attributed to an increase in heated volume and a resulting
increase thermal capacitance and thermal time constant of
the switch. Therefore, to melt the PC material in the same
100-ns-long voltage pulse, the power delivered to the heater
must be increased.
The de-embedded insertion loss of the switches in the
ON -state with the three different heater widths is shown in
Fig. 3(b). As expected the insertion loss is not dependent on
heater width and is about 0.1 dB up to 14 GHz. However,
the de-embedded isolation of the switches is dependent on the
4534 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 65, NO. 11, NOVEMBER 2017
Fig. 3. (a) MPA ( red dashed line) and MPC ( blue dashed line) for 20-μm
switches with header widths of 1, 2, and 3 μm. Each marker is a separate
device. (b) Insertion loss and (c) isolation 20-μm switches with header widths
of 1 (red line), 2 (blue line), and 3 μm (green line) from 45 to 14 GHz.
(d) Extracted C OFF for 20-μm switches with header widths of 1, 2, and 3 μm.
heater width [Fig. 3(c)]. The isolation at 10 GHz decreases
from 21 dB for a switch with a 1-μm-wide heater to less
than 18 dB for one with a 3-μm-wide heater. The extracted
C OFF for the different heater widths can be seen in Fig. 3(d).
As the width of the heater increases, C OFF increases by 4.3 fF
per micrometer. This can be attributed to an increase in CTH
[Fig. 2(d)]. The intercept of the fit (“0”-μm-wide heater)
in Fig. 3(d) is 9.7 fF which includes CTT and C F .
Thus, reducing heater width helps reduce switching power
as well as improving isolation, so it should be as narrow
as possible as allowed by the current density limitations of
nanoscale conductors. This is quantified within Section III-E,
Summary of Design Trade-offs. As will be shown in a
subsequent section on AlN thickness, this benefit of narrow-
ing heater width for isolation disappears when the dielectric
between the heater and the PC layer is thick enough. This
operating limit will be assumed in Section III-E.
B. Effect of RF Gap Length
Reducing the RF gap length has the obvious benefit of
reducing RON due to resistance being proportional to length.
However, the cumulative effects of RF gap length on RF
performance and power are not as obvious. One-port RF
measurement structures were fabricated to determine the effect
of RF gap length L RF . The fabricated switches had switch
width WRF = 20 μm, with heater widths W H = 1 μm.
The GeTe and AlN barrier layer thickness were 50 (tPC ) and
100 nm (t B ) thick, respectively. The RF gap lengths L RF
studied ranged from 400 to 900 nm in 100-nm increments.
Fig. 4(a) and (b) shows the MPA and MPC of the switches
with RF gap lengths ranging from 400 to 900 nm, respectively.
As L RF decreases, the MPA increases. This is due to the Au
contacts being closer together and sinking more heat out of
the PC material. To compensate for the increased heat sinking,
more power is needed to raise the temperature of the PC
material above its melting temperature.
Fig. 4. (a) MPA and (b) MPC results for 20-μm-wide switches with RF gaps
ranging from 400 to 900 nm in length. (c) Extracted RON and (d) C OFF of
switches with varying L RF from 45 MHz to 20 GHz. As the RF gap length
L RF reduces from 900 to 400 nm, RON reduces from 3 to 1.4  and C OFF
increases from 11.7 to 15.4 fF at 5 GHz.
Fig. 5. (a) FCO and (b) MPA (red dots) and MPC (blue squares) for
20-μm-wide switches with RF gaps ranging from 400 to 900 nm in length.
The FCO is the highest, 7.2 THz, for the shortest RF gap length of 400 nm.
The extracted RON and C OFF from 45 MHz to 20 GHz can
be seen in Fig. 4(c) and (d), respectively. As expected RON
decreased with reducing L RF . RON decreases from 3 to 1.4 
when L RF is reduced from 900 to 400 nm. However, as the
RON decreases, C OFF increases. C OFF increase from 11.7 to
15.4 fF as L RF is decreased.
Because RON and C OFF move in opposite directions with
RF gap length, it is not immediately obvious how the RF
performance of the switch is affected by its L RF . Fig. 5(a)
shows the FCO of each switch versus their corresponding L RF .
The FCO is the greatest, 7.2 THz, for the shortest L RF of
400 nm. As L RF is increased to 900 nm, the FCO decreases
to 4.2 THz. Compared to C OFF , RON is more dependent on
L RF . L RF is the largest contribution to the ON-state resistance
of the switch while, it only contributes to CTT , and not the
other capacitances in C OFF , CTH , and C F . There is a trade-off,
however, between RF performance and the required power to
amorphize the PC material. The MPA is 1.8 W for an L RF
of 400 nm, but reduces to less than 1.5 W for L RF longer than
700 nm [Fig. 5(b)]. The increase in RF performance outweighs
the cost of the increase in MPA even for the narrowest L RF
of 400 nm. The FCO /MPA is 4 THz/W for an L RF of 400 nm
and only 2.8 THz/W when L RF is increased to 900 nm.
SLOVIN et al.: DESIGN CRITERIA IN SIZING PC RF SWITCHES
Fig. 6. (a) Insertion loss and (b) isolation for 20-μm switches with AlN
barrier layer thickness of 100 nm (red) and 200 nm (blue) from 45 MHz
to 14 GHz. The increased insertion loss for the switch with 200-nm AlN is
due to a GeTe thickness of 50 nm compared to 100 nm for the switch with
100 nm of AlN. This resulted in an ON-state resistance of 3.4 and 1.5 ,
respectively. (c) OFF-state capacitance for switch widths of 10, 20, and 30 μm
with 100 nm (red dashed line) and 200 nm (blue dashed line) of AlN. The
fringing capacitance calculated to be 3.7 fF, with a per micrometer capacitance
of 0.51 and 0.32 fF for 100 and 200 nm of AlN, respectively. (d) OFF-state
capacitance of 20-μm-wide switches with varying AlN barrier layer thickness
from this paper and [1]. Points indicate measurement on distinct switches.
While RF performance improves with decreasing RF gap,
RF power handling will degrade as RF gap shrinks. While
power handling was not studied in detail in this paper, we can
examine this trade-off under the assumption that the power
handling is limited by the electric field that the OFF-state
device can handle. This is quantified in Section III-E.
C. Effect of AlN Thickness
The AlN barrier layer is responsible for electrically isolating
the heater from the PC material while still keeping them
thermally coupled. Despite thin-film AlN having a thermal
conductivity as high as of 130 W/m·K [21], increasing the
barrier layer thickness to reduce CTH does result in an increase
in MPA, due to an increase in thermal capacitance [1].
Two-port S-parameter measurements were made on devices
with various switch widths WRF of 10, 20, and 30 μm to
determine at what AlN thickness t B , the capacitance attributed
to the heater disappears. The switches measured all had a W H
of 1 μm with one set of switches having, tPC of 50 nm with
a t B of 200 nm and the second set of switches having a tPC
of 100 nm with a t B of 100 nm.
Fig. 6(a) and (b) shows the insertion loss and isolation,
respectively, from 45 MHz to 14 GHz for W H of 20 μm with a
t B of 100 (red) and 200 nm (blue). The difference in insertion
loss, 0.1 dB for t B of 100 nm and 0.26 dB for t B of 200 nm at
10 GHz, is due to the difference in tPC . The isolation increased
from 21 dB to 25 dB at 10 GHz when t B is increased from
100 to 200 nm. The OFF-state capacitance was extracted for
WRF of 10, 20, and 30 μm for both t B of 100 and 200 nm
are shown in Fig. 6(c).
4535
Fig. 7. ON -state electrical model of switch when the heater terminal
is (a) floating and (b) grounded. When the heater terminal is grounded some
of the RF signal propagates though CTH down to the grounded heater.
The switches with t B of 100 nm have an increase in
capacitance of 0.51 fF per micrometer of switch width, while
switches with a t B of 200 nm only have an increase in
capacitance of 0.32 fF per micrometer of switch width. Both
the t B of 100 and 200 nm have a fringing capacitance
(the intercept at “0-μm wide” switch) of 3.7 fF [Fig. 6(c)].
Earlier we showed that for a WRF of 20 μm, C F plus
CTT was equal to 9.7 fF. Therefore, CTT must be equal
to 6 or 0.3 fF per micrometer of switch width. The slope of
the fits in Fig. 6(c) is equivalent to CTT + CTH /2. When CTT ,
0.3 fF/μm, is subtracted from the slope of the fits, CTH /2 is
0.21 and 0.02 fF/μm for a t B of 100 and 200 nm, respectively.
This shows for a t B of 200 nm, the heater contribution to C OFF
is essentially zero. Additional increases in the AlN thickness
will not further reduce C OFF , but will increase the MPA of
the switch. Fig. 6(d) shows C OFF for WRF of 20 μm versus
the inverse of AlN thickness t B for this paper and previous
work in [1]. It becomes obvious that minimum C OFF for this
switch design is approximately 10 fF and cannot be reduced
any further by increasing the AlN barrier layer thickness. For
the summary of design trade-offs in Section III-E, we assume
that the AlN thickness is sufficient to eliminate CTH entirely.
D. Effect of Grounded Heaters
Making S-parameter measurements with the heater pads left
floating is necessary for an accurate extraction of RON and
C OFF [2]. However, this is most likely not how the switch
would be implemented in a real RF system. Ideally, when the
switch is in ON-state, all of the RF signal propagates from one
side of the switch to the other [Fig. 7(a)] and the insertion loss
of the switch is minimized. However, if one side of the heater
is grounded, this will produce a capacitive shunt to ground,
resulting in some of the RF signal propagating though CTH
to ground [Fig. 7(b)]. This results in degradation of the RF
signal propagation from one side of the switch to the other
and an increase in insertion loss at higher frequencies. This
RF “sneak” path shows the importance of minimizing CTH by
increasing t B to 200 nm.
Fig. 8(a) and (b) shows the insertion loss and isolation,
respectively, for WRF of 20 μm for both t B of 100 (red)
and 200 nm (blue) when the heater probes are up (dotted)
and landed (solid). The offset in insertion loss with the
heater probes up (dotted) between the t B of 100 and 200 nm
4536 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 65, NO. 11, NOVEMBER 2017
Fig. 8. (a) Insertion loss and (b) isolation of 20-μm switches with AlN
barrier layer thickness of 100 nm (red) and 200 nm (blue), with the heater
probes up (floating heater) and with them landed (grounded heater). It can be
seen that with a grounded heater the insertion-loss increases with frequency.
Fig. 9. (a) Zoomed-in view of switch showing the fringing capacitance (C F )
and capacitance to the heater pads (CHP ). (b) De-embedding structure used
to measure CHP . (c) Electrical model of 2-port measurement with switch in
the ON-state when the heater probes are landed. The value of C F extracted
from applying this model is 4 fF while CHP is 19 fF.
[Fig. 8(a)] is due to a difference in tPC . When the heater probes
are landed, the insertion loss increases with frequency while
when the heater probes are up it does not. For a t B of 100 nm,
the insertion loss increases from 0.1 to 0.34 dB and increases
from 0.26 to 0.45 dB for a t B of 200 nm at 10 GHz when
the heater probes are landed. The isolation is also affected by
landing the heater probes. At 10 GHz, the isolation increases
from 21 to 24 dB for a t B of 100 nm and from 25 to 28 dB
for a t B of 200 nm. The increase in both the insertion loss
and isolation is due to the switch being capacitively shunted
to ground.
In Section III-C it was shown that the CTH for a t B of
200 nm is essentially 0 fF so there should no degradation
of insertion loss though the heater, however, there are other
capacitances that can couple to the heater pads which would
result in a degradation in insertion loss even for switches
with a t B of 200 nm. Fig. 9(a) shows a zoomed-in image
of a 2-port measurement structure showing C F as well as the
capacitance to the heater pads CHP . A through de-embedding
structure [Fig. 9(b)] was measured with the heater probes
landed to extract CHP . Fig. 9(c) shows the electrical model
of a 2-port measurement structure with a switch in the ON-
state when the heater probes are landed. CHP was measured
to be 19 fF and C F was previously shown to be about 4 fF.
Fig. 10 shows the measured (solid line) and modeled (dashed
line) change in insertion loss of a through-structure (green),
a switch with a t B of 100 (red) and 200 nm (blue) when the
Fig. 10. Measured (solid line) and modeled (dashed line) insertion loss of the
through structure (green), switch with 200 nm (blue) of AlN and 100 nm (red)
of AlN. The through structure shows significant degradation in the insertion
loss despite not having a heater present. This is due to the presence of heater
probe pads (CHP ). The degradation of the of the 200-nm AlN switch is due
to CHP and C F while the 100-nm AlN switch degradation is the largest due
the additional capacitance to the heater (CTH ).
heater probes are landed. The through-structure with no heater
shows significant degradation in insertion loss just from the
capacitance to the heater pads CHP . For a switch with a t B
of 200 nm, there is additional degradation in insertion loss
due to C F despite CTH being close to 0 fF. The insertion loss
degrades even further for a switch with only a t B of 100 nm
due to CTH being approximately 8 fF. This shows that while
reducing CTH is important to maximizing RF performance
of the device, careful consideration needs to be made when
laying out traces to the heater as their coupling capacitance
to the switch can significantly degrade the RF performance of
the system if laid out incorrectly. For the summary of design
trade-offs discussed in the following section, we assume that
monolithic integration and layout adjustments have been used
to eliminate any significant RF coupling to the heater leads.
The elimination of this coupling is the appropriate way to
eliminate the RF leakage to the heater, as contrasted to trying
to introduce series resistance in to the heater driver circuit,
which will put unacceptable burden on the drive circuitry.
E. Summary of Design Trade-Offs
Taking all of the above measurements into consideration,
we arrive as some summary guidance for sizing these switches
and what performance might be expected. Fig. 11(a) shows a
summary of switch cut-off frequency and max power handling
of a switch as a function of RF gap, while Fig. 11(b) shows
the required heater power and heater current density as a
function of heater width. A series of assumptions derived from
the measurements in the previous sections have led to these
figures, which are detailed in the following list.
Assumptions of Fig. 11:
1) All coupling between the heater and the PC switch has
been eliminated. This is reasonable if the AlN layer is
200 nm or more in thickness and if monolithic integra-
tion has eliminated the coupling to the heater through
reduction of lead size and routing of lead position.
2) Per the above, the only capacitances that enter the
calculation of FCO are CTT and C F , and they are
assumed to follow the measurements shown above,
i.e., the capacitance as a function of RF gap follows
Fig. 4(d) with the appropriate subtractions.
SLOVIN et al.: DESIGN CRITERIA IN SIZING PC RF SWITCHES
TABLE II
S IMULATED MPA OF 20-μm-W IDE S WITCH FOR VARIOUS H EATER M ATERIALS
Fig. 11. Key design trade-offs in PC switches. (a) Switch performance trade-
off versus RF gap. (b) Switch power requirements and trade-off in resulting
heater current densities. In (a) two PC layer thicknesses tPC are considered
(100 nm with solid lines and 200 nm with dashed) and two switch widths
W S are labeled: 10 and 30 μm. In (b) two switch widths W S are considered
(10 μm with solid lines and 30 μm with dashed) and two heater thicknesses t H
are labeled: 100 and 500 nm.
3) PC ON-state resistivity is 3 × 10−6 ·m.
4) Series resistance for all switches is 0.25 ;
5) Contact resistance is assumed to be 0.5  for a 20-μm-
wide switch and scales inversely with switch width.
6) Breakdown electric fields of the PC in the OFF-state set
the switch power handling. Power handling for values
between 10 and 30 V/um are shown in Fig. 11(a), which
are typical
7) Heaters are W with the properties for films deposited at
400 °C (Table II).
8) Required heater power was assumed independent of
RF gap (which is an approximation of the observed
4537
behavior), with the 600-nm RF gap taken as representa-
tive.
9) Each switch requires 0.8 W as a baseline power, and for
every micrometer of switch width each heater requires
25 mW per μm of heater width [Fig. 3(a)]. (i.e.,
A 20-μm-wide switch with 2-μm-wide heater requires
1.8 W of power).
10) Heater current density is for the heater at temperature.
11) PC thickness of 100 and 200 nm are shown.
The key trade-offs in sizing PC switches with regard to
RF gap, PC layer thickness, switch width, heater width, and
heater thickness are tabulated in Fig. 11. The choice of RF
gap [Fig. 11(a)] balances raising FCO while lowering power
handling. As noted in earlier sections, reducing the RF gap
improves R more than it degrades C and this increases FCO .
Selecting heater width generally means making it as narrow
and as thick as possible, to reduce power. However, processing
constraints and acceptable current densities will limit how
narrow it can be. While this dependence was not studied in this
paper, Fig. 11(b) shows how heater current density scales with
heater width for several structures. The mean time to failure
for high current densities typically scales inversely with the
square of the current density [24]. Demonstrations of heater
cycling for over 1000 cycles have been shown with current
densities of 100 MA/cm2 [18].
IV. S CALING FOR CMOS I NTEGRATION
These RF switches have already shown promising results
when combined with RF CMOS circuits using flip chip
technology [9]. In [9], the RF switch was integrated into the
RF signal path, but the heater was still driven from off-chip
using a pulse generator. For true CMOS integration, the heater
would also have to be driven by on-chip CMOS circuitry.
This presents some challenges, as the incident voltages across
the heater in this paper and in [9] are as high as 8 V. This
would be too high for typical CMOS processes, though more
expensive high-voltage processes are available. Reducing the
voltage needed to turn the switch OFF is desirable for the most
general CMOS integration, as most CMOS circuits will be
limited by breakdown voltage rather than the ability to source
current.
A. Choosing a Heater Material
In the switch design presented above, the heaters are fab-
ricated using W that is deposited at 850 °C, which would
4538 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 65, NO. 11, NOVEMBER 2017
Fig. 12. Measured (a) resistivity and (b) TCR for W sheet films deposited
at 25 °C, 400 °C and 850 °C. As deposition temperature increase to 850 °C,
the resistivity decreases to 75.2 n-m and the TCR increases to 3.3 ppt/°C.
The sheet films were measured using four-wire resistance measurement of
Van der Pauw structures (insert).
not be back-end CMOS compatible. Using W deposited at
elevated temperature is shown to improve heater reliability,
and it also reduces the resistance of the heater allowing
for a lower actuation voltage. On the other hand, elevated
temperature deposition also increases the temperature coef-
ficient of resistance (TCR) of the heater. W heaters were
fabricated with deposition temperatures of 25 °C, 400 °C and
850 °C with resistivity and TCR measured using Van der Pauw
structures (Fig. 12). The deposition temperature of 400 °C
was chosen because it is back-end CMOS compatible [25].
As the deposition temperature increases, the resistivity of the
W decreases from 230 to 75.2 n·m while the TCR increases
from 1.3 to 3.3 ppt.
Using COMSOL Multiphysics software, the MPA was sim-
ulated using the same setup as in [1] for the three different
W deposition temperatures as well as NiCr, which has been
used as a heater in other studies [2]–[4]. Table I shows the
results of simulated MPA operating points, using the measured
resistivity and TCR of the measured W films and NiCr [23].
The initial incident voltage across the heater, VI , as well
as MPA decreases with decreasing heater resistivity. Even at
the end of the 100-ns-long voltage pulse, the resistance of
the heater at elevated temperature is still the lowest for W
deposited at 850 °C despite having the highest TCR. VI at
100 ns (at temperature) is also the lowest for the 850 °C.
This shows that minimizing the resistivity of the W heater
reduces actual switching power, and is desirable when trying
to integrate the switch with a CMOS driver despite the increase
in TCR. Much of the benefit in W behavior is available when
deposition is limited to the back-end compatible temperature
of 400 °C.
B. Segmenting the Switch
While lower resistivity helps reduce required heater volt-
ages, it is still not enough to get to the voltage levels in
standard CMOS processes. There are additional approaches:
1) increasing the thickness of the heater and 2) segmenting
the switch into multiple parts. While increasing the thickness
of heater seems like the most straightforward solution to
reducing the heater resistance, the heater would have to be over
1-μm thick for CMOS compatibility. As a result, the thermal
capacitance of the heater and thermal time constant of the
switch would increase significantly. This would result in an
Fig. 13. 2-port measurement structure of two parallel 10-μm-wide switches.
The heaters are also in parallel and are pulsed simultaneously using a GSG
probe across H1, H2, and H3. S-parameter measurements from port-1 across
S1 and G to port-2 across S2 and G using an SGS probe.
increase in MPA and may also prevent the PC material from
quenching fast enough to switch to the amorphous state.
Segmenting the switch into multiple sections and placing the
heaters in parallel would allow a reduction in effective heater
resistance seen by the CMOS driver as well assure that the
PC material can still be amorphized.
Fig. 13 shows a fabricated 2-port measurement structure that
has two 10-μm-wide switches, (WRF = 10 μm), in parallel
instead of a single switch with WRF = 20 μm. In this switch,
the heater pulses are applied using a GSG RF probe where
H1 and H3 are grounded and H2 is the applied voltage from
the pulse generator. The current from each heater pulse is thus
divided into both heaters in parallel. Table III highlights the
difference in the MPA heater pulses applied in this manner
as compared to those applied to the single heater in the
20-μm-wide switch.
Breaking the switch into two sections clearly reduces
the required voltage drop, VI , required across each heater,
by around a factor of 2 (slightly more than 2× at room temp
and slightly less than 2× at maximum heater temperature). The
VApp does not appear to change much, but that is because of
the very different impedances of the two heater configurations.
The 10 μm (×2) configuration is much closer to being a
matched load to the pulse generator at temperature, while the
20 μm (×1) has an impedance of almost 140  while hot.
Interestingly, breaking the switch into two pieces results in
a lower apparent MPA, but actually consumes more power
when the heater is at its maximum temperature (100 ns). The
lower MPA value is, again, due to the difference between room
temperature and maximum heater temperature. It is expected
that dividing the heaters will have some power penalty as heat
flows out of the ends as well as along the heater length and
the divided heater has twice as many ends, so this behavior is
reasonable and is supported by simulation.
For comparison of RF performance, S-parameter measure-
ments on the 10 μm (×2) switch are made between S1 and
S2 using a common ground pad (G). The insertion loss
and isolation of two switches with a WRF of 10 μm in
parallel (green line) is compared to a single switch with a
WRF of 20 μm (blue line) in Fig. 14. The insertion loss of
both switches is about 0.12 dB [Fig. 14(a)] but the ON-state
resistance of the parallel switches is slightly higher (2.1 )
SLOVIN et al.: DESIGN CRITERIA IN SIZING PC RF SWITCHES
TABLE III
S INGLE V ERSUS PARALLEL S WITCHES
Fig. 14. (a) Insertion loss and (b) isolation of a single 20-μm-wide switch
(blue line) and two 10-μm-wide switches in parallel (green line). The parallel
switches have a slightly higher RON of 2.1  compared to 1.8  for a single
switch. C OFF is also higher for the parallel switches 14.4 fF compared to
12.8 fF for a single switch. This is due to the presence of additional fringing
capacitance from the extra heater.
compared to the single switch (1.8 ). The isolation is lower
for switches in parallel, 21 dB, compared to 23 dB for a single
switch at 10 GHz. This results in an extracted C OFF of 14.4 fF
for the parallel switches and 12.8 fF for the single switches.
This is due to the presence of additional fringing capacitance
from the additional heater.
Thus, the strategy of breaking the switch into segments does
indeed reduce the actuation voltage. Of course, the division
into two segments, here, is simply a proof of concept. It did
not go far enough to be CMOS compatible as the heater
voltage was still over 7 V at temperature in the two-segment
topology. This suggests that CMOS driven designs would need
4–6 heater segments. Moreover, the approach shown where
the heaters were driven simultaneously and in parallel was
simply a matter of convenience. In a fully integrated design,
it heaters might be driven in parallel banks or singly as circuits
and routing constraints dictate. Finally, the difference between
VApp and VI is likely to be much smaller and much less
sensitive to heater resistance as one would likely seek to make
the source resistance, R S , of an integrated CMOS current
driver well below the heater resistance, so as not to waste
voltage headroom.
V. C ONCLUSION
In summary, then, we observe a number of trends and
correlations that can guide the design and sizing of PC RF
switches. The general rules are as follows. First, heaters should
be as narrow as possible as limited by current density. This
makes more efficient use of power and reduces capacitive
coupling to the heater. The RF gap should also be as narrow
as possible to get high switch FCO , as resistance is reduced
by more than capacitance is increased. However, more power
will be needed to actuate a switch with a narrower RF gap
4539
due to the heat sinking of the more closely spaced leads and
the power handling will drop due to increased electric fields
(leading to breakdown) across the switch for the same standoff
voltage.
The dielectric barrier between the heater and the PC layer
needs to be 200 nm thick, but not much more. At this
thickness, the capacitive coupling to the heater stripe is negli-
gible. Making this layer thicker will simply increase the heat
capacity of the switch, slow it down, and require more power,
without significantly improving its RF performance. Coupling
to the heater pads, however, may not be negligible at this
200-nm thickness, and must be considered in the switch layout.
In particular, the coupling path to the heater pads will induce
significant insertion loss if the heater is grounded. Of course,
a fully integrated CMOS solution, as suggested in [19] would
eliminate the need for bond pads and would greatly ameliorate
this issue.
Finally, it has been demonstrated that division of the switch
into multiple parallel segments is a viable approach for low-
ering the heater actuation voltage in anticipation of CMOS
integration. A 12.5% increase in C OFF (∼1.5 fF) and 35%
increase in actuation power at temperature was needed in a
two-segment parallel switch, as compared to a single segment
switch. This is due to the additional fringing capacitance and
the additional heat sinking associated with the end of each
heater. The number of switch segments likely needed for
CMOS integration would be 4–6, based on typical CMOS
voltages.
R EFERENCES
[1] G. Slovin, M. Xu, J. Paramesh, T. E. Schlesinger, and J. A. Bain, “AlN
barriers for capacitance reduction in phase-change RF switches,” IEEE
Electron Device Lett., vol. 37, no. 5, pp. 568–571, May 2016.
[2] N. El-Hinnawy et al., “12.5 THz Fco GeTe inline phase-change
switch technology for reconfigurable RF and switching applications,”
in Proc. IEEE Compound Semiconductor Integr. Circuit Symp. (CSICS),
Oct. 2014, pp. 1–3.
[3] N. El-Hinnawy et al., “A 7.3 THz cut-off frequency, inline, chalcogenide
phase-change RF switch using an independent resistive heater for
thermal actuation,” in Proc. IEEE Compound Semiconductor Integr.
Circuit Symp. (CSICS), Oct. 2013, pp. 1–4.
[4] N. El-Hinnawy et al., “A four-terminal, inline, chalcogenide phase-
change RF switch using an independent resistive heater for thermal
actuation,” IEEE Electron Device Lett., vol. 34, no. 10, pp. 1313–1315,
Oct. 2013.
[5] C.-Y. Wen, G. Slovin, J. A. Bain, T. E. Schlesinger, L. T. Pileggi,
and J. Paramesh, “A phase-change via-reconfigurable CMOS LC
VCO,” IEEE Trans. Electron Devices, vol. 60, no. 12, pp. 3979–3988,
Dec. 2013.
[6] M. Wang, F. Lin, and M. Rais-Zadeh, “Performance measurements and
non-linearity modeling of GeTe phase change RF switches with direct
and indirect heating schemes,” in IEEE MTT-S Int. Microw. Symp. Dig.,
May 2015, pp. 1–4.
4540 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 65, NO. 11, NOVEMBER 2017
[7] M. Wang, Y. Shim, and M. Rais-Zadeh, “A low-loss directly heated
two-port RF phase change switch,” IEEE Electron Device Lett., vol. 35,
no. 4, pp. 491–493, Apr. 2014.
[8] C.-Y. Wen et al., “A phase-change via-reconfigurable on-chip inductor,”
in IEDM Tech. Dig., Dec. 2010, pp. 10.3.1–10.3.4.
[9] R. Singh et al., “A 3/5 GHz reconfigurable CMOS low-noise amplifier
integrated with a four-terminal phase-change RF switch,” in IEDM Tech.
Dig., Dec. 2015, pp. 25.3.1–25.3.4.
[10] S. Raoux and M. Wuttig, Eds., Phase Change Materials: Science and
Applications. Boston, MA, USA: Springer, 2009.
[11] A. Botula et al., “A thin-film SOI 180 nm CMOS RF switch technology,”
in Proc. IEEE Top. Meeting Silicon Monolithic Integr. Circuits RF Syst.,
Jan. 2009, pp. 1–4.
[12] P. Hurwitz, R. Kanawati, K. Moen, E. Preisler, S. Chaudhry, and
M. Racanelli, “Advances in RF foundry technology for wireless and
wireline communications,” in Proc. IEEE 16th Top. Meeting Silicon
Monolithic Integr. Circuits RF Syst. (SiRF), Jan. 2016, pp. 5–8.
[13] T. Boles, J. Brogle, D. Hoag, and D. Curcio, “AlGaAs PIN diode multi-
octave, mmW switches,” in Proc. IEEE Int. Conf. Microw., Commun.
Antennas Electron. Syst. (COMCAS), Nov. 2011, pp. 1–5.
[14] R. S. Howell et al., “Low loss, high performance 1-18 GHz SPDT based
on the novel super-lattice castellated field effect transistor (SLCFET),” in
Proc. IEEE Compound Semiconductor Integr. Circuit Symp., Oct. 2014,
pp. 3–6.
[15] R. Stefanini, M. Chatras, P. Blondy, and G. M. Rebeiz, “Miniature
MEMS switches for RF applications,” J. Microelectromech. Syst.,
vol. 20, no. 6, pp. 1324–1335, Dec. 2011.
[16] A. Fruehling, R. Pimpinella, R. Nordin, and D. Peroulis, “A single-
crystal silicon DC-40 GHz RF MEMS switch,” in IEEE MTT-S Int.
Microw. Symp. Dig., Jun. 2009, pp. 1633–1636.
[17] M. Wang and M. Rais-Zadeh, “Development and evaluation of ger-
manium telluride phase change material based ohmic switches for RF
applications,” J. Micromech. Microeng., vol. 27, no. 1, p. 13001, 2017.
[18] M. Xu, “Study of the crystallization dynamics and threshold voltage
of phase change materials for use in reconfigurable RF switches and
non-volatile memories,” Ph.D. dissertation, Dept. Elect. Comput. Eng.,
Carnegie Mellon Univ., Pittsburgh, PA, USA, 2017.
[19] N. El-Hinnawy et al., “Substrate agnostic monolithic integration of the
inline phase-change switch technology,” in IEEE MTT-S Int. Microw.
Symp. Dig., May 2016, pp. 1–4.
[20] T. E. Kolding, “On-wafer calibration techniques for giga-hertz CMOS
measurements,” in Proc. Int. Conf. Microelectron. Test Struct. (ICMTS),
Mar. 1999, pp. 105–110.
[21] C. Duquenne, M.-P. Besland, P. Y. Tessier, E. Gautron, Y. Scudeller,
and D. Averty, “Thermal conductivity of aluminium nitride thin films
prepared by reactive magnetron sputtering,” J. Phys. D, Appl. Phys.,
vol. 45, no. 1, p. 15301, Jan. 2012.
[22] D. Yu, S. Brittman, J. S. Lee, A. L. Falk, and H. Park, “Minimum voltage
for threshold switching in nanoscale phase-change memory,” Nano Lett.,
vol. 8, no. 10, pp. 3429–3433, Oct. 2008.
[23] J. S. Faughn and R. A. Serway, College Physics. Pacific Grove, CA,
USA: Thomson Learning Inc, 2003.
[24] J. R. Black, “Electromigration—A brief survey and some recent results,”
IEEE Trans. Electron Devices, vol. 16, no. 4, pp. 338–347, Apr. 1969.
[25] S. Sedky, A. Witvrouw, H. Bender, and K. Baert, “Experimental
determination of the maximum post-process annealing temperature for
standard CMOS wafers,” IEEE Trans. Electron Devices, vol. 48, no. 2,
pp. 377–385, Feb. 2001.
Gregory Slovin received the B.S. and M.S. degrees
in electrical engineering from the University of
Florida, Gainesville, FL, USA, and the Ph.D. degree
in electrical engineering from Carnegie Mellon Uni-
versity, Pittsburgh, PA, USA, in 2015.
He is currently a Process Development Engineer
with TowerJazz, Newport Beach, CA, USA.
Min Xu received the B.S. degree in electrical engi-
neering from the Huazhong University of Science
and Technology, Wuhan, China, in 2012, and the
M.S. and Ph.D. degrees in electrical and com-
puter engineering from Carnegie Mellon University,
Pittsburgh, PA, USA, in 2017.
He is currently a Software Engineer with Oracle,
Cupertino, CA, USA.
Rahul Singh (M’14) received the B.Tech. degree in
electronics engineering from the IIT-Banaras Hindu
University (IIT-BHU), Varanasi, India, in 2008, and
the M.S. degree from Seoul National University,
Seoul, South Korea, in 2011. He is currently pur-
suing the Ph.D. degree in electrical engineering at
Carnegie Mellon University, Pittsburgh, PA, USA.
From 2011 to 2013, he was with the Proces-
sor Development Group, Samsung Electronics,
Gyeonggi-do, South Korea. His current research
interests include RF/mm-wave IC design.
T. E. Schlesinger (F’11) received the B.Sc. degree
from the University of Toronto, Toronto, ON,
Canada, and the M.S. and Ph.D. degrees in applied
physics from the California Institute of Technology,
Pasadena, CA, USA.
He was the David Edward Schramm Memorial
Professor and the Head of Electrical and Computer
Engineering with Carnegie Mellon University, Pitts-
burgh, PA, USA. He was the Director of the Data
Storage Systems Center, founding Co-Director of the
General Motors Collaborative Research Laboratory,
and Director of the DARPA MISCIC Center. He is currently the Benjamin
T. Rome Dean of the Whiting School of Engineering, The Johns Hopkins
University, Baltimore, MD, USA, and a Professor of electrical and computer
engineering. He has authored or co-authored over 250 papers and conference
proceedings and has 13 patents.
Dr. Schlesinger is a Fellow of the SPIE.
Jeyanandh Paramesh (SM’10) received the
B.Tech. degree in electrical engineering from IIT
Madras, Chennai, India, the M.S. degree in electrical
engineering from Oregon State University, Corvallis,
OR, USA, and the Ph.D. degree from the University
of Washington, Seattle, WA, USA.
He was with AKM Semiconductor (Analog
Devices), Motorola, and Intel, where he was
involved with product development and research
positions. He is currently an Associate Professor of
electrical and computer engineering with Carnegie
Mellon University Pittsburgh, PA, USA. His current research interests include
the design of RF and mixed-signal integrated circuits and systems for a wide
variety of applications.
James A. Bain (SM’12) received the B.S. degree
in materials science and engineering from the Uni-
versity of Pennsylvania, Philadelphia, PA, USA,
in 1988, and the M.S. and Ph.D. degrees in materials
science and engineering from Stanford University,
Stanford, CA, USA, in 1991 and 1993, respectively.
He is a Professor with the Electrical and Computer
Engineering Department, Carnegie Mellon Univer-
sity, Pittsburgh, PA, USA, where he also has a
courtesy appointment with the Department of Mate-
rials Science and Engineering and is the Associate
Director of the Data Storage Systems Center. He has authored or co-authored
more than 225 papers in the field of magnetic, optical, electrical, thermal, and
mechanical devices and materials for information technology. He currently has
active research programs in heat assisted magnetic recording, and resistive
switches for memory and reconfigurable electronics.
Dr. Bain is a Senior Member of the IEEE Magnetics Society, the IEEE
Electron Devices Society, and the IEEE Photonics Society.
IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 65, NO. 11, NOVEMBER 2017 4531

Design Criteria in Sizing Phase-Change

RF Switches
Gregory Slovin , Min Xu, Rahul Singh , Member, IEEE, T. E. Schlesinger, Fellow, IEEE,
Jeyanandh Paramesh, Senior Member, IEEE, and James A. Bain , Senior Member, IEEE

Abstract— This paper presents design criteria for four-terminal TABLE I

phase-change (PC) RF switches as a function of their dimensions, C OMPARISON OF RF S WITCH T ECHNOLOGIES
(heater width, RF gap, and barrier thickness), materials, and
connection topology. Reducing heater width is shown to reduce
switch actuation power, with this reduction ultimately being
limited by the maximum allowed heater current density, as set
by projected reliability. Narrower RF gaps are shown to increase
switch cut-off frequency FCO , because they decrease on-state
resistance more than they increase off-state capacitance. The
balance of this increased FCO versus the estimated reduction
in power handling as the RF gap shrinks is quantified. Barrier
layers of AlN thicker than 200 nm are shown to be sufficient to
electrically decouple the switch from the heater trace. Coupling to
heater pads can also be significant unless pad sizes are reduced as
in a monolithically integrated switch. Finally, it is demonstrated
that division of the switch into multiple parallel segments is a
viable approach for lowering the heater actuation voltage for
CMOS integration. A 12.5% increase in COFF (∼1.5 fF) was associated technologies in which they have been previously
observed and 35% increase in actuation power at temperature deployed can be found in [10].
was needed in a two-segment parallel switch, as compared to a For RF switches, the most promising topology uses a four-
single segment switch. terminal configuration: 1) two actuator terminals connected to
Index Terms— Germanium telluride, GeTe, PCM switch, a resistive heater and 2) two RF terminals that are connected
phase-change (PC) material, PC switch, reconfigurable circuits,
RF switch, tungsten microheaters. by the PC element, as demonstrated in [1]–[4] and [9] and
discussed in detail in the following section. The strength of
I. I NTRODUCTION this approach is the electrical isolation between the electrical

R F SWITCHES using phase change (PC) materials have

been demonstrated in several device topologies [1]–[9].
These devices exploit the fact that PC materials can be
toggled back and forth between high conductivity (ON) and
low conductivity (OFF) states through appropriate high-speed
heating events. An excellent overview of PC materials and the
Manuscript received April 1, 2017; revised June 21, 2017 and
August 29, 2017; accepted September 20, 2017. Date of publication
October 26, 2017; date of current version November 3, 2017. This work
was supported in part by the Defense Advanced Research Projects Agency
through Radio Frequency-Field Programmable Gate Arrays Program under
Grant HR0011-12-C-0095 and in part by the Office of Naval Research
through Defense University Research Instrumentation Program under Equip-
ment Grant N000141310874. (Corresponding author: James A. Bain.)
G. Slovin was with the Department of Electrical and Computer Engineering,
Carnegie Mellon University, Pittsburgh, PA 15213 USA. He is currently with
TowerJazz, Newport Beach, CA USA (e-mail: slovin.greg@gmail.com).
M. Xu was with the Department of Electrical and Computer Engineering,
Carnegie Mellon University, Pittsburgh, PA 15213 USA. He is currently
with the Oracle Corporation, Cupertino, CA 92660 USA (e-mail:
xumin9096@gmail.com).
R. Singh, J. Paramesh, and J. A. Bain are with the Department of
ECE, Carnegie Mellon University, Pittsburgh, PA 15213 USA (e-mail:
rsingh1@andrew.cmu.edu; paramesh@ece.cmu.edu; jbain@cmu.edu).
T. E. Schlesinger is with the Whiting School of Engineering, The Johns
Hopkins University, Baltimore, MD 21218 USA (e-mail: tschles4@jhu.edu).
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/TMTT.2017.2759735
0018-9480 © 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
actuation path (the heater) and the RF signal path, and it
achieves an attractive switch figure of merit (FOM). This FOM
is the cut-off frequency FCO defined as FCO = 1/2π RON C OFF ,
where RON is the ON-state resistance while C OFF is the
OFF -state capacitance. The aforementioned isolation results in
a reduction of C OFF to as low as 10 fF. This excellent electrical
isolation compares favorably to competing technologies, some
of which are summarized in Table I.
Within Table I, only the PC switches and the MEMS tech-
nologies offer the very high FOMs. These two technologies
have contrasting strengths. Electrostatic MEMS devices, typi-
cally have a very low holding power, are volatile and require
relatively large actuation voltages and specialized packaging.
PC switches require large actuation powers (∼1 W) for short
times (100 ns, for actuation energies of 100 nJ), but are
nonvolatile when not being actuated. They operate on CMOS
compatible voltages and can be integrated rather easily into a
CMOS process flow.
An important challenge in implementing these four-terminal
devices is in managing the required temperatures and powers
without compromising this electrical isolation. The use of
a heater external to the PC material adds heated volume,
and because the heater must be hotter than the PC element,
the actuation energy is substantially higher in the four-terminal
4532 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 65, NO. 11, NOVEMBER 2017
Fig. 1. (a) Cartoon cross section of the fabricated switches. A W heater
is built on top of a sapphire substrate and is used to heat the GeTe.
An AlN barrier layer is used to minimize electrical coupling between the
RF signal path and the programming path while still maintaining the thermal
coupling needed to efficiently heat the PC material. (b) SEM of a fabricated
four-terminal RF PC switch. Transformation of GeTe in the RF gap from
the crystalline state to the amorphous state prevents the RF signal from
propagating from one Cu trace to the other.
configuration than if the heating current could flow through the
PC directly [17]. If one seeks to ameliorate this situation by
putting the PC element closer to the heater, electrical isolation
suffers. This paper seeks to quantify this trade-off and identify
optimal design points which balance these considerations.
Other switch figures of merit, like linearity are outside the
scope of this paper, but may also be affected by the sizing
of components considered here. Importantly, very promising
linearity (IIP3 > 30 dBm) has been measured by others on
similar switches to demonstrate the overall viability of this
switch technology [6].
II. D EVICE D ESIGN
A cartoon cross section and SEM top-down view of the
four-terminal RF PC switch can be seen in Fig. 1(a) and (b),
respectively. In this device, the RF signal path and program-
ming path are electrically isolated from each other. The RF
current passes though the Cu traces to the W/Au contacts and
down to the PC material, GeTe. If the PC material is in the
ON -state (low resistance) the RF current flows though the GeTe
and back up though the contact and Cu trace. If the PC material
is in the OFF-state (high resistance), RF current flow is blocked
by the GeTe. The PC material is thermally actuated between
the crystalline ON-state and amorphous OFF-state with the use
of a W heater that runs transversely to the direction of the
RF current under the PC material in the RF gap. To switch
the PC material to the amorphous state, a 100-ns-long voltage
pulse is applied to the heater whose power is enough to raise
the temperature to the GeTe above its melting temperature
723 °C [10]. The pulse is then turned off and the PC element
quickly cools and freezes in the disordered amorphous state.
To switch the PC material to the crystalline state, a 1-μs-long
voltage pulse is applied to the heater to raise the temperature
of the GeTe to its crystallization temperature, approximately
200 °C.
The heater is electricity isolated from the RF signal path
using an AlN barrier layer. This barrier layer is necessary
to minimize the electrical coupling of the heater to the RF
switch. In the OFF-state, the heater can act as an additional
capacitance across the switch and degrade its RF performance.
While other studies have used SiN as the barrier layer [2]–[4],
in this paper, AlN was integrated as the barrier layer due to
its higher thermal conductivity among candidate dielectrics.
The use of AlN allows for the heater to be thermally well-
connected to the PC layer, but electrically isolated from it.
As a result, the barrier layer thickness can be increased to
further reduce the electrical coupling of the heater the RF
switch [1].
In this paper, switches with varying dimensions (as noted
in Fig. 1) were fabricated and tested to achieve an RON
from 1–2  and C OFF < 30 fF. The heater width W H was
varied between 1 and 3 μm, the switch RF path width WRF
between the 10 and 30 μm, and RF gap length L RF between
400 and 900 nm. The thickness of the GeTe layer, tPC , was
varied between 50 and 100 nm while the AlN barrier layer
thickness, t B , was varied between 100 and 200 nm.
A. Device Fabrication Overview
The switches were fabricated using an in-house fabrication
process flow. Devices were fabricated on sapphire substrates
because they are electrically insulating while still having a rea-
sonably high thermal conductivity of 35 W/m·K. A thermally
conductive substrate is important for achieving an adequate
quenching speed of the PC material [18], [19]. As noted
in [19], this can be achieved on a substrate of low thermal
conductivity if a sufficiently thick heat spreading layer of
high thermal conductivity is present immediately under the
switch. Heaters were fabricated using dc magnetron sput-
tering of W (70 nm thick) to deposit films at a substrate
temperature of 850 °C, which produced W with low elec-
trical resistivity (75 n·m), compared to films deposited at
room temperature (typ 230 n·m). W heaters were patterned
using contact photolithography and etched in a fluorine based
parallel plate reactive ion etch (RIE) system. An AlN barrier
layer (100 or 200 nm) was then reactively sputtered over the
W heater.
The PC material, GeTe, was then deposited (50 and 100 nm)
using co-sputtering from a Ge and Te target at a substrate
temperature of 170 °C. The resistivity of the resulting crys-
talline GeTe film was 1.68 μ·m, due to precise control of
composition. The GeTe was patterned and etched using an
Ar+ ion mill. The contacts to the GeTe were made using a lift-
off process, patterned using e-beam lithography. The contacts
consisted of a 10 nm W interface layer followed by 130 nm of
Au. The GeTe and contacts were then passivated with 100 nm
of sputtered SiO2 . Vias were etched down to the Au contacts
in a Fl-based parallel-plate RIE. Next, vias were etched down
to the W heater though the AlN barrier layer using a Cl-based
ICP RIE. The traces and pads were made to the contacts and
heater using a Cu plating process. A 10-nm Ta interface layer
and 100 nm of Cu were sputtered to form the seed layer for
the Cu plating. The traces and pads were patterned and then
Cu plated to thickness of 2 μm. The Cu and Ta seed layer
were etched in an Ar ion mill. The devices were then capped
with another 100 nm of sputtered SiO2 . Openings in the SiO2
over the pads were made with a Fl-based parallel plate RIE.
SLOVIN et al.: DESIGN CRITERIA IN SIZING PC RF SWITCHES
Fig. 2. (a) 1-port and (b) 2-port measurement structures. Heater pulses are
applied using an SG probe across H1 and H2. 1-port RF measurements were
made using an SG probe across S1 and G in (a) and 2-port RF measurements
were made using an SGS probe across S1 and G as port 1 and S2 and G
as port 2 in (b). (c) Heater pulse profile used to turn the switch OFF (red
line) and ON (blue line). (d) Capacitive model (partial) of the switch in the
OFF state.
B. Test Setup
Both 1-port [Fig. 2(a)] and 2-port [Fig. 2(b)] RF test
structures were fabricated to measure the ON-state resistance
(RON ) and OFF-state capacitance (C OFF ) of the various devices.
As fabricated, the devices are in the low resistance ON-state
and were first tested to determine their minimum power to
amorphize (MPA). The dc resistances of the switches were
measured using a Keithley 2400 source meter connected across
S1 and G for the 1-ports devices [Fig. 2(a)] and across
S1 and S2 for the 2-port devices [Fig. 2(b)]. To measure MPA,
100-ns-long voltage pulses of increasing amplitude [Fig. 2(c)]
were applied using an Agilent 8114A pulse generator across
H1 and H2 [Fig. 2(a) and (b)] until the measured dc resistance
of the switch increased above 10 k. Pulses were widely
separated in time (milliseconds or greater) such that the switch
returned to room temperature each time. Moreover, in this
simple protocol, no pulses are applied to return the switch to
the ON state between amorphization pulses.
The applied power (PApp) is calculated using (1) where
VApp is the voltage applied by the pulse generator to a
50- load, R H is the room temperature resistance of the
heater, and R S is the source impedance of the pulse generator
(R S = 50 )
4V 2App R H
PApp = . (1)
(R S +R H )2
It should be noted that the applied power is a convenient
parameter to tabulate, but it differs from the actual time-
varying power absorbed by the heater, because R H is a
function of temperature and time. For clarity in the definition
of PApp and VApp, we note that in the simple case of R H = R S ,
the total power supplied by the signal generator is 2V 2App /R S
and the PApp would be half of this (with the other half being
dissipated in the source). It should also be noted that we
also define VI , the incident voltage, which is equal to the
actual voltage that appears across the heater, since it is not
a matched load. Unlike VApp, VI is time and temperature
dependent and is given by VI = 2VApp R H /(R H + R S ). In all
4533
of the analysis of the voltage waveforms applied to the heater,
electrical transients are ignored as the electrical time constant
of the heater is orders of magnitude smaller than the thermal
time constant.
The minimum power to crystalize (MPC) is determined in
a manner similar to the MPA procedure, by applying, in this
latter case, 1-μs-long voltage pulses of increasing amplitude
[Fig. 2(c)] across the heater until the resistance of the switch
is within 20% of its initial as fabricated RON value.
RF measurements were made using an Agilent 8364A
PNA Network Analyzer. Using the device layout in Fig. 2(a),
1-port S-parameter measurements were made from 45 MHz to
20 GHz across S1 and G. Two-port S-parameter measurements
were made on the device layout in Fig. 2(b) from 45 MHz
to 14 GHz across S1 and S2 while sharing a common
ground (G). All S-parameter measurements were de-embedded
using standard de-embedding techniques [20] to extract the
RON and C OFF of the switches. The switch can be modeled as
a resistance and capacitance, C OFF , in parallel. The total C OFF
can be further broken down into three parallel capacitances:
1) CTT , the capacitance across the RF gap from one trace
to the other; 2) the series combination of two instances of
CTH , the series capacitances between the trace and heater
[Fig. 2(d)]; and 3) series combinations of non de-embedded
fringing capacitance (C F ) to the heater pads. This last term
is discussed in detail in Section III-D on grounded heaters.
All RF-measurements were made with the heater pads left
floating to allow for more accurate measurement of the RON
and C OFF [2] unless otherwise noted.
III. E FFECTS OF D EVICE S IZING AND
C ONNECTION T OPOLOGY
Correctly sizing the switch dimensions is important to
maximize RF performance. This entails trade-offs between
performance and power consumption. This section will discuss
the effects of heater width, RF gap length and AlN barrier layer
thickness on FCO and MPA of the RF switches.
A. Effect of Heater Width
Two-port measurement structures were fabricated to deter-
mine the effect of heater width. Fabricated switches were
20-μm-wide WRF , with heaters widths W H of 1, 2, and 3 μm.
The GeTe and AlN barrier were both 100 nm thick in these
devices. The MPA and MPC were measured for these switches
and are shown in Fig. 3(a). The MPA increases by 0.5 W per
micrometer increase in heater width, while the MPC increases
by 0.1 W per micrometer. This increase in power can be
attributed to an increase in heated volume and a resulting
increase thermal capacitance and thermal time constant of
the switch. Therefore, to melt the PC material in the same
100-ns-long voltage pulse, the power delivered to the heater
must be increased.
The de-embedded insertion loss of the switches in the
ON -state with the three different heater widths is shown in
Fig. 3(b). As expected the insertion loss is not dependent on
heater width and is about 0.1 dB up to 14 GHz. However,
the de-embedded isolation of the switches is dependent on the
4534 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 65, NO. 11, NOVEMBER 2017
Fig. 3. (a) MPA ( red dashed line) and MPC ( blue dashed line) for 20-μm
switches with header widths of 1, 2, and 3 μm. Each marker is a separate
device. (b) Insertion loss and (c) isolation 20-μm switches with header widths
of 1 (red line), 2 (blue line), and 3 μm (green line) from 45 to 14 GHz.
(d) Extracted C OFF for 20-μm switches with header widths of 1, 2, and 3 μm.
heater width [Fig. 3(c)]. The isolation at 10 GHz decreases
from 21 dB for a switch with a 1-μm-wide heater to less
than 18 dB for one with a 3-μm-wide heater. The extracted
C OFF for the different heater widths can be seen in Fig. 3(d).
As the width of the heater increases, C OFF increases by 4.3 fF
per micrometer. This can be attributed to an increase in CTH
[Fig. 2(d)]. The intercept of the fit (“0”-μm-wide heater)
in Fig. 3(d) is 9.7 fF which includes CTT and C F .
Thus, reducing heater width helps reduce switching power
as well as improving isolation, so it should be as narrow
as possible as allowed by the current density limitations of
nanoscale conductors. This is quantified within Section III-E,
Summary of Design Trade-offs. As will be shown in a
subsequent section on AlN thickness, this benefit of narrow-
ing heater width for isolation disappears when the dielectric
between the heater and the PC layer is thick enough. This
operating limit will be assumed in Section III-E.
B. Effect of RF Gap Length
Reducing the RF gap length has the obvious benefit of
reducing RON due to resistance being proportional to length.
However, the cumulative effects of RF gap length on RF
performance and power are not as obvious. One-port RF
measurement structures were fabricated to determine the effect
of RF gap length L RF . The fabricated switches had switch
width WRF = 20 μm, with heater widths W H = 1 μm.
The GeTe and AlN barrier layer thickness were 50 (tPC ) and
100 nm (t B ) thick, respectively. The RF gap lengths L RF
studied ranged from 400 to 900 nm in 100-nm increments.
Fig. 4(a) and (b) shows the MPA and MPC of the switches
with RF gap lengths ranging from 400 to 900 nm, respectively.
As L RF decreases, the MPA increases. This is due to the Au
contacts being closer together and sinking more heat out of
the PC material. To compensate for the increased heat sinking,
more power is needed to raise the temperature of the PC
material above its melting temperature.
Fig. 4. (a) MPA and (b) MPC results for 20-μm-wide switches with RF gaps
ranging from 400 to 900 nm in length. (c) Extracted RON and (d) C OFF of
switches with varying L RF from 45 MHz to 20 GHz. As the RF gap length
L RF reduces from 900 to 400 nm, RON reduces from 3 to 1.4  and C OFF
increases from 11.7 to 15.4 fF at 5 GHz.
Fig. 5. (a) FCO and (b) MPA (red dots) and MPC (blue squares) for
20-μm-wide switches with RF gaps ranging from 400 to 900 nm in length.
The FCO is the highest, 7.2 THz, for the shortest RF gap length of 400 nm.
The extracted RON and C OFF from 45 MHz to 20 GHz can
be seen in Fig. 4(c) and (d), respectively. As expected RON
decreased with reducing L RF . RON decreases from 3 to 1.4 
when L RF is reduced from 900 to 400 nm. However, as the
RON decreases, C OFF increases. C OFF increase from 11.7 to
15.4 fF as L RF is decreased.
Because RON and C OFF move in opposite directions with
RF gap length, it is not immediately obvious how the RF
performance of the switch is affected by its L RF . Fig. 5(a)
shows the FCO of each switch versus their corresponding L RF .
The FCO is the greatest, 7.2 THz, for the shortest L RF of
400 nm. As L RF is increased to 900 nm, the FCO decreases
to 4.2 THz. Compared to C OFF , RON is more dependent on
L RF . L RF is the largest contribution to the ON-state resistance
of the switch while, it only contributes to CTT , and not the
other capacitances in C OFF , CTH , and C F . There is a trade-off,
however, between RF performance and the required power to
amorphize the PC material. The MPA is 1.8 W for an L RF
of 400 nm, but reduces to less than 1.5 W for L RF longer than
700 nm [Fig. 5(b)]. The increase in RF performance outweighs
the cost of the increase in MPA even for the narrowest L RF
of 400 nm. The FCO /MPA is 4 THz/W for an L RF of 400 nm
and only 2.8 THz/W when L RF is increased to 900 nm.
SLOVIN et al.: DESIGN CRITERIA IN SIZING PC RF SWITCHES
Fig. 6. (a) Insertion loss and (b) isolation for 20-μm switches with AlN
barrier layer thickness of 100 nm (red) and 200 nm (blue) from 45 MHz
to 14 GHz. The increased insertion loss for the switch with 200-nm AlN is
due to a GeTe thickness of 50 nm compared to 100 nm for the switch with
100 nm of AlN. This resulted in an ON-state resistance of 3.4 and 1.5 ,
respectively. (c) OFF-state capacitance for switch widths of 10, 20, and 30 μm
with 100 nm (red dashed line) and 200 nm (blue dashed line) of AlN. The
fringing capacitance calculated to be 3.7 fF, with a per micrometer capacitance
of 0.51 and 0.32 fF for 100 and 200 nm of AlN, respectively. (d) OFF-state
capacitance of 20-μm-wide switches with varying AlN barrier layer thickness
from this paper and [1]. Points indicate measurement on distinct switches.
While RF performance improves with decreasing RF gap,
RF power handling will degrade as RF gap shrinks. While
power handling was not studied in detail in this paper, we can
examine this trade-off under the assumption that the power
handling is limited by the electric field that the OFF-state
device can handle. This is quantified in Section III-E.
C. Effect of AlN Thickness
The AlN barrier layer is responsible for electrically isolating
the heater from the PC material while still keeping them
thermally coupled. Despite thin-film AlN having a thermal
conductivity as high as of 130 W/m·K [21], increasing the
barrier layer thickness to reduce CTH does result in an increase
in MPA, due to an increase in thermal capacitance [1].
Two-port S-parameter measurements were made on devices
with various switch widths WRF of 10, 20, and 30 μm to
determine at what AlN thickness t B , the capacitance attributed
to the heater disappears. The switches measured all had a W H
of 1 μm with one set of switches having, tPC of 50 nm with
a t B of 200 nm and the second set of switches having a tPC
of 100 nm with a t B of 100 nm.
Fig. 6(a) and (b) shows the insertion loss and isolation,
respectively, from 45 MHz to 14 GHz for W H of 20 μm with a
t B of 100 (red) and 200 nm (blue). The difference in insertion
loss, 0.1 dB for t B of 100 nm and 0.26 dB for t B of 200 nm at
10 GHz, is due to the difference in tPC . The isolation increased
from 21 dB to 25 dB at 10 GHz when t B is increased from
100 to 200 nm. The OFF-state capacitance was extracted for
WRF of 10, 20, and 30 μm for both t B of 100 and 200 nm
are shown in Fig. 6(c).
4535
Fig. 7. ON -state electrical model of switch when the heater terminal
is (a) floating and (b) grounded. When the heater terminal is grounded some
of the RF signal propagates though CTH down to the grounded heater.
The switches with t B of 100 nm have an increase in
capacitance of 0.51 fF per micrometer of switch width, while
switches with a t B of 200 nm only have an increase in
capacitance of 0.32 fF per micrometer of switch width. Both
the t B of 100 and 200 nm have a fringing capacitance
(the intercept at “0-μm wide” switch) of 3.7 fF [Fig. 6(c)].
Earlier we showed that for a WRF of 20 μm, C F plus
CTT was equal to 9.7 fF. Therefore, CTT must be equal
to 6 or 0.3 fF per micrometer of switch width. The slope of
the fits in Fig. 6(c) is equivalent to CTT + CTH /2. When CTT ,
0.3 fF/μm, is subtracted from the slope of the fits, CTH /2 is
0.21 and 0.02 fF/μm for a t B of 100 and 200 nm, respectively.
This shows for a t B of 200 nm, the heater contribution to C OFF
is essentially zero. Additional increases in the AlN thickness
will not further reduce C OFF , but will increase the MPA of
the switch. Fig. 6(d) shows C OFF for WRF of 20 μm versus
the inverse of AlN thickness t B for this paper and previous
work in [1]. It becomes obvious that minimum C OFF for this
switch design is approximately 10 fF and cannot be reduced
any further by increasing the AlN barrier layer thickness. For
the summary of design trade-offs in Section III-E, we assume
that the AlN thickness is sufficient to eliminate CTH entirely.
D. Effect of Grounded Heaters
Making S-parameter measurements with the heater pads left
floating is necessary for an accurate extraction of RON and
C OFF [2]. However, this is most likely not how the switch
would be implemented in a real RF system. Ideally, when the
switch is in ON-state, all of the RF signal propagates from one
side of the switch to the other [Fig. 7(a)] and the insertion loss
of the switch is minimized. However, if one side of the heater
is grounded, this will produce a capacitive shunt to ground,
resulting in some of the RF signal propagating though CTH
to ground [Fig. 7(b)]. This results in degradation of the RF
signal propagation from one side of the switch to the other
and an increase in insertion loss at higher frequencies. This
RF “sneak” path shows the importance of minimizing CTH by
increasing t B to 200 nm.
Fig. 8(a) and (b) shows the insertion loss and isolation,
respectively, for WRF of 20 μm for both t B of 100 (red)
and 200 nm (blue) when the heater probes are up (dotted)
and landed (solid). The offset in insertion loss with the
heater probes up (dotted) between the t B of 100 and 200 nm
4536 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 65, NO. 11, NOVEMBER 2017
Fig. 8. (a) Insertion loss and (b) isolation of 20-μm switches with AlN
barrier layer thickness of 100 nm (red) and 200 nm (blue), with the heater
probes up (floating heater) and with them landed (grounded heater). It can be
seen that with a grounded heater the insertion-loss increases with frequency.
Fig. 9. (a) Zoomed-in view of switch showing the fringing capacitance (C F )
and capacitance to the heater pads (CHP ). (b) De-embedding structure used
to measure CHP . (c) Electrical model of 2-port measurement with switch in
the ON-state when the heater probes are landed. The value of C F extracted
from applying this model is 4 fF while CHP is 19 fF.
[Fig. 8(a)] is due to a difference in tPC . When the heater probes
are landed, the insertion loss increases with frequency while
when the heater probes are up it does not. For a t B of 100 nm,
the insertion loss increases from 0.1 to 0.34 dB and increases
from 0.26 to 0.45 dB for a t B of 200 nm at 10 GHz when
the heater probes are landed. The isolation is also affected by
landing the heater probes. At 10 GHz, the isolation increases
from 21 to 24 dB for a t B of 100 nm and from 25 to 28 dB
for a t B of 200 nm. The increase in both the insertion loss
and isolation is due to the switch being capacitively shunted
to ground.
In Section III-C it was shown that the CTH for a t B of
200 nm is essentially 0 fF so there should no degradation
of insertion loss though the heater, however, there are other
capacitances that can couple to the heater pads which would
result in a degradation in insertion loss even for switches
with a t B of 200 nm. Fig. 9(a) shows a zoomed-in image
of a 2-port measurement structure showing C F as well as the
capacitance to the heater pads CHP . A through de-embedding
structure [Fig. 9(b)] was measured with the heater probes
landed to extract CHP . Fig. 9(c) shows the electrical model
of a 2-port measurement structure with a switch in the ON-
state when the heater probes are landed. CHP was measured
to be 19 fF and C F was previously shown to be about 4 fF.
Fig. 10 shows the measured (solid line) and modeled (dashed
line) change in insertion loss of a through-structure (green),
a switch with a t B of 100 (red) and 200 nm (blue) when the
Fig. 10. Measured (solid line) and modeled (dashed line) insertion loss of the
through structure (green), switch with 200 nm (blue) of AlN and 100 nm (red)
of AlN. The through structure shows significant degradation in the insertion
loss despite not having a heater present. This is due to the presence of heater
probe pads (CHP ). The degradation of the of the 200-nm AlN switch is due
to CHP and C F while the 100-nm AlN switch degradation is the largest due
the additional capacitance to the heater (CTH ).
heater probes are landed. The through-structure with no heater
shows significant degradation in insertion loss just from the
capacitance to the heater pads CHP . For a switch with a t B
of 200 nm, there is additional degradation in insertion loss
due to C F despite CTH being close to 0 fF. The insertion loss
degrades even further for a switch with only a t B of 100 nm
due to CTH being approximately 8 fF. This shows that while
reducing CTH is important to maximizing RF performance
of the device, careful consideration needs to be made when
laying out traces to the heater as their coupling capacitance
to the switch can significantly degrade the RF performance of
the system if laid out incorrectly. For the summary of design
trade-offs discussed in the following section, we assume that
monolithic integration and layout adjustments have been used
to eliminate any significant RF coupling to the heater leads.
The elimination of this coupling is the appropriate way to
eliminate the RF leakage to the heater, as contrasted to trying
to introduce series resistance in to the heater driver circuit,
which will put unacceptable burden on the drive circuitry.
E. Summary of Design Trade-Offs
Taking all of the above measurements into consideration,
we arrive as some summary guidance for sizing these switches
and what performance might be expected. Fig. 11(a) shows a
summary of switch cut-off frequency and max power handling
of a switch as a function of RF gap, while Fig. 11(b) shows
the required heater power and heater current density as a
function of heater width. A series of assumptions derived from
the measurements in the previous sections have led to these
figures, which are detailed in the following list.
Assumptions of Fig. 11:
1) All coupling between the heater and the PC switch has
been eliminated. This is reasonable if the AlN layer is
200 nm or more in thickness and if monolithic integra-
tion has eliminated the coupling to the heater through
reduction of lead size and routing of lead position.
2) Per the above, the only capacitances that enter the
calculation of FCO are CTT and C F , and they are
assumed to follow the measurements shown above,
i.e., the capacitance as a function of RF gap follows
Fig. 4(d) with the appropriate subtractions.
SLOVIN et al.: DESIGN CRITERIA IN SIZING PC RF SWITCHES
TABLE II
S IMULATED MPA OF 20-μm-W IDE S WITCH FOR VARIOUS H EATER M ATERIALS
Fig. 11. Key design trade-offs in PC switches. (a) Switch performance trade-
off versus RF gap. (b) Switch power requirements and trade-off in resulting
heater current densities. In (a) two PC layer thicknesses tPC are considered
(100 nm with solid lines and 200 nm with dashed) and two switch widths
W S are labeled: 10 and 30 μm. In (b) two switch widths W S are considered
(10 μm with solid lines and 30 μm with dashed) and two heater thicknesses t H
are labeled: 100 and 500 nm.
3) PC ON-state resistivity is 3 × 10−6 ·m.
4) Series resistance for all switches is 0.25 ;
5) Contact resistance is assumed to be 0.5  for a 20-μm-
wide switch and scales inversely with switch width.
6) Breakdown electric fields of the PC in the OFF-state set
the switch power handling. Power handling for values
between 10 and 30 V/um are shown in Fig. 11(a), which
are typical
7) Heaters are W with the properties for films deposited at
400 °C (Table II).
8) Required heater power was assumed independent of
RF gap (which is an approximation of the observed
4537
behavior), with the 600-nm RF gap taken as representa-
tive.
9) Each switch requires 0.8 W as a baseline power, and for
every micrometer of switch width each heater requires
25 mW per μm of heater width [Fig. 3(a)]. (i.e.,
A 20-μm-wide switch with 2-μm-wide heater requires
1.8 W of power).
10) Heater current density is for the heater at temperature.
11) PC thickness of 100 and 200 nm are shown.
The key trade-offs in sizing PC switches with regard to
RF gap, PC layer thickness, switch width, heater width, and
heater thickness are tabulated in Fig. 11. The choice of RF
gap [Fig. 11(a)] balances raising FCO while lowering power
handling. As noted in earlier sections, reducing the RF gap
improves R more than it degrades C and this increases FCO .
Selecting heater width generally means making it as narrow
and as thick as possible, to reduce power. However, processing
constraints and acceptable current densities will limit how
narrow it can be. While this dependence was not studied in this
paper, Fig. 11(b) shows how heater current density scales with
heater width for several structures. The mean time to failure
for high current densities typically scales inversely with the
square of the current density [24]. Demonstrations of heater
cycling for over 1000 cycles have been shown with current
densities of 100 MA/cm2 [18].
IV. S CALING FOR CMOS I NTEGRATION
These RF switches have already shown promising results
when combined with RF CMOS circuits using flip chip
technology [9]. In [9], the RF switch was integrated into the
RF signal path, but the heater was still driven from off-chip
using a pulse generator. For true CMOS integration, the heater
would also have to be driven by on-chip CMOS circuitry.
This presents some challenges, as the incident voltages across
the heater in this paper and in [9] are as high as 8 V. This
would be too high for typical CMOS processes, though more
expensive high-voltage processes are available. Reducing the
voltage needed to turn the switch OFF is desirable for the most
general CMOS integration, as most CMOS circuits will be
limited by breakdown voltage rather than the ability to source
current.
A. Choosing a Heater Material
In the switch design presented above, the heaters are fab-
ricated using W that is deposited at 850 °C, which would
4538 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 65, NO. 11, NOVEMBER 2017
Fig. 12. Measured (a) resistivity and (b) TCR for W sheet films deposited
at 25 °C, 400 °C and 850 °C. As deposition temperature increase to 850 °C,
the resistivity decreases to 75.2 n-m and the TCR increases to 3.3 ppt/°C.
The sheet films were measured using four-wire resistance measurement of
Van der Pauw structures (insert).
not be back-end CMOS compatible. Using W deposited at
elevated temperature is shown to improve heater reliability,
and it also reduces the resistance of the heater allowing
for a lower actuation voltage. On the other hand, elevated
temperature deposition also increases the temperature coef-
ficient of resistance (TCR) of the heater. W heaters were
fabricated with deposition temperatures of 25 °C, 400 °C and
850 °C with resistivity and TCR measured using Van der Pauw
structures (Fig. 12). The deposition temperature of 400 °C
was chosen because it is back-end CMOS compatible [25].
As the deposition temperature increases, the resistivity of the
W decreases from 230 to 75.2 n·m while the TCR increases
from 1.3 to 3.3 ppt.
Using COMSOL Multiphysics software, the MPA was sim-
ulated using the same setup as in [1] for the three different
W deposition temperatures as well as NiCr, which has been
used as a heater in other studies [2]–[4]. Table I shows the
results of simulated MPA operating points, using the measured
resistivity and TCR of the measured W films and NiCr [23].
The initial incident voltage across the heater, VI , as well
as MPA decreases with decreasing heater resistivity. Even at
the end of the 100-ns-long voltage pulse, the resistance of
the heater at elevated temperature is still the lowest for W
deposited at 850 °C despite having the highest TCR. VI at
100 ns (at temperature) is also the lowest for the 850 °C.
This shows that minimizing the resistivity of the W heater
reduces actual switching power, and is desirable when trying
to integrate the switch with a CMOS driver despite the increase
in TCR. Much of the benefit in W behavior is available when
deposition is limited to the back-end compatible temperature
of 400 °C.
B. Segmenting the Switch
While lower resistivity helps reduce required heater volt-
ages, it is still not enough to get to the voltage levels in
standard CMOS processes. There are additional approaches:
1) increasing the thickness of the heater and 2) segmenting
the switch into multiple parts. While increasing the thickness
of heater seems like the most straightforward solution to
reducing the heater resistance, the heater would have to be over
1-μm thick for CMOS compatibility. As a result, the thermal
capacitance of the heater and thermal time constant of the
switch would increase significantly. This would result in an
Fig. 13. 2-port measurement structure of two parallel 10-μm-wide switches.
The heaters are also in parallel and are pulsed simultaneously using a GSG
probe across H1, H2, and H3. S-parameter measurements from port-1 across
S1 and G to port-2 across S2 and G using an SGS probe.
increase in MPA and may also prevent the PC material from
quenching fast enough to switch to the amorphous state.
Segmenting the switch into multiple sections and placing the
heaters in parallel would allow a reduction in effective heater
resistance seen by the CMOS driver as well assure that the
PC material can still be amorphized.
Fig. 13 shows a fabricated 2-port measurement structure that
has two 10-μm-wide switches, (WRF = 10 μm), in parallel
instead of a single switch with WRF = 20 μm. In this switch,
the heater pulses are applied using a GSG RF probe where
H1 and H3 are grounded and H2 is the applied voltage from
the pulse generator. The current from each heater pulse is thus
divided into both heaters in parallel. Table III highlights the
difference in the MPA heater pulses applied in this manner
as compared to those applied to the single heater in the
20-μm-wide switch.
Breaking the switch into two sections clearly reduces
the required voltage drop, VI , required across each heater,
by around a factor of 2 (slightly more than 2× at room temp
and slightly less than 2× at maximum heater temperature). The
VApp does not appear to change much, but that is because of
the very different impedances of the two heater configurations.
The 10 μm (×2) configuration is much closer to being a
matched load to the pulse generator at temperature, while the
20 μm (×1) has an impedance of almost 140  while hot.
Interestingly, breaking the switch into two pieces results in
a lower apparent MPA, but actually consumes more power
when the heater is at its maximum temperature (100 ns). The
lower MPA value is, again, due to the difference between room
temperature and maximum heater temperature. It is expected
that dividing the heaters will have some power penalty as heat
flows out of the ends as well as along the heater length and
the divided heater has twice as many ends, so this behavior is
reasonable and is supported by simulation.
For comparison of RF performance, S-parameter measure-
ments on the 10 μm (×2) switch are made between S1 and
S2 using a common ground pad (G). The insertion loss
and isolation of two switches with a WRF of 10 μm in
parallel (green line) is compared to a single switch with a
WRF of 20 μm (blue line) in Fig. 14. The insertion loss of
both switches is about 0.12 dB [Fig. 14(a)] but the ON-state
resistance of the parallel switches is slightly higher (2.1 )
SLOVIN et al.: DESIGN CRITERIA IN SIZING PC RF SWITCHES
TABLE III
S INGLE V ERSUS PARALLEL S WITCHES
Fig. 14. (a) Insertion loss and (b) isolation of a single 20-μm-wide switch
(blue line) and two 10-μm-wide switches in parallel (green line). The parallel
switches have a slightly higher RON of 2.1  compared to 1.8  for a single
switch. C OFF is also higher for the parallel switches 14.4 fF compared to
12.8 fF for a single switch. This is due to the presence of additional fringing
capacitance from the extra heater.
compared to the single switch (1.8 ). The isolation is lower
for switches in parallel, 21 dB, compared to 23 dB for a single
switch at 10 GHz. This results in an extracted C OFF of 14.4 fF
for the parallel switches and 12.8 fF for the single switches.
This is due to the presence of additional fringing capacitance
from the additional heater.
Thus, the strategy of breaking the switch into segments does
indeed reduce the actuation voltage. Of course, the division
into two segments, here, is simply a proof of concept. It did
not go far enough to be CMOS compatible as the heater
voltage was still over 7 V at temperature in the two-segment
topology. This suggests that CMOS driven designs would need
4–6 heater segments. Moreover, the approach shown where
the heaters were driven simultaneously and in parallel was
simply a matter of convenience. In a fully integrated design,
it heaters might be driven in parallel banks or singly as circuits
and routing constraints dictate. Finally, the difference between
VApp and VI is likely to be much smaller and much less
sensitive to heater resistance as one would likely seek to make
the source resistance, R S , of an integrated CMOS current
driver well below the heater resistance, so as not to waste
voltage headroom.
V. C ONCLUSION
In summary, then, we observe a number of trends and
correlations that can guide the design and sizing of PC RF
switches. The general rules are as follows. First, heaters should
be as narrow as possible as limited by current density. This
makes more efficient use of power and reduces capacitive
coupling to the heater. The RF gap should also be as narrow
as possible to get high switch FCO , as resistance is reduced
by more than capacitance is increased. However, more power
will be needed to actuate a switch with a narrower RF gap
4539
due to the heat sinking of the more closely spaced leads and
the power handling will drop due to increased electric fields
(leading to breakdown) across the switch for the same standoff
voltage.
The dielectric barrier between the heater and the PC layer
needs to be 200 nm thick, but not much more. At this
thickness, the capacitive coupling to the heater stripe is negli-
gible. Making this layer thicker will simply increase the heat
capacity of the switch, slow it down, and require more power,
without significantly improving its RF performance. Coupling
to the heater pads, however, may not be negligible at this
200-nm thickness, and must be considered in the switch layout.
In particular, the coupling path to the heater pads will induce
significant insertion loss if the heater is grounded. Of course,
a fully integrated CMOS solution, as suggested in [19] would
eliminate the need for bond pads and would greatly ameliorate
this issue.
Finally, it has been demonstrated that division of the switch
into multiple parallel segments is a viable approach for low-
ering the heater actuation voltage in anticipation of CMOS
integration. A 12.5% increase in C OFF (∼1.5 fF) and 35%
increase in actuation power at temperature was needed in a
two-segment parallel switch, as compared to a single segment
switch. This is due to the additional fringing capacitance and
the additional heat sinking associated with the end of each
heater. The number of switch segments likely needed for
CMOS integration would be 4–6, based on typical CMOS
voltages.
R EFERENCES
[1] G. Slovin, M. Xu, J. Paramesh, T. E. Schlesinger, and J. A. Bain, “AlN
barriers for capacitance reduction in phase-change RF switches,” IEEE
Electron Device Lett., vol. 37, no. 5, pp. 568–571, May 2016.
[2] N. El-Hinnawy et al., “12.5 THz Fco GeTe inline phase-change
switch technology for reconfigurable RF and switching applications,”
in Proc. IEEE Compound Semiconductor Integr. Circuit Symp. (CSICS),
Oct. 2014, pp. 1–3.
[3] N. El-Hinnawy et al., “A 7.3 THz cut-off frequency, inline, chalcogenide
phase-change RF switch using an independent resistive heater for
thermal actuation,” in Proc. IEEE Compound Semiconductor Integr.
Circuit Symp. (CSICS), Oct. 2013, pp. 1–4.
[4] N. El-Hinnawy et al., “A four-terminal, inline, chalcogenide phase-
change RF switch using an independent resistive heater for thermal
actuation,” IEEE Electron Device Lett., vol. 34, no. 10, pp. 1313–1315,
Oct. 2013.
[5] C.-Y. Wen, G. Slovin, J. A. Bain, T. E. Schlesinger, L. T. Pileggi,
and J. Paramesh, “A phase-change via-reconfigurable CMOS LC
VCO,” IEEE Trans. Electron Devices, vol. 60, no. 12, pp. 3979–3988,
Dec. 2013.
[6] M. Wang, F. Lin, and M. Rais-Zadeh, “Performance measurements and
non-linearity modeling of GeTe phase change RF switches with direct
and indirect heating schemes,” in IEEE MTT-S Int. Microw. Symp. Dig.,
May 2015, pp. 1–4.
4540 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 65, NO. 11, NOVEMBER 2017
[7] M. Wang, Y. Shim, and M. Rais-Zadeh, “A low-loss directly heated
two-port RF phase change switch,” IEEE Electron Device Lett., vol. 35,
no. 4, pp. 491–493, Apr. 2014.
[8] C.-Y. Wen et al., “A phase-change via-reconfigurable on-chip inductor,”
in IEDM Tech. Dig., Dec. 2010, pp. 10.3.1–10.3.4.
[9] R. Singh et al., “A 3/5 GHz reconfigurable CMOS low-noise amplifier
integrated with a four-terminal phase-change RF switch,” in IEDM Tech.
Dig., Dec. 2015, pp. 25.3.1–25.3.4.
[10] S. Raoux and M. Wuttig, Eds., Phase Change Materials: Science and
Applications. Boston, MA, USA: Springer, 2009.
[11] A. Botula et al., “A thin-film SOI 180 nm CMOS RF switch technology,”
in Proc. IEEE Top. Meeting Silicon Monolithic Integr. Circuits RF Syst.,
Jan. 2009, pp. 1–4.
[12] P. Hurwitz, R. Kanawati, K. Moen, E. Preisler, S. Chaudhry, and
M. Racanelli, “Advances in RF foundry technology for wireless and
wireline communications,” in Proc. IEEE 16th Top. Meeting Silicon
Monolithic Integr. Circuits RF Syst. (SiRF), Jan. 2016, pp. 5–8.
[13] T. Boles, J. Brogle, D. Hoag, and D. Curcio, “AlGaAs PIN diode multi-
octave, mmW switches,” in Proc. IEEE Int. Conf. Microw., Commun.
Antennas Electron. Syst. (COMCAS), Nov. 2011, pp. 1–5.
[14] R. S. Howell et al., “Low loss, high performance 1-18 GHz SPDT based
on the novel super-lattice castellated field effect transistor (SLCFET),” in
Proc. IEEE Compound Semiconductor Integr. Circuit Symp., Oct. 2014,
pp. 3–6.
[15] R. Stefanini, M. Chatras, P. Blondy, and G. M. Rebeiz, “Miniature
MEMS switches for RF applications,” J. Microelectromech. Syst.,
vol. 20, no. 6, pp. 1324–1335, Dec. 2011.
[16] A. Fruehling, R. Pimpinella, R. Nordin, and D. Peroulis, “A single-
crystal silicon DC-40 GHz RF MEMS switch,” in IEEE MTT-S Int.
Microw. Symp. Dig., Jun. 2009, pp. 1633–1636.
[17] M. Wang and M. Rais-Zadeh, “Development and evaluation of ger-
manium telluride phase change material based ohmic switches for RF
applications,” J. Micromech. Microeng., vol. 27, no. 1, p. 13001, 2017.
[18] M. Xu, “Study of the crystallization dynamics and threshold voltage
of phase change materials for use in reconfigurable RF switches and
non-volatile memories,” Ph.D. dissertation, Dept. Elect. Comput. Eng.,
Carnegie Mellon Univ., Pittsburgh, PA, USA, 2017.
[19] N. El-Hinnawy et al., “Substrate agnostic monolithic integration of the
inline phase-change switch technology,” in IEEE MTT-S Int. Microw.
Symp. Dig., May 2016, pp. 1–4.
[20] T. E. Kolding, “On-wafer calibration techniques for giga-hertz CMOS
measurements,” in Proc. Int. Conf. Microelectron. Test Struct. (ICMTS),
Mar. 1999, pp. 105–110.
[21] C. Duquenne, M.-P. Besland, P. Y. Tessier, E. Gautron, Y. Scudeller,
and D. Averty, “Thermal conductivity of aluminium nitride thin films
prepared by reactive magnetron sputtering,” J. Phys. D, Appl. Phys.,
vol. 45, no. 1, p. 15301, Jan. 2012.
[22] D. Yu, S. Brittman, J. S. Lee, A. L. Falk, and H. Park, “Minimum voltage
for threshold switching in nanoscale phase-change memory,” Nano Lett.,
vol. 8, no. 10, pp. 3429–3433, Oct. 2008.
[23] J. S. Faughn and R. A. Serway, College Physics. Pacific Grove, CA,
USA: Thomson Learning Inc, 2003.
[24] J. R. Black, “Electromigration—A brief survey and some recent results,”
IEEE Trans. Electron Devices, vol. 16, no. 4, pp. 338–347, Apr. 1969.
[25] S. Sedky, A. Witvrouw, H. Bender, and K. Baert, “Experimental
determination of the maximum post-process annealing temperature for
standard CMOS wafers,” IEEE Trans. Electron Devices, vol. 48, no. 2,
pp. 377–385, Feb. 2001.
Gregory Slovin received the B.S. and M.S. degrees
in electrical engineering from the University of
Florida, Gainesville, FL, USA, and the Ph.D. degree
in electrical engineering from Carnegie Mellon Uni-
versity, Pittsburgh, PA, USA, in 2015.
He is currently a Process Development Engineer
with TowerJazz, Newport Beach, CA, USA.
Min Xu received the B.S. degree in electrical engi-
neering from the Huazhong University of Science
and Technology, Wuhan, China, in 2012, and the
M.S. and Ph.D. degrees in electrical and com-
puter engineering from Carnegie Mellon University,
Pittsburgh, PA, USA, in 2017.
He is currently a Software Engineer with Oracle,
Cupertino, CA, USA.
Rahul Singh (M’14) received the B.Tech. degree in
electronics engineering from the IIT-Banaras Hindu
University (IIT-BHU), Varanasi, India, in 2008, and
the M.S. degree from Seoul National University,
Seoul, South Korea, in 2011. He is currently pur-
suing the Ph.D. degree in electrical engineering at
Carnegie Mellon University, Pittsburgh, PA, USA.
From 2011 to 2013, he was with the Proces-
sor Development Group, Samsung Electronics,
Gyeonggi-do, South Korea. His current research
interests include RF/mm-wave IC design.
T. E. Schlesinger (F’11) received the B.Sc. degree
from the University of Toronto, Toronto, ON,
Canada, and the M.S. and Ph.D. degrees in applied
physics from the California Institute of Technology,
Pasadena, CA, USA.
He was the David Edward Schramm Memorial
Professor and the Head of Electrical and Computer
Engineering with Carnegie Mellon University, Pitts-
burgh, PA, USA. He was the Director of the Data
Storage Systems Center, founding Co-Director of the
General Motors Collaborative Research Laboratory,
and Director of the DARPA MISCIC Center. He is currently the Benjamin
T. Rome Dean of the Whiting School of Engineering, The Johns Hopkins
University, Baltimore, MD, USA, and a Professor of electrical and computer
engineering. He has authored or co-authored over 250 papers and conference
proceedings and has 13 patents.
Dr. Schlesinger is a Fellow of the SPIE.
Jeyanandh Paramesh (SM’10) received the
B.Tech. degree in electrical engineering from IIT
Madras, Chennai, India, the M.S. degree in electrical
engineering from Oregon State University, Corvallis,
OR, USA, and the Ph.D. degree from the University
of Washington, Seattle, WA, USA.
He was with AKM Semiconductor (Analog
Devices), Motorola, and Intel, where he was
involved with product development and research
positions. He is currently an Associate Professor of
electrical and computer engineering with Carnegie
Mellon University Pittsburgh, PA, USA. His current research interests include
the design of RF and mixed-signal integrated circuits and systems for a wide
variety of applications.
James A. Bain (SM’12) received the B.S. degree
in materials science and engineering from the Uni-
versity of Pennsylvania, Philadelphia, PA, USA,
in 1988, and the M.S. and Ph.D. degrees in materials
science and engineering from Stanford University,
Stanford, CA, USA, in 1991 and 1993, respectively.
He is a Professor with the Electrical and Computer
Engineering Department, Carnegie Mellon Univer-
sity, Pittsburgh, PA, USA, where he also has a
courtesy appointment with the Department of Mate-
rials Science and Engineering and is the Associate
Director of the Data Storage Systems Center. He has authored or co-authored
more than 225 papers in the field of magnetic, optical, electrical, thermal, and
mechanical devices and materials for information technology. He currently has
active research programs in heat assisted magnetic recording, and resistive
switches for memory and reconfigurable electronics.
Dr. Bain is a Senior Member of the IEEE Magnetics Society, the IEEE
Electron Devices Society, and the IEEE Photonics Society.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES
Prediction of Nonlinear Distortion in Wideband
Active Antenna Arrays
Katharina Hausmair, Sebastian Gustafsson, Student Member, IEEE, César Sánchez-Pérez,
Per N. Landin, Ulf Gustavsson, Thomas Eriksson, Member, IEEE, and Christian Fager, Senior Member, IEEE
Abstract— In this paper, we propose a technique for com-
prehensive analysis of nonlinear and dynamic characteristics
of multi-antenna transmitters (TXs). The analysis technique is
enabled by the development of a Volterra series-based dual-
input model for power amplifiers (PAs), which is capable of
considering the joint effects of PA nonlinearity, antenna crosstalk,
and mismatch for wideband modulated signals. By combining
multiple instances of the PA model with linear dynamic antenna
simulations, we develop the analysis technique. The proposed
method allows the prediction of the output signal of every
antenna in an arbitrarily sized TX array, as well as the total
far-field radiated wave of the TX for any input signal with low
computational effort. A 2.12-GHz four-element TX demonstrator
based on GaAs PAs is implemented to verify simulation results
with measurements. The proposed technique is a powerful tool
to study hardware characteristics as, for example, the effects
of antenna design and element spacing. It can be used in cases
where experiments are not feasible, and thus aid the development
of next generation wireless systems.
Index Terms— Active antenna array, antenna crosstalk, mis-
match, multiple-input multiple-output (MIMO) transmitter (TX),
power-amplifier (PA) modeling.
I. I NTRODUCTION
W IRELESS communication systems face a steadily
growing demand for higher data rates. However,
the radio spectrum is a limited resource. Multiple-input
multiple-output (MIMO) systems can be utilized to increase
spectral efficiency [1]. For this reason, modern wireless
telecommunication standards, such as LTE and Wi-Fi, include
Manuscript received February 7, 2017; revised April 15, 2017; accepted
April 20, 2017. This was supported in part by the GigaHertz Center in a
joint project financed by the Swedish Governmental Agency for Innovation
Systems (VINNOVA), Chalmers University of Technology, Ericsson, Infineon
Technologies Austria, Ampleon, National Instruments, and Saab, in part by
the EMPIR Program co-financed by the Participating States, and in part
by the European Union’s Horizon 2020 Research and Innovation Program.
(Corresponding author: Katharina Hausmair.)
K. Hausmair and T. Eriksson are with the Communication Systems
Group, Department of Signals and Systems, Chalmers University of
Technology, 412 96 Gothenburg, Sweden (e-mail: hausmair@chalmers.se;
thomase@chalmers.se).
S. Gustafsson and C. Fager are with the Microwave Electronics Labora-
tory, Department of Microtechnology and Nanoscience, Chalmers University
of Technology, 412 96 Gothenburg, Sweden (e-mail: sebgus@chalmers.se;
christian.fager@chalmers.se).
C. Sánchez-Pérez was with the Microwave Electronics Laboratory,
Department of Microtechnology and Nanoscience, Chalmers University
of Technology, 412 96 Gothenburg, Sweden. He is now with Qam-
com Research Technology AB, SE-412 85 Gothenburg, Sweden (e-mail:
cesar.sanchez@qamcom.se).
P. N. Landin and U. Gustavsson are with Ericsson AB, 417 56 Gothenburg,
Sweden (e-mail: per.landin@ericsson.com; ulf.gustavsson@ericsson.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/TMTT.2017.2699962
0018-9480 © 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
1
the use of multiple antennas. Large-scale antenna systems,
which comprise hundreds of antennas, have become a hot topic
in the research community [2].
The use of several transmit paths in a transmitter (TX)
increases system complexity and cost [3]. Therefore, inte-
grated solutions, as have been used in radar applications
for many years, are preferred. Such integrated designs avoid
costly components, such as bulky isolators between power
amplifiers (PAs) and antennas [4]. However, such designs are
vulnerable to crosstalk due to mutual coupling between the
antennas, and antenna mismatches. These effects, together
with the nonlinear behavior of the PA, cause nonlinear
distortion at the TX outputs and thus undesired radiated
field properties. Predicting the output of a multi-antenna
TX suffering from such distortion is essential for assessing
its overall performance. It is also necessary for identify-
ing the need for techniques to compensate for undesired
effects and for the design of such techniques, such as digital
predistortion (DPD) [5], [6].
Farsi et al. [7] present a low-complexity technique to model
the nonlinear characteristics of the different PAs in an active
antenna array system. Each PA is modeled by a combination of
a core model that is common to all the PAs in the array and an
individual model. The models are based on single-input single-
output (SISO) wideband PA models. However, the interaction
between the nonlinear behavior of the PA and the effects
of antenna coupling and mismatches under wideband signal
conditions cannot be described by conventional PA models
used for SISO TXs. Qi and Aïssa [8] take a system level-
oriented approach to investigate the effects of PA nonlinearity,
in-phase/quadrature-phase (I/Q) imbalance, and crosstalk in
MIMO beamforming systems. They even propose a compen-
sation method for the undesired effects. Rather than analyzing
the performance of the TX hardware, the technique proposed
in [8] allows to estimate the overall system performance,
including the channel and receiver, in terms of average symbol
error probability. However, a memoryless model is used for
the PA. Hence, the analysis is not suitable for wideband
signals, which require dynamic effects to be considered [9].
In [4] and [10], a hardware-oriented approach is used to
predict radiation patterns of active antenna arrays with direct
connections between PAs and antennas. In both papers, dual-
input PA models based on polyharmonic distortion (PHD)
models (nonlinear scattering functions) [11] and antenna
S-parameters are used to investigate the effects of mutual
antenna coupling and mismatches on the behavior of PAs
and on the overall performance of a TX antenna array.
The proposed methods are frequency-domain-based and
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES

quasi-static. Therefore, they are not suitable for analyzing

multi-antenna TXs with wideband signals as used in modern
wireless systems.
Recent work by Zargar et al. [12] presents a dual-input
PA model that is capable of modeling large reflections at both
input and output PA ports while also considering dynamic
effects. However, multi-antenna systems are not investigated.
In [13], we introduced a powerful technique to predict and
analyze the performance of multi-antenna TXs. It was evalu-
ated with spectrum measurements for a two-path TX. In this
paper, we extend our work in [13] by presenting the derivations
for the equations. We give detailed explanations of all the steps
that are required to implement our technique. We have also
extended our technique to account for nonconstant frequency
response of the antennas, thereby extending its accuracy for
wideband signal excitations. Furthermore, we now include a
comprehensive experimental evaluation, where we evaluate our
technique using both time- and frequency-domain measure-
ments of a four-path TX.
With our method, it is possible to predict the output of
every single antenna of an arbitrarily sized TX array for
any MIMO or digital beamforming input signal scenario.
Dynamic effects in multi-antenna systems can be predicted by
incorporating a time-domain dual-input PA model into antenna
array simulations. The PA model, which is similar to the mod-
eling approach presented in [12], considers PA nonlinearity,
antenna crosstalk, and mismatch at the same time. While our
approach is also related to the PHD-based approaches, this
paper includes both a dual-input PA model that is capable of
considering nonlinear dynamic effects and its incorporation
into multi-antenna systems. Therefore, the proposed analysis
technique enables completely new possibilities to analyze
hardware effects in integrated multi-antenna TXs with wide-
band signals. This makes it a convenient and valuable tool for
the design and development of future wireless systems.
This paper is organized as follows. Section II gives an
overview of multi-antenna TXs utilizing active antenna arrays
and the effects that are present in such TXs. In Section III,
we give the full derivation of the Volterra series-based dual-
input PA model that we proposed in [13]. Then, in Section IV,
we show how the proposed PA model can be used to predict
the output of a TX with an arbitrary number of antenna
elements. After that, we present the design of an experi-
mental four-element TX for measurements and simulations in
Section V. Measurement results are used to validate the
simulation results in Section VI. An outlook of how our
work could be utilized and continued in the future is given
in Section VII together with our conclusions.
II. M ULTI -A NTENNA TX S YSTEM M ODEL
The model is formulated in the equivalent discrete-time
low-pass description, as is commonly done when modeling
RF PAs [14]. A block diagram of the multi-antenna TX
is shown in Fig. 1. The multi-antenna TX has L parallel
transmit paths. Each path operates in the same frequency
band and consists of an RF PA, which is connected to one
antenna element in the transmit array. In our system model,
we refer only to PAs and antennas. However, if there is no
Fig. 1. Multi-antenna TX system model with L transmit paths. Each path
consists of one PA connected to an antenna element. All PAs are operated in
the same frequency.
I/Q imbalance present, this system model may represent the
full multi-antenna TX chain from digital-to-analog converter
to antenna. The signal b2i (n) describes the PA output voltage
wave of the i th TX path at time step n. The incident wave
a1i (n) is the input signal to the PA of the i th branch. The signal
a2i (n) is a wave incident to the output of the i th PA. It arrives
from the antenna and contains contributions from antenna
crosstalk and mismatches. Finite impulse response (FIR) filters
can be used to describe the relation between an incident wave
a2i (n) and the output signals b2i (n) as

L 
K
a2i (n) = λil (k)b2l (n − k)
l=1 k=0
K
= (λi (k))T b2 (n − k) (1)
k=0
where λil (k) is the kth of K + 1 filter coefficients of the
FIR filter that describes the contribution of the lth antenna
to the incident wave a2i (n) of the i th antenna, λi (k) =
[λi1 (k), . . . , λi L (k)]T and b2 (n − k) = [b21(n − k), . . . ,
b2L (n − k)]T . The array scattering parameters (S-parameters)
Sant describe the characteristics of an antenna array in fre-
quency domain. The time-domain FIR filters in (1) can be
extracted from antenna array S-parameters given over a range
of frequencies. If the antennas are wideband compared with
the signal bandwidth, the single-frequency S-parameters of the
antenna array can be used to describe the relation between the
incident waves a2i (n) and the output signals b2i (n) as
a2i (n) = (λi (0))T b2 (n) (2)
where λi (0) is an L × 1 vector containing the i th column
of the L × L S-parameter matrix at center frequency. This
description is equivalent to the FIR filter representation for a
filter with K = 0.
In a TX without any distortion, crosstalk, or mismatch,
a2i (n) equals zero, while a1i (n) is the signal that, amplified,
would result in b2i (n), radiated by the i th antenna. However,
realistic PAs show dynamic nonlinear behavior, and in multi-
antenna systems also crosstalk and mismatches are present.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

HAUSMAIR et al.: PREDICTION OF NONLINEAR DISTORTION IN WIDEBAND ACTIVE ANTENNA ARRAYS 3

The effects of this combined behavior result in dynamic where α, β, and γ are the model coefficients, and M is the
nonlinear distortion. memory depth. Memory effects are introduced to make the
In order to predict the output of a TX with multiple model suitable for wideband signals, where dynamic effects
antennas, the joint effects of PA nonlinearity, crosstalk, and need to be considered [9]. The terms described by (3a)
mismatches have to be considered. Therefore, in Section III, are linear dynamic effects of the PA on the input a1i (n),
a time-domain dynamic nonlinear PA model with two inputs, while (3b) describes linear dynamic antenna reflection and
corresponding to a1i (n) and a2i (n), is developed. The pre- coupling effects. In (3c), the third-order nonlinear dynamic
sented model is suitable for analysis of multi-antenna TX sys- effects of the PA on a1i (n) are described. Finally, (3d) contains
tems with wideband input signals. joint third-order nonlinear effects, which arise from mixing
of mutual antenna coupling, antenna mismatches, and PA
III. PA M ODELS FOR M ULTI -A NTENNA TX S YSTEMS nonlinearity. As explained before, only linear terms of the
First, the baseband description of a dual-input RF PA that signal a2i (n) occur with significant power. While the effects
considers two input signals around the same carrier frequency described in (3a) and (3c) are similar to the effects experienced
f c is derived. Similar to the model presented in [12], the result- by a PA in an SISO TX, the terms in (3b) and (3d) appear
ing model is based on the Volterra series approach [15]. only in systems with multiple antennas.1
However, the structure of the model here is adapted to fit It can be seen that (3) contains only odd-order combinations
the description of the output signal of a PA in the presence of signals, where in each combination there is exactly one
of antenna mismatch and crosstalk. To avoid complexity more nonconjugate term than conjugate terms. This is due
issues, the presented model is then reduced, following the to the fact that only these specific combinations will result
memory polynomial approach [9]. We also show a reduced in signal components located in the frequency band that is
quasi-static form of the proposed model that is equivalent to relevant to the description of the nonlinear system [17].
PHD models [11]. B. Reduced Dual-Input PA Models
Introducing memory according to the full Volterra series
A. Dynamic Dual-Input PA Model leads to extremely high model complexity, as is demonstrated
As shown in Fig. 1, the two inputs to the i th PA of a multi- in (3), and in Appendix A. Because of the high complex-
antenna TX are a1i (n) and a2i (n), while the output is denoted ity, a full Volterra-based model is infeasible. We, therefore,
by b2i (n). Like in a conventional single-input PA, nonlinear propose to reduce (3) to a memory polynomial structure [9].
terms and memory effects depending on the input signal a1i (n) In this structure, cross terms between a signal and terms of
are expected at the output of the PA. If the second input a2i (n), the same signal with different delays are not considered. For
which depends on crosstalk and mismatches, can be considered example, after reducing (3c), following the memory poly-
relatively small in power, only linear terms of a2i (n) need to be nomial approach, only terms where m 1 = m 2 = m 3 are
considered [16]. Due to the dynamic behavior of the system, considered. Hence, pruning the Volterra series-based model
also past values of a2i (n) may have an effect. In addition to a memory polynomial structure results in
to that, the signal a2i (n) mixes with the PA output in the ⎫
M1 ( P1
 −1)/2
(2 p+1) ⎪
⎬
nonlinear PA. These mixing terms also need to be considered αm 1 a1i (n − m 1 )
b2i (n) = (4a)
in the model. m 1 =0 p=0 ⎪
⎭
A Volterra series-based model that fits this structure is given × |a1i (n − m 1 )| 2 p
⎫
in Appendix A. The resulting model for the i th branch of the M2 ⎪
⎪
+
(1)
β0 m 2 a2i (n − m 2 ) ⎪
⎪
TX up to a nonlinear order of 3 is given by ⎪
⎪
m 2 =0
⎪
⎬

M
  
M 3 M 4 ( P2 −1)/2 (4b)
(1)
b2i (n) = αm a (n − m 1 ) (3a) +
(2 p+1)
βm 4 m 3 ⎪
⎪
1 1i ⎪
⎪
m 1 =0 ⎪
⎪
m 3 =0 m 4 =0 p=1 ⎪
⎭

M × a2i (n − m 3 )|a1i (n − m 4 )| 2 p
⎫
+ βm(1)1 a2i (n − m 1 ) (3b) M6 ( P3 −1)/2 ⎪
M5 
⎪
(n − m 5 ) ⎬
m 1 =0 (2 p+1) ∗
⎫ + γm 6 m 5 a2i
. (4c)

M 
M 
M
⎪
⎬ m 5 =0 m 6 =0 p=1   p−1 ⎪ ⎪
+ (3)
αm × (a (n − m )) p+1 ∗
a (n − m ) ⎭
1 m2 m3 (3c) 1i 6 1i 6
m 1 =0 m 2 =m 1 m 3 =0 ⎪
⎭ In (4a), the terms containing only the signal a1i (n) are
∗
× a1i (n − m 1 )a1i (n − m 2 )a1i (n − m 3 )
⎫ combined. These terms describe the behavior of the PA due
 M  M  M ⎪
⎪ to the amplification of a1i (n) and are the same as in a
+ (3)
βm 1 m 2 m 3 ⎪
⎪
⎪
⎪ SISO memory polynomial model. In (4b) and (4c), the effects
m 1 =0 m 2 =0 m 3 =0 ⎪
⎪
∗
⎪
⎬ of coupling and mismatch and the mixing of these effects
× a1i (n − m 1 )a1i (n − m 2 )a2i (n − m 3 ) with PA nonlinearity are described, where in (4b) all terms
(3d)
 M  M M ⎪
⎪
+ (3)
γm 1 m 2 m 3 ⎪
⎪
⎪
⎪
1 Note that mismatch is also present in SISO TXs. The mismatch in SISO

m 1 =0 m 2 =m 1 m 3 =0
⎪
⎪ TXs is a function of the PA input a1i (n), i.e., a2i (n) is a function of a1i (n).
⎪
× a1i (n − m 1 )a1i (n − m 2 )a2i ∗ (n − m ) ⎭ Hence, for an SISO TX with mismatch, (3) inherently reduces to a single-input
3
model depending only on a1i (n).
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES

containing a2i (n) are combined, and in (4c) all terms con-
∗ (n), are combined. Note that the
taining its conjugate, i.e., a2i
nonlinear orders P1 , P2 , and P3 , and the memory tap lengths
M1 , M2 , M3 , M4 , M5 , and M6 that are necessary to obtain a
good model accuracy can be different from each other.
This reduced version of the model given in (3) has lower
complexity, while still considering memory effects. Just as a
single-input memory polynomial model, the model in (4) is
linear in the coefficients. This means that the linear least-
squares method can be used for identification of the model
coefficients from measurement data. However, for the dual-
input model, two input signals, a1i (n) and a2i (n), as well as
the output signal b2i (n) need to be measured at the same time.
A suitable measurement technique to obtain data for coefficient
identification is presented in Section V-B1.
Note that while we chose to prune the full Volterra series-
based model in (3) to a memory polynomial structure, any
other pruning scheme, e.g., the generalized memory polyno-
mial structure [17] or the dynamic deviation reduction-based
pruning approach [18], can be applied just as well.
In order to compare the presented dual-input model to
related research, it is worth mentioning a special case of the
presented model, the quasi-static version. The model in (3)
can be further pruned to a memoryless model, given by
( P1
−1)/2
b2i (n) = a1i (n) α (2 p+1) |a1i (n)|2 p
p=0
=S21
⎛ ⎞
( P2
−1)/2
+ a2i (n) ⎝β (1) + β (2 p+1)|a1i (n)|2 p ⎠
p=1
IV. P REDICTION OF M ULTI -A NTENNA TX O UTPUT
Our goal is to predict the output signals b2i (n) of a multi-
antenna TX, which are combined in b2 (n). Each signal a2i (n)
is a function of the signals b2 (n), as can be seen in (1). Hence,
via the relation given in (1), both sides of the dual-input PA
model in (4) contain current and past samples of the output
signals b2 (n). While the input signals a1i (n) are known for
all time samples, only past samples of the signals b2 (n) can
be known at the current time step n. Therefore, the output
signals b2 (n) can only be computed in a time-stepped manner,
i.e., step by step for each sample time n from n = 0 to N − 1.
In this section, we first present such a time-stepped solution of
the output at each antenna of the transmit array. We show how
the results can be used to also compute the far-field radiation
pattern of the full TX with minimum computational effort.
A. Stepwise Solution of Multi-Antenna TX Output
In order to compute the samples of the output vec-
tor b2 (n), (4) is transformed, such that all current time samples
of b2 (n), i.e., the terms for m = 0, are contained on one side of
the equation. All past samples are combined on the other side
together with all samples of a1i (n). The detailed derivation is
given in Appendix C. Then, (4) can be rewritten as
b2i (n) = (λi (0))T b2 (n)S22,i (|a1i (n)|)
+ (λ∗i (0))T b2 ∗ (n)T22,i (a1i (n))
(5a)
+ f i (a1i (n), b2 (npast )) (7)
where n = [n, . . . , n − max (M1 , M4 , M6 and npast = )]T
[n − 1, . . . , n − max (M2 , M3 , M5 ) − K ]T . Furthermore,
−1)/2
( P2

M4
(5b) S22,i (|a1i (n)|) =
(1)
β0 0 +
(2 p+1)
βm 4 0
m 4 =0 p=1

=S22
× |a1i (n − m 4 )|2 p (8)
M6 ( P3−1)/2
( P3
−1)/2
∗
 ∗  p−1 T22,i (a1i (n)) =
(2 p+1)
γm 6 0
+ a2i (n) γ (2 p+1)(a1i (n)) p+1 a1i (n) (5c)
m 6 =0 p=1
p=1  ∗  p−1
× (a1i (n − m 6 )) p+1 a1i (n − m 6 ) . (9)
=T22
The remaining terms of (4), i.e., all past samples of b2 (n) and
where (5a) relates to (4a), (5b) to (4b), and (5c) to (4c). all samples of a1i (n), are contained by
As indicated by the braces in (5), this reduced version of the f i (a1i (n), b2 (npast ))
proposed model can be directly compared with the quasi-static −1)/2
( P1
PHD models [11], which, for the fundamental frequency, are 
M1
(2 p+1)
= αm 1 a1i (n − m 1 )|a1i (n − m 1 )|2 p
described by
m 1 =0 p=0

B21 (n) = A1i (n)S21 (|A1i (n)|) + A2i (n)S22 (|A1i (n)|) 
K
+ (λi (k))T b2 (n − k)
+ A∗2i (n)T22 (A1i (n)) (6) k=1
⎛ ⎞

M4 −1)/2
( P2
where A1i (n) and A2i (n) are the two incident wave phasors (1) (2 p+1)
× ⎝ β0 0 + βm 4 0 |a1i (n − m 4 )|2 p ⎠
and B21 (n) is the corresponding scattered wave. PHD models
m 4 =0 p=1
have been used to predict radiation patterns of active antenna
arrays [4], [10]. However, since they are quasi-static and, 
M2
(1)

K

as such, do not consider the history of the input signals, + β0 m 2 (λi (k))T b2 (n − k − m 2 )
they are not suitable for wideband signals. As illustrated, m 2 =1 k=0


M3  −1)/2
M4 ( P2
the dynamic models proposed in (3) and (4) can, therefore, (2 p+1)
be seen as a generalization of the PHD models, to include + βm 4 m 3 |a1i (n − m 4 )|2 p
memory effects. m 3 =1 m 4 =0 p=1
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

HAUSMAIR et al.: PREDICTION OF NONLINEAR DISTORTION IN WIDEBAND ACTIVE ANTENNA ARRAYS 5


K
B. Prediction of TX Radiation Pattern
× (λi (k))T b2 (n − k − m 3 )
k=0 By including knowledge about the antenna element radi-
−1)/2
M6 ( P3 ation patterns, it is possible to predict the total far-field
 (2 p+1)  ∗  p−1
+ γm 6 0 (a1i (n − m 6 )) p+1 a1i (n − m 6 ) electromagnetic (EM) wave generated by the TX. In most
m 6 =0 p=1 commercial EM software, the single antenna embedded far-
field data are conveniently available by postprocessing of the

K
× (λ∗i (k))T b∗2 (n − k) same simulation results that are used to compute the antenna
k=1
characteristics. Using this data, the total far-field EM wave
−1)/2
( P3 E tot (θ, φ) is calculated as the superposition of the single

M5 
M6
+
(2 p+1)
γm 6 antenna output signals b2i (n), scaled by the far-field radiation
m5
m 5 =1 m 6 =0 p=1
pattern E i (θ, φ) of the corresponding antenna element. This
 ∗  p−1 is given by
× (a1i (n − m 6 )) p+1 a1i (n − m 6 )
E tot (θ, φ, n) = (b2 (n))T E(θ, φ) (14)

K
× (λ∗i (k))T b∗2 (n − k − m 5 ). (10) where E(θ, φ) is an L × 1 vector containing the unity ampli-
k=0
tude far-field pattern of each antenna element with all other
All transmit paths of the TX described separately by (7) are elements terminated with the reference impedance (50 ).
combined to obtain the output of all TX antennas in Even though the far-field is not experimentally evaluated in
Section VI, the theory opens up several interesting possibilities
b2 (n) = S22 (|a1 (n)|)(0)b2 (n) to study how the radiation pattern and far-field distortion is
+ T22 (a1 (n))∗(0)b∗2 (n)+f(a1(n), b2 (npast )) influenced by interactions between PAs and antennas. As an
(11) example, we have used it to investigate far-field distortion
effects in phased-array TXs [19].
with
C. Implementation of the Simulation Technique
S22 (|a1 (n)|)
  In order to use the presented theory to predict the output
= diag S22,1 (|a11 (n)|), . . . , S22,L (|a1L (n)|)
of a multi-antenna TX, several steps are necessary. First,
T22 (a1 (n)) the individual components of the TX, i.e., PAs, antenna ele-
 
= diag T22,1(a11 (n)), . . . , T22,L (a1L (n)) ments, and array configuration, have to be chosen or designed.
f(a1(n), b2 (npast )) Then, the PA model coefficients and the antenna array char-
acteristics have to be identified. The identification is done
= [ f 1 (a11 (n), b2 (npast )), . . . , f L (a1L (n), b2 (npast ))]T
through individual measurements of the components, or from
(0) = [λ1 (0), . . . , λ L (0)]T . (12) circuit/antenna simulation. For example, the antenna array
characteristics can be identified from measurements of the
In (11), all currently unknown samples of the output signals
array scattering parameters versus frequency. The PA model
are included in b2 (n), while all past and known signal samples
coefficients can be identified using active load pull measure-
are contained in f(a1 (n), b2 (npast )). Equation (11) can be
ments [20], [21]. All identified coefficients can then be used
solved analytically for b2 (n). First, (11) is split into real
in (13) to predict the output of the multi-antenna TX for known
and imaginary parts, denoted by {·} and {·}. The real and
input signals.
imaginary parts of the output b2 (n) are obtained by (13),
Using separate, rather simple measurements or simulations
shown at the bottom of this page, and b2 (n) is computed by
for the characterization of the different components, the system
b2 (n) = {b2 (n)} + j {b2 (n)}.
performance can, therefore, be evaluated for different types of
The solution presented in (13) allows the output of all
components without great effort. An implementation example
TX paths to be determined for any input signal combination
for a four-path TX, including all identification procedures,
step by step for each time sample n. Since the mutual cou-
is explained in detail in Section V.
pling is usually dominated by neighboring antenna elements,
the problem is sparse and well-conditioned. Well established
numerical techniques can, therefore, be employed to enable V. MIMO S YSTEM -BASED TX D EMONSTRATOR D ESIGN
simulations of very large multi-antenna TX systems with In the remainder of this paper, the theory presented in
limited computational effort. Sections II–IV is evaluated using a wireless multiantenna TX,

   −1
 {b2 (n)} I − {S22 (|a1 (n)|)(0)} − {T22 (a1 (n))∗(0)} {S22 (|a1 (n)|)(0)} − {T22 (a1 (n))∗(0)}
=
 {b2 (n)} −{S22 (|a1 (n)|)(0)} − {T22 (a1 (n))∗(0)} I − {S22 (|a1 (n)|)(0)} + {T22 (a1 (n))∗(0)}
 
{f(a1 (n), b2 (npast ))}
× (13)
{f(a1(n), b2 (npast ))}
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
6 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES
Fig. 2. Photograph of the antenna arrays, and the dimensions of an antenna
element.
where each TX path is driven by different, independent input
signals. Operating a TX in such a way is commonly done
in wireless communications-based MIMO systems. A four-
element TX demonstrator was designed for simulations and
measurements.
In this section, we explain the design and characterization
of the two main components that are necessary to build such
a TX, i.e., the TX antenna array and the PAs. The measurement
and simulation results are later presented and compared in
Section VI.
A. Antenna Design
A microstrip patch antenna is selected as radiation element
of the antenna arrays. Two rectangular four-element antenna
arrays were designed to be able to observe different coupling
intensities. In Fig. 2, a photograph of the arrays is shown,
where also the dimensions of a single antenna element are
given. For the array with higher coupling, distance between
antenna element centers is 49 mm, and for the array with
lower coupling, the distance is 70 mm. The antenna arrays
were designed using the Keysight Momentum EM simulator
for a resonant frequency of f c = 2.14 GHz, and manufactured
using FR4 substrate [ r = 4.4, tan δ = 0.02, and thickness =
62 mil (1.57 mm)].
The characteristics of the manufactured arrays were deter-
mined in measurements with a two-port vector network ana-
lyzer (VNA) using pairwise measurements with the other
ports terminated in 50 . The measured array scattering
parameters Sant versus frequency are shown in Fig. 3. The
array scattering parameters are shown for only one antenna of
each array, since they are similar for each of the antennas due
to reciprocity. The resonant frequency of the antenna elements
was measured around f c = 2.12 GHz, as opposed to the
targeted 2.14 GHz. This is most likely due to small deviations
of the actual substrate characteristics from the characteris-
tics provided by the manufacturer, which were used in the
design process. However, a small change in antenna resonant
frequency is not critical to our experiments. All subsequent
measurements and simulations are, therefore, referring to a
center frequency of f c = 2.12 GHz. The highest coupling
factor is between two directly adjacent antenna elements, and
it was measured as around −12 dB for the high-coupling array,
and around −24 dB for the low-coupling array. In order to use
Fig. 3. Antenna array characteristics Sant versus frequency for antenna 1
for (a) high-coupling array and (b) low-coupling array. The figures show the
scattering parameters for reflection (S11), adjacent element (S14), opposite
element (S12), and diagonally opposite element (S13). The results for the
corresponding extracted FIR filters with five filter taps are shown in gray
dotted line.
the measured antenna characteristics in the simulations, the
single-frequency S-parameters at center frequency f c =
2.12 GHz, as well as FIR filters describing the antenna char-
acteristics were extracted from the measured array scattering
parameters. The FIR filters were designed using linear least-
squares estimation. For each FIR filter, a test signal x(n)
was filtered in frequency domain using the corresponding
measured S-parameters Sant , resulting in a signal y(n). The
FIR filter coefficients λ = [λ(0), . . . , λ(K )]T were found in
time domain as the least-squares estimate
λ = X+ y (15)
where the pseudoinverse = X+ (X H X)−1 X H
with X =
[x0 , . . . , x K ] and xk = [x(0 − k), . . . , x(N − 1 − k)]T , and
y = [y(0), . . . , y(N − 1)]T . The demonstrator is designed for
an input signal bandwidth of 20 MHz. The bandwidth of the
test signal x(n) was 60 MHz, resulting in an accurate FIR filter
description over the same bandwidth, which also includes the
first sidebands. The results for five-tap filters, i.e., K = 4, are
shown together with the measured S-parameters in Fig. 3.
B. PA Characterization and Modeling
The demonstrator is based on four GaAs PA evaluation
boards from Skyworks with identical design (SKY66001-11).
The frequency range of the PAs is 2.1–2.2 GHz. The PAs
have integrated couplers at their outputs. Each PA in the
demonstrator was supplied with 3.3 V.
In order to model and predict the output of the multi-antenna
TX in simulations, the coefficients α, β, and γ of the dual-
input PA model described in (4) must be identified.
1) Active Load–Pull Measurements: In order to characterize
dual-input PAs as in a multi-antenna TX, it is necessary
to synchronously inject and measure signals both at the PA
input and output at well-defined reference planes [20]. Hence,
a mixed-mode active load–pull measurement setup [21] is used
for the experimental extraction of the dual-input PA model
coefficients. The active load–pull setup is calibrated using a
short-open-load-through technique in a similar way as it is
done with VNAs. Additionally, power and phase reference
calibration is performed using techniques described in [22].
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
HAUSMAIR et al.: PREDICTION OF NONLINEAR DISTORTION IN WIDEBAND ACTIVE ANTENNA ARRAYS
Fig. 4. Block diagram of the active load–pull measurement system that is
used to extract the model coefficients to characterize the PAs.
A block diagram of the measurement setup is shown in Fig. 4.
This setup allows the injection of different multisine signals
at the PA input and output, respectively. The incident and
reflected waves at the PA input/output calibrated reference
plane, i.e., a1 (n), b1 (n), a2 (n), and b2 (n), can be accurately
measured. Hence, this measurement system can be used to
emulate a dual-input PA fitting the described model. The
model coefficients can be extracted from the measured data
following the process explained in Appendix B.
2) Results of Model Coefficient Extraction: Each of the four
PAs was measured and characterized individually. Measure-
ments were taken for an input signal bandwidth of Ba =
20 MHz for both a1 (n) and a2 (n), where the calibration and
measurement bandwidth was set to f s = 5 Ba = 100 MHz
to allow capturing nonlinear effects in sidebands. Both a1 (n)
and a2 (n) were multisine signals with randomly chosen phases
and an amplitude probability density function that matched
that of the signals, which are later used in the multi-antenna
TX demonstrator in Section VI. The average power of the
input signal a1 (n) was set to −7 dBm, and the average power
of the signal a2 (n) was chosen to emulate a coupling factor
of −12 dB between antennas. Using the measured data for
a1 (n), a2 (n), and b2 (n) and (4), the least-squares method was
used to extract model coefficients for each of the four PAs.
The accuracy of the model is evaluated using the normalized
mean square error (NMSE) and the adjacent channel error
power ratio (ACEPR) [23]. The NMSE between the model
output bmod (n) and the measured data bmeas (n) is calculated
as
 N−1
|bmeas (n) − bmod (n)|2
NMSE = n=0 N−1 (16)
n=0 |bmeas (n)|
2
with N being the total number of time samples. The ACEPR
is calculated as
 i.e., the highest coupling. We then show that our technique
f ad j |Bmeas ( f ) − Bmod ( f )|
2
ACEPR =  (17)
f ch |Bmeas ( f )|
2
7
Fig. 5. (a) NMSE versus number of coefficients. (b) ACEPR ver-
sus number of coefficients. The results are for one of the PAs (PA 1).
Each light gray dot indicates a separate combination of model parameters
P1 , P2 , P3 , M1 , M2 , M3 , M4 , M5 , and M6 in (4). The diamonds indicate
results for a single-input model, and the squares results for a quasi-static
model. The solid lines indicate the lowest achievable NMSE and ACEPR
for a specific number of coefficients. The stars indicate the result given in
Table I where the NMSE is −43.9 dB for 29 coefficients with P1 = 9, P2 =
P3 = 5, M1 = 2, M2 = M3 = M4 = M6 = 1, and M5 = 0, and the
corresponding ACEPR is −49.8 dB.
where Bmod ( f ) and Bmeas ( f ) are the Fourier transform of
the model output and the measured data, f ch denotes inband
frequencies, and f ad j denotes frequencies in the adjacent
channel. The ACEPR is calculated separately for both the
upper and the lower adjacent channels, with the maximum
used for evaluation.
Model coefficients were extracted for various combina-
tions of parameters P1 , P2 , P3 , M1 , M2 , M3 , M4 , M5 , and M6 .
Fig. 5(a) shows an example of the resulting NMSEs, and
Fig. 5(b) shows an example of the resulting ACEPRs for
different numbers of coefficients for one of the PAs. Fig. 5 also
indicates the NMSEs and ACEPRs obtained for a single-input
PA model, i.e., a model where a2 (n) is not considered, and
for a quasi-static model, i.e., M1 = M2 = M3 = M4 = M5 =
M6 = 0. As can be seen from Fig. 5, the proposed PA model
outperforms quasi-static and single-input PA models in terms
of NMSE. The ACEPR for the proposed model and the quasi-
static model are very similar, while the single-input model
performs worse. Fig. 5 also shows that there is a tradeoff
between the number of coefficients, i.e., complexity, and
model accuracy. Hence, for each PA, a parameter combination
that leads to a low number of coefficients while achieving
satisfying results for both NMSE and ACEPR has to be found.
The chosen parameter combinations for the four PAs and the
resulting NMSEs and ACEPRs are given in Table I. Results
are given for the proposed PA model as well as a single-input
PA model and a quasi-static model.
VI. R ESULTS
In this section, the simulation results of the proposed tech-
nique are validated against measurement results of the four-
element TX for the array with smallest antenna separation,
can be used to predict the amount of distortion introduced by
crosstalk and mismatch separately from the distortion intro-
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
8 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES
TABLE I
PA I DENTIFICATION R ESULTS FOR D IFFERENT PA M ODELS
Fig. 6. Block diagram of the setup for the evaluation of the four-element
TX demonstrator.
duced by amplification of a1i (n). Finally, measurements were
performed to investigate whether conventional SISO DPD is a
sufficient linearization technique, or if dedicated multi-antenna
TX DPD, often called MIMO DPD, is necessary.
A. Validation for High-Coupling Four-Element Array
A block diagram of the measurement setup of the four-
element TX is shown in Fig. 6, and a photograph of the
setup in the laboratory is shown in Fig. 7. Four differ-
ent and independent orthogonal frequency-division multiplex-
ing (OFDM) signals with bandwidths Ba = 20 MHz and
peak-to-average power ratios of around 8.5 dB were gener-
ated in MATLAB. Two synchronized high-speed dual-channel
arbitrary waveform generators (Agilent M8190A) synthesized
the four driving signals for the PAs. The integrated couplers
of the PA test boards were used to measure the individual
PA output signals. The manufactured four-element antenna
Fig. 7. Photograph of the measurement setup in the laboratory.
arrays were used as TX array. The individual PA output signals
were connected to an RF switchbox with multiple inputs and
one output. Only one signal at a time was switched through to
the output of the switchbox, which was connected to a vector
signal analyzer (VSA; Agilent PXA N9030a). This way, each
signal was individually captured by the VSA. Processing was
done in MATLAB.
For the simulations, the same four OFDM signals as in the
measurements were used. Simulations were performed for the
following techniques and settings:
1) the proposed technique using the proposed dynamic
PA models, given in (4), in combination with a single-
frequency S-parameter description of the antenna array,
given in (2);
2) the proposed technique using dynamic PA models, given
in (4), in combination with a five-tap filter descrip-
tion of the antenna array derived from multifrequency
S-parameters, given in (1);
3) the proposed technique using the quasi-static PA models,
given in (5), in combination with a single-frequency
S-parameter description of the antenna array, given
in (2);
4) conventional single-input PA models for each transmit
path.
A comparison of the spectrum of each individual simulated
PA output to the spectrum of the respective measured PA out-
put is shown in Fig. 8 for the proposed technique using the
dynamic PA models. Fig. 8(a) shows the simulation results
for a single-frequency S-parameters description, and Fig. 8(b)
shows the simulation results for a filter description of the
antenna array. The error spectra are also shown. As can be
seen in Fig. 8, the simulated spectra match well with the
measured spectra for all presented cases. Fig. 9 shows the
normalized error spectra. The error spectra are shown for
the proposed technique using dynamic PA models, as well as
quasi-static PA models. The error spectra for simulations using
single-input PA models are given as well. It can clearly be
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

HAUSMAIR et al.: PREDICTION OF NONLINEAR DISTORTION IN WIDEBAND ACTIVE ANTENNA ARRAYS 9

Fig. 8. PA output spectra of measurement (meas), simulation with the proposed technique (sim), and error (err) for (a) dynamic PA models and single-frequency
S-parameters, and (b) dynamic PA models and filters derived from multifrequency S-parameters.

Fig. 9. Error spectra of PA output. The plot shows the errors for the
proposed technique with the proposed dynamic PA models and single-
frequency S-parameter antenna array description (SP), the proposed tech-
nique with dynamic PA models and filter antenna array description (FIR5),
the proposed technique with quasi-static PA models and single-frequency
S-parameter antenna array description (static), and simulations with single-
input PA models (SI).
Fig. 10. NMSE (left) and ACEPR (right) of PA output. Results are shown
for the proposed technique with the proposed dynamic PA models and
single-frequency S-parameter antenna array description (SP), the proposed
technique with dynamic PA models and filter antenna array description (FIR5),
the proposed technique with quasi-static PA models and single-frequency
S-parameter antenna array description (static), and simulations with single-
input PA models (SI).

noticed that the single-input PA model performs significantly
worse than the proposed technique in all cases. It seems that
the performance of the proposed technique is slightly better
when using the dynamic PA models with an FIR antenna
description performs than when using quasi-static PA models
and dynamic PA models with a single-frequency S-parameters
antenna description.
To give a more exact measure of performance, the NMSEs
and ACEPRs for all simulations are given in Fig. 10. In all
cases, the best performance in terms of ACEPR was obtained
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES

with the proposed technique with the dynamic PA models

in combination with an antenna filter description, followed
by the dynamic PA models in combination with an antenna
single-frequency S-parameter description, and the quasi-static
PA models. With one exception (PA3), the same is true for the
performance in terms of NMSE. It is clear that the simulation
using single-input PA models is not suitable to predict the
output of the four-element TX. The proposed technique with
quasi-static PA models performs slightly worse than with the
proposed dynamic models. This is because of the dynamic
effects of the PA and antenna that are not considered in
the quasi-static PA models. The best overall performance
of the proposed technique using the dynamic PA models with
the antenna filter description can be explained when looking at
the measurements of the antenna scattering parameters versus
frequency in Fig. 3, where it can be seen that the coupling
between the antennas is not perfectly frequency flat within
the measurement bandwidth. In addition to that, the reflection
shows rather strong frequency dependent behavior, though it
contributes only little power compared with the coupling from
adjacent antennas. For these reasons, the single-frequency
S-parameter description does not sufficiently describe the
behavior of the antenna array, and the filter description leads
to a slightly better result.
Overall, it should be noted that the accuracy of our evalua-
tion is limited by several factors. The model coefficients and Fig. 11. Spectra of PA1 for (a) high-coupling array and (b) low-coupling
array. Left: PA is operated in a single-path scenario. Right: MIMO scenario.
parameters are extracted from different, separate measurement Measurements (meas) without SISO DPD and with SISO DPD are compared
setups rather than the setup of the full TX. Ideally, behavioral with simulations (sim) without SISO DPD and with SISO DPD.
model extraction is done in the same setup and environment
as the evaluation measurements, since every difference in the
conditions, e.g., measurement instrument imprecision, cables, amount of additional distortion that has to be expected in
temperature, and so on, can cause small uncertainties that in a multi-antenna TX as compared with an SISO TX. It will,
combination influence the outcome of the result. However, per- therefore, help determine if it is necessary to design advanced
forming such measurements requires the full implementation DPD techniques, or if conventional SISO DPD can be enough
of the TX. Building a full multi-antenna TX for large antenna to reach desired signal quality requirements, such as adjacent
arrays is very costly and time-consuming. In addition to that, channel power ratio, even in a multi-antenna TX.
the measurements that are necessary to extract models from a By compensating for the effects that are introduced during
full multi-antenna TX are extremely complicated, difficult to amplification of a1i (n), our technique can be used to investi-
calibrate and synchronize, and require expensive equipment, gate only mismatch and crosstalk effects. To this end, a SISO
which is often not available. In fact, one of the benefits of our DPD was identified and applied in both measurements and
technique is that it enables the prediction of the output of a simulations. SISO DPDs for each of the paths were designed
multi-antenna TX by measuring its individual components, and separately. This was done by driving each PA in a single-path
by doing so getting an estimate of the performance without the scenario, i.e., by applying a signal to only one path of the TX,
need to implement the whole TX. This allows for investigating while for the other path the signal was set to zero, where in the
many design options, and making design changes in early measurements biasing was on for both amplifiers. A vector-
development phases, and gives insights into the nonlinear switched DPD as proposed in [24] was used. The obtained
interactions between circuits, antennas, and signals. SISO DPD will only compensate for distortion caused by
amplification of a1i (n), while not eliminating the crosstalk and
mismatch effects. Hence, the remaining out-of-band distortion
B. Analysis of Distortion Due to Crosstalk and Mismatch is due to crosstalk and mismatch effects.
An interesting application of our work is the possibility Fig. 11 shows the measured and simulated spectra of the
to investigate mismatch and crosstalk effects in multi-antenna PA output of TX path 1. In Fig. 11(a), the results for the
systems for different antenna arrays. For this purpose, we want high-coupling array are given, and in Fig. 11(b), the results
to be able to observe these effects separately from the effects for the low-coupling array are given. In the plots on the
that are introduced by the nonlinear amplification of a1i (n) in left, the spectra of the PA driven in single-path scenario with
the PA. Investigating only the effects specific to multi-antenna and without SISO DPD are shown. In this scenario, all out-
enables a convenient comparison of the crosstalk effects for of-band distortion is due to amplification of a1i (n). It can
different antenna arrays. It can also give an idea about the be seen that the SISO DPD compensates for the distortion.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

HAUSMAIR et al.: PREDICTION OF NONLINEAR DISTORTION IN WIDEBAND ACTIVE ANTENNA ARRAYS 11

In the plots on the right, the spectra of the PA driven in The proposed integration of characterization and modeling
MIMO scenario with and without SISO DPD are shown. of subcircuits, passive interconnects, and antenna elements as
Two things can be noticed: first, without DPD, the difference described in this paper could, therefore, lay the foundation
between the amount of distortion in single-path scenarios and for the design of such circuits. Hence, applications whose
MIMO scenarios for the different arrays is very small. The design process could benefit from our work range from
distortion due to amplification is higher than the distortion high-performance low-cost wireless communication systems
due to crosstalk, such that the crosstalk distortion is masked. employing (massive) MIMO to radar applications.
Second, by application of SISO DPD, we eliminated the effect
of amplitude distortion. Yet, for the MIMO scenario, there is A PPENDIX
a large amount of out-of-band distortion visible. This means
that SISO DPD cannot compensate for the distortion created A. Volterra-Series-Based Dual-Input PA Model for
by crosstalk. As is expected, the distortion due to crosstalk Multi-Antenna TXs
is clearly worse for the high-coupling array. As can be seen A single-input low-pass equivalent Volterra model up to
in Fig. 11, the simulation results agree with the measurements, nonlinear order P with input a1 (n) and output b2 (n) is given
which shows that the proposed technique can be used to by [25]
analyze the effects of crosstalk and mismatch.

M ( P−1)/2+1

(1)
b2 (n) = αm a (n
1 1
− mk ) +
VII. C ONCLUSION m 1 =0 p=2
⎡
In this paper, we present the derivation of a wideband

M 
M 
M 
M
dual-input PA model, which is then utilized in combination ⎣ ··· ···
with linear dynamic antenna array simulations to predict the m 1 =0 m p =m p−1 m p+1 =0 m 2 p−1 =m 2 p−2
characteristics of a multi-antenna TX. The proposed technique
(2 p−1)
allows the output at every antenna of an arbitrarily sized αm 1 ,m 2 ,...,m 2 p−1
array, as well as the total radiated far-field of the array, ⎤
to be predicted with only low computational effort. Results 
p 2
p−1

are validated in measurements with a four-element TX. The × a1 (n − m k ) a1∗ (n − m l )⎦ (18)

20-MHz signals used in the validation cause dynamic effects, k=1 l= p+1
which define a wideband system, in the PAs as well as the where P is odd and M is the memory depth. As can be seen,
antenna arrays. Hence, our technique can be used as a reliable only odd-order combinations of the input signal a1 (n) need to
analysis tool for wideband multi-antenna TXs. be considered in the baseband model, where each combination
The presented analysis tool can be implemented by design- contains exactly one less conjugate term than nonconjugate
ing and characterizing only two main components: the antenna terms.
array and the PAs. For our evaluation, we use VNA measure- We want to obtain a baseband dual-input Volterra series-
ments to determine the characteristics of the antenna arrays. based model suitable for multi-antenna TXs, where only
However, the antenna array can be designed in dedicated linear terms of the second input occur. Assuming that the
software to obtain the array scattering matrices and the far- two RF input signals are located around the same carrier
field pattern. An actual fabrication of the antenna array is not frequency, (18) can be generalized to the low-pass equivalent
necessarily required. For the PA characterization, we employ of such a dual-input Volterra model with inputs a1 (n) and
a mixed-mode active load–pull measurement setup to emulate a2 (n). This is done by adding all necessary combinations of
a PA in a multi-antenna TX scenario. With this setup, it is a1 (n) and a2 (n) and their conjugates to the model in (18).
possible to acquire the data that is required to identify the These are all odd-order combinations where a2 (n) occurs only
dual-input PA model coefficients. Also in this case, the PA data in linear terms, and where the total number of conjugate terms
could be obtained from CAD simulations in the TX design is one less than the total number of nonconjugate terms. The
stage. dual-input Volterra model for multi-antenna TX is given by
With our technique, it is possible to investigate the effects
of different antenna arrays on system performance without 
M ( P−1)/2+1

(1)
complicated and costly, sometimes even infeasible, experi- b2 (n) = αm a (n − m k ) +
1 1
ments. While the presented demonstrator results are for a m 1 =0 p=2
multi-antenna TX operated as in wireless communications- ⎡
based MIMO systems, where each TX path is driven with 
M 
M 
M 
M
⎣ ··· ···
independent input signals, the presented analysis can equally
m 1 =0 m p =m p−1 m p+1 =0 m 2 p−1 =m 2 p−2
well be applied for any input signal combination. For example,
a very important application of the proposed tool could be (2 p−1)
αm 1 ,m 2 ,...,m 2 p−1
for the analysis of highly integrated millimeter-wave MIMO ⎤
and phased-array radar TX chips. The complexity, density, 
p 2
p−1
and interconnect challenges in such applications prevent any × a1 (n − m k ) a1∗ (n − m l )⎦
in-circuit measurements of the full chip to be performed. k=1 l= p+1
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

12 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES


M ( P−1)/2+1
 C. Derivations for Stepwise Solution of
+ βm(1)1 a2 (n − m k ) + Multi-Antenna TX Output
m 1 =0 p=2
⎡ In order to compute the samples of the output vector b2 (n),

M 
M 
M 
M 
M first, (1) is introduced in (4). Then, all current samples of
⎣ ··· ··· b2 (n) are factored out. Introducing (1) in (4) yields
m 1 =0 m 2 =0 m p =m p−1 m p+1 =0 m 2 p−1 =m 2 p−2
⎫
M1 ( P1
 −1)/2 ⎪
⎪
(2 p−1)
βm 1 ,m 2 ,...,m 2 p−1 a2 (n − m 1 ) (2 p+1) ⎬
αm 1 a1i (n − m 1 )
⎤ b2i (n) = (22a)
m 1 =0 p=0 ⎪
⎪

p 2
p−1 ( P−1)/2+1
 × |a (n − m )|2 p ⎭
a1∗ (n −m l )⎦ +
1i 1
× a1 (n −m k )
k=2 l= p+1 p=2 ⎫

M2 
K
⎪
⎪
⎡ + β0(1)m 2 (λi (k))T b2 (n − k − m 2 ) ⎪
⎪
⎪

M 
M 
M
⎪
⎪
⎣ ··· m 2 =0 k=0 ⎪
⎪
−1)/2
M4 ( P2 ⎪
⎬
m 1 =0 m 2 =0 m p+1 =m p M3 
(2 p+1)
+ βm 4 m 3 |a1i (n −m 4 )|
2p (22b)
⎪
⎪

M 
M m 3 =0 m 4 =0 p=1 ⎪
⎪
(2 p−1) ⎪
⎪
··· γm 1 ,m 2 ,...,m 2 p−1 K ⎪
⎪
× (λi (k))T b2 (n − k − m 3 ) ⎪
⎪
m p+2 =0 m 2 p−1 =m 2 p−2 ⎭
⎤ k=0

p+1 2
p−1
⎫
× a2∗ (n −m 1) a1 (n −m k ) a1∗ (n −m l )⎦. (19) 
M5 
M6 ( P3
−1)/2 ⎪
⎪
+
(2 p+1)
γm 6 m 5 ⎪
⎪
k=2 l= p+2 ⎪
⎪
⎪
⎪
5m =0 m =0
6
 ∗
p=1
 p−1 ⎪
⎬
× (a1i (n − m 6 )) p+1 a1i (n − m 6 ) (22c)
⎪
⎪
⎪
⎪
B. Least-Squares Identification of Model Coefficients 
K ⎪
⎪
× (λ∗i (k))T b∗2 (n − k − m 5 ). ⎪
⎪
⎪
⎭
As explained in Section V-B1, the linear least-squares k=0
method can be used to estimate the model coefficients
α, β, and γ from measurement data, i.e., the mea-
sured data vectors a1 = [a1 (0), . . . , a1 (N − 1)]T , a2 = In (22a), b2 (n) does not occur. Hence, first, (22b) is trans-
[a2 (0), . . . , a2 (N − 1)]T , and b2 = [b2 (0), . . . , b2 (N − 1)]T . formed into
The linear least-squares method can be used for all dual-input 
K
models given in this paper, i.e., the full Volterra series-based β0(1)0 (λi (0))T b2 (n) + β0(1)0 (λi (k))T b2 (n − k)
model in (3) and (19), the model with memory polynomial k=1
structure in (4), and the memoryless model in (5). Using the
measured signals, the model output is written in a matrix 
M2
(1)

K
form as + β0 m 2 (λi (k))T b2 (n − k − m 2 )
m 2 =1 k=0
b2 = [Hα Hβ Hγ ][α T β T γ T ]T . (20)

M4 −1)/2
( P2
(2 p+1)
Each row of the matrix Hα comprises all terms that contain + (λi (0))T b2 (n) βm 4 0 |a1i (n − m 4 )|2 p
m 4 =0
combinations of a1 (n) and a1∗ (n), e.g., a1 (n), a1 (n)|a1 (n)|2 , p=1

a1 (n − 1), a1 (n − 1)|a1 (n − 1)|2 , and so on, where each −1)/2

( P2
column comprises these values for one specific n with 
M4
(2 p+1)
+ βm 4 0 |a1i (n − m 4 )|2 p
n = 0, . . . , N − 1. In the same manner, the matrix Hβ
m 4 =0 p=1
contains all combinations, which include a2 (n), e.g., a2 (n),
a2 (n − 1), a2 (n)|a1 (n)|2 , a2 (n)|a1 (n − 1)|2 , and so on. In the 
K
matrix Hγ , the terms where a2∗ (n) occurs, e.g., a2∗ (n)(a1 (n))2 , × (λi (k))T b2 (n − k)
a2∗ (n)(a1 (n − 1))2 , a2∗ (n)(a1 (n))3 a1∗ (n), and so forth, are con- k=1
tained. The vectors α, β, and γ contain the model coefficients,

M3 
M4 −1)/2
( P2
in the sequence that matches the order of the entries in the (2 p+1)
matrices according to the model structure. + βm 4 m 3 |a1i (n − m 4 )|2 p
The model coefficients are estimated by transforming (20) m 3 =1 m 4 =0 p=1
using the pseudoinverse with

K
× (λi (k))T b2 (n − k − m 3 ). (23)
[α T β T γ T ]T = [Hα Hβ Hγ ]+ b2 . (21) k=0
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

HAUSMAIR et al.: PREDICTION OF NONLINEAR DISTORTION IN WIDEBAND ACTIVE ANTENNA ARRAYS 13

Then, (22c) is transformed into [13] C. Fager, X. Bland, K. Hausmair, J. C. Cahuana, and T. Eriksson,
“Prediction of smart antenna transmitter characteristics using a new
 ∗ T ∗
λi (0) b2 (n) behavioral modeling approach,” in IEEE MTT-S Int. Microw. Symp. Dig.,
Jun. 2014, pp. 1–4.
[14] J. C. Pedro and S. A. Maas, “A comparative overview of microwave and

M6 −1)/2
( P3
(2 p+1)  ∗  p−1 wireless power-amplifier behavioral modeling approaches,” IEEE Trans.
× γm 6 0 (a1i (n − m 6 )) p+1 a1i (n − m 6 ) Microw. Theory Techn., vol. 53, no. 4, pp. 1150–1163, Apr. 2005.
m 6 =0 p=1
[15] M. Schetzen, The Volterra and Wiener Theories of Nonlinear Systems,
2nd ed. Malabar, FL, USA: Krieger Publishing Company, 2006.
( P3
−1)/2 [16] G. Z. E. Nashef et al., “Second order extension of power amplifiers

M6
(2 p+1)  ∗  p−1 behavioral models for accuracy improvements,” in Proc. Eur. Microw.
+ γm 6 0 (a1i (n − m 6 )) p+1 a1i (n − m 6 ) Conf. (EuMC), Sep. 2010, pp. 1030–1033.
m 6 =0 p=1 [17] D. R. Morgan, Z. Ma, J. Kim, M. G. Zierdt, and J. Pastalan, “A gener-
alized memory polynomial model for digital predistortion of RF power

K 
M5 
M6 −1)/2
( P3
(2 p+1) amplifiers,” IEEE Trans. Signal Process., vol. 54, no. 10, pp. 3852–3860,
× (λ∗i (k))T b∗2 (n − k) + γm 6 m5 Oct. 2006.
k=1 m 5 =1 m 6 =0 p=1 [18] A. Zhu, J. C. Pedro, and T. J. Brazil, “Dynamic deviation reduction-
p+1
 ∗
 p−1 based volterra behavioral modeling of RF power amplifiers,” IEEE
× (a1i (n − m 6 )) a1i (n − m 6) Trans. Microw. Theory Techn., vol. 54, no. 12, pp. 4323–4332,
Dec. 2006.

K
[19] C. Fager, K. Hausmair, K. Buisman, K. Andersson, E. Sienkiewicz,
× (λ∗i (k))T b∗2 (n − k − m 5 ). (24) and D. Gustafsson, “Analysis of nonlinear distortion in phased array
k=0 transmitters,” in Proc. Workshop Integr. Nonlinear Microw. Millim. Wave
Circuits (INMMiC), Apr. 2017, pp. 1–4.
[20] H. Zargar, A. Banai, and J. C. Pedro, “DIDO behavioral model extrac-
ACKNOWLEDGMENT tion setup using uncorrelated envelope signals,” in Proc. Eur. Microw.
Conf. (EuMC), Sep. 2015, pp. 646–649.
The authors would like to thank Skyworks Solutions, Inc., [21] S. Gustafsson, M. Thorsell, and C. Fager, “A novel active load–pull
for donating the PA test boards used in the experiments. system with multi-band capabilities,” in Proc. ARFTG Microw. Meas.
Conf. (ARFTG), Jun. 2013, pp. 1–4.
[22] K. Andersson and C. Fager, “Oscilloscope based two-port measurement
system using error-corrected modulated signals,” in Proc. Workshop
R EFERENCES Integr. Nonlinear Microw. Millim. Wave Circuits (INMMIC), Sep. 2012,
pp. 1–3.
[1] H. Inanoglu, “Multiple-input multiple-output system capacity: Antenna [23] M. Isaksson, D. Wisell, and D. Ronnow, “A comparative analysis of
and propagation aspects,” IEEE Antennas Propag. Mag., vol. 55, no. 1, behavioral models for RF power amplifiers,” IEEE Trans. Microw.
pp. 253–273, Feb. 2013. Theory Techn., vol. 54, no. 1, pp. 348–359, Jan. 2006.
[2] E. G. Larsson, O. Edfors, F. Tufvesson, and T. L. Marzetta, “Massive [24] S. Afsardoost, T. Eriksson, and C. Fager, “Digital predistortion using a
MIMO for next generation wireless systems,” IEEE Commun. Mag., vector-switched model,” IEEE Trans. Microw. Theory Techn., vol. 60,
vol. 52, no. 2, pp. 186–195, Feb. 2014. no. 4, pp. 1166–1174, Apr. 2012.
[3] F. Rusek et al., “Scaling up MIMO: Opportunities and challenges with [25] E. G. Lima, T. R. Cunha, H. M. Teixeira, M. Pirola, and J. C. Pedro,
very large arrays,” IEEE Signal Process. Mag., vol. 30, no. 1, pp. 40–60, “Base-band derived volterra series for power amplifier modeling,” in
Jan. 2013. IEEE MTT-S Int. Microw. Symp. Dig., Jun. 2009, pp. 1361–1364.
[4] M. Romier, A. Barka, H. Aubert, J. P. Martinaud, and M. Soiron,
“Load–pull effect on radiation characteristics of active antennas,” IEEE
Antennas Wireless Propag. Lett., vol. 7, pp. 550–552, Sep. 2008.
[5] S. A. Bassam, M. Helaoui, and F. M. Ghannouchi, “Crossover digital
predistorter for the compensation of crosstalk and nonlinearity in MIMO
transmitters,” IEEE Trans. Microw. Theory Techn., vol. 57, no. 5, Katharina Hausmair received the Dipl.Ing.
pp. 1119–1128, May 2009. degree in electrical and information engineering
[6] S. Amin, P. N. Landin, P. Händel, and D. Rönnow, “Behavioral modeling from the Graz University of Technology, Graz,
and linearization of crosstalk and memory effects in RF MIMO transmit- Austria, in 2010. She is currently pursuing the
ters,” IEEE Trans. Microw. Theory Techn., vol. 62, no. 4, pp. 810–823, Ph.D. degree at the Department of Signals and
Apr. 2014. Systems, Chalmers University of Technology,
[7] S. Farsi, J. Dooley, K. Finnerty, D. Schreurs, B. Nauwelaers, and Gothenburg, Sweden.
R. Farrell, “Behavioral modeling approach for array of amplifiers in From 2010 to 2013, she was a Researcher with
active antenna array system,” in Proc. IEEE Topical Conf. Power Amplif. the Signal Processing and Speech Communication
Wireless Radio Appl. (PAWR), Jan. 2013, pp. 73–75. Laboratory, Graz University of Technology. Her
[8] J. Qi and S. Aïssa, “Analysis and compensation for the joint effects of current research interests include signal processing
hpa nonlinearity, I/Q imbalance and crosstalk in MIMO beamforming for communication systems with an emphasis on modeling and compensation
systems,” in Proc. IEEE Wireless Commun. Netw. Conf. (WCNC), of undesired effects occurring in analog circuits.
Mar. 2011, pp. 1562–1567.
[9] J. Kim and K. Konstantinou, “Digital predistortion of wideband signals
based on power amplifier model with memory,” Electron. Lett., vol. 37,
no. 23, pp. 1417–1418, Nov. 2001.
[10] G. Z. E. Nashef et al., “EM/circuit mixed simulation technique for
an active antenna,” IEEE Antennas Wireless Propag. Lett., vol. 10, Sebastian Gustafsson (S’12) received the M.Sc.
pp. 354–357, May 2011. degree in electrical engineering from the Chalmers
[11] D. E. Root, J. Verspecht, D. Sharrit, J. Wood, and A. Cognata, “Broad- University of Technology, Gothenburg, Sweden,
band poly-harmonic distortion (PHD) behavioral models from fast auto- in 2013, where he is currently pursuing the Ph.D.
mated simulations and large-signal vectorial network measurements,” degree at the Microwave Electronics Laboratory
IEEE Trans. Microw. Theory Techn., vol. 53, no. 11, pp. 3656–3664, Group.
Nov. 2005. His current research interests include RF metrol-
[12] H. Zargar, A. Banai, and J. C. Pedro, “A new double input-double output ogy and GaN HEMT characterization.
complex envelope amplifier behavioral model taking into account source
and load mismatch effects,” IEEE Trans. Microw. Theory Techn., vol. 63,
no. 2, pp. 766–774, Feb. 2015.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
14
César Sánchez-Pérez received the Ph.D. degree
from the University of Zaragoza, Zaragoza, Spain,
in 2012.
From 2012 to 2014, he held a post-doctoral posi-
tion with the Microwave Electronics Laboratory,
Chalmers University of Technology, Gothenburg,
Sweden. Since 2015, he has been an RF/Microwave
Specialist with Qamcom Research Technology AB,
Gothenburg. His current research interests include
wireless communications systems with an emphasis
on tunable matching networks and high-efficiency
transmitters.
Per N. Landin received the M.Sc. degree from
Uppsala University, Uppsala, Sweden, in 2007, and
the joint Ph.D. degree from the KTH Royal Institute
of Technology, Stockholm, Sweden, and Vrije Uni-
versiteit Brussel, Brussels, Belgium, in 2012.
From 2013 to 2014 he was a Post-Doctoral
Researcher with the Chalmers University of Tech-
nology, Gothenburg, Sweden. Since 2015, he has
been with Ericsson AB, Gothenburg. His current
research interests include RF measurements and
signal processing, over-the-air measurements, and
system identification applied to power amplifier modeling and linearization.
Ulf Gustavsson received the M.Sc. degree in electrical engineering from
Örebro University, Örebro, Sweden, in 2006, and the Ph.D. degree from the
Chalmers University of Technology, Gothenburg, Sweden, in 2011.
He is currently a Senior Researcher with Ericsson AB, Gothenburg, where
he is also the Lead Scientist with the Marie Skłodowska-Curie Innovative
Training Network, SILIKA. His current research interests include radio signal
processing and behavioral modeling of radio hardware for advanced antenna
systems.
IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES
Thomas Eriksson (M’15) received the Ph.D. degree
in information theory from the Chalmers University
of Technology, Gothenburg, Sweden, in 1996.
From 1990 to 1996, he was with the Chalmers
University of Technology. In 1997 and 1998, he
was with AT&T Labs, Murray Hill, NJ, USA.
In 1998 and 1999, he was with Ericsson Radio
Systems AB, Kista, Sweden. Since 1999, he has
been with the Chalmers University of Technology,
where he is currently a Professor of communication
systems. He was a Guest Professor with Yonsei
University, Seoul, South Korea, from 2003 to 2004. He is leading the research
on hardware-constrained communications with the Chalmers University of
Technology. He is leading several projects on massive multiple-input multiple-
output (MIMO) communications with imperfect hardware, MIMO commu-
nication taken to its limits: 100-Gb/s link demonstration, massive MIMO
test-bed design, satellite communication with phase noise limitations, and
efficient and linear transceivers. He is also the Vice Head of the Department
of Signals and Systems, Chalmers University of Technology, where he
is responsible for bachelor’s and master’s education. He has authored or
co-authored over 200 journal and conference papers and holds 11 patents.
His current research interests include communication, data compression,
and modeling and compensation of nonideal hardware components, such
as amplifiers, oscillators, and modulators in communication transmitters and
receivers, including massive MIMOs.
Christian Fager (S’98–M’03–SM’15) received the
Ph.D. degree from the Chalmers University of Tech-
nology, Gothenburg, Sweden, in 2003.
Since 2015, he has been a Professor with the
Microwave Electronics Laboratory, Chalmers Uni-
versity of Technology, where he is also the Deputy
Director of the GHz Centre for industrial collabora-
tions. He has authored or co-authored over 120 pub-
lications in international journals and conferences.
His current research interests include the area of
energy efficient and linear transmitters for future
wireless communication systems.
Dr. Fager serves as a TPC member of the IEEE IMS and INMMiC Technical
Conferences. He received the Best Student Paper Award at the IEEE MTT-S
International Microwave Symposium in 2003. He is currently an Associate
Editor of IEEE Microwave Magazine.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES 1

Quasi-Lumped-Element Filter-Integrated
Single-Pole Double-Throw Switch
Jin Xu, Qi-Hang Cai, Zhi-Yu Chen, and Yong-Qian Du

Abstract— In this paper, the capacitively coupled LC resonators

with loaded p-i-n diodes are used to develop fixed-response
filter-integrated single-pole double-throw (SPDT) switch. The
filter-integrated on-state channel with reverse-biased p-i-n
diodes can be designed and synthesized using the coupled-
resonator filter theory, and the off-state channel with forward-
biased p-i-n diodes exhibits high suppression. A 1-GHz
filter-integrated SPDT switch with ripple bandwidth of 195 MHz
is designed to demonstrate the proposed design concept.
Measured results show that it has a low on-state insertion loss (IL)
of 1.03 dB, a wide passband, high off-state suppression (OSS),
Fig. 1. Circuit model of proposed filter-integrated SPDT switch.
high port-to-port isolation, and compact circuit size. Since both
the properly misaligned resonant frequencies of LC resonators
and weakly coupling between LC resonators also result in
high suppression, a frequency-agile filter-integrated switchless bandpass filter (BPF) [9], standing-wave filtering struc-
SPDT switch without p-i-n diodes can be designed by using ture [10], capacitively loaded multicoupled line with loaded
the varactor diodes to realize the corresponding capacitors. p-i-n diodes [11], distributed coupling asymmetrical quarter-
A frequency-agile filter-integrated switchless SPDT switch with
constant 1-dB fractional bandwidth of 19.5% is also designed. wavelength resonator with loaded p-i-n diodes [12], and
Measured results show that its operating frequency can be tuned switchable impedance matching circuit connected BPFs [13],
from 0.612 to 1.088 GHz with corresponding frequency tuning have been done to develop filter-integrated switches. However,
range of 58% and IL varying from 2.8 to 2.3 dB. Good return all of these reported filter-integrated switches are based on dis-
loss, high OSS, and port-to-port isolation can be also observed. tributed resonators [7]–[13], and most of them have relatively
Index Terms— Filter integrated, frequency agile, single-pole narrow bandwidth (less than 10%) [7], [8], [10]–[12] which
double throw (SPDT), switch, switchless. cannot meet the requirement of high data-rate communication.
Moreover, these reported filter-integrated switches also suffer
I. I NTRODUCTION from high insertion loss (IL) [7], [12], poor port-to-port
isolation [9], or poor passband selectivity [10]. In addition,
I NTEGRATING multiple functions into one device is an
effective method to reduce circuit size, mismatching loss,
and fabrication cost. Radio frequency (RF) and microwave
to the authors’ best knowledge, there is few literature on
frequency-agile filter-integrated switch. Thus, it is significant
device with integrated filtering response have become a popu- to do some further work on the design of filter-integrated
lar design concept in the recent years, i.e., integrated filtering switch.
antenna array [1], bandpass frequency-selective structure [2], Based on the authors’ previous work [14], this paper
filter-based Wilkinson power divider [3], filtering rat-race presents a novel approach to design filter-integrated single-
hybrid [4], and power amplifier with filtering response [5], [6]. pole double-throw (SPDT) switch by using capacitively cou-
Filter-integrated switch is another multifunction device. Many pling LC resonators. In Section II, a fixed response 1-GHz
efforts, i.e., p-i-n diodes or HEMT transistor loaded stepped- filter-integrated SPDT switch with ripple bandwidth (RBW)
impedance resonator [7], [8], FET loaded quarter-wavelength of 195 MHz is designed. In Section III, a 0.612–1.088 GHz
frequency-agile filter-integrated switchless SPDT switch with
Manuscript received April 5, 2017; revised May 30, 2017; accepted constant 1-dB fractional bandwidth (FBW) of 19.5% is
June 18, 2017. This work was supported in part by the State Key Laboratory of
Millimeter Waves Open Research Program under Grant K201614, in part by designed. Two SPDT switches are designed and fabricated
the National Natural Science Foundation of China under Grant 61401358, on the substrate Rogers RT/duroid 5880 (h = 0.508 mm,
in part by the Young Talent Fund of University Association for Science εre = 2.2, and tanδ = 0.0009). Conclusions are followed
and Technology in Shaanxi under Grant 20160202, and in part by the
Natural Science Foundation of Shaanxi Province under Grant 2017JM6034. in Section V.
(Corresponding author: Jin Xu.)
J. Xu is with the School of Electronics and Information, Northwestern
Polytechnical University, Xi’an 710072, China, and also with the State II. F IXED -R ESPONSE F ILTER -I NTEGRATED SPDT S WITCH
Key Laboratory of Millimeter Waves, Southeast University, Nanjing 210096,
China (e-mail: xujin227@nwpu.edu.cn). Fig. 1 shows the circuit model of proposed filter-integrated
Q.-H. Cai, Z.-Y. Chen, and Y.-Q. Du are with the School of Electronics and SPDT switch which consists of five LC resonators with
Information, Northwestern Polytechnical University, Xi’an 710072, China. capacitively coupling and four sets of p-i-n diodes (DSW11 ,
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. DSW12 , DSW21 , and DSW22 ). One 50  feeding line is directly
Digital Object Identifier 10.1109/TMTT.2017.2722404 connected to the middle-stage LC p resonator to serve as the
0018-9480 © 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES

Fig. 4. Calculated variation of Q e versus L it .

which is independent on the tapped position. Therefore,

the natural angular frequency of the middle-stage LC res-
onator (ωms ) is determined by

ωms = 1/ L i (Cim + 2Cie ). (3)
Fig. 2. (a) Circuit model and (b) coupling routing scheme of ideal third-
order BPF. The external quality factor Q e in Fig. 3 can be calculated
by

ωes ∂(Yin,LC) 
Qe = Z 0 (4)
2 ∂ω  ω=ωes

Fig. 3. End-stage LC resonator.
common port (Port 1), and another two 50  feeding lines
are directly connected to two end-stage LC resonators to serve
as the input–output ports (Ports 2 and 3). Murata capacitors
Cb = 100 pF in Fig. 1 is used to block the direct current (dc)
voltage and bypass RF signal. The dc voltages (VSW1 and
VSW2 ) together with many resistors RSW = 1 k are used
to switch on/off four p-i-n diodes. When the reverse-biased
DSW11 and DSW12 and the forward-biased DSW21 and DSW22
are set, the Port 2 is switched on and the Port 3 switched
off, and vice versa. If the p-i-n diodes has ideal performance,
the on-state port should be equivalent to a third-order BPF,
as shown in Fig. 2(a), where the subscript “i ” denotes the ideal
third-order BPF to distinguish with the components in Fig. 1.
It should be noted in Fig. 2(a) that two negative capacitances
in each J -inverter can be absorbed by the adjacent resonators.
This means that the practical resonant frequencies of LC res-
onators in Fig. 2(a) should be chosen higher than the central
frequency (CF) of this filter f 0 . Fig. 2(b) gives the coupling
routing scheme of the ideal third-order BPF. The design of
the proposed filter-integrated SPDT switch can start with the
ideal third-order BPF, when temporarily not considering the
nonideal factors.
A. Ideal Third-Order BPF
Fig. 3 shows the circuit model of end-stage LC res-
onator which considers the negative capacitance of J -inverter
in Fig. 2(a). Its one-port input admittance Yin,LC can be
derived as
1 − ω2 L i (Ci + Cie )
Yin,LC = . (1)
j ωL it [1 − ω2 (L i − L it )(Ci + Cie )]
The resonance of LC resonator is determined by
Im(Yin,LC) = 0, so that one can find that the natural
angular frequency of the end-stage LC resonator (ωes ) is
determined by

ωes = 1/ L i (Ci + Cie ) (2)
where Z 0 = 50  is the port impedance. And the coupling
coefficients M = M12 = M23 in Fig. 2(b) can be calculated
by

M = J/ bes bms (5)
where
J = ω0 Cie = 2π f 0 Cie
bes = ωes (Ci + Cie )
bms = ωms (Cim + 2Cie ).
The Chebyshev filtering response is designed with the CF of
f 0 = 1.0 GHz, RBW of 195 MHz, and in-band return
loss (RL) of 20 dB. Therefore, Q e = 4.3767 and M =
0.2009 can be found [15]. There are many values of lumped-
element components that can meet this design specification.
L i = 11 nH is first preset in this design for simplicity.
According to f 0 = 1 GHz and (5), Ci = 2.08 pF, Cim =
1.45 pF, and Cie = 0.471 pF are acquired by using the
numerical calculation. Fig. 4 shows the calculated variation of
Q e versus L it by using (4), and L it = 4.79 nH is chosen. The
dashed line in Fig. 5 gives the simulated frequency response
of this ideal third-order BPF.
B. Modification
The discussion in the above texts does not consider the
effect of p-i-n diode switches in on-state channel. Fig. 6 shows
the p-i-n diode switch and its RF equivalent circuit model.
The resistor RSW often has a very large value, and it is open
circuited in the RF equivalent circuit. L s is the parasitic series
inductor, C OFF is the reversed-biased capacitor and RON is the
forward-biased resistor. The input impedance of reverse-biased
p-i-n diode switch Z in,rb can be derived as
Cb + C OFF
Z in.rb = j ωL s + . (6)
j ωCb C OFF
Since Cb  C OFF and ωL s  1/(ωC OFF ) around ω0 are built
in this design, the above (6) can be simplified as
Z in,rb ≈ 1/( j ωC OFF ). (7)
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

XU et al.: QUASI-LUMPED-ELEMENT FILTER-INTEGRATED SPDT SWITCH 3

Fig. 5. Frequency response of the ideal third-order BPF and proposed filter-
integrated SPDT switch.

Fig. 7. (a) Circuit model of off-state channel. (b) Simulated Z in,OFF . (L s =

0.7 nH, RON = 2 , and Cb = 100 pF).

Fig. 6. p-i-n diode switch and its RF equivalent circuit model.

This means that the following relationships:

C ≈ Ci − C OFF
Cm ≈ Cim − C OFF
can be built, after L = L i , L t = L it , and Ce = Cie are
chosen. The input impedance of forward-biased p-i-n diode
switch Z in,fb can be derived as
1
Z in.fb = RON + j ωL s + . (9)
j ωCb
Obviously, Z in,fb is not equal to zero, which means that the
redundancy reactance from the off-state channel may need to
be considered. Fig. 7(a) gives the circuit model of off-state
channel, and Fig. 7(b) plots its simulated real part of input
impedance Re(Z in,off ) and imaginary part of input impedance
Imag(Z in,off ). The small absolute value of Re(Z in,off ) will
increase the IL of proposed filter-integrated SPDT switch, and
the Imag(Z in,off ) is with capacitive property. Thus, the follow-
ing relationship:
1
C p ≈ Ci −
Imag(Z in,OFF )ω0
can be built.
Skyworks product of SMP1345-079LF (L s = 0.7 nH,
RON = 2 , and C OFF = 0.15 pF) which has been validated
in [16] is chosen to realize four sets of p-i-n diodes. According
to (8) and (10), L = 11 nH, L t = 4.79 nH, C = 1.93 pF, Cm =
1.3 pF, C p = 1.6 pF, and Ce = 0.471 pF are determined for
proposed filter-integrated SPDT switch design. Moreover, it is
(8a)
(8b)
Fig. 8. Layout (not to scale) of designed filter-integrated SPDT switch.
interesting found that 1/Imag(Z in,OFF )ω0 is 0.48 pF very close
to Ce . This means that the off-sate channel can be simply
equivalent to a capacitor Ce . The solid line in Fig. 5 also plots
the simulated frequency response of proposed filter-integrated
SPDT switch, when the reverse-biased DSW11 and DSW12 and
the forward-biased DSW21 and DSW22 are set. Good agreement
between ideal third-order BPF and proposed filter-integrated
(10) SPDT switch validates the above derivation.
C. Simulated and Measured Results
Fig. 8 shows the layout of filter-integrated SPDT switch.
The inductors in Fig. 1 are realized by high-impedance
microstrip lines (HIMLs) with fixed width of 0.2 mm, the
shunt capacitors are realized by using radial stubs (RSs) with
fixed angle of 90°, and the series capacitor are realized by
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
4 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES
Fig. 9. Simulated and measured results of fabricated filter-integrated SPDT
switch. (The inset is the photograph of fabricated filter-integrated SPDT
switch.)
interdigital microstrip lines (IDMLs) with fixed finger width
of 0.2 mm. The inductance of the HIML inductor can be
controlled by its length, the capacitance of RS capacitor can be
controlled by its radii, and the capacitance of IDML capacitor
can be controlled by its length, the space between the fingers
and the number of fingers. The detailed physical dimensions
of these lumped-element components can be acquired by
using the iterative optimization method discussed in [17],
and Fig. 8 labels the final physical dimensions. The inset
in Fig. 9 shows the photograph of fabricated filter-integrated
SPDT switch, which occupies a circuit area of 0.12 λg1 ×
0.15 λg1 including bias circuit but excluding the feeding lines.
λg1 is the guided wavelength of 50  microstrip line on the
used substrate at 1 GHz.
Fig. 9 shows the simulated and measured results of fabri-
cated filter-integrated SPDT switch, when one port (Port 2) is
turned on and the other port (Port 3) is turned off (VSW1 =
−10 V and VSW2 = 10 V). The results with Port 2 off and
Port 3 on are similar with no significant differences. Its total dc
power consumption is 0.2 W. The bandpass response on-state
channels is measured with f 0 = 0.99 GHz and 3 dB FBW
of 30.22%. The measured minimum IL of on-state channel is
1.03 dB with output RL (|S22 |) better than 17.1 dB. The on-
state channel also exhibits a 20 dB rejection wide upper
stopband from 1.37 to 7.01 GHz. The common port RL (|S11 |)
better than 15.8 dB is also measured. The measured off-
state suppression (OSS, |S31 |) is better than 52.8 dB over
the whole on-state passband bandwidth, and is better than
14 dB from dc to greater than 8 GHz. The spurious response
around 2.5 GHz deteriorates the OSS, which mainly due to the
nonideal property of used p-i-n diodes. The measured port-to-
port isolation between Port 2 and Port 3 (|S23 |) is better than
50.8 dB over the whole on-state passband bandwidth, and is
better than 36.2 dB from dc to greater than 5.87 GHz. Its
measured input 1-dB compression point (IP1dB ) is greater than
26 dBm at 0.99 GHz. Its measured input third-order intercept
point (IIP3 ) is greater than 34 dBm at 0.99 GHz.
Fig. 10. Circuit model of ideal frequency-agile third-order BPF.
III. F REQUENCY-AGILE F ILTER -I NTEGRATED
S WITCHLESS SPDT S WITCH
When the capacitors in Fig. 1 are realized by the corre-
sponding varactor diodes, C, Cm , and C p can be tuned to
achieve frequency-agile property, and Ce can be tuned to
meet the required coupling coefficients. Thus, a frequency-
agile filter-integrated SPDT switch can be designed with
constant RL or constant bandwidth. The SPDT switch func-
tion in this design concept is still realized by p-i-n diodes.
The IL of such type of frequency-agile filter-integrated SPDT
switch is the sum of BPF, p-i-n diodes, and varactor diodes.
Obviously, the p-i-n diodes with low quality factors will
introduce extra ILs, and also cause the spurious response due
to its nonideal property. Therefore, it is benefit to remove
the p-i-n diodes in the frequency-agile filter-integrated SPDT
switch if it is possible. In some literatures, the intrinsic-
switching capability is used to realize this design goal, i.e., the
cancelation of mixed electric and magnetic couplings existing
between adjacent resonators in [18] and the tunable transmis-
sion zeros to configure a bandpass or bandstop filter response
in [19]. The intrinsic-switching capability realized by using
misaligned resonant frequencies of LC resonators and weakly
coupling between LC resonators is employed here to design a
frequency-agile filter-integrated switchless SPDT switch. This
design concept is impossible in the conventional SPDT switch,
but it is indeed possible in frequency-agile filter-integrated
SPDT switch. When one channel is with bandpass response
passband and the other channel is under detuning state,
a frequency-agile filter-integrated switchless SPDT switch can
be designed. At current case, the IL of proposed frequency-
agile filter-integrated switchless SPDT switch is the sum of
BPF and varactor diodes, and the IL from p-i-n diodes can be
removed. The detailed explanation and design procedure are
discussed in the following texts, and the design of proposed
frequency-agile filter-integrated switchless SPDT switch with
constant 1-dB FBW of 19.5% start with the frequency-agile
third-order BPF, after temporarily not considering the other
factors.
A. Frequency-Agile Third-Order BPF
Fig. 10 shows the circuit model of ideal frequency-agile
third-order BPF, in which L s = 0.7 nH is the series parasitic
inductor of used varactor diodes. L i = 8.2 nH is chosen in
this design. According to Fig. 4, Q e of end-stage LC resonator
in Fig. 10 will decrease as Ci increases. Therefore, to main-
tain a constant 1-dB FBW for the frequency-agile passband,
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
XU et al.: QUASI-LUMPED-ELEMENT FILTER-INTEGRATED SPDT SWITCH
Fig. 11. Variation of 1-dB FBW of ideal frequency-agile third-order BPF
versus C D1 with different L t when RL = 10 dB or 40 dB. (a) L it = 3.2 nH.
(b) L it = 4.2 nH, and (c) L it = 5.2 nH.
Fig. 12. Variation of f c of ideal frequency-agile third-BPF with constant
1-dB FBW of 19.5% versus Ci . (L t = 4.2 nH.)
the corresponding RL should be met by tuning Cim and Cie
when Ci varies. Fig. 11 shows the variation of 1-dB FBW
of this ideal frequency-agile third-order BPF versus Ci with
different values of L t when RL = 10 or 40 dB. For the
specified value of Ci , a wider passband can be acquired when
RL = 1 dB, and a narrower passband can be acquired when
RL = 40 dB. As shown in Fig. 11, the different values of
L it result in the different tuning range of 1-dB FBW when
Ci varies. The desired constant 1-dB FBW of 19.5% can be
acquired when Ci is smaller than about 3 pF for L it = 3.2 nH,
between about 3 to 10 pF for L it = 4.2 nH, and greater than
about 7 pF for L it = 5.2 nH. L it = 4.2 nH is chosen in this
design. Fig. 12 shows the variation of passband frequency ( f c )
of ideal frequency-agile third-order BPF with constant 1-dB
FBW of 19.5% as Ci varies, and the required Cim and Cie are
also given in Fig. 12. As shown in Fig. 12, f c varies from
5
Fig. 13. Circuit model of actual frequency-agile third-order BPF.
Fig. 14. Effect of C0 to frequency tunability of actual frequency-agile third-
order BPF with constant 1-dB FBW of 19.5%.
0.933 to 0.503 GHz as Ci varies from 3 to 10 pF, and the
required RL tuned by Cim and Cie varies from 35.5 to 10.8 dB
to meet the desired constant 1-dB FBW of 19.5%.
Different varactor diodes employed to realize Ci , Cim ,
and Cie results in different frequency tuning range (FTR).
The Skyworks product of SMV1234-079LF (Rs = 0.8 ,
C j = 11.878 pF − 1.238 pF at 850 MHz) is first chosen to
realize Ci . According to Fig. 12, SMV1234-079LF is also a
good candidate to realize Cim , but the Skyworks product of
SMV1233-079LF (Rs = 1.2 , C j = 5.699 pF − 0.831 pF at
850 MHz) is actually chosen in this design after considering
the parasitic component of varactor diodes and the next actual
design circuit. The Skyworks product of SMV1232-079LF
(Rs = 1.5 , C j = 4.521 pF − 0.701 pF at 850 MHz)
is chosen to realize Cie . Fig. 13 shows the circuit model
of actual frequency-agile third-order BPF, when the tunable
capacitor is realized by the corresponding varactor diodes. The
dc voltages (Vi , Vmi , and Vei ) together with many resistors
R S = 5.6 k are used to tune the capacitance of varactor
diodes (Di , Dmi , and Dei ). To provide the dc bias for varactor
diodes, many Murata capacitors Cb = 100 pF and C0 in
Fig. 13 are introduced to block the dc voltage and bypass
RF signal. It is known that the coupling capacitance in Fig. 13
will be determined by the varactor diode SMV1232-079LF
when C0 has a very large value. However, the coupling
capacitance in Fig. 13 will be changed when C0 has a
relatively small value. Fig. 14 shows the effect of C0 to
frequency tunability of actual frequency-agile third-order BPF
with constant 1-dB FBW of 19.5%. The SPICE model of
employed varactor diodes is used in the simulation. f cmin and
f cmax in Fig. 14 represent the lowest passband frequency and
the highest passband frequency of actual frequency-agile third-
order BPF, respectively. Obviously, C0 with small values will
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
6 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES
Fig. 15. Circuit model of proposed frequency-agile filter-integrated switchless
SPDT switch.
TABLE I
D ETAILED M EASURED R ESULTS OF FABRICATED F REQUENCY-A GILE
F ILTER -I NTEGRATED S WITCHLESS SPDT S WITCH
affect f cmin and f cmax . C0 = 2.7 pF is chosen in this design
to improve f cmax but has minor effect on f cmin , so that the
passband frequency can be tuned from 0.62 to 1.01 GHz.
B. Design Mechanism
Based on the discussion in Sections II and III-A, Fig. 15
gives the circuit model of proposed frequency-agile filter-
integrated switchless SPDT switch, where L = 8.2 nH, L t =
4.2 nH, and Murata capacitors Cb = 100 pF and C0 = 2.7 pF
are chosen. D1 , D2 , and dc are realized by SMV1234-079LF,
Dm1 and Dm2 are realized by SMV1233-079LF, and
De1 and De2 are realized by SMV1232-079LF. Thus, the dis-
cussion in Section III-A is still useful in this design.
Fig. 16 is given to illustrate the design mechanism of proposed
frequency-agile filter-integrated switchless SPDT switch. The
dashed line in Fig. 16 shows that it is failed to constitute a
frequency-agile filter-integrated switchless SPDT when V1 =
V2 , Vm1 , and Vm2 , and Ve1 = Ve2 are chosen. When V2 =
Vm2 = 0 V is chosen, the LC resonators in Port 3 is
under the detuning state which results in the suppressed |S31 |
and |S23 |, so that the dashed–dotted line in Fig. 16 shows
that a frequency-agile filter-integrated switchless SPDT can
be successfully constituted. |S21 | and |S23 | can be further
improved by increasing the dc voltage Ve2 to 16 V. The above
result is also effective when the bandpass response passband
also operates at a lower frequency or a higher frequency, since
the larger value of Ve2 results in the weak loading effect from
Port 3 for the lower operating frequency, and the stronger
detuning effect results in a higher suppression for the higher
operating frequency. The next texts will give the detailed
results. The above analysis is also effective for the off-sate
Port 2 and the on-state Port 3.
According to the above discussion, the following setups can
be done, i.e., the on-state Port 2 and off-state Port 3, to achieve
the frequency-agile filter-integrated switchless SPDT switch
function.
Fig. 16. Illustration of frequency-agile filter-integrated switchless SPDT
switch with f c = 0.85 GHz and 1-dB FBW of 19.5%. (V1 = 3.53 V,
Vm1 = 2.17 V, and Ve1 = 4.85 V.)
1) V2 and Vm2 connect to very low dc voltages, so that
two varactor diodes D2 and Dm2 will have large capaci-
tances. The LC resonators in the off-state channel will be
detuned with the LC resonators in the on-state channel.
2) V2e connects to a very high dc voltage so that two
varactor diodes De2 will have a very low capacitance,
resulting in low loading effect from off-state channel.
3) V1 , Vm1 , VC , and Ve1 can now be tuned to achieve a
frequency-agile bandpass response passband.
C. Simulated and Measured Results
Fig. 17(a) gives the layout of fabricated frequency-agile
filter-integrated switchless SPDT switch. The inductors in
Fig. 15 are still realized by HIML inductors. The optimized
physical dimensions are also labeled in Fig. 17(a). The sim-
ulated unloaded quality factor (Q u ) of the end-stage and
common LC resonators is 64.7–325.8 at 0.612–1.088 GHz,
and the simulated Q u of middle-stage LC resonator is
42.8–440.2 at 0.612–1.088 GHz. Fig. 17(b) gives the photo-
graph of fabricated frequency-agile filter-integrated switchless
SPDT switch, which occupies a circuit area of 0.072 λg2 ×
0.09 λg2 including bias circuit but excluding the feeding lines.
λg2 is the guided wavelength of 50- microstrip line on the
used substrate at 0.85 GHz.
Fig. 18 shows the simulated and measured results of
fabricated frequency-agile filter-integrated switchless SPDT
switch with constant 1-dB FBW of 19.5% ± 0.5%, when one
port (Port 2) is tuned and the other port (Port 3) is turned
off. The results with Port 2 off and Port 3 on are similar
with no significant differences. Five sets of simulated and
measured curves numbered 1–5 from left to right are given
Another random document with
no related content on Scribd:
perhaps, when one considers the difficulties of his art; for these
particular forms are, as we shall try to explain immediately below,
easy to copy and trace. As in caligraphy, unfixed and diverse in its
qualities and peculiarities, so with every individual in pen drawing,
certain traits occur in the strokes. In pen-and-ink drawing the more
individual and distinctive the style, the harder it will be to copy or
counterfeit it. But if the markings approach the geometric, definite
and precise, then they are easily copied and imitated. This is why the
little circles and similar curved markings are so frequently used in
animated cartoons. There is nothing ambiguous in the lineaments of
a face made with saucer-like eyes, and a nose like a circle. Its
peculiarities are quickly noticed, easily remembered, and traced with
facility.
As has been explained, an artist rarely finishes an entire set of
drawings for a film without help, but has a staff of helpers. It can be
well understood, then, that an essential to success is that the
members of this staff keep the same quality of line in all the
drawings. One of the difficulties in a staff of helpers is that of keeping
a uniformity of portraiture in the characters. And because the circular
lineaments are easy to trace that is the reason why they are chosen
to form the basis for the details of a face.
 Easily drawn circular forms and curves make for
speed in animated cartoon work.
There is a tendency in every one, even on the part of the author of
the original model, to depart from the first-planned type of face. The
approved way of avoiding this is to have a set of sketches of the
characters drawn on special sheets of paper that are to be used by
all the workers to trace from. In a studio with numerous workers, all
rushing to finish a five-hundred-foot reel in every week, it is the
custom to have plates engraved from the original sketches and a
number of copies printed, so that all may have a set. With these
printed copies it will then be merely a matter of having a steady hand
and an ability to trace accurately from the copy on to a fresh sheet of
paper placed over the illuminated glass of the drawing-board.
No doubt, as it has been referred to so many times, it is clearly
understood now what an important part transparent celluloid plays in
this art. It is employed not only to save the labor of reproducing a
number of times the details of a scene, but also to help keep these
details coincident, or uniform. In a face, there is a certainty that its
lineaments will be the same if it is drawn but once on celluloid; but if
it is copied each time on a long string of successive sheets of paper,
there is a likelihood that it will vary and so give the lines on the
screen an effect of wiggling about.
 There are many little matters of technic and rendering that arise in
this art. For example, in making certain parts of a figure, say a coat,
in solid black, it has been found best, instead of making it an
absolute silhouette, to indicate by the thinnest of white lines the
contours of the details. A sleeve, for instance, should be outlined
with such a white line. This seems to be a lot of trouble for so little,
but, judged by the result on the screen, has been shown to be worth
while.
At this point we can touch upon the question of what is meant by
“animation.” An artist with little experience may make a series of
movement phases for an action, but when the drawings are tested it
is found that they do not animate; that is, give in synthesis the
illusion of easy motion. It may be a matter of incorrect drawing,
perhaps, or he may have the drawings nearly correct, but he has
failed to make use of certain little tricks, or, shall we say, failed to
observe certain dexterous points in the technic of the art?
We will cite one little trick—humoring the vision, if one may put it
this way: have a spot, or patch, of black repeated relatively in the
same position throughout the series of a movement. An example is
that of having the boots of a figure of a solid black. The eye catching
the two black spots as they alternately go back and forth is deluded
with respect to the forcibleness of the animation even if the walking
action is not as correctly drawn as it should be. An added effect is
given to this illusory ruse if a tiny high light is left on the toe of each
black boot.
The final test for drawings for animation is, it stands to reason, the
result on the screen. One may, though, approximately find out
whether or not any sequence of drawings animate by flapping them
in a sort of way akin to the book-form kineograph novelty noted in a
preceding chapter. Two immediately following drawings can be
tested this way: with one hand they are held near one corner
pressed against the drawing-board, then with the other hand the top
drawing is moved rapidly up and down. In this way the two drawings
are synthesized somewhat, and if the action is delineated correctly
there will be some notion of the appearance on the screen.
 This little experiment crudely demonstrates the phenomenon of
after-images and the operation typifies a simple synthesizing
apparatus.
A significant addition to a scene, if it is suited to the story and
consistent with the general plan, is to have some foreground detail in
front of the moving figure, or figures. This sometimes consists of a
rock, a clump of foliage, or a tree trunk. The contrast of the inertness
in these details gives an added force to the animating that takes
place back of their mass.

Foreground details of a pictorial composition help the

animator in several ways. Their inertness, for one thing,
affords a contrast to the moving figure.
This feature of a picture is drawn on celluloid that is placed on top
of the rest of the set having to do with the particular animation. It is
possible, though, for an artist, if he is dexterous, to fasten this inert
foreground to the under-side of the glass in the frame which is
pressed down over the drawings during the photography. The
foreground feature, of course, is cut out in silhouette and fastened
with an adhesive like rubber cement. This cement is an article of
great usefulness in a photographic studio; especially for temporary
use over drawings, as it can be easily rubbed off afterward by the
friction of the finger-tips.
Radically opposite in method to the scheme described above, in
which an inert object helps the animation, is the panorama. In this
screen illusion the figure, which is thought of as moving, occupies
the same position; while the landscape, normally quiet, is in motion.
Certainly we have all experienced the sensation, when seated in a
railway-train waiting for it to go, of suddenly imagining that it has
started; when, in fact, it has not budged. This simply has happened:
while occupied with thoughts not pertaining to our surroundings—
perhaps reading—we casually caught sight of a moving train on an
adjacent track, and as we were in the state of expectancy of at any
moment being on the move, we immediately thought that our
anticipation had been fulfilled. Even if, in a moment or two, we
realize that our senses have deceived us, it is hard to shake off the
first-formed delusion of being in motion.
Now the screen panorama is a similar delusion. We see near the
centre of the screen a figure going through the motions of
progression, but we know perfectly well that he is in the same place
all the time. And we know that the landscape is drawn on a band of
paper that is pushed along back of the figure. All our knowing does
not help us. In spite of it the little figure spectrally advances and the
landscape deceptively passes by as we know it does (visionally)
when we ourselves are running very fast.
 MAKING AN ANIMATED CARTOON PANORAMA.
The figure is depicted in a series of movement phases drawn on
separate sheets of celluloid. These are used continuously, one
at a time, and in their proper order during the photography. The
landscape, drawn on a strip of paper, moves along under the
celluloid little by little in the direction of the arrow.

The manner in which a panorama is produced is this: the

landscape is drawn on a long strip of paper; this is to be moved little
by little and photographed at each place to which it has been moved.
The figure that is to walk, or run, is drawn in the different phases of
action on sheets of celluloid. These are placed in their order over the
landscape during the photography. The separate drawings of the
actions of the figure were drawn so that the bodies remained
relatively in the same place, but the limbs, or heads, varied in
attitudes. The planning of the action in a figure for a panorama is
proceeded with in the same way as that for producing a regular walk
or run. One special care in the work, however, is this: the limbs as
they are sketched in their appropriate attitudes in the several
drawings must not have identical outlines. That is, explaining it in
another way, if all of the set are placed together over the illuminated
tracing glass, no two drawings should correspond with respect to the
positions of the limbs. The bodies in the drawings should exactly
concur in position, but if some attention is given to the rise and fall of
the trunk, as in a typical walk, the screen illusion will be very much
better. Slightly shifting it up and down on a vertical would effect this.
The band of paper with the
landscape is moved in the direction
opposite to that in which the figure is
supposed to go.
The photographer has many things
to think of while he is putting this
panorama effect on a film. He must
move the landscape strip; sometimes
as little as one-sixteenth of an inch at a
time; put a celluloid sheet with one of
the phases of the action in place, get it
in its proper order, and then turn the
camera gearing to make the exposure.
In some special cases he will have
another matter to think of; namely, a
second panorama strip to move, and
at a different speed.
This is when he wishes to give a
little better representation of
verisimilitude than that produced by
the single panorama strip.
Far-off moving objects, as we know,
appear to go slower than those that ILLUSTRATING THE
APPARENT SLOWNESS
are close to us. We are aware of this in OF A DISTANT MOVING
looking at a distant airplane high up in OBJECT COMPARED TO
the sky that we know is going very fast ONE PASSING CLOSE TO
but seems as though it is going very THE EYE.
slowly. And at night an illuminated
railway-train in the valley below us, when we are on an elevation,
seems to creep along like a snail.
To bring it to pass that a panorama have the effect of near objects
going faster than those that are distant, it is necessary to have two
strips of panorama details. One strip will represent the foreground,
which is to be moved much quicker, one-eighth of an inch, or so. A
second strip will answer for the distance, which is moved, about one-
sixteenth of an inch, or even less. If the foreground strip is moved at
rather wide intervals, the effect on the screen will be a little like that
which we see from the window of a railway-coach when telegraph-
poles and near objects seem to fly by.
The panorama strip for the foreground is designed with simple
elements so that it can be cut out in silhouette and laid over the other
one. With reference to the quality of the details of a scene on a
panorama; although it is usual to fill up the whole length with items of
interest, there must be observed some degree of simplicity. Perhaps
it might be best to say that there should be a subordination in the
details, even if they are numerous, and then have some striking
feature or object occurring every once in a while, to catch the eye
and so help the movement.
Objects, too, automobiles and other vehicles, are combined with
these panoramas. This brings us to the consideration of the matter of
animating wheels, or making them turn in the screen illusion.

Some distinguishing mark on a wheel is needed to give it

the screen illusion of turning.
A wheel true and accurately adjusted and going rapidly gives—
with the exception of a blurring of spokes, if there are any—very little
evidence of rotation. It is only when it turns unsteadily, or when there
is some distinguishing mark found on or near the rim, that we see
plainly that the wheel turns. Sometimes it is a stain, a spot on the
tire, a temporary repair, or a piece of paper that has caught in the
spokes that indicates a turning of the wheel. Further amplification is
needless, as a glance at the vehicles, as they pass in the roadway,
will make clear. So the animator, when he wishes to show a wheel
turning, simply copies actuality by drawing a wheel with some such
feature as noted above. A mere black spot on a wheel near the
circumference is sometimes sufficient. It is usual to have the wheels
drawn on thin cardboard and cut out and fastened in their proper
places so that they can be turned. They are turned a little at a time
and photographed after each turn.

To represent the hunter in sketch A suddenly trembling with

fear as in sketch B, two drawings, 1 and 2, with varying
wavy lines are used alternately during the photography.
PHOTOGRAPHY AND OTHER TECHNICAL
MATTERS
 CHAPTER IX

PHOTOGRAPHY AND OTHER TECHNICAL

MATTERS

RESPECTING adaptability and results, the same motion-picture

camera that is used in the field, or the studio, can be used to make
films for animated cartoons. In making cartoons, however, two
particulars at variance with the usual procedure first must be noted:
(1) The camera is pointed downward and not horizontally, as is
ordinarily the case, and (2) with each turn of the camera handle only
one frame—one-sixteenth of a foot of film—is photographed, and not
eight, as is commonly the case.
The camera in making animated cartoons is held, pointing
downward, by a firmly built framework. The artist, having decided on
the dimension of the field for his drawings, determines the height
approximately of the camera above the table top, where the
drawings are placed. Naturally it will be high enough so that when he
works at the table while disposing the drawings, adjusting the
dummies, or in some cases making drawings, his head will not come
in contact with the front of the lens. The particular distance between
the lens and the table top is dependent upon the kind of lens in the
camera. It is a common practice to equip a camera with a two-inch
(fifty-millimetre) lens. It is possible to use a lens of this focus for
cartoons.
There is no special type of structure for supporting the camera
above the board upon which the drawings are placed for
photography. An artist contemplating embarking upon this line of
work, and intending to carry on the whole process from the
beginning to the time when he hands the exposed film to the
laboratory for development, will have a chance to put any inventive
ability that he may have into practice in designing a framework for
the purpose. In building such a structure these things must be
thought of: (1) The structure must be firmly built so that the likelihood
of the camera being jarred is lessened; (2) the distance between the
camera and board to be ascertained, approximately at first; (3) an
arrangement for fixing the camera in a grooved sliding section so
that its exact height can be adjusted when the field and focus are
definitely fixed or there is to be any later readjustment. The camera,
for instance, may get jarred and put out of focus, or get set obliquely
with respect to the lines defining the field.

TYPICAL ARRANGEMENT OF CAMERA AND LIGHTS TO

PHOTOGRAPH DRAWINGS FOR ANIMATED CARTOONS.
C. Camera. L. Lights. M. Mechanism to turn camera
shutter. F. Hinged frame with glass to press down on the
drawings. B. Board holding the registering pegs.
Some animators have mounted their camera so that the same
framework can be used for a small field as well as a larger one. This
necessitates, each time that the size of field is changed, a
troublesome setting of the camera in order again. It is wisdom to
keep to one size of field for all work, so that when the camera is
once in position it need not be changed.
The frame that holds the glass, and which is hinged to the board
where the drawings are placed, and the registering pegs have
already been described. It is an excellent plan to have this board
with the above-named adjuncts separate but screwed down upon the
table top. By having it this way it is possible to have another means
of getting the camera and the field lines adjusted. Then if the outline
of the field on the board and those defining the field in the camera do
not fit each other exactly, the board can be unscrewed, shifted until it
is right, and fastened again.
In any film where there is a preponderance of straight lines—
horizontal ones, especially—it is a serious fault to have the slightest
obliquity. It will be emphasized on the screen. The outlines of the
little rectangular area, where the pictures are taken in the camera,
must coincide with the outlines of the field on the board. When the
field is fixed and permanently marked with ink lines, it is a good plan
to draw a smaller rectangle, one-half inch all around, within the outer
one. The idea of this is to have a limiting area within which all
important matters of the drawing are kept.
If the animator has had any experience with the ordinary still
camera, the practical knowledge gained then will help him in the
matter of focussing, or regulating the diaphragm of the lens, so that
all the details of the picture are sharply defined. This comes next, or
rather in conjunction with the determining of the field and the
permanent fixing of the camera. In a still camera—that is to say, an
ordinary portrait or view apparatus—the focussing is on a ground
glass, while in a cinematographic instrument it is usual to place a
piece of celluloid with a grained surface somewhat like ground glass
into the place where the film passes. The picture is focussed on this
celluloid. Some, however, find a piece of blank film answers the
purpose.
 To the above consideration of setting up the camera and
ascertaining the correctness of the field and the sharpness of the
image, the worker wise in perception will, before beginning any
important work, make a test. This is merely a matter of
photographing a drawing, or two, on a short length of film, taking it
out of the camera, and developing it. Here, again, any knowledge of
photographic processes previously learned will be found useful.
There are in all metropolitan centres film laboratories to which the
animator can send his exposed films to be developed and printed.
But for a test before beginning the work it is prudent and expeditious
to keep a supply of chemicals on hand, and then, in a few minutes, it
will be possible to tell how matters stand in any particular that is in
doubt.
The next step, after the camera has been fixed in place, is to
construct a mechanism by which it can be turned conveniently by the
photographer, as he is seated below at the board where the
drawings are placed. This is contrived by a system of sprocket-
wheels and chain-belts coming from the camera and carried down to
the side of the table top, where it ends in a wheel with a turning
handle. For the average individual this would not be a difficult
construction to put up; but it would be an altogether different problem
if the animator wished to equip his camera with an electric motor to
turn the camera mechanism. In this case he would have many things
to consider, getting the particular type of motor, for instance, that will
operate with the continual turning on and off of the power. Here
certainly the best course is to have an expert install the motor and fix
the intermediary mechanism connecting it with the camera-working
parts.
Electric motors to drive camera mechanisms are in general use
among those who make titles for moving-picture films. For this
particular branch of the industry they are an indispensable adjunct.
It would seem to the spectator in the theatre, unfamiliar with the
technic of cinematography, that when he sees a title held on the
screen for any lengthy period, the practical way of effecting this
would be to have a single picture of this title kept stationary during
 the period. But this is not the way the
matter is worked out. A title in a screen
story is given a certain length of film,
with every frame in this length
containing the same words. The
particular length—footage—allowed for
a title depends upon the amount of its
reading-matter. Some titles are very
long. One such, requiring, say, fifteen
feet, makes it necessary to turn the
camera handle two hundred and forty
times, if the operation is by hand. A
very monotonous job. So title studios
attach a motor and appropriate
mechanism to a camera, and with it,
too, an automatic counter. Then in
photographing a title it is a simple
matter of starting the mechanically
driven shutter, watching the figures on
the counter dial, and when the
required exposures have been
registered, pulling the lever that stops
the mechanism. Where a camera,
however, is used for animated
drawings exclusively, a motor is not
absolutely necessary.
An automatic counter would be a
very useful addition to a camera in
PART OF A LENGTH OF FILM making dissolves. One form of these
FOR A TITLE. fantasies is that in which the screen is
For every second that the perfectly black at first and then a small
wording is viewed on the spot of light appears, which grows
screen, sixteen of these larger by degrees, to reveal at the full
frames pass through the
opening the scene or subject of the
projector.
film. This is produced by a vignetter, or
iris dissolve. A vignetter is a device,
fixed generally in front of a lens, that consists of a number of
crescent-shaped segments of thin metal pivoted on a circumference.
When these segments move in unison toward the centre, they
gradually decrease the aperture in the lens tube. But when the
movement is in a contrary direction, they cause the aperture to open
by degrees. Those who have used an ordinary snap-shot camera no
doubt are familiar with a similar device—the iris diaphragm, or lens
stop. But in the diaphragm the segments do not completely close,
and there is always a tiny opening left in the centre. The iris dissolve,
or vignetter, is made to close completely.
The way by which pictures are “faded on” is to start with the
vignetter closed and then open it while the camera handle is turned
to take the picture. To “fade off” a picture, the process is simply
reversed; i. e., gradually closing the vignetter while the last part of
the picture is being taken.
The most frequent application that an animated cartoon artist
makes of a vignetter is making cross dissolves, or causing one
picture to blend into another. Imagine now that the idea to be
expressed, through the medium of one of these cross dissolves, is
that of a character standing in an attitude of reflection and supposed
to be thinking of how he would look in a complete suit of armor.
There will be two drawings: one with the figure in ordinary dress, and
the other with him clad in the armor. First the picture with ordinary
dress is photographed. During this operation the vignetter is closed
by degrees. When it is closed, the film that was just photographed
upon is wound back again into the magazine. Now, as we know,
during this procedure the light, which was getting weaker and
weaker, proportionately lessened its effect on the sensitized
emulsion of the film, so that its picture-forming property was not all
used up. There is still a certain proportion of photographic potency
left for the next exposure. The next step is to replace the first
drawing with the one showing the character in armor.
 VIGNETTER, OR IRIS DISSOLVE.
Below: Three stages during the movement of the pivoted
segments.
We left the vignetter completely closed, and the same length of
film that had just passed back of the lens has been wound back into
the magazine and is ready to cross the exposure field again and be
photographed upon the second time. Now the vignetter is gradually
opened, the new picture is being taken and blended with the image
of the first picture.
These two procedures in their method of operating and their
effects compensate one another. The gradual closing of the vignetter
has its reciprocal part in the gradual opening; the lessening of the
light strength is reciprocal to the increase of the light strength; then
the fading of definiteness in one picture is made up by the gradually
increasing clearness in the other.
In trick work of this kind a mechanical counter would be very
useful in measuring the length of film as it is turned into the
magazine and then out again. It is understood, of course, that our
particular counter also counts backward. And, again, with reference
to cameras: an animator when he selects his camera should be
certain that he gets one with which it is possible to turn the camera
backward for making these dissolves and any other trick work
involving like manipulation.
Immediately above we gave certain reasons for the making of
tests on a small piece of film before photographing. Another matter
for which tests should be made is the question of illumination. It is
important that the field should be evenly illuminated. All this is an
affair of adjusting the lights; that is, getting them one on each side of
the camera in their proper positions with reference to the lens
opening and the distance away from the drawing-board.
 DIAGRAM TO EXPLAIN THE DISTRIBUTION OF LIGHT IN A CROSS
DISSOLVE.
A. When the vignetter is gradually closed during the taking of the
first picture. (The film having been wound back is ready to be
photographed upon again for the second part of the procedure.)
B. While the vignetter is gradually opened during the taking of
the second picture. C. The percentages of light in the two
exposures combined and giving the complete exposure time.
The mercury vapor-lamp which, as has been mentioned, is in
general use for cartoon films, has besides its illuminating qualities
another great merit. It is this: it does not emit heat rays. When it is
remembered that an artist sometimes spends hours at a stretch
photographing his numerous drawings for a cartoon film, and that all
this time his head is but a few inches from the lights, this absence of
heat is a desirable feature.
The manner of going about the photography, which is the next
stage of the work, has been touched upon in another part of the
book.
 There are many more minute particulars in the making of an
animated film to be considered. Take, for instance, the technical
questions respecting the preparation of the drawings. In the process
where most of the drawings are made on paper, the paper should be
a fair quality of white linen ledger paper—but not too thick, as
transparency is a thing to think of, and it is preferable, too, that there
be no water-mark. The design of a water-mark would be a disturbing
element in tracing from one drawing to another. Ordinary black
drawing ink is used for the line work, but when a large area is to be
solid black, it has been found best to employ one of the black varnish
stains that are mixed with turpentine. In spite of the turpentine
medium it is possible to apply it to paper. These black stains are an
intense black and do not lose their strength when viewed through the
celluloid sheets.
It is not usual to obliterate a mistake in drawing with white
pigment, as it is an uncertain quantity in photography. Whether or not
it will come out as a patch of gray, or photograph correctly as white,
is difficult to judge beforehand. It is best to take out ink lines that are
not wanted with a sharp-bladed penknife and then smooth the
surface of the paper with an ink eraser (of rubber).
In drawing over the smooth surface of the celluloid a preliminary
cleaning with weak ammonia water will make the ink flow evenly. It is
of course understood that the celluloid sheets can be used again
after any particular film is finished. Ink or pigment can very easily be
washed off with water.
In drawing on celluloid with a pen it is well to select one that will
not scratch the surface. Scratches will hold, in their shallow depths,
enough ink or pigment to break the evenness of a uniform
background. They will come out as spots on the film. A well-worn
pen, one that has been “broken in,” as the pen draftsmen say, is the
best.
The scheme of employing celluloid sheets to hold simple ink
drawings, which scheme is in common usage in the art, has been
adapted to the purpose of holding intricate drawings in distemper
pigment. Before drawing any series of movements on celluloid it is