This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/LAWP.2015.2418336, IEEE Antennas and Wireless Propagation Letters

Fast EM-Driven Size Reduction of Antenna

Structures By Means of Adjoint Sensitivities
and Trust Regions
Slawomir Koziel, Senior Member, IEEE, and Adrian Bekasiewicz

Abstract—In this letter, a simple yet robust and includes surrogate-based optimization (SBO) methods [10],
computationally efficient optimization technique for explicit size such as space mapping [11], or adaptively adjusted design
reduction of antenna structures is presented. Our approach specifications [12]. Despite their huge potential, SBO methods
directly handles the antenna size as the main design objective, still suffer from convergence problems and are sensitive to the
while ensuring satisfactory electrical performance by means of
setup of an underlying low-fidelity model (typically, a
suitably defined penalty functions. For the sake of accuracy, the
antenna structure is evaluated using high-fidelity EM simulation. coarsely-discretized EM one) [13].
In order to maintain computational efficiency of the design In this work, we propose a technique for explicit optimization-
optimization process, it is carried out using cheap adjoint based miniaturization of antenna structures using gradient-based
sensitivities and trust region framework utilized as convergence search with the antenna size being the main objective and
safeguard. Our technique is illustrated through a design of a electrical performance parameters controlled by means of
compact quasi-isotropic dielectric resonator antenna and a UWB
appropriately defined penalty functions [14]. Computational
monopole. Numerical results are supported by physical
measurements of a fabricated prototype. efficiency of the design process is maintained by utilization of
cheap adjoint sensitivities [15]-[17]. For the sake of robustness, in
Index Terms—EM-driven optimization, adjoint sensitivity, particular guaranteed convergence, it is embedded in the trust
trust regions, antenna design, compact antennas. region framework [18]. Our approach is demonstrated using two
examples: a compact quasi-isotropic dielectric resonator antenna,
I. INTRODUCTION and a planar UWB monopole. Numerical results for the latter are
confirmed by experimental validation.
S MALL size has become an important prerequisite for
contemporary antennas, especially for handheld and
wearable applications [1]. On the other hand, design of II. ANTENNA SIZE REDUCTION THROUGH OPTIMIZATION.
miniaturized antennas poses considerable challenges due to ADJOINT-BASED ALGORITHM
the necessity of satisfying strict requirements imposed on In this section, we formulate the antenna design problem as
electrical performance parameters (concerning, e.g., reflection a nonlinear optimization task. We also define an objective
response, gain, radiation pattern, etc.). Consequently, the function that allows us to control both the antenna size and its
design process normally seeks for acceptable trade-offs reflection characteristics, as well as provide a rigorous
between these conflicting objectives. At the same time, description of the adjoint-based optimization algorithm.
reliable antenna evaluation requires high-fidelity EM analysis Operation and performance of the algorithm is demonstrated
[2], which is computationally expensive. in Sections III and IV.
Perhaps the most common approach to simulation-driven
antenna design is primarily based on repetitive parameter A. Design Problem Formulation and Objective Function
sweeps (usually, one parameter at a time with the overall Let R(x) be the EM-simulated antenna model where x is a
procedure guided by engineering insight), unfortunately, it vector of geometry parameters to be adjusted. We distinguish
does not permit efficient control of multiple objectives, two components of the model response R(x) = [A(x) E(x)],
leaving alone the antenna size, especially when the number of where A is the antenna size, whereas E(x) represents electrical
geometry variables is large. Reliable search for truly optimum performance parameters (e.g., reflection vs. frequency, etc.).
design requires automated numerical optimization. However, The antenna design problem to be solved can be formulated
conventional optimization techniques, both gradient-based as a nonlinear minimization task of the form
with numerical derivatives [3] and derivative-free [4] x*  arg min U  R( x)   arg min U  A( x), E ( x)  (1)
x x
(including local methods such as pattern search algorithms [5],
where U is an objective function encoding given design
and global ones, e.g., population-based metaheuristics [6], [7],
specifications.
[8], [9]) are normally prohibitive in computational terms.
In this work, our primary objective is to minimize the antenna
A promising class of techniques for fast antenna design
size A(x), however, one needs to ensure that relevant electrical
performance criteria are satisfied at the same time. Without loss
Manuscript submitted on February 6, 2015.
S. Koziel and A. Bekasiewicz are with the Faculty of Electronics, of generality, let us assume that we are interested in keeping the
Telecommunications and Informatics, Gdansk University of Technology, reflection |S11| below Sm = –10 dB for certain frequency band of
Poland (e-mail: [email protected]; [email protected])

1536-1225 (c) 2015 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 publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/LAWP.2015.2418336, IEEE Antennas and Wireless Propagation Letters

interest. We use the notation Smax = max{|S11|} in that frequency differentiable). This guaranteed the algorithm convergence to at
range. In order to take this requirement into account, the least local optimum of the EM antenna model R.
objective function is defined as follows: We utilize two types of termination conditions for the
U  A( x), E ( x)   A( x)    max ( Smax  Sm )/ | Sm |, 0 (2)
2
algorithm: (i) convergence in argument, i.e., ||x(k+1) – x(k)|| < 1
The penalty function is defined so that it only contributes to (with 1 being a user-defined threshold), or (ii) reduction of the
U if the aforementioned performance condition is violated, trust-region radius below a user-defined value 2, i.e., (k) < 2.
i.e., Smax > Sm. The value of  has to be sufficiently large, so
that the penalty function takes noticeable values (compared to III. CASE STUDIES
A(x)) for relative violation larger than a few percent. In our In this section we present numerical results that demonstrate
experiments,  = 1000, however, this specific value is not operation and efficiency of the proposed optimization scheme.
critical. In practice, when close to optimum, the maximum in- Our case studies include a compact dielectric resonator
band reflection is expected to be close to –10 dB (assuming antenna and a UWB monopole antenna. Experimental
that no other constraints are active). validation for the latter is shown in Section IV.

B. Optimization Algorithm A. Compact Quasi-Isotropic DRA

Our main optimization engine is a gradient-based algorithm Our first example is a quasi-isotropic dielectric resonator
embedded in the trust-region framework [18]. The algorithm antenna shown in Fig. 1. It is constructed using a cuboid shape
produces a sequence of approximate solutions to (1), x(k), k = dielectric resonator (Taconic CER-10, εr = 10, tanδ = 0.0035),
0, 1, …, (x(0) is the initial solution) as follows excited through a coaxial probe localized within the material. A
x ( k 1)  arg min U (G ( k ) ( x)) (3) driven element is fed from the bottom through a coaxial
x:|| x  x ( k ) || ( k ) transmission line. The input impedance is 50 Ohm. The antenna
(k)
The linear model G is a first-order Taylor expansion of R is based on a reference design of [21], and slightly modified to
at x(k), defined as introduce additional degrees of freedom for the probe location,
G ( k ) ( x)  R( x ( k ) )  J R ( x ( k ) )  ( x  x ( k ) ) (4) so that a better control of the structure behavior during
where JR is the Jacobian of R obtained through adjoint optimization process is ensured. The design variables are: x = [a
sensitivity. Adjoint technology is available through certain b c o1 o2 l]T. The solution space is defined by the lower and
commercial EM solvers (here, we use both Ansys HFSS [19] upper bounds l/u: l = [3 3 3 –0.45∙a –0.45∙b 0]T and u = [30 30
and CST Microwave Studio [20] for evaluating the model 30 0.45∙a 0.45∙b 0.9∙c]T. The linear constraints are necessary to
response and its derivatives). The use of adjoints allows ensure that the probe is allocated within the resonator.
obtaining complete information about response sensitivities at Parameters d = 1.26 and g = 0.82 are fixed to ensure 50 ohm
the expense of some small extra computational overhead. The input impedance (all dimensions in mm).
benefits of using adjoint technology increase with the The EM model of the antenna is implemented in Ansys
dimensionality (the number of parameters) of the optimization HFSS [19] (~3000 tetrahedral mesh cells, average simulation
problem at hand, because the CPU cost of evaluating response time is 250 seconds). The design objective is to minimize
gradients increases slowly with the number of design variables. antenna volume A = a × b × c, while maintaining reflection
The process of finding x(k+1) (cf. (3)) and updating the trust below the acceptable level of –10 dB within 2.4 GHz to 2.5
region (TR) radius  is the following (x(k) and (k) are the GHz frequency band. The initial design (reference structure
current design and search radius, respectively): [13]) is x0 = [27 27 14.5 10 0 9]T and its corresponding volume
is 10571 mm3. The optimized design x = [33.17 5 14.94 3.55
1. Set (k+1) = (k);
0.41 6.43]T has been obtained using the algorithm of Section II.
2. Solve x tmp  arg min U (G ( k ) ( x)) ;
x:|| x  x
(k )
||
(k ) A volume of optimized antenna is only 2478 mm3, i.e.,
miniaturization rate is 77% compared to the reference structure.
3. Calculate the gain ratio   U ( R( x ))  U ( R( x )) ;
tmp (k )

Figure 2(a) compares the reflection characteristics of a

U (G ( x ))  U (G ( x ( k ) ))
(k ) tmp (k )
reference antenna (cf. [13]) and the optimized one. It is
4. Update trust-region radius: If  < 0.25 then  (k+1) =  (k)/3; concluded that a considerable volume reduction has no negative
else if  > 0.75 then  (k+1) = 2 (k); end influence on the reflection performance of the optimized
5. If  < 0 go to 2; structure. The maximum reflection of the DRA for the defined
6. Return x(k+1) = xtmp and  (k); band is below –11.6 dB, which is almost 14% lower than the
The candidate solution xtmp is found by optimizing G(k) in the minimal acceptable |S11| of –10 dB. Also, the –10 dB band of
vicinity of x(k) defined by the current TR radius, the latter being the miniaturized DRA is from 2.39 GHz to 2.57 GHz, thus the
further updated based on the gain ratio . The design xtmp is only structure offers 38% broader bandwidth than the reference
accepted if  > 0 (so that the actual improvement of the objective antenna (–10 dB frequency from 2.38 GHz to 2.51 GHz) and
function U(R(x)) has been observed). Because G(k)(x(k)) = R(x(k)) 80% broader than the assumed one.
and JG(k)(x(k)) = JR(x(k)), solving (3) will result in the improvement The radiation patterns of both designs are shown in Fig. 2(b).
of U(R()) for sufficiently small (k) (assuming that the functions The obtained plots indicate that DRA miniaturization has minor
involved are sufficiently smooth, here, first-order continuously influence on the radiation characteristics. The front-to-back
ratio of the reference and miniaturized structure is 2.75 dB and

1536-1225 (c) 2015 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 publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/LAWP.2015.2418336, IEEE Antennas and Wireless Propagation Letters

1.5 dB, respectively. However, 46% better result is achieved at In more general terms, it should be mentioned that this way of
the expense of a slight degradation of the quasi-isotropic terminating the trust region algorithm is typical for noisy
radiation capability, i.e., it differs by 9.8 dB for the compact and functions, and the EM-simulation-based objective function is
5.7 dB for the reference structure. inherently noise (especially when the structure discretization is
The final design has been obtained at a cost of only 17 EM relatively coarse and/or adaptive meshing techniques are utilized).
antenna model evaluations, which corresponds to just 1.2 h of B. UWB Monopole Antenna
the overall simulation time (see Fig. 3 for the convergence plot As a second example, consider a planar monopole antenna
of the antenna size evolution). The convergence analysis composed of two trapezoids and a rectangle, stacked together
reveals that majority of unsuccessful iterations occurred at the into a radiator (see Fig. 4). The structure is constructed on a
very end of the algorithm run so that the reason for Taconic RF-35 dielectric substrate (h = 0.762 mm; εr = 3.5;
termination was actually reduction of the TR radius below the tanδ = 0.0018). The set of adjustable dimensions is x = [a1 a2
prescribed threshold (here, 10–1). a3 b1 b2 b3 l w]T, whereas the parameters d = 0.6 and w0 = 1.7
are fixed (all dimensions in mm). The input impedance is 50
Ohm. The EM model of the structure is implemented in CST
Microwave Studio [20] (transient solver, ~2,500,000 mesh
cells, simulation time is ~15 min). The initial design is x0 =
[11 22 11 4.5 3.5 14 12 40]T.
For the sake of comparison, the antenna is initially optimized
z to minimize maximum reflection within the UWB band (3.1 GHz
to 10.6 GHz). We use the trust-region-based gradient search
(a) x y
algorithm (cf. (3), (4)) but with the objective function defined
b using reflection only. The final design is x*.1 = [11.38 23.38 20.96
4.49 4.79 10.23 13.31 39.49]T mm (|S11|  –15.8 dB for 3.1 GHz
b
to 10.6 GHz). The corresponding antenna footprint is 1710 mm2.
o1 d
a The optimization cost was 24 EM antenna simulations.
o2
g
l o1
c In the next step, the antenna was optimized to reduce its size,
here, starting from x*.1. This order is generally recommended (in
d y t z case the S11-optimized design is available) because size
g
x x reduction is more complex to solve as a constrained problem.
(b) (c)
Fig. 1. Compact DRA: (a) visualization; (b) side view; (c) bottom view. The final design is x*.2 = [8.96 21.05 24.09 7.13 5.64 8.74 8.58
White, light and dark gray represents vacuum, dielectric resonator and 20.02]T mm, and the corresponding antenna footprint is 1038
metallization (copper), respectively. mm2. Figure 5 shows the reflection response at x0, x*.1 and at
0 0o0 dB x*.2, respectively. The optimization cost here is 25 EM antenna
45o -45o simulations. It can be observed that the reflection response still
satisfied the condition |S11| < –10 dB for 3.1 GHz to 10.6 GHz.
Magnitude [dB]

-20
-10
a2
90o -90o
-20 b3
-20
a3
b2 a2
-30 135o 0 dB -135o b1 a1
2
2.5 3 d
Frequency [GHz] 180o l
GND
(a) (b)
w
Fig. 2. Comparison of reflection characteristics (a) and radiation patterns for 
Fig. 4. Geometry of a planar monopole antenna composed of a stack of a
= 90 (b) of reference (– – –) and miniaturized (——) DRA antenna.
rectangle and two trapezoids (ground plane shown as darker gray shade).
0
10000
1 -5
Antenna Size [mm]

10
3

8000
|S11| [dB]

-10
||
(i) (i+1)
||x -x

10
0 6000 -15

4000
-20
-1
10 -25
2 4 6 8 10 12 2 4 6 8 10 12 2 4 6 8 10 12
Iteration Index Iteration Index Frequency [GHz]
(a) (b) Fig. 5. Comparison of UWB monopole |S11| responses: () initial design, (- -
Fig. 3. Optimization history: (a) convergence, (b) evolution of the antenna -) design optimized for minimum in-band reflection, and (—) design
size. optimized for minimum size.

1536-1225 (c) 2015 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 publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/LAWP.2015.2418336, IEEE Antennas and Wireless Propagation Letters

IV. EXPERIMENTAL VERIFICATION ACKNOWLEDGMENT

The UWB antenna structures of Section III.B have been The authors would like to thank Computer Simulation
fabricated and measured. A photograph of both designs is Technology AG, Darmstadt, Germany, for making CST
shown in Fig. 6. Measured reflection characteristics of Microwave Studio available. This work is partially supported by
obtained designs are shown in Fig. 7. The structure optimized the Icelandic Centre for Research (RANNIS) Grant 141272051.
for minimum size slightly violates the assumed design
specifications, which is mostly due to the fact that the EM REFERENCES
antenna model utilized in the optimization process did not [1] M. Bod, H.R. Hassani, and M.M.S. Taheri, “Compact UWB Printed Slot
contain a connector. Antenna With Extra Bluetooth, GSM, and GPS Bands,” IEEE Ant.
Wireless Prop. Lett., vol. 11, pp. 531-534, 2012.
[2] N. Chahat, M. Zhadobov, R. Sauleau, and K. Ito, “A Compact UWB
V. CONCLUSION Antenna for On-Body Applications,” IEEE Trans. Ant. Prop., vol. 59,
The letter presents a computationally efficient technique for no. 4, pp. 1123-1131, 2011.
[3] J. Nocedal, and S.J. Wright, Numerical Optimization, Springer Series in
explicit size reduction of antenna structures. Our approach Operations Research, Springer, 2000.
exploits adjoint-based gradient search, as well as appropriately [4] A.R. Conn, K. Scheinberg, and L.N. Vicente, Introduction to Derivative-
defined objective function that allows for controlling both the Free Optimization, MPS-SIAM Series on Optimization, MPS-SIAM,
antenna size and its electrical performance parameters (in 2009.
particular, a reflection response). As demonstrated through a [5] T.G. Kolda, R.M. Lewis, and V. Torczon, “Optimization by direct
search: new perspectives on some classical and modern methods,” SIAM
design of two antenna structures (a compact quasi-isotropic Rev., 45, pp. 385—482, 2003.
DRA, and a UWB monopole), optimum designs can be [6] M. John and M.J. Ammann, “Antenna optimization with a
obtained at a low computational cost, with guaranteed computationally efficient multiobjective evolutionary algorithm,” IEEE
Tran. Antennas Propag., vol. 57, no. 1, Jan 2007, pp. 260-263.
convergence, and ensuring required reflection bandwidth. The
[7] S. Chamaani, M.S. Abrishamian, S.A. Mirtaheri, “Time-domain design
future research will be focused on simultaneous handling of of UWB Vivaldi antenna array using multiobjective particle swarm
the antenna size, and both its reflection response and radiation optimization,” IEEE Antennas and Wireless Prop. Lett., vol. 9, pp. 666-
figures directly in the optimization process. 669, 2010.
[8] S. Koziel, A. Bekasiewicz, I. Couckuyt, and T. Dhaene, “Efficient
Multi-Objective Simulation-Driven Antenna Design Using Co-Kriging,”
IEEE Trans. Antennas Prop., vol. 62, no. 11, pp. 5900-5905, 2014.
[9] S. Koziel, A. Bekasiewicz, and W. Zieniutycz, “Expedite EM-Driven
Multi-Objective Antenna Design in Highly-Dimensional Parameter
Spaces,” IEEE Antennas and Wireless Prop. Lett., vol. 13, pp. 631-634,
2014.
[10] N.V. Queipo, R.T. Haftka, W. Shyy, T. Goel, R. Vaidynathan, and P.K.
Tucker, “Surrogate-based analysis and optimization,” Progress in
Aerospace Sciences, vol. 41, no. 1, pp. 1-28, Jan. 2005.
[11] S. Koziel, Q.S. Cheng, and J.W. Bandler, “Space mapping,” IEEE
Microwave Magazine, vol. 9, no. 6, pp. 105-122, Dec. 2008.
[12] S. Koziel and S. Ogurtsov, “Rapid optimization of omnidirectional
antennas using adaptively adjusted design specifications and kriging
surrogates,” IET Microwaves, Ant. Prop., vol. 7, no. 15, pp. 1194-1200,
2013.
[13] S. Koziel and S. Ogurtsov, “Antenna design by simulation-driven
optimization. Surrogate-based approach,” Springer, 2014.
[14] S. Koziel and X.S. Yang (Eds.), “Computational optimization, methods
and algorithms,” Series: Studies in Computational Intelligence, vol. 356,
Springer, 2011.
[15] J.W. Bandler and R.E. Seviora, “Direct method for evaluating scattering
matrix sensitivities,” Electr. Lett., vol. 6, no. 24, pp. 773-774, Nov.
1970.
Fig. 6. Photographs of the fabricated UWB monopole antennas: design
[16] M.H. Bakr and N.K. Nikolova, “An adjoint variable method for time
optimized for reflection (left), and design optimized for minimum size (right). domain TLM with wideband Johns matrix boundaries,” IEEE Trans.
0 Microwave Theory Tech., vol. 52, no. 2, pp. 678-685, Feb. 2004.
[17] P.A.W. Basl, M.H. Bakr, and N.K. Nikolova, “Theory of self-adjoint S-
-5 parameter sensitivities for lossless nonhomogeneous transmission-line
modeling problems,” IET Microwaves, Antennas and Propagation, vol.
|S11| [dB]

-10 2, no. 3, pp. 211-220, Apr. 2008.

[18] A.R. Conn, N.I.M. Gould, and P.L. Toint, Trust Region Methods, 2000
-15 MPS-SIAM Series on Optimization, 2000.
[19] Ansys HFSS, ver. 14.0 (2012), ANSYS, Inc., Southpointe 275
-20 Technology Dr., Canonsburg, PA.
[20] CST Microwave Studio, ver. 2014, CST AG, Bad Nauheimer Str. 19, D-
-25
2 3 4 5 6 7 8 9 10 11 12 64289 Darmstadt, Germany, 2014.
Frequency [GHz] [21] Y.M. Pan, K.W. Leung, K. Lu, “Compact quasi-isotropic dielectric
Fig. 7. Measured reflection characteristics of monopole antenna optimized resonator antenna with small ground plane,” IEEE Trans. Ant. Prop.,
w.r.t. size (- - -) and minimum in-band reflection (—), respectively. vol. 62, no. 2, pp. 577-585, 2014.

1536-1225 (c) 2015 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.