 Research
 Open Access
 Published:
Fast design of multiband fractal antennas through a systembydesign approach for NBIoT applications
EURASIP Journal on Wireless Communications and Networking volume 2019, Article number: 68 (2019)
Abstract
The efficient design of compact antennas operating over multiple bands suitable for the Internet of Things (IoT) is addressed by means of an instance of the systembydesign (SbD) paradigm. More specifically, an iterative strategy that combines different software modules for the search spaceexploration, the fast physicalmodeling of the radiators, and the quality evaluation of the guess solutions is proposed. To enable such a SbD instance, an innovative strategy that exploits an orthogonal array (OA) scheme to determine the training set of a LearningbyExample (LBE) algorithm based on a support vector regressor (SVR) is introduced for the efficient physicalmodeling of the layout to be optimized. The features and the potentialities of the proposed methodological approach are assessed in different applicative scenarios by considering representative numerical and experimental validation examples.
Introduction and rationale
Nowadays, the evergrowing development of the Internet of Things/Internet of Everything (IoT/IoE) is rapidly pushing the concept of “pervasive intelligence” [1–3] to an unprecedented level, in which several small and relatively cheap objects of our daily lives will be densely interconnected through wireless machinetomachine (M2M) interactions [4]. Within this context, 3GPP recently released in June 2016 the first version of the NarrowBandIoT (NBIoT) [5, 6]. NBIoT is an emerging new wireless access technology, which will exist together with the other existing cellular networks like GSM, UMTS, and LTE. The main concept from 3GPP standards is the integration of NBIoT with current LTE networks. Clearly, several heterogeneous communication (e.g., WLAN [7], WIMAX [8], Bluetooth [9], UMTS [10], LTE [11], 5G [12, 13]), tagging (RFID, NFC) [14], and localization (GPS, GALILEO) [15] services will have to be suitably hosted/integrated onto a single device to make the IoT/IoE vision feasible in the next few years. Multiband antennas are a brilliant technological solution to reach this goal, while minimizing the size of the RF frontend [16]. However, standard antennas (e.g., dipoles, loops, square patches) cannot be easily tuned to simultaneously support wireless standards arbitrarily located in the frequency spectrum [16, 17]. Moreover, the constraints on the size, the fabrication technology, the costs, and the antenna robustness further complicate the design problems at hand [17]. To properly address these issues, several different technological/algorithmic approaches for the synthesis of multiband radiators for mobile terminals have been proposed and they are still under development. In this framework, microstrip printed antennas with perturbed fractal shapes have been considered as a viable solution to yield lowprofile and lowcost antennas operating over multiple independent bands [18–24]. As a matter of fact, although selfsimilarity properties of standard/unperturbed fractal shapes enable antennas to exhibit a multiband behavior [16, 25–33], fractals with perturbed geometry proved to be more flexible since they can effectively break the fixed relationships among working resonances [23]. Despite its effectiveness, the underlying design procedure is often extremely cumbersome from the numerical viewpoint. As a matter of fact, nowadays synthesis techniques are based on iterative optimization techniques derived from Evolutionary Algorithms (EAs) [34–37] combined with fullwave antenna simulators. However, it cannot be neglected that in such schemes, each guess solution (i.e., an antenna configuration) must be evaluated through a fullwave simulator [18–23] and the computation time to yield the final design can easily diverge. Moreover, any new antenna design is obtained as the result of an independent synthesis process.
To overcome the above limitations, this work presents a new multiband antenna design concept as a complement and generalization of existing design methodologies. An instance of the systembydesign (SbD) paradigm [38–40] is introduced for the numerically efficient and scalable synthesis of multiband microstrip printed antennas based on perturbed fractal shapes. The key motivation for this choice is that the SbD paradigm, which is defined as “a functional ecosystem to handle complexity in the design of large systems,” enables the formulation of arbitrary design problems in a modular way so that each module (block) of the synthesis process solves a simple task and it can be replaced by a functionally equivalent one on the basis of the design needs, objectives, constraints, and degreesoffreedom (DoFs). More specifically, the SbD paradigm is here applied to define a design procedure (Fig. 1) that integrates (a) an effective solutionspace search (SSS) functional block aimed at iteratively generating a sequence of guess solutions (i.e., perturbed fractal geometries) that converges towards an optimal antenna geometry fitting the problem constraints; (b) a physical response emulator (PRE) block devoted to the fast computation/modelling of the electrical/radiating behavior of each guess radiator; and (c) a physical objective assessment (POA) block responsible for the evaluation of the “quality” (i.e., the matching of the design objectives) of each trial design. Accordingly, the aim of this paper is to present a novel methodology based on the suitable customization of the SbD functional blocks towards a flexible and computationally efficient synthesis of perturbed microstrip fractal antennas.
As for the former functional block (SSS block; Fig. 1), iterative global optimization methods turn out to be suitable tools for handling the DoFs of the antenna design (i.e., its geometrical descriptors) and the nonconvex synthesis problem at hand [34–37]. The POA block is mathematically defined by specifying a “cost function” quantifying the mismatch between the response of each trial solution and the project constraints and requirements. Designing a suitable PRE block is a more challenging task. Indeed, the most natural choice of using a fullwave EM solver for predicting the electromagnetic behavior of each guess solution is generally avoided because of the resulting computational costs. Otherwise, LearningbyExample (LBE) strategies [41] based on support vector regressors (SVRs) [42] represent a suitable candidate alternative. These methods, which have solid mathematical foundations in statistical learning theory [42], are twostep procedures based on an (offline) “training phase,” where a set of inputoutput examples are provided to let the SVR “learn” the corresponding physical relations, and a “testing phase,” where the SVR is used to emulate the fullwave solver predicting in realtime the physical response of the system [41, 42]. However, the effective application of such approaches requires the offline generation of a “representative” training set whose cardinality is proportional to the number of DoFs of the problem at hand. Of course, building the training set by enumerating every possible combination of the control parameters turns out numerically unfeasible since the number of different training configurations would grow exponentially with the number of DoFs [43]. To overcome such a bottleneck, orthogonal arrays (OAs) are here chosen as an enabling tool for exploiting LBE techniques since they allow to enforce the training ensemble to span the search space uniformly (i.e., sampling each design parameter with the same accuracy), while minimizing its size [43–45].
Accordingly, the main novelty of this paper consists in (i) the introduction of a scalable synthesis tool, which can be efficiently reused for several different objectives/designs, since the offline training phase for the electromagnetic prediction is done only once for each class of reference antenna geometry (e.g., Sierpinski, Hilbert, Koch fractal patches); (ii) the first attempt to exploit the OA features to build an optimal training set with reduced cardinality to enhance and speedup the SVR training phase; and (iii) the first customization of the SbD building blocks to efficiently and effectively deal with the synthesis of multiband microstrip antennas.
The paper is organized as follows. The multiband antenna synthesis problem at hand is mathematically stated in Section 2, while the SbD design approach is detailed in Section 3. A set of numerical and experimental assessments is then reported to give the interested reader some insights on the key features and the potentialities of the proposed design approach (Section 4). Some concluding remarks will follow (Section 5).
Synthesis problem formulation
Let us consider a generic structure composed of a perturbed fractal metallic layer printed on a dielectric substrate of thickness h characterized by a relative dielectric permittivity ε_{r} and a dielectric loss tangent tanδ, and backed by a metallic groundplane. By describing such a geometry by means of the feature vector g={g_{p}, p=1,...,P} containing all the P (realvalued) design parameters or problem DoFs (i.e., fractal dimensions, feed shape, substrate width and height), the antenna synthesis problem at hand can be stated as follows:
Multiband antennasynthesis problem—set the values of the unknown entries of g within the userdefined DoF boundaries such that \(s_{11}\left (f;\mathbf {g}\right)\leq s_{11}^{th}\) for all f∈{f_{n}, n=1,...,N}
s_{11}(f;g) being the antenna scattering parameter at the frequency f and \(s_{11}^{th}\) the corresponding userdefined constraint/requirement. Moreover, N is the number of bands of interest and f_{n} is the central frequency of the nth (n=1,...,N) band. Of course, such a statement as well as the corresponding problem can be easily extended to include additional constraints (e.g., mechanical, thermal, and chemical features) by adding suitable constraint terms. However, the generalization of the antenna synthesis to multiphysics formulations is beyond the scope of the current work and it will not be discussed in the following.
SbD synthesis procedure
Statement of SbD problem
Following the SbD approach [38–40] pictorially summarized in Fig. 1, the multiband antenna synthesis problem is reformulated as the following optimization one:
Systembydesign multiband antenna synthesis problem—find g^{opt}= ming[Φ(g)]such that \(\mathbf {g}\in \mathcal {G}\)
where \(\mathcal {G}=\left \{\left [g_{p}^{\min },g_{p}^{\max }\right ],\, p=1,...,P\right \} \) is the DoF feasibility region accounting for technological and/or physical constraints and Φ(g) is the cost function of the synthesis problem:
quantifying the mismatch with the project requirements, \(\widehat {f}_{n}\left (\mathbf {g}\right)\) being the nth resonant frequency [i.e., the central frequency of the nth band for which \(s_{11}\left (f_{n};\mathbf {g}\right)\leq s_{11}^{th}\)] of the multiband antenna with geometry descriptors g.
To solve such a problem, the iterative SbD procedure in Fig. 1 is exploited. It comprises the following elementary functional blocks:

A SSS block that generates, according to a suitable global optimization strategy, the guess solutions (i.e., perturbed fractal shapes coded into trial vectors g) convergent to the optimal antenna configuration g^{opt}

A PRE block aimed at estimating the resonant frequencies \(\widehat {f}_{n}\left (\mathbf {g}\right)\), n=1,...,N of the SSSblockgenerated guess geometry g by means of a computationally effective “forward” strategy

a POA block that evaluates the matching between the desired and the resonant frequencies of the trial multiband antenna, g, by computing the value of Φ(g), which takes into account also the physical/application constraints
It is worth pointing out that the modularity of the scheme in Fig. 1 represents one of the most important features of the SbD paradigm. As a matter of fact, it enables the choice of the most suitable tool to perform each subtask of the SbD loop by fully exploiting the interrelations among the blocks [38–40]. As a consequence, the different functionalities can act as mutual enablers for the selection of the most suitable techniques to be adopted, depending on the applicative constraints and objectives [38–40]. As concerns the multiband antenna synthesis, the POA block is straightforwardly implemented since its definition requires the evaluation of (1) and the assessment of the membership of the trial solution to the DoF feasibility region \(\mathcal {G}\). Conversely, the implementations of the PRE block (Section 3.2) and the SSS block (Section 3.3) need to be carefully addressed, and they will be detailed in the following subsections.
PRE block
Due to the computational constraints in effectively using an iterative global optimization, a LBE approach is adopted for the PRE block implementation [41]. More specifically, an algorithm derived from SVR [42] is considered. After an offline training phase performed to learn the inputoutput relations of the system to be emulated (Section 3.2.1) starting from a set of T training set couples \(\left [\mathbf {g}^{(t)},\,\widehat {f}_{n}\left (\mathbf {g}^{(t)}\right)\right ], t=1,...,T, n=1,...,N\), a fast online testing phase is then carried out to emulate the physical system itself (Section 3.2.2). Towards this end, the relation between the estimated resonant frequencies, \(\widehat {f}_{n}\left (\mathbf {g}\right)\), n=1,...,N, and the antenna parameters g is modelled as [46]:
where \(\mathcal {K}\left (\cdot,\cdot \right)\) is the socalled kernel function^{Footnote 1}, \(\boldsymbol {\alpha }_{n}\triangleq \left \{ \alpha _{n}^{(t)};\, t=1,...,T\right \} \) and \(\boldsymbol {\beta }_{n}\triangleq \left \{ \beta _{n}^{(t)};\, t=1,...,T\right \} \), n=1,...,N, are the SVR weights, and b_{n} is the offset for the nth frequency.
Offline SVR phase
In order to define the training set (i.e., \(\left \{\left [\mathbf {g}^{(t)},\,\widehat {f}_{n}\left (\mathbf {g}^{(t)}\right)\right ];\right. \left. t=1,...,T; n=1,...,N\text {l}{\vphantom {\left [\mathbf {g}^{(t)},\,\widehat {f}_{n}\left (\mathbf {g}^{(t)}\right)\right ]}}\right \}\)), an OA approach [43, 44, 47] is adopted to determine the representative sample points, {g^{(t)},t=1,...,T}, of the domain of the functional space, {\(\widehat {f}_{n}\left (\mathbf {g}\right)\); n=1,...,N} to be estimated through LBE. Indeed, it is well known that such a method allows one to statistically sample the space \(\mathcal {G}\) in a uniform way, while also minimizing the size T of the training set [43, 44, 47]. More in detail, the tth (t=1,...,T) input training sample is defined as:
where
and ω_{tp} is the t,pelement of a (L,T,P)OA
The entries of Ω are built so that they can assume an integer value within the range [0,L−1], L being the number of quantization steps (i.e., levels) chosen to discretize each parameter g_{p}, p=1,...,P, in the sampling space \(\mathcal {G}\). Moreover, the following properties hold true [47]:

1
Any (discrete) value of ω_{tp} appears \(\frac {T}{L}\) times in any column t of Ω (t=1,...,T).

2
Any couple of (discrete) values (ω_{tp},ω_{rp}) appears \(\frac {T}{L^{2}}\) times in any couple of columns (t,r) of Ω (t=1,...,T; r=1,...,T).

3
A matrix \(\widetilde {\Omega }\) obtained by swapping the columns of Ω is still an OA.

4
A matrix \(\widetilde {\Omega }\) obtained by taking away some of the columns of Ω is still an OA.
For illustrative purposes, let us consider the (L,T,P)OA Ω with L=3,P=4,T=9 reported in Table 1 [48]. As it can be noticed, the entries are integers belonging to the range [0,L−1]=[0,2]. Moreover, the matrix Ω satisfies the OA properties: (Property 1) the value “0” appears \(\frac {T}{L}=3\) times in any column of Ω and the same holds true for the values “1” and “2”; (Property 2) the couple “22” appear \(\frac {T}{L^{2}}=1\) times for each couple of columns of Ω and the same holds true for each couple of values; the same properties are satisfied also by swapping the columns of Ω (Property 3) and by removing some of them (Property 4).
Since the OA sampling is best rule for representing a L^{P}size feasibility space with T uniformly located points, the optimal SVR training samples are computed through (3) by referring to an (L,T,P)OA Ω matrix either available in online OA repositories [48] or built according to the procedures described in [47] (see the Appendix).
Once the set of input training samples, { g^{(t)},t=1,...,T}, is determined by means of (3), the computation of the associated resonance frequencies, {\(\widehat {f}_{n}\left (\mathbf {g}^{(t)}\right); t=1,...,T, n=1,...,N\)} (i.e., the output training samples), is at hand to complete the definition of the training set {\(\left [\mathbf {g}^{(t)},\,\widehat {f}_{n}\left (\mathbf {g}^{(t)}\right)\right ]\); t=1,...,T, n=1,...,N}. Towards this end, T fullwave electromagnetic simulations are carried out with a MethodofMoments (MoM) technique to determine the values of s_{11}(f;g^{(t)}), t=1,...,T.
The training phase is completed by determining the optimal values of the εSVR parameters (i.e., \(\boldsymbol {\widetilde {\alpha }}_{n}\triangleq \left \{\widetilde {\alpha }_{n}^{(t)};\, t=1,...,T\right \} \), \(\boldsymbol {\widetilde {\beta }}_{n}\triangleq \left \{\widetilde {\beta }_{n}^{(t)};\, t=1,...,T\right \} \), and \(\widetilde {b}_{n}, n=1,...,N\) [46]). Accordingly, the following minimization problem is solved:
by means of a local minimization strategy based on the Sequential Minimal Optimization approach proposed in [46]. Afterwards, the KarushKuhnTucker condition is exploited to estimate \(\widetilde {b}_{n}\) (n=1,...,N) (see [46], Section 4.1.5). In (6), Q= {\(q_{tr}=\mathcal {K}\left (\mathbf {g}^{(t)},\mathbf {g}^{(r)}\right)\); t=1,...,T; r=1,..,T} is the kernel matrix, ε and C are userdefined control parameters, and ·^{′} stands for the transpose operator.
Online SVR phase
Once \(\left (\boldsymbol {\widetilde {\alpha }}_{n},\,\boldsymbol {\widetilde {\beta }}_{n},\,\widetilde {b}_{n}\right)\), n=1,...,N, have been computed by means of (6), the online SVR phase is carried out by simply substituting the desired g in (2) and enforcing \(\boldsymbol {\alpha }_{n}\leftarrow \boldsymbol {\widetilde {\alpha }}_{n}\), \(\boldsymbol {\beta }_{n}\leftarrow \boldsymbol {\widetilde {\beta }}_{n}\), and \(b_{n}\leftarrow \widetilde {b}_{n}\) [46]. Thus, a simple and very fast function evaluation is needed to emulate the physical response (resonance frequencies) of each guess antenna configuration instead of recurring to a computationally heavy fullwave MoM simulation. On the other hand, it is worth pointing out that such a feature is a key asset for the integration of the PRE block within the SbD design loop (Fig. 1). Indeed, since the physical response of an antenna can be evaluated in a very efficient manner, the PRE block acts as an enabler for SSS blocks based on iterative search strategies [e.g., multiagents global optimizers based on EAs (Section 3.3)] requiring the evaluation of a huge number of guess solutions. Moreover, the offline training phase (6) needs to be performed only once for each antenna typology (i.e., fractal shape); thus, such a PRE block can be reused in various SbD procedures concerned with different applicative scenarios, even though still dealing with fractal shapes, as proved in Section 4 through some representative experiments.
SSS block
As for the SSS block, it is required that the underlying methodology complies with the applicative constraints and the features of the other functional tools involved in the SbD loop (Fig. 1) [38–40]. More specifically, it must be able to (i) properly handle the type of design variables of interest (i.e., realvalued g), (ii) effectively sample the solution space by avoiding to be trapped in local minima during the iterative searching procedure, and (iii) enable the definition of arbitrary cost functions, Φ(g), and physical constraints, \(\mathcal {G}\), for the design of interest. Because of the features of the multiband antenna problem at hand, a global optimization based on EAs [35] is adopted. More specifically, the realvalued nature of g and the (potentially) large number of parameters P describing the antenna geometry suggest the use of a particle swarm (PS) optimizer [35]. Accordingly, the inertialweight version of the PS with reflecting boundary condition is considered [35]. The iterative procedure, which generates an optimal guess solution \(\mathbf {g}_{i}^{opt}\) at each iteration i, is stopped when one of the following conditions is met:

A maximum number of iterations I is reached (i.e., i=I,I being an userdefined iteration number)

The fitness reaches stationarity
$$ I_{win}\Phi\left(\mathbf{g}_{i}^{opt}\right)\sum_{v=1}^{I_{win}}\Phi\left(\mathbf{g}_{iv}^{opt}\right) \leq\eta\Phi\left(\mathbf{g}_{i}^{opt}\right) $$(7)η and I_{win} being a userdefined stagnation threshold and an iteration window, respectively

The fitness value is smaller than a userdefined threshold value, \(\zeta \left [\text {i.e.}, \Phi \left (\mathbf {g}_{i}^{opt}\right)\leq \zeta \right ]\)
It is worth remarking that the proposed SbD approach can be applied without restrictions on the multifrequency antenna geometry, constraints, and objectives, provided that \(\mathbf {g}, \mathcal {G}\), and Φ(·) are properly defined and the T training samples \(\left [\mathbf {g}^{(t)},\,\widehat {f}_{n}\left (\mathbf {g}^{(t)}\right)\right ]\), t=1,...,T are computed. This is a key advantage of the SbD paradigm over standard design methodologies since it also guarantees the reusability of the majority of the SbD modules across different synthesis problems.
Numerical and experimental results
This section is aimed at illustrating both the offline (Section 4.1) and the online (Section 4.2) phases of the proposed SbD approach, as well as assessing its effectiveness through an experimental validation (Section 4.3). Towards this end, the design of a dualband (N=2) Sierpinski Gasket fractal shape [27] shown in Fig. 2 is assumed as a benchmark example. In such a case, the feature vector g is characterized by P=9 entries encoding the geometric descriptors in Fig. 2: \(g_{1}=w_{a}, g_{2}=h_{a}, g_{3}=w_{1}, g_{4}=w_{2}, g_{5}=h_{f}, g_{6}=h_{g}, g_{7}=\frac {w_{3}}{W_{3}}, g_{8}=\frac {w_{4}}{W_{4}}\), and \(g_{9}=\frac {w_{5}}{W_{5}}\). Concerning the substrate properties, an Arlon layer with ε_{r}=3.38, tanδ=0.0025, and h=7.6×10^{−4} [m] has been considered. For comparison purposes, all the computational costs refer to a nonoptimized Matlab implementation of the methodology running on a desktop PC with a 2.6 GHz singlecore processor.
Offline training of the PRE block
As for the offline phase of the PRE block, the definition of the training set \(\left [\mathbf {g}^{(t)},\,\widehat {f}_{n}\left (\mathbf {g}^{(t)}\right)\right ], t=1,...,T, n=1,...,N\) requires the choice of the suitable OA Ω for the problem at hand. Towards this end, L=61 quantization levels for the antenna descriptors have been considered to carefully discretize the entire search space \(\mathcal {G}\). The auxiliary OA \(\Omega ' \triangleq \) {\(\omega _{tp}'\in \mathbb {N}\cap \left [0,L1\right ]\); t=1,...,T,p=1,...,P^{′}} has been determined by following the procedure in [47] yielding J=2, T=L^{J}=3721, \(P'=\frac {L^{J}1}{L1}=62\) (see the Appendix). The training OA Ω has been successively identified by deleting the last P^{′}−P=53 columns of Ω^{′} by enforcing:
Once Ω has been defined, Eq. (3) has been used to compute the T input training samples { g^{(t)}; t=1,...,T} by setting the range of the antenna descriptors (i.e., \(\mathcal {G}\); Table 2) according to the guidelines discussed in [21] so that the resonating frequencies fall within the band of interest for IoT/IoE communications. The resonant frequencies, {\(\widehat {f}_{n}\left (\mathbf {g}^{(t)}\right)\); t=1,...,T; n=1,...,N} in correspondence with the input set { g^{(t)}; t=1,...,T}, have been computed through a fullwave solver ^{Footnote 2} by setting \(s_{11}^{th}=10\) [dB]. As it can be observed (Fig. 3), there are several training samples characterized by a dualband behavior (i.e., two resonances) in the [1.0−6.0] GHz range [i.e., \(s_{11}\left (f_{n}\left (\mathbf {g}^{(t)}\right);\,\mathbf {g}^{(t)}\right)\leq s_{11}^{th}\), n=1,2]. Of course, different bandwidths could be easily covered by suitably scaling \(\mathcal {G} \left [\text {i.e.}, f_{n}\left (\mathbf {g}\right)\to \frac {f_{n}\left (\mathbf {g}\right)}{10} \text {then}\ \mathcal {G}\to \frac {\mathcal {G}}{10}\right ]\).
Concerning the SVR implementation, a radial basis function (RBF) kernel has been adopted [46]:
fixing to γ=0.1 [46] the value of the RBF control parameter γ. To complete the offline phase of the PRE block, Eq. (6) (i.e., the SVR training) has been solved by using the procedure in Section 3.2.1 and enforcing the control parameters to ε=0.01 and C=1 as suggested in [46]. It is worth pointing out that such a PRE block allows one to estimate the output values {\(\widehat {f}_{n}\left (\mathbf {g}\right)\); n=1,..,N} through (2) in about:
while the same result would require about Δt_{MoM}≈2.7×10^{2}[s] using a standard MoM code for simulating the same antenna configuration, g.
Numerical assessment
The first numerical experiment is aimed at illustrating, on a stepbystep basis, the exploitation of the proposed integrated SbD procedure (Fig. 1) for designing a dualband antenna working at the LTE2100 (f_{1}=2.045 GHz) and the LTE3500 (f_{2}=3.5 GHz) channels. Towards this end, the PRE block (Section 4.1) has been combined with the POA block (1) and the PSbased SSS block configured by following the guidelines in [21, 35]: S=8, ζ=10^{−10}, I=200, η=10^{−4}, and I_{win}=30. The plot of the “global best” value of the cost function, \(\Phi _{i}\triangleq \Phi \left (\mathbf {g}_{i}^{opt}\right)\), versus the iteration index, i, in Fig. 4(a) shows that the method converges in about I_{conv}≈100 iterations ^{Footnote 3}, which correspond to I_{conv}×S=800 cost function evaluations requiring Δt_{SbD}≈800×Δt_{PRE}=0.8 [s]. The comparison with the computational cost of a standard optimization loop remarks the efficiency of the PRE block implementation and the whole SbD design procedure. As a matter of fact, the same process adopting a simulation block based on a MoM solver would have taken about Δt_{standard}≈800×Δt_{MoM}≈2.16×10^{5} [s]. On the other hand, the accuracy of the PRE emulation is assessed by the plots of the simulated scattering parameter ^{Footnote 4} of the optimized layout (Fig. 4b). As a matter of fact, despite the emulation of the physical response of the antenna, the synthesized antenna profile (Fig. 5a–b) perfectly matches the design objectives expressed in terms of impedance matching in the operating bands (i.e., \(s_{11}\left (f;\mathbf {g}\right)\leq s_{11}^{th}=10\) dB, f∈{f_{n}, n=1,...,N}={2.045, 3.5} GHz; Fig. 4b). As for the radiation properties, the simulated 3D gain patterns computed at f_{n}, n=1,...,N indicate that the antenna exhibits a dipolelike behavior (e.g., f_{n}=2.045 GHz; Fig. 6a), which is only slightly perturbed at the higher frequency (e.g., f_{n}=3.5 GHz; Fig. 6b). Such a behavior is actually not surprising since it has been typically observed in antennas based on such a fractal shape [21].
Experimental validation
In order to assess the effectiveness and the reliability of the SbD synthesis technique, a set of design examples concerned with different objectives/constraints has been carried out next. The optimized layouts have been fabricated with a photolithographic printing circuit technology using an Arlon substrate with ε_{r}=3.38, tanδ=0.0025, and h=7.6×10^{−4} [m], and successively measured in an anechoic chamber at the ELEDIA Research Center. More specifically, each prototype has been designed to be fed by a single 50 Ohm RF port connected to the bottom of the Sierpinski antenna for measuring its impedance matching properties and the corresponding gain pattern.
The case of a dualband antenna resonating in the WCDMA1500 (f_{1}=1.470 GHz) and the LTE2600 (f_{2}=2.595 GHz) bands has been considered first (Table 3). Thanks to the reusability property of the SbD blocks, the design problem at hand does not require a new SVR training and the same PRE block deduced in Section 4.1 has been directly exploited without any modification. Because of the same type of unknowns and cost function, also the SSS block has been kept equal to the previous one with its parameter setup, as well. Thus, only the list of the resonance frequencies { f_{n}, n=1,...,N} in (1) of the POA block has been changed to derive the antenna profile in Fig. 7a (for completeness, the complete list of optimized geometrical descriptors has been reported in Table 4). As expected, both measured and simulated s_{11} frequency behaviors (Fig. 7b) indicate that the antenna resonances are located at the center of the WCDMA1500 band (≈1.470 GHz) and in correspondence with the LTE2600 (≈2.595 GHz) channel. Moreover, a good (measured) impedance matching holds true in the whole WCDMA1500 [i.e., \(\left s_{11}^{meas}(f)\right \leq 13.2\) dB for f∈[1.430, 1.508] GHz; Fig. 7b] and LTE2600 uplink/downlink bands [i.e., \(\left s_{11}^{meas}(f)\right \leq 11.1\) dB for f∈[2.500, 2.690] GHz; Fig. 7b]. With reference to the radiation properties, Fig. 7c–d show the plots of the simulated and measured gain patterns along the vertical [ φ=90 [deg]; Fig. 7c] and the horizontal [ θ=90 [deg]; Fig. 7d] planes at f=f_{1} and f=f_{2}. The antenna acts as a dipole in both frequency bands, with only a nonperfect omnidirectionality along the horizontal plan when f=3.5 GHz (Fig. 7d). However, it is worth pointing out that the radiation features were not optimized in the synthesis process and that a straightforward method extension is possible by simply adding to (1) another constraint on the userdesired radiation performances (e.g., a gain pattern mask) and performing the PRE offline/online procedure accordingly.
The last example deals with the design of a dualband antenna able to support the LTE1800 (f_{1}=1.795 GHz) and the LTE3500 (f_{2}=3.500 GHz) standards (Table 3). Just adapting the POA block to the design objectives at hand, the SbDoptimized layout in Fig. 8a has been obtained (Table 4), whose corresponding impedance matching properties are reported in Fig. 8b. Once again measured and simulated quantities turn out to be in quite close agreement.
They assess that the optimized scattering parameters comply with the design requirements (i.e., LTE1800: \(\left s_{11}^{meas}(f)\right \leq 10.3\) dB for f∈[1.710, 1.879] GHz; LTE3500:\(\left s_{11}^{meas}(f)\right \leq 10.1\) dB for f∈[3.410, 3.590] GHz; Fig. 8b). Moreover, the gain plots (φ=90[deg], Fig. 8c; θ=90[deg], Fig. 8d) confirm the reliability of the synthesized radiation device for IoT applications because of its almost perfect omnidirectional pattern along the horizontal plane θ=90[deg] in both frequency bands (Fig. 8c–d).
For completeness, Table 3 gives the CPUtime of the SbD process for synthesizing the antennas in Figs. 7 and 8 (Δt_{SbD}) in comparison with the time required when the PRE block is substituted by a standard fullwave MoM solver (Δt_{standard}). As it can be noticed, there is a time saving of about 5 orders of magnitude (Table 3).
Conclusions
The problem of efficiently designing multiband antennas has been addressed by means of an instance of the SbD paradigm. The synthesis problem has been firstly recast within the SbD framework, and each SbD building block has been then implemented. A suitable combination of an SSS block (aimed at the exploration of the search space), a PRE block (concerned with the fast prediction of the antenna electromagnetic response), and a POA block (to evaluate the guess antenna quality) has been integrated by exploiting an innovative LBE strategy based on SVR and trained through effective OAs.
To the best of the authors’ knowledge, the novelty of this work consists in (i) the derivation of a scalable, flexible, and reusable synthesis tool; (ii) the innovative exploitation of the OA features to enhance and speedup the SVR training phase; and (iii) the first instance of the SbD aimed at enabling an effective and computationally efficient synthesis of multiband microstrip antennas.
A set of numerical and experimental results has been reported to illustrate the features and the potentialities of the proposed synthesis approach in different applicative scenarios, as well. More precisely, it has been shown that both numerical (Section 4.2) and experimental (Section 4.3) designs exhibit a very good matching of the design objectives expressed in terms of impedance matching in the desired operative bands, as well as a dipolelike behavior in terms of radiation features with only minor deformations occurring at the highest resonant frequency.
Furthermore, the numerical and experimental results have pointed out that the proposed SbD loop (i) provides extremely efficient and effective performances with a time saving of about 5 orders of magnitude with respect to a standard antenna optimization exploiting fullwave solvers (Table 3); (ii) can be straightforwardly applied/extended to various applicative scenarios differing for the operative bands and the userdefined requirements (Section 4.3) thanks to its reusability and modularity; and (iii) synthesizes antenna layouts useful for IoT applications thanks to their effective impedance and radiation features.
Future works beyond the scope of this paper will be devoted to analyze the features of the SbD paradigm with different antenna geometries (e.g., different fractal profiles) and frequency bands (e.g., falling within the 6–300 [GHz] spectrum announced by the Federal Communications Commission (FCC) for nextgeneration 5G services [13]). As for the methodological point of view, different architectural tradeoffs in selecting the theoretical implementation of the SSS, the POA, and the PRE blocks will be investigated.
Appendix
(L,T,P)OA construction procedure [47]
The construction of an (L,T,P)OA complying with the properties discussed in Section 3.2 can be carried out according to the following procedure [47]:

Step 1. Select the smallest integer J such that \(\frac {L^{J}1}{L1}\geq P\);

Step 2. Set \(P'=\frac {L^{J}1}{L1}\) and T=L^{J};

Step 3. Construction of auxiliary OA Ω^{′}:

Step 3.1. Set \(\psi _{j}=\frac {L^{j1}1}{L1}+1, j=1,...,J\);

Step 3.2. Set \(\omega _{t\psi _{j}}'=\left.\frac {t1}{L^{Jj}}\right \rfloor _{mod\, L}, j=1,...,J, t=1,...,T\);

Step 3.3. Set \(\omega _{t\left (\psi _{j}+\left (v1\right)\left (L1\right)+l\right)}'=\left.l\times \omega _{tv}'+\omega _{t\psi _{j}}'\right \rfloor _{mod\, L}, t=1,...,T, j=2,...,J, v=1,...,\psi _{j}1, l=1,...,L1\);

Step 3.4. Set ωtp′←ωtp′+1, t=1,...,T, p=1,...,P^{′};


Step 4. Delete the last P^{′}−P columns of Ω^{′} and obtain Ω.
Accordingly, such a construction procedure (i) receives as inputs P and L and (ii) provides in output the computation of T and Ω.
Notes
 1.
Basic functions that can be adopted in the SVR model include linear, polynomial, Gaussian, and sigmoid kernels [49].
 2.
It must be remarked that, according to the OA construction, the training phase required T=3721 antenna simulations through a MoM tool. However, this procedure has been done offline and it has been carried out only once for each problem class (i.e., fractal geometry of interest).
 3.
In the following, I_{conv} will denote the SbD convergence iteration [i.e., at least one of the convergence criteria in Section 3.3 is met] when the SSS block stops and it gives the output solution.
 4.
For validation purposes, all the s_{11} plots have been computed by a fullwave MoMbased solver.
Abbreviations
 3GPP:

3rd Generation Partnership Project
 5G:

5th (Fifth) generation
 CPU:

Central processing unit
 DoFs:

Degreesoffreedom
 EAs:

Evolutionary algorithms
 EM:

Electromagnetic
 FCC:

Federal Communications Commission
 GPS:

Global Positioning System
 GSM:

Global System for Mobile Communications
 IoE:

Internet of Everything
 IoT:

Internet of Things
 LBE:

Learningbyexample
 LTE:

LongTerm Evolution
 M2M:

Machinetomachine
 MoM:

MethodofMoments
 NFC:

Nearfield communication
 NBIoT:

NarrowbandIoT
 OA:

Orthogonal array
 PC:

Personal computer
 POA:

Physical objective assessment
 PRE:

Physical response emulator
 PS:

Particle swarm
 RBF:

Radial basis function
 RF:

Radio frequency
 RFID:

Radiofrequency identification
 SbD:

Systembydesign
 SSS:

Solutionspace search
 SVR:

Support vector regressor
 UMTS:

Universal Mobile Telecommunications System
 WCDMA:

Wideband Code Division Multiple Access
 WIMAX:

Worldwide Interoperability for Microwave Access
 WLAN:

Wireless local area network
References
 1
M. R. Ebling, Pervasive computing and the Internet of Things. IEEE Pervasive Comput. 15(1), 2–4 (2016).
 2
A. Paraskevopoulos, P. Dallas, K. Siakavara, S. Goudos, Cognitive radio engine design for IoT using realcoded biogeographybased optimization and fuzzy decision making. Wirel. Pers. Commun.97(2), 1813–1833 (2017).
 3
S. Goudos, P. Dallas, S. Chatziefthymiou, S. Kyriazakos, A survey of IoT key enabling and future technologies: 5G, mobile IoT, sematic web and applications. Wirel. Pers. Commun.97(2), 1645–1675 (2017).
 4
Y. Mehmood, C. Gorg, M. Muehleisen, A. TimmGiel, Mobile M2M communication architectures, upcoming challenges, applications, and future directions. EURASIP J. Wirel. Commun. Netw.2015:, 1–37 (2015).
 5
3GPP, Tr 45.820 Cellular system support for ultralow complexity and low throughput internet of things (CIoT) (relase 13). Tech. Rep., 1–495 (2015).
 6
3GPP, Ts 36.802, Evolved universal terrestrial radio access (EUTRA); NBIoT; technical report for BS and UE radio transmission and reception (release 13). Tech. Rep., 1–59 (2016).
 7
U. Chakraborty, A. Kundu, S. K. Chowdhury, A. K. Bhattacharjee, Compact dualband microstrip antenna for IEEE 802.11a WLAN application. IEEE Antennas Wirel. Propag. Lett. 13:, 407–410 (2014).
 8
W. Ting, S. XiaoWei, L. Ping, B. Hao, Triband microstripfed monopole antenna with dualpolarisation characteristics for WLAN and WiMAX applications. Electron. Lett.49:, 1597–1598 (2013).
 9
S. K. Mishra, R. K. Gupta, A. Vaidya, J. Mukherjee, A compact dualband forkshaped monopole antenna for Bluetooth and UWB applications. IEEE Antennas Wirel. Propag. Lett.10:, 627–630 (2011).
 10
A. Diallo, C. Luxey, P. Le Thuc, R. Staraj, G. Kossiavas, Efficient twoport antenna system for GSMDCSUMTS multimode mobile phones. Electron. Lett.43:, 369–370 (2007).
 11
Y. J. Ren, Ceramic based small LTE MIMO handset antenna. IEEE Trans. Antennas Propag.61(2), 934–938 (2013).
 12
L. Le, V. Lau, E. Jorswieck, N. Dao, A. Haghighat, Kim D., T. LeNgoc, Enabling 5G mobile wireless technologies. EURASIP J Wirel. Commun. Netw.2015:, 1–14 (2015).
 13
G. Oliveri, G. Gottardi, F. Robol, A. Polo, L. Poli, M. Salucci, M. Chuan, C. Massagrande, P. Vinetti, M. Mattivi, R. Lombardi, A. Massa, Codesign of unconventional array architectures and antenna elements for 5G base station. IEEE Trans. Antennas Propag.65(12), 6752–6767 (2017).
 14
C. M. Kruesi, R. J. Vyas, M. M. Tentzeris, Design and development of a novel 3D cubic antenna for wireless sensor networks (WSNs) and RFID applications. IEEE Trans. Antennas Propag. 57(10), 3293–3299 (2009).
 15
Y. Zhou, C. C. Chen, J. L. Volakis, Singlefed circularly polarized antenna element with reduced coupling for GPS arrays. IEEE Trans. Antennas Propag.56(5), 1469–1472 (2008).
 16
C. PuenteBaliarda, J. Romeu, R. Pous, A. Cardama, On the behavior of the Sierpinski multiband fractal antenna. IEEE Trans. Antennas Propag.46(4), 517–524 (1998).
 17
M. MartinezVazquez, O. Litschke, M. Geissler, D. Heberling, A. M. MartinezGonzalez, D. SanchezHernandez, Integrated planar multiband antennas for personal communication handsets. IEEE Trans. Antennas Propag.54(2), 384–391 (2006).
 18
E. Zeni, R. Azaro, P. Rocca, A. Massa, Quadband patch antenna for Galileo and WiMax services. Electron. Lett.43:, 960–962 (2007).
 19
R. Azaro, E. Zeni, P. Rocca, A. Massa, Synthesis of a Galileo and WiMax threeband fractaleroded patch antenna. IEEE Antennas Wirel. Propag. Lett.6:, 510–514 (2007).
 20
R. Azaro, F. Viani, L. Lizzi, E. Zeni, A. Massa, A monopolar quadband antenna based on a Hilbert selfaffine prefractal geometry. IEEE Antennas Wirel. Propag. Lett.8:, 177–180 (2009).
 21
L. Lizzi, R. Azaro, G. Oliveri, A. Massa, Multiband fractal antenna for wireless communication systems for emergency management. J Electromagn. Waves Applicat.26(1), 1–11 (2012).
 22
F. Viani, M. Salucci, F. Robol, A. Massa, Multiband fractal Zigbee/WLAN antenna for ubiquitous wireless environments. J Electromagn. Waves Appl.26:, 1554–1562 (2012).
 23
F. Viani, M. Salucci, F. Robol, G. Oliveri, A. Massa, Design of a UHF RFID/GPS fractal antenna for logistics management. J Electromagn. Waves Appl.26(4), 480–492 (2012).
 24
E. L. Chuma, Y. Iano, Roger L.L.B., in 2017 SBMO/IEEE MTTS International Microwave and Optoelectronics Conference (IMOC), Aguas de Lindoia. Compact antenna based on fractal for IoT subGHz wireless communications, (2017), pp. 1–5. https://doi.org/10.1109/IMOC.2017.8121016.
 25
D. H. Werner, S. Ganguly, An overview of fractal antenna engineering research. IEEE Antennas Propag. Mag.45(1), 38–57 (2003).
 26
S. R. Best, A discussion on the significance of geometry in determining the resonant behavior of fractal and other nonEuclidean wire antennas. IEEE Antennas Propag. Mag.45(3), 9–28 (2003).
 27
S. R. Best, Operating band comparison of the Perturbated Sierpinski and modified Parany gasket antennas. IEEE Antennas Wirel. Propag. Lett.1:, 35–38 (2002).
 28
R. C. Guido, A note on the practical relationship between filters coefficients and the scaling and wavelets functions of the discrete wavelet transform. Appl. Math. Lett.24(7), 1257–1259 (2011).
 29
R. C. Guido, S. Barbon, L. S. Vieira, F. L Sanchez, C. D Maciel, J. C Pereira, P. R Scalassara, E. S Fonseca, in 2008 IEEE International Symposium on Circuits and Systems, Seattle, WA. Introduction to the Discrete Shapelet Transform and a new paradigm: Joint timefrequencyshape analysis, (2008), pp. 2893–2896. https://doi.org/10.1109/ISCAS.2008.4542062.
 30
E. Guaraglia, Entropy and fractal antennas. Entropy. 18(3), 84 (2016).
 31
E. Guariglia, Harmonic Sierpinski gasket and applications. Entropy. 20(9), 714 (2018).
 32
W. J. Krzysztofik, Fractal geometry in electromagnetics applications  From antenna to metamaterials. Microw. Rev.19(2), 3–14 (2013).
 33
R. G. Hohlfeld, N. L. Cohen, Selfsimilarity and the geometric requirements for frequency independence in antennae. Fractals. 7(1), 79–84 (1999).
 34
R. Haupt, D. Werner, Genetic algorithms in electromagnetics (Wiley, New Jersey, 2007).
 35
P. Rocca, M. Benedetti, M. Donelli, D. Franceschini, A. Massa, Evolutionary optimization as applied to inverse scattering problems. Inverse Probl.25(12), 1–41 (2009).
 36
P. Rocca, G. Oliveri, A. Massa, Differential evolution as applied to electromagnetics. IEEE Antennas Propag. Mag.53(1), 38–49 (2011).
 37
S. Goudos, C. Kalialakis, R. Mittra, Evolutionary algorithms applied to antennas and propagation: a review of state of the art. Int. J. Antennas Propag.2016:, 1–12 (2016).
 38
G. Oliveri, F. Viani, N. Anselmi, A. Massa, Synthesis of multilayer WAIM coatings for planar phased arrays within the systembydesign framework. IEEE Trans. Antennas Propag.63(6), 2482–2496 (2015).
 39
G. Oliveri, M. Salucci, N. Anselmi, A. Massa, Multiscale SystembyDesign synthesis of printed WAIMs for waveguide array enhancement. IEEE J. Multiscale. Multiphysics Computat. Techn.2:, 84–96 (2017).
 40
G. Oliveri, L. Tenuti, E. Bekele, M. Carlin, A. Massa, An SbDQCTO approach to the synthesis of isotropic metamaterial lenses. IEEE Antennas Wirel. Propag. Lett.13:, 1783–1786 (2014).
 41
A. Massa, G. Oliveri, M. Salucci, N. Anselmi, P. Rocca, Learningbyexamples techniques as applied to electromagnetics, (2017). https://doi.org/10.1080/09205071.2017.1402713.
 42
A. J. Smola, B. Scholkopf, A tutorial on support vector regression. Stat. Comput.14:, 199–222 (2004).
 43
S. H. Park, Robust design and analysis for quality engineering (Chapman and Hall, London, 1996).
 44
A. S. Hedayat, N. J. A. Sloane, J. Stufken, Orthogonal arrays (Springer, New York, 2013).
 45
Z. Bayraktar, D. H. Werner, P. L. Werner, Miniature meanderline dipole antenna arrays, designed via an orthogonalarrayinitialized hybrid particleswarm optimizer. IEEE Antennas Propagat. Mag.53(3), 42–59 (2011).
 46
C. C. Chang, C. J. Lin, LIBSVM: a library for support vector machines. ACM Trans. Intell. Systems and Tech. 2:, 1–39 (2011).
 47
Y. W. Leung, Y. Wang, An orthogonal genetic algorithm with quantization for global numerical optimization. IEEE Trans. Evolution. Comput.5(1), 41–53 (2001).
 48
N. J. A Sloane, A library of Orthogonal Arrays (2007). http://neilsloane.com/oadir. Accessed 09 Jan 2018.
 49
C. W. Hsu, C. C. Chang, Lin C.J., A practical guide to support vector classification (National Taiwan University, 2003).
Acknowledgments
Not applicable.
Funding
This work benefited from the networking activities carried out within the SNATCH Project (2017–2019) funded by the Italian Ministry of Foreign Affairs and International Cooperation, Directorate General for Cultural and Economic Promotion and Innovation, the Project “WATERTECH  Smart Community per lo Sviluppo e l’Applicazione di Tecnologie di Monitoraggio Innovative per le Reti di Distribuzione Idrica negli usi idropotabili ed agricoli” (Grant no. SCN_00489) funded by the Italian Ministry of Education, University, and Research within the Program “Smart cities and communities and Social Innovation” (CUP: E44G14000060008), and the Project “SMARTOUR  Piattaforma Intelligente per il Turismo” (Grant no. SCN_00166) funded by the Italian Ministry of Education, University, and Research within the Program “Smart cities and communities and Social Innovation”.
Availability of data and materials
Not applicable.
Author information
Affiliations
Contributions
All authors read and approved the final manuscript.
Corresponding author
Correspondence to Andrea Massa.
Ethics declarations
Competing interests
The authors declare that they have no competing interests.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Additional information
Authors’ information
Not applicable.
Rights and permissions
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.
About this article
Received
Accepted
Published
DOI
Keywords
 Multiband antennas
 Systembydesign
 Antenna synthesis
 Fractals
 Optimization
 Internet of Things (IoT)