Broadband 400 GHz On-Chip Antenna with a Metastructured Ground Plane and Dielectric Resonator

The analysis, modeling, design, simulation, and experimental evaluation of a 400GHz on-chip antenna is presented, with a novel combination of metastructures, a microstrip patch, a quartz-based dielectric resonator, and a diamond-based anti-reflex layer—all integrated on a 35nm InGaAs metamorphic high-electron-mobility transistor (mHEMT) technology. Said combination represents a first-time implementation for all submillimeter-wave-capable semiconductor technologies. Circumventing a substrate-thickness limitation of 4.98 μm, a state-of-the-art broadband, efficient, and to-the-broadside radiating on-chip antenna solution is realized. It achieves a measured impedance bandwidth of 100 GHz, 25.6 %, spanning from 340 GHz to 440GHz. A consistent pattern bandwidth of 75GHz is recorded, with an efficiency of 50% to 66 %, and a directivity of up to 10.4 dBi—or 27 dBi, with the utilization of a polypropylene-based dielectric lens. The theoretical analysis of the proposed on-chip antenna is presented, as well as two modeling approaches are shown and compared. Between the analytical and the 3D electromagnetic model, the latter is chosen as it offers a greater precision at defining the metastructure unit cell and enables the inclusion of the remaining components of the proposed antenna setup. The measured reflection coefficient and far-field patterns are compared to simulations via the utilized model, and a strong agreement is observed. These far-field patterns are acquired with the on-chip antenna inserted within a broadband 400GHz transmitter submillimeter-wave monolithic integrated circuit, processed on the 35nm mHEMT technology.


I. INTRODUCTION
S UBMILLIMETER-WAVE frequency bands allow for broadband transmit and receive windows, serviceable to both communications- [1]- [5] and radar-based [6]- [15] applications-increasing data-rates and imaging resolutions, respectively. The submillimeter-wave operational frequency allows for miniaturized integrated-circuit (IC) elements. This promotes singular-chip solutions which integrate complete active-chains with corresponding on-chip antennas. A majority of such fully-integrated solutions are based on state-of-the-art silicon-on-insulator (SOI) complementary metal-oxide semiconductor (CMOS) and SiGe bipolar CMOS (BiCMOS) technology nodes [16]- [26]. While there is a benefit in the available dielectric thickness in the back-end-of-line (BEOL) of SOI CMOS and SiGe BiCMOS, the corresponding maximum oscillation frequency restricts applications in terms of the realizable bandwidth [27]. Fraunhofer IAF's 35 nm In 0.52 Al 0. 48 As/In 0.80 Ga 0.20 As metamorphic high-electron-mobility transistor (mHEMT) IC technology boasts a transition and maximum oscillation frequency of f T > 500 GHz and f max > 1000 GHz [28], [29]. Processed on said technology, complete active-chains are capable of supporting a relative bandwidth of up to 30 % [30]- [33]. Yet, a major drawback for realizing broadband and to-the-broadside radiating on-chip antennas is the limiting thickness of the combined benzocyclobutene (BCB) dielectric layers, as seen in Fig. 1 it is 4.98 µm. The latter is defined between the ground plane, metal layer MET1, and the top metal layer METG. Henceforth, the term "substrate" is used to define the combination of BCB3, SiN and BCB2. Regarding a center frequency (f c ) of 400 GHz, the ratio between the substrate-thickness and the nominal-wavelength h/λ 0 is below 0.006. This restricts a microstrip-patch antenna approach to a relative bandwidth of below 4 %. In this regard, the on-chip microstrip antenna can be viewed as a lossy capacitor [34].
The intent of this paper is to present an efficient, broadband, and to-the-broadside radiating on-chip antenna for substrate-limited submillimeter-wave-capable semiconductor technologies. In order to achieve said goal, the first-time combination of metastructures, a dielectric resonator, an anti-reflex layer, and a microstrip patch antenna is presented. These are all unified in a broadband transmitter submillimeter-wave monolithic integrated circuit (S-MMIC), processed on the aforementioned 35 nm mHEMT technology. No alteration in the BEOL is required. The restricted substrate thickness is electrically elongated via the utilization of a metastructured ground plane. The reflected substrate-incident waves are utilized to excite standing waves in the quartz-based dielectric resonator positioned on top. The efficient decoupling of the resulting standing waves is achieved via the addition of a thin anti-reflex diamond layer. The S-MMIC is assembled on a submount carrier board and a laser-sintered polypropylene-based dielectric lens is included. The mentioned components are designed for operation in the WR-2.2 waveguide frequency band (325 GHz to 500 GHz). Submillimeter-wave approaches that share similarities with the presented on-chip antenna solution can be found in: [35], where dielectric resonators are paired with a substrate-integrated waveguide cavity-backed on-chip antenna processed in SiGe BiCMOS. [36], where a metastructure-based antenna array is realized on a GaAs substrate. The array is fed by through-the-substrate vias connecting to slot lines processed on the backside. [37], where a dielectric resonator is combined with a microstrip patch processed in SOI CMOS. [38], where a metastructure-based antenna array is realized in SOI CMOS. The radiating elements are fed from the backside of the wafer via a combination of slot lines and open-circuited stubs. [16], where a thin superstrate is combined with an one-dimensional microstrip-antenna-based phased array processed in SOI CMOS. The unique difference between the aforementioned and this work lies in the utilization of a metastructured ground plane as an artificial magnetic conductor that reflects the substrate-incident waves and excites multiple modes within a quartz-based dielectric resonator. Not only is the bandwidth addressed, rather the coapplication contributes towards consistent radiation patterns.
This article is divided into four further sections, excluding the introduction. Section II contains the theoretical analysis, modeling and simulation of the proposed antenna solution. Sections III and IV focus on the characterization and interpretation of the reflection coefficient (S 11 ) and far-field patterns, respectively. Measurements and simulations are included and scrutinized. Lastly, a brief conclusion is provided in Section V.

II. ON-CHIP ANTENNA
The three-dimensional (3D) electromagnetic (EM) model of the proposed on-chip antenna solution can be seen in Fig. 2a. It consists of the metastructured ground plane and microstrip patch, both processed in the mHEMT stack. The quartz dielectric resonator and diamond anti-reflex layer are included as well. The design environment is that of CST Studio Suite (CST). A closer look of the metastructure and patch is provided in Fig. 2b, where the BEOL of the utilized process is reflected. The input-matching network, composed of an inductor-capacitor-inductor arrangement, is highlighted as well. It makes use of the SiN layer to implement the desired parallel capacitance. This "lumped" approach is more compact than a microstrip-line-based radial stub. It induces less losses and allows for a direct integration with the active chain.
The implemented antenna concept includes a polypropylene-based dielectric elliptical lens, which is depicted in Fig. 2c. The parametric values utilized in the design process in terms of the dielectric permittivity (ϵ r ) and loss tangent (tan δ) are 2.2 and 0.00055, respectively [39]. The lens has a diameter of 10 mm and an offset to the phase center of the dielectric resonator of 1.5 mm. The latter is acquired via EM simulation. Each component of the proposed on-chip antenna is individually discussed in the following subsections. The analytical and EM modeling of the metastructure and dielectric resonator are presented.

A. Metastructures
The utilization of metastructures in tandem with planar antennas is discussed prior in [40]- [42]. These are predominantly deployed to suppress surface currents via the synthesis of bandgap-material-like properties. The modeling methods utilized in this work are an adaptation of what is initially presented in [43], adjusted to fit the 35 nm mHEMT stack.
The metastructured ground plane allows for a controlled reflection of the substrate-incident waves, where the resulting phase delay is withing a range from -90°to 90°. The phase reference point is set at the top metal layer METG located 4.98 µm above MET1. Within the operating bandwidth-the frequency span in which the aforementioned phase delay is present-the reflected substrate-incident waves and the radiated waves undergo constructive interference. Furthermore, the metastructured ground plane presents a high surface impedance. This ensures that the percentile of the substrate-incident waves entering the 50 µm GaAs substrate is as low as possible. As such, through the utilization of metastructures, a high-impedance ground plane is realized with artificial-magnetic conductor properties-despite a large electric field along the surface, the tangential magnetic field is small. This implementation increases the bandwidth and radiation efficiency within the operating frequency. It improves the far-filed patterns of the complete setup as well [41], [42]. The metastructure unit cell represents a defined repeating pattern with its unit dimensions below a tenth of the operating wavelength. For this work, the pattern of choice is presented in Fig. 3. The design makes use of a uni-planar metastructure surface, with its unit cells resembling heavily altered Jerusalem-cross-like structures. The latter are expanded upon with additional axes every 45°, allowing for a wider range of incidence angles for the in-substrate waves. Furthermore, the gaps between these structures are filled with singular segments that alternate in their vertical and horizontal orientation. These provide a further resonance and allow for a denser metastructure layout. The implemented metastructure omits the usage of through-substrate vias and takes into consideration the co-and cross-polarization. The desired inductance is obtained from the current loop along the length of the segments and axes, whilst the capacitance is derived from the gap and edge-length between two adjacent metallic structures. The unit cell as a whole is symmetric. The dimensions of the microstrip patch and metastructure unit cell are included in Fig 3a-the patch is discussed later in the text. The processed stand-alone on-chip antenna can be seen in Fig. 3a. It contains the respective radio-frequency (RF) pad and extension for RF-probing. A close view of the processed metastructure, 4.98 µm below the top metal, can be seen in Fig. 3b. The capacitive portion is composed of an edge length of 4.5 µm and a gap width of 1 µm. The inductive portion is composed of either an axis length of 13 µm or 11 µm. The uniform thickness of all the segments is 1 µm.
Such a metastructure can be approximated via the application of the loaded transmission-line model for plane-wave incidences, as initially established in [43], [44] and highlighted in [42]. This analytical model is altered and adjusted specifically for the 35 nm mHEMT technology. A graphical interpretation is presented in Fig. 4. The main idea lies in the evaluation of the phase of the reflection coefficient for which the complex surface impedance of the metastructured ground is required. The evaluation of the latter is achieved via the separation of the model into a frequency selective surface with an impedance Z G , the respective transition from the reference height of 4.98 µm to the metastructured layer with an impedance Z L , and the transition from the latter to the ground plane with an impedance Z D . The resulting metastructure surface impedance Z S and the respective reflection coefficient for both transverse electric (TE) and transverse magnetic (TM) waves, are: The incident plane wave originates from the microstrip patch in the METG layer, thus the formulation slightly deviates from what is presented in [43]. The wave impedance η L does not refer to free-space, but rather to the impedance seen in the medium created through the combination of the BCB3, SiN, and the BCB2 layers. The angle of incidence is denoted by θ in . At the resonance angular frequency ω 0 , the following holds true [42], [43]: Considering the influence of the mHEMT stack and the geometry of the unit cell, the respective TE and TM complex impedances of the three main layers can be approximated as follows [43]: In the above, the relative permeability-dependent on which portion of the mHEMT stack is taken into account either µ (BCB3, SiN, BCB2) or µ (GaAs, BCB1) -can be expressed as a product of the relative permittivity (dielectric constant) and the square of the relative wave impedance, µ L = ϵ L η 2 L . These parameters are acquired via an EM simulation of the 35 nm mHEMT stack. Dependent on which portion of the stack is considered, the height h varies between 4.98 µm and 51.4 µm. Equations (9) and (10), are dependent on the shape of the metastructure unit cell, thus the effective capacitance C G -originating from the capacitive coupling of adjacent structures-and effective inductance L G -originating from the surface currents on the straight portions of the repeating unit cell. The following equations represent slight alterations to what can be found in [43] for expressing the L G and C G of these heavily-altered Jerusalem-star-like structures: where η ′ is the wave impedance of the effective medium (GaAs and BCB1), which is extracted via an EM simulation of the 35 nm mHEMT stack. α 1 and α 2 are the grid parameters for the expanded Jerusalem-star-like structure and the alternating segments. L 1 , L 2 , and t denote the dimensions of the unit cell and can be found in Fig. 3b. The k ′ is the wavenumber of the effective medium and is expressed as follows [43]: The effective capacitance is expressed as [43]: where λ ′ = 2π/k ′ is the wavelength of the effective medium, and [43]: l and g represent dimensions of the metastructure, which can be found in Fig. 3b. These analytical formulations provide an understanding as to what affects the total impedance Z S , thus the reflection coefficient. As aforementioned, the dielectric properties of the respective layers are extracted via an EM simulation of the 35 nm mHEMT stack. Due to the difficulties in characterizing such a multi-layer stack at the operating frequency of 400 GHz, a dependency on EM-simulated results is present. Taking into consideration that the proposed antenna includes a microstrip patch, a dielectric resonator, and an anti-reflex layer, a full 3D EM numerical simulation approach via CST is chosen in favor of the analytical model. The resulting phase responses of both approaches are discussed in the next paragraph. A 3D EM model of the unit cell, including periodic boundaries and a plane-wave source, is displayed in Fig. 5a. The model contains the complete BEOL, partially concealed for an improved visual recognition. A field probe is placed 4.98 µm from the metastructured layer, and the phase of the reflected signal is analyzed. Since the periodic boundaries ensure an infinitely-expanding and repeating structure, a feature that is not found in practice, a larger segment of the metastructure is simulated via a waveguide-excitation port and respective symmetry boundaries, as presented in Fig. 5b.
The unwrapped phase response of the reflected signal can be viewed in Fig. 5c. It is acquired via both the analytical (transmission-line) model and the 3D EM model. A normal incidence angle is selected for θ in . Since the metastructure is symmetric, only the TE case is plotted. With regard to the EM simulation, the desired phase shift of -90°to 90°is present within a range of 350 GHz to 460 GHz. large operating bandwidth dictates the usage of more involved structures, which due to the high operating frequency have a miniature size. Nonetheless, once the dielectric properties of the respective layers are a known quantity, the analytical model allows for a more computing-time efficient manner of estimating the phase response. The presented results indicated an operating bandwidth of 110 GHz. This allows for an efficient, broadband, and to-the-broadside radiating on-chip antenna to be realized in the 35 nm mHEMT technology, as at least half of the reflected substrate-incident waves undergo a constructive interference with the radiated and evanescent ones.
The respective experimental validation is performed as part of the complete antenna setup. Individually characterizing such a layer is impossible, due to the miniature distances that have to be considered in a transmission-and-reflection measurement. Rather, the characterization is performed as part of the reflection-coefficient and far-field measurements of the complete antenna solution.

B. Dielectric Resonator and Anti-Reflex Layer
The metastructured ground plane allows for the virtual elongation of the substrate height resulting in an increased impedance bandwidth. However, the combined effect of it and a quartz dielectric resonator is required to increase the pattern bandwidth. Making use of the microstrip patch on METG as a feeding point, the 3D EM model is extended to contain a quartz rectangular resonator. As highlighted in Fig. 3a, the microstrip patch has an edge length of 73 µm and a corresponding width of 102 µm. These dimensions differ from what is expected by standard formulations [45], due to the influence of the multi-layer stack, the quartz on top, and the metastructure underneath. Since the latter is not a perfect-magnetic conductor, the GaAs substrate is partially opened. Thus, it influences the final dimensions of the patch. In such a use-case, the implementation of a 3D EM simulation is essential in designing the microstrip patch.
As can be seen in Fig. 3a, the microstrip patch has a grounded frame connected through vias to the backside metal. The resonator functions if a strong magnetic field is excited on a dielectric slab placed on top of a grounded plane. The frame around the microstrip patch represents the latter. The distance of this frame to the edges of the microstrip patch is approximately a quarter-wave in the respective dielectric medium-namely, an edge length of 200 µm and width of 236 µm, optimized via an EM simulation. This allows for the reflected surface currents to be in-phase with the fringing fields.
With regard to the dielectric resonator, charged particles passing through the microstrip-patch-induced magnetic field cause fringing fields. Reflections from its side walls induce standing waves and store electrical energy. The various resonant modes, or field states, that persist within its semi-permeable dielectric walls define the surface-current density distributions, thus the radiation patterns. The aforementioned effect of the metastructured ground plane to reduce unwanted and sporadic surface currents, inherently supports this particular type of integration and contributes towards pattern stability. Furthermore, the transition from the BCB3 to the quartz is advantageous for the EM waves.
The mode propagation within the dielectric resonator depends mainly on its dimensions, relative permittivity, and point of excitation [46]. In this work, quartz was selected as the desired material with a permittivity (ϵ r-SiO 2 ) of 3.75. It represents a favorable transition from the BCB3, which has a permittivity (ϵ r-BCB3 ) of 3.2. Due to mechanical restrictions (as of the writing of this paper), the lateral dimensions of the quartz slab are fixed to 500 µm. The point of excitation is chosen to be the center of the slab's side making contact with the BEOL, since it yields a uniform field distribution. The remaining variable is the vertical z-dimension. Under the assumption of perfect-electrical-conductor boundaries on the four sides of the rectangular slab, with perfect-magnetic-boundary conditions on the bottom and top, the resonant frequency is given by [46]: where z 1 , z 2 , and z 3 are integers, whilst a, b and d are the dimensions of the rectangular slab.
With regard to the height d, (19) is utilized to define an initial starting point for the design process. It is in the range of 375 µm to 475 µm. Since the boundary conditions required for (19) are not present in the implemented design, a full 3D EM simulation and accompanying parameterized sweep is required. The final dielectric slab height of 425 µm is chosen as it results in the desired far-field response. Furthermore, it retains On-chip antenna for S meas. a height-to-lateral-dimensions ratio compatible for mechanical dicing, and excites the TE δ,1,4 mode-as seen Fig. 6a. The far-field pattern at 400 GHz, that results due to the modes induced in the dielectric resonator, is depicted in Fig. 6b. A simulated realized gain of 7.8 dBi is present. The radiation direction is to the broadside.
To improve the decoupling of the wave in the propagating z-direction, an anti-reflex layer is added to the top of the dielectric resonator. This layer comes in the form of a 500 × 500 × 50 µm 3 polycrystaline diamond piece with a permittivity (ϵ r-C ) of 5.68. The EM waves are pulled towards the diamond plate, which is sufficiently thin as to not induce standing waves, thus effectively decoupling the EM waves towards free space. A high frequency capable, thin dielectric material with a respective dielectric constant between the values of 5 to 6, deduced from EM simulations, is desired. A material with a higher ϵ r may result in a mismatched transition, thus in the reflection of the EM waves and subsequent hampering of the metastructured layer.

III. REFLECTION-COEFFICIENT CHARACTERIZATION
In order to characterize the reflection coefficient of the proposed antenna, a stand-alone variant is processed as seen in Fig. 7a. It contains the required RF-contacting pad and microstrip-line extension. To the left, the processed chip is presented with the quartz dielectric resonator on top. To the right, the additional anti-reflex diamond layer is added on top and the side view of the complete setup can be seen. Figure 7b presents the measured and simulated S 11 . The measurements are acquired via an Agilent N5224A PNA network analyzer in combination with a Virginia Diodes, Inc. WR-2.2 extension. The utilized T-Wave T500 RF-probe is calibrated via a through-reflect-line procedure on a standard-impedance substrate. The plotted results do not include the dielectric lens.
The measured impedance bandwidth spans from 340 GHz to 440 GHz, or a 25.6 % relative bandwidth. The simulations consist of multiple curves. Namely, the initial design, which ignores any BEOL deformations of the mHEMT process. The simulated impedance bandwidth of this initial design spans from 330 GHz to 460 GHz. The second iteration, labeled "True-to-Processed-BEOL", updates the EM model to include the overgrowth of the METG layer and the shrinking of the capacitor-contacting via. Since, these systematic deviations from the designed sizes were not considered in the initial model. The simulated impedance bandwidth ranges from 340 GHz to 425 GHz. The third iteration, labeled "True-to-Processed-BEOL and RF-Pad Compensation", takes into account the influence of the RF-pad and microstrip-line extension. The EM model is updated accordingly. The simulated impedance bandwidth ranges from 330 GHz to 425 GHz. The resonance peaks and overall behavior across the operating bandwidth fit well with the measurement.
In order to gain a better understanding of the measured relative bandwidth of 25.6 %, the proposed solution is compared to a hypothetical microstrip-patch-antenna implementation within an "ideal" BEOL that supports a BCB-substrate thickness of 105 µm. This would ensure that the reflected substrate-incident waves are in-phase with the radiated ones. Such a dielectric thickness within the BEOL of the 35 nm process is beyond any realm of possibility. Nonetheless, it presents a good comparison to reveal the advancements made with the coapplication of the metastructured ground plane and dielectric resonator.
To calculate the relative impedance bandwidth of this hypothetical "ideal" solution, the radiated power from the fringing fields P Rad and the surface-wave power P SW must be considered, yielding a bandwidth formulation as follows [47]: with ϵ r-BCB denoting the dielectric constant of the "ideal" BCB substrate with a height of h ideal =105 µm. The surface-wave radiation efficiency η SW is [47]: (22) Taking into account the ratio between the radiated power from a rectangular patch and that of a Hertzian dipole, as well as the patch length L P , the width W P , and the propagation constant k = 2π/λ 0 [47]: To utilize (20), the L P and W P are required. These can be found via models which rely on radiating fringing fields, which extend the effective electrical size of the microstrip patch by ∆ [48]: where ϵ eff-BCB is the effective-dielectric constant [49]:   for which the patch width is defined as: Knowing (24)- (26), the patch length can be devised by: Calculating the required parameters via (21)- (27), allows for the relative impedance bandwidth of the hypothetical microstrip-patch antenna to be calculated via the usage of (20), yielding a BW = 31.5 %. Since the actual substrate thickness is only 4.98 µm, a factor of 21 smaller than the "ideal" case, the achieved value of 25.6 %, when compared to the hypothetical, shows how successful the presented metastructure-and-dielectric-resonator combination truly is.

IV. FAR-FIELD CHARACTERIZATION
Due to the miniature size of the on-chip antenna, the direct RF-probing of it introduces a much larger metallic object within the near-field region. This degrades the far-field patterns and makes such a characterization impossible. Thus, in order to solve said issue, the proposed antenna is integrated into a transmitter S-MMIC as seen in Fig. 8. The analysis of the transmitter-S-MMIC active chain is presented in [31]. It is composed of a multiplier-by-four (x4)-upconverting the input signal from the WR-10 (75 GHz to 110 GHz) to the WR-2.2 frequency band-and a subsequent high power amplifier (PA) [31]. The corresponding on-wafer characterization is depicted in Fig. 9. The output power and transducer gain of the stand-alone active chain is plotted across the measurement-frequency range from 330 GHz to 440 GHz. A peak output power and peak transducer gain of 8 dBm and 10 dB is recorded. A 3 dB bandwidth of 95 GHz spans from 340 GHz to 435 GHz. The output-power and gain curves are flat across the operating bandwidth. A relative bandwidth of 24.5 % is achieved. Said transmitter S-MMIC is assembled on a carrier board, which includes an accompanying in-house multiplier-by-twelve [50] as seen in Fig. 10. This setup no longer requires an RF-probe, rather a baseband signal is fed to the carrier board. This baseband input signal's frequency ranges from 7.3 GHz to 9.4 GHz, the combined 48 th harmonic of which enables the acquisition of the far-field patterns from 350 GHz to 450 GHz.
The complete measurement setup can be viewed in Fig 11, wherein the polypropylene-based dielectric lens is included in the antenna-carrier board setup. The device under test is mounted on a KUKA KR 10 R1100 six-axis maneuverable robotic arm. The transmitter is characterized via a WR-2.2 ZRX500 Receiver by Radiometer Physics (RPG) and a respective WR-2.2 horn antenna with 23 dB of gain (also by RPG). The rest of the setup includes two signal sources for the transmitter and receiver, DC-bias power supplies, and a spectrum analyzer.
The far-field patterns are acquired at a distance of 90 mm to the receiver. The transmitter board is tilted in an angle range from -45°to 45°for both the Pitch and Yaw-along the respective axes and with regard to a fixed phase center-as highlighted in Fig. 11. The measured data, with and without the utilization of a dielectric lens, are plotted in Fig. 12. The principal planes are indicated via the Phi (ϕ) and Theta (θ) angles. The plotted data correspond to a ϕ = 90 • -for the electric-field plane (E-plane)-and ϕ = 0 • -for the magnetic-field plane (H-plane). The θ is swept across the Pitchand Yaw-angle range. The conversion gain of the receiver is compensated according to the data supplied by the manufacturer. The path-loss is included in the plotted data. A 3 dB pattern bandwidth spanning from 365 GHz to 440 GHz can be deduced. The received power ranges from -12 dBm to -9 dBm. Without the lens, the received power spans from -29 dBm to -26 dBm. The resulting difference of 17 dB is attributed to the lens, which collimates the antenna beam. It is dependent on the corresponding aperture illumination, which is affected by the radiation characteristics of the on-chip antenna across the operating bandwidth.
The flat response of the received power is in-part due to the performance of the active chain, yet most significantly it is affected by the operating region of the metastructured ground plane. This is visible in the measured curves, as the drop-off in received signal coincides with the frequency points where the metastructured layer no longer serves its purpose. However, the sharp drop-off in the lower 3 dB cutoff point is purely affected by the utilized multiplier-by-twelve chip. The latter is only able to provide a sufficient twelfth-harmonic drive for the frequency range from 90 GHz to 112.5 GHz-beyond which the signal  is barely strong enough for the pattern shape to be captured. The truncation of the bandwidth, initially reported in [50], is due to the added bondwires and run-to-run statistical variations in the processing of the wafers. Included in Fig. 12 are the 3 dB main-lobe angular width and the main-lobe direction-for the with-lens case. These are consistent over the operating bandwidth and result in a broadside radiating on-chip antenna with an angular width of 7°.
The measured and normalized far-field patterns, for the with-lens case, in both ϕ = 0 • (H-plane) and ϕ = 90 • (E-plane) are presented in Fig. 13a, over a bandwidth from 350 GHz to 450 GHz. The θ ranges from -45°to 45°, limited by the measurement setup. As intended in the design, the patterns are consistent over the measured bandwidth-meaning they hold the same main-lobe direction throughout. A side-lobe level (SLL) suppression of 10 dB or better is present for each instance of frequency measured. The simulated far-field patterns are included as well. The coherence of the measurement to the simulation-in terms of the angular width, main-lobe direction and SLL-supports the functionality of the proposed antenna setup.
The measured and normalized far-field patterns for the without-lens case are plotted in Fig. 13b. Similar to the with-lens case, the measurements span a bandwidth from 350 GHz to 450 GHz. The patterns are consistent throughout the operating bandwidth, with regard to the main-lobe direction and SLL. The angular width is in the range of 50°. Different from the with-lens case, the angular width drifts from 45°to 55°across the operating bandwidth. With the increase in frequency, the angular width decreases. This represents a behavior common to dielectric-resonator based approaches, as over the complete bandwidth various modes overlap. The ripple on the measurements is due to the accuracy limitations of the spectrum analyzer.
In terms of the simulation, included in Fig. 13b, the overlap with the measurement is satisfactory. This is the case for all frequency points, namely 350 GHz, 370 GHz, 400 GHz, and 430 GHz-except for the case of 450 GHz, where the 3D EM model deviates in terms of the SLL. The later becomes apparent when the lens is absent. To be noted is that a complete one-to-one overlap is highly unlikely for the operating frequencies at hand. This is due to the challenges in creating a 3D EM model that completely mimics the processed BEOL structuring (systematic deviations from design sizes). Furthermore, the expected load targets dependent on the biasing of the active chain (meaning the output impedance that is presented to the input of the antenna), the variations in quartz and diamond placement, as well as the submount-carrier-board deviations in terms of the lens positioning (since the latter attaches to the former), are all factors that present a great difficulty when attempting an inclusion in a 3D EM simulation.
In order to derive the directivity from the presented data, a clear understanding of each component in the measurement setup needs to be obtained. The known quantities include the on-wafer performance of the active transmitter chain, the difference in the received power due to the lens, the standard-gain horn, the conversion gain of the receiver, and the distance to it. The contemporary interpretation of Friis' transmission formula, originally presented in [51], is utilized to derive the antenna directivity from the measurements: P R denotes the received power at the output terminal of the respective antenna, in this case the received power seen in Fig. 12. P T is the power fed into the transmitting-antenna input port, in this case the on-wafer measured output power of the active-transmitter chain presented in Fig. 9. The directivity of the transmitting device is labeled D T . In this case, it is composed of the contribution via collimation by the lens ∆G lens and the directivity of the on-chip antenna D ant. . The former can be extracted from Fig. 12. The latter is the desired unknown. The wavelength and distance between the transmitting and receiving antenna are represented by λ and d, respectively. The plot of the directivity and antenna efficiency is presented in Fig. 14. With regard to the directivity, a good coherence between the measurement and simulation is present. Both align to a large degree over the complete operating bandwidth, which is truncated in the upper frequency band due to the active chain only having undergone large-signal characterization up to 440 GHz. As a consequence, a P T reference is absent for the frequencies beyond. Considering the received power in Fig. 12, the directivity extracted from the measurements ranges from 4.4 dBi to 10.4 dBi. It satisfies the simulated variant of 6 dBi to 10 dBi. Including the lens, it spans from 22 dBi to 27 dBi. The difference is dependent on the aperture illumination of the lens. Since the on-chip antenna angular width varies across the operating bandwidth, so does the degree of illumination of the aperture-thus the effectiveness of the lens across bandwidth. The EIRP is plotted as well, and it ranges from 27.5 dBm to 32.9 dBm. It is calculated via the summation of P T , D ant. , and ∆G lens .
The plotted efficiency is that of the simulation, since measuring a radiation efficiency is impossible in the proposed setup. First of all, the antenna load-pulling on the output of the active chain, when operated under large-signal driving conditions, is in this current arrangement not characterizable. The on-wafer characterization of said chip, when the output is terminated with the on-chip antenna rather than an electrically short RF-pad, is not reliable. This is due to the presence of the on-chip antenna resonance within the output characteristics of the active chain, since it is seen in parallel to the RF-contacting probes. Another aspect is the effective bias condition observed during the on-wafer characterization, which cannot be one-to-one replicated in the far-field measurements. This is due to the presence of a biasing board and discrete power supplies, instead of a source and measure unit. Furthermore, the active chain of the chip is characterized via direct probing, different from the final setup, where the submount contains a multiplier-by-twelve up front. Directly probing the antenna is not a solution either, since the RF-probe is always within the near-field of it. The absence of a reference measurement is also significant, as this is an integrated antenna solution rather than a waveguide split-block-module based one. The latter may be mounted on a reference setup, calibrated with known standard-gain horn antennas.
Considering the characterization difficulties listed above and the effectiveness of the simulation model with regard to the far-field patterns, the S 11 , and the directivity, it is reasonable to assume that the radiation efficiency will be comparable to the simulation plotted in Fig. 14-within a reasoning framework that takes into account a certain degree of simulation uncertainty. Within the operating bandwidth, the simulated efficiency varies from 50 % to 66 %. Despite the sub-optimal substrate height, through the utilization of the metastructured ground plane, it is possible to achieve efficiencies commonly expected from patch-antenna-based implementation with a closer-to-ideal substrate thickness.
A comparison of the most significant and recently published results for on-chip antenna solutions operating in the frequency range from 300 GHz to 500 GHz is presented in Fig. 15. There, the directivity is plotted across the pattern bandwidth. Both parameters are depicted via an elliptical shape, with the directivity encoded on the minor axis, whilst the bandwidth is encoded on the major axis. The included on-chip antenna examples contain arrays, as well as singular elements, with and without the implementation of a dielectric lens. Both, to-the-broadside and through-the-substrate radiating concepts are considered. Whether the provided information is measured or simulated, can be deduced by the type of line composing the circumference of the ellipses.
With regard to the measured pattern bandwidth, the proposed solution sets the state-of-the-art spanning from 365 GHz to 440 GHz. This constitutes a relative pattern bandwidth of 19.6 %. With regard to the measured directivity, the achieved peak value of 10.4 dBi, without the lens, is comparable to array-based solutions presented in [16], [36] and singular antenna implementations reported in [23], [35]. In combination with the utilized dielectric lens, the achieved value of 27 dBi sets the absolute state-of-the-art.
V. CONCLUSION This paper presents the analysis, modeling, design, simulation, and characterization of a novel combination   Fig. 15: Directivity-versus frequency-range with respect to the pattern bandwidth for on-chip antenna solutions operating in the frequency range from 300 GHz to 500 GHz. The directivity is encoded in the minor axes of the ellipses and the pattern bandwidth in the major axes.
between a metastructured ground plane, a microstrip patch, a quartz-based dielectric resonator, and a diamond-based anti-reflex layer to realize an efficient, broadband, and to-the-broadside radiating 400 GHz on-chip antenna in a 35 nm InGaAs mHEMT process. The proposed solution represents a first-time implementation for all submillimeter-wave-capable semiconductor technologies. It achieves a measured impedance bandwidth of 100 GHz, a pattern bandwidth of 75 GHz, an efficiency of 50 % to 66 %, and a directivity of up to 27 dBi-with the utilization of a polypropylene-based dielectric lens.
Implemented in the selected semiconductor technology, any to-the-broadside radiating on-chip antenna concept has to overcome a substrate-thickness limitation of 4.98 µm, which favors a leaky capacitor instead of broadband antenna. To solve said issue, a metastructured ground plane is implemented that reflects the substrate-incident waves with a desired phase delay, ensuring that the reflected and the radiated waves undergo constructive interference. An analysis on the design of the metastructure is presented, as well as two modeling approaches are revealed and compared. Between the analytical and 3D EM model, the latter is selected as it allows for the inclusion of the microstrip patch, the dielectric resonator, the anti-reflex layer, and the lens. Whilst the metastructure improves the efficiency and impedance bandwidth, the coapplication in-tandem with a quartz-based dielectric resonator allows for the generation of consistent far-field patterns via the superposition of multiple modes across the operating bandwidth. The utilization of a diamond-based anti-reflex layer aids in the decoupling of the EM waves from the quartz to free space.
Due to the minuscule size of the proposed antenna concept, the only means of far-field characterization, where no RF-probe is within the near-field, is through the integration into a complete transmitter S-MMIC-750 × 2750 µm 2 in size. Said chip is placed on a carrier board, upon which the respective polypropylene-based dielectric lens is mounted. The pattern characterization is performed with the aid of a maneuverable robotic arm. Simulation and measurement results of both with and without the lens are shown, and a good agreement between theoretical and experimental data is observed.

ACKNOWLEDGMENT
The authors would like to acknowledge and thank their colleagues at the Fraunhofer IAF. Oliver Göhlich, for his excellent work in coat-spinning and positioning of the quartz dielectric resonators. Birgit Weismann-Thaden for dicing the quartz wafers. Ralf Schmidt and Martin Zink for laser cutting the diamond plates, and Dirk Meder for assembling the evaluation boards.