Interface-engineered barium magnetoplumbite–wide-bandgap semiconductor integration enabling 5G system-on-wafer solutions for full-duplexing phased arrays

Modern cellular technology is in the midst of a transition from 4G, operating at 0.6 < f < 2.3 GHz, to 5G, proposed to operate at 28 and 39 GHz with plans to extend to 86 GHz depending upon the commercial carrier. An initial phase-in period would operate at 3.3 < f < 3.8 GHz in order to support legacy systems during product transition. The move to higher frequencies offers consumers much higher data handling capacity with 10–100 times faster transfer rates of up to 10 Gbps at latency less than 10 ms.¹ 

This Ka-band technology, having characteristic shorter wavelengths, experiences higher attenuation and provides localized coverage over short distances requiring denser concentrations of base stations than legacy 3G and 4G technologies. To address this need, massive multiple input–multiple output (MIMO) antenna systems having large numbers of elements that simultaneously transmit and receive (STAR) data are being developed. STAR technologies provide full duplexing while concomitantly reducing form factor and weight and rely upon active electronically scanned array (AESA) technology first developed for military systems. Massive MIMO multi-element base stations will provide greater capacity to many more users at faster data rates.

STAR is made possible by the circulator, a three-terminal passive device. In its conventional manifestation, the circulator employs a ferrite substrate magnetically biased by a rare earth magnet [typically of the Nd(Tb,Dy)₂Fe₁₄B variety] to break time reversal symmetry and provide the necessary nonreciprocal flow of electromagnetic (EM) energy for the simultaneous transmission and reception of communication signals. Key device performance metrics include low insertion loss (IL) through the device (port 1 $→$ 3 and 3 $→$ 2), low reflected loss (RL) at the high-power amplifier channel (port 1 $↔$ 1), and high isolation (IS) between transmit and receive channels (port 1 $→$ 3), the latter is needed to protect the low-noise amplifier network.² 

As with any new or emerging technology, innovation in the form of new material compositions, structures, or processing technologies often proves singularly enabling. As in the case of the modern circulator, the most disruptive advance over the last few decades has come from the successful development of ferrite materials possessing high magnetic retention, i.e., high magnetic remanence, M_r, also represented as hysteresis loop squareness, $Sq=MrMs$⁠. This material property defines the ability of the ferrite to retain magnetic polarization upon the removal of an applied magnetic field. Hence, poling with a magnetic field, i.e., application and removal of a saturation field (H_sat), allows these ferrites to retain more than 90% of their saturation magnetization (M_s) unless the ferrite experiences a degaussing event, for example, exposure to a temperature approaching its Curie temperature (T_C). We refer to such ferrites as being self-biased (SB).³ 

If one employs a self-biased ferrite as the substrate for a circulator, the self-biased (SB) circulator no longer requires a biasing magnet to align magnetic spins to achieve time reversal symmetry breaking and device reciprocity; these are provided intrinsically by the ferrite without the bias magnetic field. Without the affixed bias magnet, the traditional three-dimensional (3D) device topology becomes two-dimensional (2D), $∼$90% smaller, lighter, substantially less costly, with a robust and secure supply chain. The first commercial sales of SB circulators were made by Metamagnetics Inc. circa 2015.⁴ Those products are surface mounted devices that employ hexaferrite materials as the SB magnetic media in below resonance circulators operating at Ka-band offering competitive performance such as isolation >15 dB and an insertion loss <1 dB over a bandwidth greater than 15%.⁴ 

The continued development of 2D SB circulators is very relevant to the successful implementation of 5G and, in particular, its Ka MIMO base stations. Conventional circulators, due to their high aspect ratio, introduced by the bias magnets, preclude the design and construction of Ka-band MIMO arrays whose array lattice does not allow for key element spacings. In order to overcome this challenge, SB ferrites must be integrated with common wide-bandgap (WBG) heterostructures enabling the integration of active circuitry (e.g., amplifier networks, filter banks, limiters, and digital signal processing subsystems). All these systems by virtue of gains made by very large scale integration (VLSI) have over time experienced dramatic size reduction, while the form factors of circulators, whose dimensionalities are often dictated by operational wavelengths, have not experienced commensurate volumetric efficiencies.

To this end, a central tenet of this research is to demonstrate the growth of device quality ferrite films on WBG heterostructures without degradation to the heterostructure chemistry, surface, structure, nor prefabricated circuitry, which are suitable for system on wafer (SoW) solutions that would enable Ka-band MIMO arrays and ultimately the timely realization of 5G implementation.

As early as the mid-1980s, the U.S. government, by way of DARPA, sponsored research to explore the integration of ferrites with semiconductor substrates. At that time, the leading semiconductor materials were silicon (Si) and the III–V gallium arsenic (GaAs) systems. Unfortunately, the high substrate temperatures required to grow high crystal quality pure phase ferrite films with ultralow RF losses, led to alloying at the (ferrite film)/(Si substrate) interface, or dissociation of the GaAs substrate, leading to untenable increases to device insertion loss.

During the 1990s to the present, wide-bandgap semiconductor substrates, such as GaN and SiC, were refined for high power–high frequency (HP–HF) applications. Years of research and development led to the successful large-scale processing of high-purity single crystal GaN and SiC having low defect concentrations, with and without the introduction of p- and n-dopants. To date, GaN has assumed a dominant market position for HP-HF electronic devices, whereas SiC has led for its exceptional thermal conductivity. The fact that the two share a common crystal symmetry, and near lattice match, allows for ready heteroepitaxy, i.e., GaN/SiC, and optimization of their respective properties.

Because of the successful development of GaN and SiC technologies (and other WBGs, such as AlN, ZnO, and h- and c-BN), the integration of ferrites with semiconductors is being re-examined. Although the overall challenges remain the same, the boundary conditions have been altered, namely: (i) WBG semiconductors under study, unlike Si and GaAs, can withstand high temperature growth conditions without degradation to chemistry nor structure; (ii) GaN and SiC have hexagonal structures that have close, but not ideal, lattice matches to promising hexaferrite phases; and finally, (iii) WBG semiconductors under study, together with hexaferrite phases best suited for heterogeneous epitaxy, have near-ideal magnetic and RF properties to address the needs of Ka-band MIMO arrays for 5G applications. The attractive and synergistic properties afforded by the integration of barium hexaferrite with WBG materials as candidate mm-wave HP–HF heteroepitaxy systems were realized over 2007–2011.^5–9 

In 2007, Chen et al. grew magnetoplumbite barium hexaferrite (i.e., BaM, BaFe₁₂O₁₉) by pulsed laser deposition (PLD), on 10 nm MgO-coated 6H SiC single crystal substrates. This represented the first successful growth of device-quality ferrite films on semiconductor substrates. The films had clear (00l ) crystal orientation and ferromagnetic resonance (FMR) linewidths of 96 Oe (at 53 GHz).⁶ [We choose to express the Miller indices of diffracted planes from the hexagonal phase as (h k l ) as opposed to the (h k i l ) convention.] This group of researchers followed with important studies of the aging affects upon chemical, structural, and magnetic properties,⁷ and reactive ion etching behavior,⁸ all essential to advancing materials and device processing technologies.

In 2010, Chen et al. followed with the first PLD demonstration of BaFe₁₂O₁₉ films on MgO-coated GaN/Al₂O₃ single crystal substrates. Similarly, these had strong (00l ) crystal orientation and FMR linewidths of 86 Oe (at 53 GHz).⁹ 

Another important contribution to ferrite film development that focused on semiconductor integration was the work of Kulik et al. who demonstrated the PLD growth of high quality BaM films on SOI substrates.¹⁰ Alternative deposition techniques, such as sputter deposition, MBE, metalorganic chemical vapor deposition (MOCVD), and screen printing, among others, have also been explored. Zhang et al., using RF magnetron sputtering, deposited strongly textured BaM films on Pt coated Si wafers showing promise for this alternative path.¹¹ Cai et al. employed molecular beam epitaxy to deposit BaM on 6H-SiC. The BaM film crystal quality was confirmed by in situ RHEED, but RF magnetic properties measured post-deposition fell short of industrial requirements.¹² Nie et al. employed a metal-organic decomposition technique to produce BaM films on Pt coated Si substrates that showed high retention and attractive RF properties.¹³ An unconventional approach to high quality ferrite film processing was taken by Chen et al. in the use of screen printing. In a notable series of publications, these authors demonstrated thick film processing of strongly textured BaM films on a variety of substrates, including Si and SOI.^14–16 

A detailed review of these and other related works has been presented by Chen and Harris as reference.¹⁷ 

Although the previous work of Chen et al.^6–9 confirmed the ability to grow high quality ferrite films on surfaces of WBG heterostructures, the substrates employed did not represent WBG substrate systems employed by industry, which typically include GaN/SiC and AlN/SiC, among others. As such, they did not serve as suitable models in either surface quality, defect concentration, chemical stability, thermal conductivity, nor thermal expansion behavior. For these reasons, we revisit the challenge of growing BaM films on GaN and AlN surfaces on SiC substrates.

Nucleation layers (NLs) in many cases are necessary to provide a smooth, pinhole free, diffusion barrier, and atomic template for growth that minimizes strain induced from the film to substrate interface arising from mismatches in surface atomic positions (i.e., the growth template) and thermal expansion coefficients. Interface roughness, manifested in composition or structure, can lead to detrimental increases in device insertion losses. Here, a variety of nucleation layers were explored.

All substrate heterostructures reported here are of the form BaM/NL/WBG-CL/SiC(X), where NL denotes nucleation layer, WBG-CL is the wide-bandgap capping layer (i.e., here GaN or AlN), and X indicates that the 6H-SiC substrate is a single crystal. In some cases, the termination of the SiC, be it C-polar vs Si-polar, was also explored. The growth of the WBG capping layers was performed using metalorganic chemical vapor deposition (MOCVD) following the protocols of Ref. 18, while nucleation layers of MgO and Pt were grown by atomic layer deposition (ALD, Ultratech Fiji plasma-assisted ALD reactor). Pt nucleation layers of 20 nm, with an average surface roughness of 0.43 nm, were deposited at 250 °C with (111) crystal texture. Similarly, MgO nucleation layers of 32 nm, with an average surface roughness of 0.31 nm, were deposited at 325 °C with (111) crystal texture. As representative of film quality, the FWHM of the MgO (222) was 1000 arc sec (or 0.278°).

BaM films were pulsed laser deposited using a KrF excimer laser operating at 248 nm generating 10 ns pulses at a rate of 10 Hz. Laser pulses were incident upon a pure phase 25 mm diameter BaM target at a 45° angle in a dynamic vacuum of 26.6 Pa (or 200 mTorr) of oxygen partial pressure. The laser plume was intercepted by a substrate holder held at 920 °C. The substrate holder, which also served as a resistive heater, was aligned parallel to the target at a distance of 70 mm. The laser beam remained fixed, while the target was rotated at 30 rpm and rastered relative to the incident laser beam in order to ensure a homogeneous plasma and uniform films. The substrate surfaces were cleaned by sonication in separate baths of acetone, alcohol, and de-ionized water and affixed with silver paint to the substrate holder/heater stage. The laser output power was controlled by varying the voltage output of its power supply achieving a power density at the target surface of $∼$4.8 J/cm². This power density has been shown previously to allow accurate stoichiometric transfer from target to substrate.¹⁹ Each sample included 36 000 laser pulses resulting in a $∼$500 nm BaM film.

Samples were characterized for phase purity and crystallographic texture, microstructure, atomic structure, surface morphology, and composition. Cross-sectional imaging and analyses included evaluation of chemistry and phase and structure of interface regions. DC magnetic and RF properties were measured and correlated with data gleaned from the above measurements. Investigative techniques included x-ray diffractometry (XRD) including pole figure analyses, atomic force microscopy (AFM), x-ray reflectivity, scanning electron microscopy (SEM), and high resolution transmission electron microscopy (HR-TEM), and energy dispersive x-ray spectroscopy (EDXS). Magnetic characterization was performed using vibrating sample magnetometry (VSM), whereas RF characterization was performed by the measurement of ferromagnetic resonance (FMR) following the methodology of Ref. 20. FMR measurements were carried out using a vector network analyzer (VNA) at a frequency of 48 GHz. High frequency losses are presented as the FMR linewidth $(ΔHFMR)$⁠, sometimes referred to as the gyromagnetic resonance linewidth, that has a direct impact upon the insertion loss of circulators and isolators.

Figure 1 is a representative room temperature $θ−2θ$ x-ray diffraction pattern from a BaM film grown on MgO/AlN/SiC(Si-polar). The spectrum, in which the diffracted intensity is plotted on the logarithmic scale, illustrates (00l ) reflections from the BaM, AlN, and SiC phases with additional low-intensity reflections from the MgO (111) nucleation layer. No other diffraction features rise above the $∼$1% intensity threshold relative to the BaM (008) peak. The inset to Fig. 1 contains pole figures acquired along the ⟨008⟩ and ⟨110⟩ directions confirming the epitaxial nature of the BaM film having c-axis-oriented grains and sixfold in-plane registry expected from hexagonal crystals, respectively.

Figures 2(a) and 2(b) shows AFM and SEM images, respectively, of the surface morphology of the sample whose XRD data are presented in Fig. 1. Here, one observes 0.5–1 $μ$m diameter grains with visible hexagonal facets oriented with c-axes normal to the sample plane. This is also observed in the panel (b) SEM image. Close inspection of both images reveals the presence of terraces.

PLD film growth typically takes place by step-flow and/or Stranski–Krastanov (SK) mechanisms.²¹ Step-flow growth occurs when substrate temperatures are sufficiently high to allow migration of surface adatoms to form terraces. Such features are typically smooth and of high crystalline quality. Continued growth often leads to the stacking of terraces that are clearly visible in Fig. 2. Alternatively, SK growth consists of an initial 2D growth, in this case likely associated with a step-flow mechanism, before transitioning to 3D island growth. 2D terraces of the growing film act as templates for the nucleation of 3D islands. Nucleation is triggered by the accumulation of stacking faults that form to accommodate stress resulting from the atomic lattice mismatch between the BaM film and the heterostructure template. The images of Fig. 2, together with the x-ray diffraction spectrum and pole figures of Fig. 1, confirm growth mechanisms and the epitaxial nature of BaM films on select WBG heterostructures presented here.

Figures 3(a)–3(e) represent a composite of transmission electron microscopy images and fast Fourier transforms (FFT) of the heterostructure cross section. Panel (a) is illustrative in that it depicts five distinct regions: (i) the AlN capping layer (⁠$∼$50 nm) on the Si-polar SiC crystal substrate, (ii) an amorphous-like diffusion region of Al₂O₃ (⁠$∼$45 nm) that forms between AlN and MgO, (iii) a crystalline nucleation layer of MgO (32 nm) with (111) crystallographic orientation, (iv) a diffusion region between the MgO and BaM films forming the spinel MgFe₂O₄ (⁠$∼$40 nm), and (v) a highly crystalline thick film of BaM. Panel (b) is an expanded HR-TEM view of the BaM film structure showing a near perfect arrangement of parallel atomic planes defining the length of M-blocks to be 1.186 nm. This value is consistent with values reported by others [e.g., 1.16 nm (Ref. 22)]. From these data, BaM lattice parameters of a = b = 0.601 nm and c = 2.366 nm were deduced. Based on these calculations, a and c lattice parameters experience a $∼$2% dilation relative to published bulk BaM values of a = b = 0.588 76 nm and c = 2.318 85 nm.²³ However, these values reflect the atomic structure nearest the interface region where strain is most pronounced. The in-plane lattice parameter of the BaM film >50 nm above the interface region indicates lattice parameters of a = b = 0.5873 nm reflecting a slight compressive strain of $∼$0.3% and a constant c parameter of 2.366 nm, which maintains the aforementioned tensile strain of 2%. Figure 4 illustrates the transition from the MgFe₂O₄ spinel layer to the hexaferrite phase that includes a large mismatch in lattice parameters (⁠$aMgFe2O4=0.836 nm$⁠)²⁴ providing a considerable strain field. This transition triggers the formation of stacking faults. The accumulation of copious stacking faults in the immediate region above the interface of the spinel interdiffusion layer and the BaM film is clearly visible. The stacking fault region extends to about 30 nm above the interface. Beyond this region, stacking faults are no longer detected in large numbers.

These atomic characteristics for BaM film growth suggest that the in-plane strain is relieved as the BaM grows thicker, a common observation in heteroepitaxial growth. Alternatively, the out-of-plane strain remains unchanged. This behavior has been attributed by others to the presence of oxygen deficiencies in the BaM film.²⁵ 

Figures 5(a)–5(d) are hysteresis loops for heterostructures whose comparison proves illustrative in relating sample structure and chemistry to magnetic properties. In Fig. 5(a) are hysteresis loops for the BaM/MgO/AlN/SiC(X) sample that show clear magnetic anisotropy with an anisotropy field (H_a) of $∼$16.5 kOe and a saturation magnetization (4 $π$M_s) of 4.2 kOe. These values are very near published values for bulk BaM materials, i.e., 17 kOe and 4.3 kG, respectively.²⁶ Similar loops are presented for heterostructures without the AlN capping layer and for a BaM/Pt/SiC(X) heterostructure. For the BaM/MgO/SiC(X) sample [Fig. 5(b)], loops collected for both in-plane and out-of-plane directions nearly overlap indicating very little magnetic anisotropy and a low H_a that is difficult to quantify merely by inspection. Also the saturation magnetization, as 4 $π$M_s, is 1.7 kG, or 62% lower than the BaM/MgO/AlN/SiC(X) sample. Electron microscopy characterization of this sample shows a lack of epitaxy and a microstructure dominated by misoriented grains that did not coalesce during growth, indicating the important role of the capping layer in passivating the SiC surface. Figure 5(c) shows a comparison of loops collected from a sample similar to that presented in panel (b) where the nucleation layer has been replaced by Pt. In this sample, the anisotropy is smaller than (a), and the magnetization reflects porosity that was confirmed by SEM and TEM studies. Finally, panel (d) illustrates hysteresis loops collected from the BaM/MgO/GaN/sapphire(X) heterostructure in which the saturation magnetization is near the bulk value. However, electron microscopy reveals a nonuniform film surface with grain outgrowth that induces surface pinning of domain walls resulting in a high coercivity (H_c) of >2 kOe. The high coercivity positively impacts squareness, $Sq=MrMs$⁠, which reaches 0.70 but degrades $Δ$ H_FMR by >60% over the BaM/MgO/GaN/SiC(X) heterostructure value of 340 Oe. Contributions to FMR linewidth in our samples by magnetic anisotropy are reflected in Schlömann's expressions as $ΔHanis=1.08JHA24πM$ and from porosity as $ΔHporos=0.502J4πM(p)$²⁷ where H_A is the anisotropy field (⁠$2KuM$⁠), p is the porosity of the material (⁠$∼$0.04 $±$ $0.02$⁠), J is a shape factor 0.89 (demagnetizing factor, N_Z = 0.95),²⁸ and 4πM_s = 4210 G.

Using these expressions, we estimate ΔH_anis = 58 Oe and ΔH_poros = 75 Oe from which ΔH_total = 133 Oe. The measured linewidth was 340 Oe. The remaining 207 Oe of dynamic loss is attributed to two-magnon scattering (TMS). TMS arises from the interaction of two spin waves with impurities or discontinuities, such as grain boundaries or triple junctions, giving rise to energy loss to the magnetic media. In our cases, this would be through interactions with one- and two-dimensional defects (e.g., stacking faults), interfacial discontinuities in structure and chemistry, and at grain boundaries and triple points. Defects giving rise to TMS are often also responsible for DC hysteretic losses manifesting as high H_c. In our samples, H_c values approach or exceed 1 kOe supporting this assertion.

Table I presents quantifiable magnetic properties where measurements permitted. The first row of Table I is data reported by Karim et al.³⁰ of BaM single crystal platelets²⁹ at microwave frequencies. A complete set of properties are provided with the exception of H_c. Additionally, data from Wu et al.³¹ on bulk polycrystalline textured BaM samples are also presented as a reference. These data reflect the bulk-like values common to polycrystalline samples. In the work of Wu et al., it is important to realize that coercivity (H_c), anisotropy field (H_A), and FMR linewidth (ΔH_FMR) are highly sensitive to microstructure and processing conditions and can vary greatly among studies.

By inspection of Table I, one observes the highest quality heterostructure sample is the BaM/MgO/AlN/SiC(X, Si-polar). An identical sample with C-polar termination of the SiC substrate, being the only difference, led to an AlN capping layer having five times the average surface roughness compared to the sample having Si-polar termination. This led to an overall degradation of crystal quality and magnetic properties (i.e., 61% increase in coercivity, 13.3% reduction in saturation magnetization, and a lack of discernable magnetic anisotropy). The causal relationship between Si- vs C-termination, and the resulting interfacial structure, relates to the chemical affinity and bonds formed between these atoms to those of the AlN capping layer. If one considers Si-termination, Si–Al (and Si–N) bonds are examined. Kiv et al. performed total-energy density-functional theory calculations using angular momentum dependent pseudopotentials to explore the nature of Si–Al bonding. Adiabatic potentials for Si–Al interaction were calculated and found that Al atoms participate in strong sp³ hybrid bonds with Si.³² Additionally, Chen et al. explored the relative stability Si–N bonds also by ab initio calculations. The local atomic structure with Si–N bonding exhibits a considerably lower total energy of 0.65 eV, indicating a strong affinity of Si–N bonds in a-SiCN thin films.³³ Alternatively, C–Al bonding was investigated by Du et al. using a total-energy density-functional theory. All electrostatic and quantum-mechanical interactions of valence electrons were accounted by angular momentum dependent pseudopotentials. Findings indicate that almost all of the 3p electrons of Al atoms transfer to C atoms providing bonds of ionic nature.³⁴ 

Finally, C–N bonding is well known to chemists since they form in cyanide compounds. These bonds also experience sp hybridization. Taken together, the Si–(AlN) bonds are strongest and are likely to maintain flat and featureless surfaces of the capping layer that is verified by AFM measurement. Alternatively, the C–(AlN) is weaker and would allow for C–C surface reconstruction giving rise to a rougher AlN surface and the deterioration of the BaM magnetic properties that has also been verified here.

The BaM/MgO/GaN/SiC(X) heterostructure is characterized as having the lowest ΔH_exp = 340 Oe and near bulk-like anisotropy field. However, its saturation magnetization is 24% lower than the best sample signally high porosity or extensive interfacial alloying. All samples had hysteresis loop squareness values below 50%. At these values, the structures lack self-bias properties. Lacking self-bias properties does not preclude their use in circulators but it does require the use of a bias magnet. An exception is the BaM/MgO/GaN/sapphire(X) heterostructure. This structure has a squareness of 70% and a saturation magnetization approaching bulk-like values. However, the static and dynamic loss properties of this structure are higher than representative industrial material standards.

In summary, presented here are results of a systematic study of PLD grown BaM films on industrial-compatible WBG semiconductor heterostructures suitable for operation in Ka-band circulators. BaM films having device quality performance in structure, epitaxy, and magnetic properties were realized for BaM/MgO/AlN/SiC(X) heterostructures. Film properties include near bulk-like values of magnetic anisotropy field, H_a$∼$16.6 kOe, and saturation magnetization, 4πM_s$∼$4.2 kG. FMR linewidth values are consistent with device design goals for insertion loss and other device performance metrics. Heterostructures where SiC substrates have Si-polar termination show particular promise, whereas C-polar surfaces give rise to high and detrimental average roughness leading to the deterioration of RF performance. Overall, hysteresis loop squareness values remain low for self-bias device applications. Prior work has shown that reducing dynamic losses (i.e., two-magnon scattering) while maintaining static losses (i.e., domain wall pinning) is one strategy to realize high squareness and low FMR linewidths.⁹ 

These results provide an important step in the integration of active and passive components on a single wafer essential to the realization of a transmit/receive system on wafer (SoW) topology necessary for the development of 5G Ka-band massive multiple input–multiple output (m-MIMO) antenna arrays with true full duplexing in time and frequency domains.

The authors acknowledge support from the Army Research Office under Contract No. W911NF-17-S-0003 (Program Manager: Dr. Chakrapani Varanasi) and the Defense Advanced Research Program Agency's M3IC program (original Program Manager: Dr. Dev Palmer, presently managed by Dr. Young-Kai Chen).