Embedded glass fan-out integration method for high-performance Ka-band RF microsystem Authors: Bohan Zhang, Lang Chen, Qi Wang, Wei Wang Published: 2026 ABSTRACT Heterogeneous radio-frequency (RF) microsystem integration is pivotal for overcoming the physical limitations of monolithic integration, enabling the low signal transmission loss and high operating frequencies demanded by next-generation communication networks. Glass interposers, characterized by inherently low dielectric loss, exceptional planarity, and highly tunable coefficient of thermal expansion, have emerged as an ideal platform for millimeter-wave (mmWave) RF microsystems. In this study, we propose a high-density, low-noise RF integration technology utilizing an embedded glass fan-out process. To address the challenges of glass micromachining, laser-induced deep etching (LIDE) was employed to fabricate high-precision cavities with superior verticality and minimal sidewall roughness for seamless die embedding. Subsequently, through an optimized chemical-mechanical polishing (CMP) process, the surface roughness of the redistribution layer (RDL) is reduced by 97%, successfully suppressing the transmission loss to below 0.25 dB/mm. A compact Ka-band microsystem integrating the low-noise amplifier and the antenna was designed and fabricated, featuring an optimized Chebyshev antenna array and a Klopfenstein taper transition for broadband impedance matching. Furthermore, we demonstrate a heterogeneously integrated Ka-band transceiver microsystem that combines high-performance GaN-based amplifiers with a cost-efficient silicon-based switch. The integrated transceiver exhibits a transmit gain of 26.08 dB and a receive noise figure of 2.73 dB at 28 GHz, within an ultra-compact footprint of 6 × 3 mm². This work provides a highly scalable and robust heterogeneous integration solution for high-performance, multi-functional mmWave microsystems in 5G-Advanced and future 6 G networks. Introduction The rapid evolution toward 5G-Advanced and 6 G communication networks has fueled an ever-growing demand for high-performance and miniaturized millimeter-wave (mmWave) RF front-ends1,2. However, as operating frequencies extend into the Ka-band and beyond, the performance limits of monolithic microwave integrated circuits (MMICs) are fundamentally constrained by the physical properties of any single material system3–5. Specifically, no single substrate can simultaneously deliver the superior RF performance of III-V compounds and the high-density logic control capabilities of silicon-based processes without compromising cost and yield currently. To overcome this bottleneck, heterogeneous integration (HI) has emerged as the defining paradigm for next-generation RF microsystems6–9. This approach enables the co-integration of best-in-class chiplets within a single advanced packaging platform. For instance, combining GaAs or GaN RF dies—noted for their high gain and low noise—with Si-based dies that excel in high-density logic control not only drastically reduces system-level costs but also significantly enhances architectural flexibility and overall RF performance. A critical challenge in mmWave HI architectures lies in establishing low-loss, impedance-continuous packaging interconnects. Conventional wire-bonding introduces substantial parasitic inductance10, radiation loss11, and impedance mismatch12, which severely degrade signal integrity. To shorten these interconnect paths, fan-out wafer-level packaging (FOWLP) based on epoxy molding compounds (EMC) has been widely adopted. However, EMC suffers from severe thermal warpage and die-shift during the curing process, heavily restricting integration density13–15. Meanwhile, although ceramic substrates like low-temperature co-fired ceramic (LTCC) possess excellent high-frequency dielectric properties, volumetric shrinkage during sintering and the resolution limits of mainstream thick-film screen printing prevent them from achieving the high-density routing required for advanced heterogeneous integration16. While mature silicon interposers can facilitate high-density, high-quality routing, the inherent semiconducting nature of silicon induces problematic substrate leakage and high-frequency parasitic capacitance17. Meanwhile, glass interposer is rapidly emerging as a superior alternative platform for mmWave heterogeneous integration18,19. Glass interposers inherently possess an ultra-low, tunable dielectric constant and an extremely low loss tangent, which significantly reduce the insertion loss of high-frequency signals20. Furthermore, the outstanding mechanical rigidity and dimensional stability of glass provide an ideal physical foundation for ultra-flat, high-precision redistribution layer (RDL) routing21. Being naturally compatible with large-area panel-level manufacturing, glass also offers a highly cost-effective mass-production route for high-density RF microsystems22. These advantages have motivated increasing research interest in glass-based fan-out packaging for high-frequency heterogeneous integration. The Georgia Institute of Technology first demonstrated glass fan-out (GFO) packages using through-glass cavities, glass-to-glass bonding, and semi-additive RDL formation, establishing the feasibility of glass fan-out for multi-chip integration22. Yu et al. at Xiamen University then extended the embedded silicon fan-out (eSiFO) concept to glass and developed embedded glass fan-out (eGFO) packaging by forming through-glass cavities, filling the chip-to-glass gaps with molding compound, and fabricating antenna arrays on the top RDL23. More recently, the Georgia Institute of Technology reported a multiple-die glass package in which partial cavities were used for thin dies and through cavities were used for thick dies, enabling heterogeneous chiplet embedding20. They further characterized high-frequency interconnects in die-embedded glass interposers, supporting the RF feasibility of glass-based die-embedding technology24. However, these studies mainly validated passive interconnect performance, rather than demonstrating functional performance of packaged active dies or the complete active RF microsystem. Moreover, previously reported glass fan-out processes were generally based on laser ablation to fabricate through-glass cavities or mixed structures combining partial cavities and through cavities, which require large die-to-cavity clearances and may introduce thermal defects25. In addition, the surface roughness of plated RDLs remains a key source of conductor loss at mmWave frequencies26. To address these limitations, this study proposes and systematically validates an embedded glass fan-out (eGFO) heterogeneous integration methodology for high-performance Ka-band RF microsystems. The proposed process employs laser-induced deep etching (LIDE) to fabricate high-precision blind cavities for chip embedding, which preserves a stable cavity floor for die attach and backside common-ground formation, while the cavity depth can be independently defined according to the height of each chiplet. By building RDLs directly atop the embedded RF chiplets, we eliminate the need for wire bonds or flip-chip bumps in the critical RF path, thereby minimizing parasitic inductance. Furthermore, through an optimized chemical-mechanical polishing (CMP) process, the surface roughness of the RDL is reduced by 97%, aggressively suppressing the transmission loss to below 0.25 dB/mm. Based on this advanced packaging platform, this work demonstrates two highly integrated mmWave microsystems to validate the proposed technology, as schematically illustrated in Fig. 1. First, a compact Ka-band antenna in package (AiP) microsystem is designed and fabricated on the glass interposer, featuring an optimized Chebyshev antenna array and a Klopfenstein taper transition for broadband impedance matching. Furthermore, we demonstrate a heterogeneously integrated Ka-band transceiver (TRX) microsystem that seamlessly combines the high-performance GaN-based low noise amplifier (LNA) and power amplifier (PA) with a cost-efficient silicon-based switch. The integrated microsystem exhibits a transmit gain of 26.08 dB and a receive noise figure of 2.73 dB at 28 GHz, achieved within an ultra-compact footprint of 6 × 3 mm². Ultimately, this work provides a robust, high-performance, and multi-functional heterogeneous integration solution tailored for next-generation 5G-Advanced and 6 G mmWave communication networks.Fig. 1Schematic of the proposed glass-based embedded fan-out technology.The diagram illustrates the heterogeneous integration of chiplets for both the LNA-AiP microsystem and the heterogeneously integrated TRX microsystem Schematic of the proposed glass-based embedded fan-out technology. The diagram illustrates the heterogeneous integration of chiplets for both the LNA-AiP microsystem and the heterogeneously integrated TRX microsystem Results Optimization of key manufacturing processes Well-defined embedded cavities can minimize the area of the microsystem, enhancing the yield and reliability of chip embedding. Currently, the patterning of glass interposers can be primarily categorized into two main methods: laser ablation and laser-induced etching. We employed both methods to process the embedded cavities and compared the resulting topography with that achieved through traditional silicon-based etching processes, as shown in Fig. 2. Laser ablation utilizes high-energy lasers to directly act on the glass surface, causing the material to vaporize or melt instantaneously through photothermal effects, forming holes or structures. However, the high temperatures during processing can easily cause micro-cracks in the patterned areas, thereby affecting chip packaging reliability. Additionally, as shown in Fig. 2a, the sidewalls after laser ablation exhibit high roughness under scanning electron microscopy (SEM), making it difficult to meet the requirements of high-precision buried fan-out processes. Laser-induced etching, on the other hand, utilizes ultra-short pulse lasers to induce chemical structural changes within the glass, followed by selective removal of the modified regions using hydrofluoric acid etching. Since the wet etching step can simultaneously remove multiple laser-modified regions across the glass panel, LIDE provides relatively higher processing efficiency and better scalability for fabricating large numbers of cavities in panel-level manufacturing. As shown in Fig. 2b, the embedded cavities created by laser-induced etching can achieve higher precision patterning, with results closely resembling those of the deep silicon etching in Fig. 2c. Additionally, the sidewalls and trench cavities are smooth and crack-free, meeting the requirements for high-precision chip embedding.Fig. 2Optimization of key manufacturing processes.a–c Comparison of cavity profiles formed by laser ablation, LIDE, and silicon etching. d Schematic representation of the skin effect. e Simulation results showing the increase in insertion loss with surface roughness at 0–40 GHz. f AFM surface measurements of the copper RDL roughness before and after CMP. g Measured transmission line insertion loss before and after CMP Optimization of key manufacturing processes. a–c Comparison of cavity profiles formed by laser ablation, LIDE, and silicon etching. d Schematic representation of the skin effect. e Simulation results showing the increase in insertion loss with surface roughness at 0–40 GHz. f AFM surface measurements of the copper RDL roughness before and after CMP. g Measured transmission line insertion loss before and after CMP In Ka-band applications, transmission loss due to the skin effect can not be ignored. With frequency increases, current concentrates in a thin layer on the conductor surface. As illustrated in Fig. 2d, increased metal surface roughness effectively lengthens the current path. The impact of roughness on RF signal transmission loss can be modeled using the Hammerstad correction factor27:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K}_{H}=1+\frac{2}{\pi }\arctan \left[1.4{\left(\frac{\Delta }{\delta }\right)}^{2}\right]$$\end{document}KH=1+2πarctan1.4Δδ2Where Δ represents the root-mean-square (RMS) roughness of the metal surface, and δ denotes the skin depth of the corresponding metal at the specific operating frequency. Consequently, the equivalent resistance Reff is formulated as:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{eff}={K}_{H}\sqrt{f}\frac{\rho L}{S}$$\end{document}Reff=KHfρLS Rough surfaces cause current path meandering, leading to increased ohmic and eddy current losses. The simulation result, as shown in Fig. 2e, indicates that increasing roughness from 0 nm to 15 nm increases insertion loss by 100% at 40 GHz for a specific transmission line geometry. To mitigate the impact of surface roughness on RF performance, we employed an optimized CMP process to the coplanar waveguide (CPW) transmission lines after the electroplating of interconnection metals and subsequent seed layer etching. Atomic force microscopy (AFM) was used to measure the surface roughness of the metal transmission lines before and after the optimized CMP process, as illustrated in Fig. 2f. Before CMP, the RMS roughness was measured to be 13.15 nm. After the CMP process, the RMS roughness was reduced to 0.338 nm, representing a remarkable reduction in surface roughness of 97%. The measured transmission loss of the CPW before and after CMP across 0 to 40 GHz is illustrated in Fig. 2g. After the implementation of the CMP process, the insertion loss of the CPW transmission line was significantly reduced to less than 0.25 dB/mm across 0–40 GHz. This represents a substantial improvement compared to the insertion loss of 0.75 dB/mm at 30 GHz observed before the CMP process, corresponding to a reduction in loss of ~67%. Design of passive components on glass interposer To achieve system miniaturization, we designed a Ka-band microstrip patch antenna array directly on the glass interposer, as shown in Fig. 3a. Using the Chebyshev synthesis method, a 4-element series-fed array was designed with dimensions detailed in Table 1. The antenna array was optimized in design to enhance directivity and gain. Through finite element simulation, the structural parameters were optimized to maximize directivity and gain, yielding a simulated bandwidth of 3.50 GHz and a peak gain of 10.10 dBi, as depicted in Fig. 3b, c.Fig. 3Design and characterization of passive components.a Layout of the 4-element series-fed microstrip patch antenna array designed with Chebyshev synthesis. b Simulated S11 of the antenna array. c Simulated radiation pattern of the antenna array. d Layout overview of the interconnection between the microscale chiplet and the antenna. e Detailed view of the Klopfenstein taper transition structure. f Simulated return loss of the transition structure. g Comparison of measured S11 and Gain between the glass interposer antenna and the Rogers substrate referenceTable 1Dimensions of the antenna array (unit: mm)ParameterValueParameterValueParameterValuew14.500l12.500w23.654l22.500w32.334l32.500w41.179l42.500w51.210l51.860wfeed10.400lfeed12.500wfeed20.400lfeed22.500wfeed30.400lfeed32.500wfeed40.400lfeed42.500 Design and characterization of passive components. a Layout of the 4-element series-fed microstrip patch antenna array designed with Chebyshev synthesis. b Simulated S11 of the antenna array. c Simulated radiation pattern of the antenna array. d Layout overview of the interconnection between the microscale chiplet and the antenna. e Detailed view of the Klopfenstein taper transition structure. f Simulated return loss of the transition structure. g Comparison of measured S11 and Gain between the glass interposer antenna and the Rogers substrate reference Table 1Dimensions of the antenna array (unit: mm)ParameterValueParameterValueParameterValuew14.500l12.500w23.654l22.500w32.334l32.500w41.179l42.500w51.210l51.860wfeed10.400lfeed12.500wfeed20.400lfeed22.500wfeed30.400lfeed32.500wfeed40.400lfeed42.500 Dimensions of the antenna array (unit: mm) A significant challenge in mmWave integration is the scale mismatch between the chiplet pads (50–100 μm) and the antenna elements (1–10 mm). Figure 3d shows the layout of the antenna-to-chiplet interconnection, where the antenna feed is substantially larger than the chiplet pads. To address this, we implemented a gradual impedance matching network based on Klopfenstein taper theory. As shown in Fig. 3e, the transition structure modulates the input impedance by varying the gap width of the coplanar waveguide. This structure connects the 50 μm × 50 μm chip pads to the 1.21 mm-wide antenna feedline, ensuring a short and smooth transition with a simulated S11 below −15 dB across the entire 0–40 GHz band, as shown in Fig. 3f. For comparison, a conventional PCB-based antenna was fabricated using the Rogers 5880 substrate. In the reference system, the same antenna was implemented on the Rogers 5880 substrate and connected to the external measurement setup through SMA connectors, representing a typical discrete RF integration solution. We fabricated the aforementioned passive components on the glass interposer and compared them with the Rogers substrate reference, with measurement results shown in Fig. 3g. The red curve labelled Glass (CPW) represents the measured response of the glass-interposer antenna including the CPW between the pad and the antenna feed. The glass-based antenna exhibited an operating bandwidth of 30.51 GHz to 36.79 GHz and a maximum gain of 7.88 dBi. Compared to the Rogers substrate, the glass-based design extended the bandwidth by 2.54 GHz and improved the peak gain by 2.21 dB, validating the superior performance of the glass-integrated approach. Further design and optimization details are provided in the Materials and Methods section. LNA-antenna integrated microsystem We integrated the 4-element antenna array and the transition structure with a GaN-based LNA chiplet using the glass embedded fan-out process, shown in Fig. 4a. For comparative analysis, a conventional discrete integration solution based on a Rogers 5880 organic substrate with SMA connectors was also fabricated. The wafer-level glass-based antenna-in-package achieved a footprint of only 19 × 4 mm2, significantly smaller than the 12 × 6 cm2 overall footprint required for the discrete Rogers-based reference system, including the PCB for chips, antenna region, and SMA connectors, as shown in Fig. 4b.Fig. 4Fabrication and measurement of the LNA-Antenna integration microsystem.a Wafer-level fabrication process flow. b Photograph of the compact glass-based AiP microsystem 19 × 4 mm2 versus the discrete Rogers reference 12 × 6 cm2. c Measured Gain and Noise Figure of the packaged LNA compared with die datasheet specifications. d Measured system gain comparison between the glass-embedded fan-out microsystem and the wire-bonded Rogers reference system Fabrication and measurement of the LNA-Antenna integration microsystem. a Wafer-level fabrication process flow. b Photograph of the compact glass-based AiP microsystem 19 × 4 mm2 versus the discrete Rogers reference 12 × 6 cm2. c Measured Gain and Noise Figure of the packaged LNA compared with die datasheet specifications. d Measured system gain comparison between the glass-embedded fan-out microsystem and the wire-bonded Rogers reference system The measurement results revealed that the overall package introduced a maximum insertion loss of 0.5 dB compared to the direct probing of the bare LNA die pads, shown in Fig. 4c. The measured gain of the integrated LNA-antenna microsystem is shown in Fig. 4d. The results demonstrate that the glass-embedded microsystem achieved higher gain than the reference wire-bonded system while maintaining consistent performance trends, demonstrating that the packaging process effectively preserves signal integrity with minimal parasitic effects. Switch-PA-LNA heterogeneous integrated transceiver microsystem Furthermore, we fabricated a Ka-band transceiver microsystem by embedding GaN-based amplifiers and a silicon-based switch into the glass interposer, using the process illustrated in Fig. 5a. The transmit and receive functions are controlled by the DC bias of the switch. In receive mode, the receive channel is activated, enabling the LNA to amplify incoming external RF signals. Conversely, the PA is enabled in transmit mode, allowing the microsystem to radiate RF signals outward.Fig. 5Integration and characterization of the multi-chip transceiver microsystem.a Fabrication process flow for the heterogeneous integration of the silicon-based switch, GaN-based LNA and PA. b Micrographs of the fabricated transceiver microsystem. c Measured S-parameter of the receive path. d Measured S-parameter of the transmit path. e Measured noise figure Integration and characterization of the multi-chip transceiver microsystem. a Fabrication process flow for the heterogeneous integration of the silicon-based switch, GaN-based LNA and PA. b Micrographs of the fabricated transceiver microsystem. c Measured S-parameter of the receive path. d Measured S-parameter of the transmit path. e Measured noise figure The fabricated Ka-band front-end microsystem is shown in Fig. 5b. The inset on the top-right highlights the critical components—the silicon-based switch, GaN-based LNA and PA—which are compactly integrated on a single glass interposer via the embedded process. The top-left inset details the input channel with the glass interposer, where signals are routed from the LNA to the switch through optimized interconnects, while the bottom-right inset details the output channel. The bottom-left inset presents an enlarged view of the through-trench interconnections with excellent morphology. Additionally, the figure reveals the ground layer, which provides a stable grounding path and effectively mitigate electromagnetic interference, reducing the system noise. Finally, the integrated wiring demonstrates high-precision alignment and micron-level dimensional control. This precise routing capability enables multi-functional integration within a reduced footprint while simultaneously guaranteeing signal integrity and overall system stability. The measured S-parameters are presented in Fig. 5c–e. For the receive chain, accounting for the insertion loss of the switch and the package, the typical receive gain is 14.91 dB at 28 GHz. The noise figure of the packaged microsystem is less than 4.5 dB, with values ranging from 2.5 to 3.05 dB across the operating band and a typical value of 2.73 dB at 28 GHz. For the transmit chain, accounting for the switch insertion loss and package loss, the typical transmit gain is 26.08 dB at 28 GHz. Discussion Advantages of glass-based AiP Despite the slight increase in loss, the performance trends of the packaged LNA die remained consistent with those of the bare die, indicating that the glass-embedded fan-out packaging process effectively preserved the intrinsic performance of the LNA die. This consistency demonstrates the robustness of the packaging technology in maintaining signal integrity and minimizing parasitic effects, which are critical for high-frequency applications. Compared to traditional Rogers substrate discrete integration, the wafer-level glass AiP offers higher integration density, a smaller footprint, and lower packaging loss. Performance comparison To further evaluate the overall system capabilities, Fig. 6 compares the measured transmitter gain and receiver noise figure of this work with recently reported Ka-band transceiver. Typically, achieving both ultra-low noise and high transmission gain in a highly integrated microsystem remains challenging due to the inherent insertion loss of integration substrates and interconnect parasitics. By leveraging the ultra-low dielectric loss of the eGFO technology, our glass-based heterogeneous integration effectively minimizes these interconnect penalties and preserves the intrinsic high-gain and low-noise characteristics of the RF dielets. Consequently, this work is positioned in the optimal high-performance region of the comparison chart, demonstrating an exceptional receiver noise figure of 2.73 dB and a remarkable transmitter gain of 26.08 dB, which proves highly competitive among existing monolithic and advanced packaging solutions.Fig. 6Performance comparison of Ka-band transceiver.Comparison of transmit gain versus receive noise figure against state-of-the-art monolithic29–31, silicon interposer32, LTCC33, and organic-based34 microsystems, highlighting the high-performance attributes of this work. * Tx gain derived from large-signal output power Performance comparison of Ka-band transceiver. Comparison of transmit gain versus receive noise figure against state-of-the-art monolithic29–31, silicon interposer32, LTCC33, and organic-based34 microsystems, highlighting the high-performance attributes of this work. * Tx gain derived from large-signal output power Materials and methods Design of the Ka-band wideband high-gain microstrip antenna array The antenna was designed based on the transmission line model, where the rectangular patch is equivalent to a transmission line resonator of width w and length l. The radiation width w is calculated by:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w=\frac{c}{2{f}_{c}}{\left(\frac{{\varepsilon }_{r}+1}{2}\right)}^{-\frac{1}{2}}$$\end{document}w=c2fcεr+12−12where fc is the center frequency, c is the speed of light, and εr is the relative dielectric constant of the glass interposer. The glass interposer used in this work exhibits a dielectric constant of 3.78, and the dry film for dielectric layer features a dielectric constant of 3.2. Based on the structural material parameters, the initial dimensions of a single patch element were analytically calculated as w = 2.0 mm and l = 1.2 mm. However, simulations revealed that single patch elements provide limited radiation gain, which is insufficient for high-performance millimeter-wave microsystems. To achieve high gain and a narrow beamwidth while strictly controlling the side lobes, we employed a 4-element series-fed array utilizing Chebyshev synthesis. This method effectively balances the main lobe width and the side lobe level. For an N-element array, the array factor S(β) is derived from Chebyshev polynomials to ensure that all side lobes remain at an equal level. We specifically designed the array for an side lobe level of −15 dB to provide a sufficient margin for potential process variations during manufacturing. Accordingly, the current distribution ratio for the radiating elements was calculated as:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${I}_{1}:{I}_{2}:{I}_{3}:{I}_{4}=1.55:1.2:1:0.55$$\end{document}I1:I2:I3:I4=1.55:1.2:1:0.55 To maintain phase consistency, the center-to-center spacing of the patches was set to λg, and the patch lengths were set to approximately 0.5λg. The specific dimensions of the feed network were calculated to ensure a 50 Ω characteristic impedance, resulting in a main feed width w5 of 1.210 mm. Based on the material parameters and analytically derived dimensions discussed above, the entire 4-element antenna array structure underwent iterative optimization using the finite element method. The structural dimensions were continuously fine-tuned until the target performance metrics, including broadband impedance matching, optimal resonant frequency, and maximum gain. This iterative optimization process ultimately yielded the final dimensional parameters for the antenna array, which are summarized in Table 1. Design of the Klopfenstein taper matching structure To bridge the significant dimensional gap between the chiplet pads (~50 μm) and the antenna feedline (1.21 mm), a Klopfenstein taper was employed. This structure provides the shortest physical length for a given reflection coefficient specification compared to linear tapers. The impedance variation follows the function:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{ln}}\,Z(z)=\frac{1}{2}\,{\mathrm{ln}}({Z}_{{0}}{Z}_{{L}})+\frac{{\mathrm{ln}}({Z}_{{L}}/{Z}_{{0}})}{\cosh \,A}{A}^{2}{\int }_{0}^{z/L}\frac{{I}_{1}(A\sqrt{1-{y}^{2}})}{A\sqrt{1-{y}^{2}}}dy$$\end{document}lnZ(z)=12ln(Z0ZL)+ln(ZL/Z0)coshAA2∫0z/LI1(A1−y2)A1−y2dywhere L is the transition length and I1 is the modified Bessel function of the first kind. In our physical design, the gap size of the coplanar waveguide structure varies according to the curve:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\rm{gap}}=158.8{{\rm{e}}}^{-0.002858{\rm{z}}}$$\end{document}gap=158.8e−0.002858z Simulations confirmed that a 450 μm long transition achieves a return loss better than −15 dB across the full DC to 40 GHz range. Glass-based embedded fan-out fabrication process The process for the proposed glass-based embedded fan-out process is summarized as follows. First, the chips are prepared by laminating a die attach film (DAF) onto their backsides to ensure a robust mechanical bond with the cavity floor upon embedding. In this study, a conductive DAF was utilized to provide the necessary electrical interconnection between the chip backside and the interposer’s common ground. Subsequently, embedding cavities are fabricated using LIDE. To ensure smooth chip placement during die bonding, the lateral dimensions of each cavity are designed to be slightly larger than those of the corresponding chiplet, resulting in an ~10 μm gap between the chiplet edge and the cavity sidewall. Benefiting from the high dimensional accuracy and vertical sidewall quality of LIDE, this design gap can be tightly controlled while maintaining sufficient tolerance for chip insertion. Given the high-frequency operating environment of RF microsystems, a global ground metal layer is deposited across the glass interposer to suppress electromagnetic interference and ensure signal integrity. Following the formation of this common ground layer, the chips are precisely placed into the cavities using a sub-micron die bonder. A vacuum lamination process using TMMF S2000 photosensitive dry film is then employed to fill the gaps surrounding the embedded chips and planarize the reconstituted wafer surface. During lamination, the dry film is softened under heating and pressure and conforms to the embedded-chip topography, while the vacuum environment removes trapped air and promotes void-free gap filling. With the optimized vacuum lamination conditions, the gaps between the embedded chips and the cavity sidewalls can be fully filled without apparent voids or surface dimples, thereby forming a continuous and flat dielectric surface for subsequent RDL routing. Even if slight local dimples occur in the gap-filling regions, they are not expected to affect RDL fabrication, as also supported by a previous silicon-embedded heterogeneous integration28. Photolithography is then performed to open vias and expose the I/O pads of the embedded dies. Finally, the RDLs for inter-chip connectivity and surface passive components are fabricated using a semi-additive copper plating process, thereby completing the integrated packaging. In this study, Z4U polishing slurry was employed and optimized for the CMP process. The polishing pressure was set to 1.2 psi, the rotation speed of the polishing platen was maintained at 35 rpm, and the polishing duration was varied between 30 s and 60 s to achieve optimal reduction in surface roughness of the metal. Specific process variations were implemented for the different microsystems. For the antenna-in-package realization, to enhance antenna bandwidth, the topside ground layer directly below the antenna elements was etched away, and a backside ground layer was deposited. The antenna array and the matching structure were patterned simultaneously with the top metal routing layer. For the integrated transceiver microsystem, the step of preparing a backside ground layer on the substrate was omitted. RF measurement setup The RF measurements were performed using calibrated probe-station and instrument setups according to the corresponding test requirements. Specifically, S-parameter characterization was performed using a Keysight N5245B vector network analyzer, calibrated with a Keysight N4693D electronic calibration module, on a Cascade Summit 200 semi-automatic probe station. RF excitation and output characterization were conducted using a Keysight N5183B signal generator, a Rohde & Schwarz FSVA3044 spectrum analyzer, and a Keysight N1912A power-meter setup. DC biasing and electrical monitoring were provided by a Keysight N6705C power supply and a Keysight 34465 A digital multimeter. Conclusion In conclusion, this work demonstrates a highly scalable embedded glass fan-out integration methodology tailored for high-performance Ka-band RF microsystems. By combining LIDE for precise die embedding and an optimized CMP process that reduces surface roughness by 97%, the platform effectively suppresses mmWave transmission losses to below 0.25 dB/mm. The successful fabrication of a compact LNA-AiP microsystem and a fully heterogeneously integrated GaN/Si transceiver microsystem validates this approach. The integrated system achieves a 26.08 dB transmit gain and a 2.73 dB receive noise figure at 28 GHz within an ultra-compact 6 × 3 mm² footprint. Ultimately, this eGFO technology overcomes the limitations of conventional substrates, providing a robust heterogeneous integration solution for next-generation 5G-Advanced and 6 G communication networks.