Leveraging the infill density variation capacity of 3D printers to create an inhomogeneous substrate, an approach based on a single-layer substrate is investigated for antenna miniaturization. Additionally, given the high level of customization offered by 3D-printing, this article explores optimal layouts and minimum material consumption required to achieve a given resonance frequency. A reduction in resonance frequency from 2.40 GHz to 1.77 GHz is demonstrated experimentally while maintaining identical patch and substrate dimensions. Due to the reduction in effective aperture, reductions in gain and bandwidth of 1 dB and 1% respectively, are observed compared with a reference prototype on a homogeneous substrate; however, bandwidth can be preserved when compared to a homogeneous substrate with a similar resonance frequency. For experimental validation the patch antennas are fabricated with the dielectric filament Preperm ABS1000 and copper tape for the metallic components.
This work presents an experimental evaluation of transmission loss introduced by copper pillar interconnects in a D-band flip-chip packaging environment. The test assembly comprises two chips fabricated using a high-performance 130 nm SiGe BiCMOS technology and interconnected through gold-to-gold thermo-compression bonded copper pillars with ENIG finish. This process enables low-loss microstrip transitions with stable high-frequency behavior, supporting accurate characterization of the copper pillar interconnects. In addition to the standard thru and reflect structures used for TRL calibration, multiple microstrip transmission lines with varying lengths were implemented to quantify the insertion loss introduced by the copper pillar interconnects. The measured results indicate a transition loss of 0.44 dB per copper pillar interconnect at 170 GHz. The use of solder-free copper pillars eliminates reflow-related variability and minimizes contact resistance and parasitic discontinuities, making this approach well-suited for high-frequency and fine-pitch millimeter-wave and sub-THz system integration.
This paper presents a high-frequency, high-accuracy electrical model for symmetric ground-signal-signal-ground (GSSG) through-silicon vias (TSVs) by integrating their physical structural features with full-wave simulation results. Based on odd-even mode analysis, the proposed model derives scalable resistance, inductance, conductance, and capacitance (RLGC) equations to characterize the electrical properties of GSSG TSVs, and a polynomial fitting method is applied to optimize the RLGC parameters with incorporated frequency and physical dimension dependencies. A large-scale modeling approach is utilized to generate 900 variable combinations across the frequency range of 0–40 GHz, TSV heights of 100–200 μm, radius of 10–25 μm, and pitches of 50–250 μm. The extracted lumped-parameters, including TSV resistance(RTSV), TSV inductance (LTSV), TSV capacitance (CTSV), silicon capacitance (CSi), silicon conductance(GSi) and oxide capacitance (Cox) are employed to reconstruct frequency-dependent formulations, and the model’s validity is rigorously verified at the component level by both simulation and experimental measurement. This work provides a scalable framework for the simulation of arbitrary symmetric GSSG TSVs, obviating the complex formulation derivation of analytical methods for different TSV processes and morphologies. The component-level validation ensures the proposed RLGC formulations can be directly integrated into circuit design flows, laying a reliable foundation for the accurate modeling of TSV-based 3-D IC interconnect systems at both component and system levels.
In this paper, physics-informed neural networks (PINNs) are developed for the efficient stochastic analysis of multiconductor transmission line (MTL) networks assuming frequency independent per-unit-length parameters. The proposed PINNs minimize a physics-informed loss function comprising of the residuals of (i) the Telegrapher’s partial differential equations representing the MTLs, (ii) the nonlinear modified nodal analysis (MNA) equations representing the circuit terminations, and (iii) the initial conditions of the network. Consequently, the PINNs can be trained in a data-free manner at very low time and memory costs while achieving superior generalizability outside the training domain compared to conventional artificial neural networks (ANNs). Moreover, the proposed PINNs can accommodate non-zero initial conditions spatially distributed along the MTLs using an auxiliary shallow network, thereby removing the need to retrain the entire PINNs at high time costs every time the initial conditions change. Finally, to address the slow convergence associated with training PINNs, a homotopy-based multi-fidelity strategy along with prior knowledge infused transfer learning is also developed in this work. The above features of the proposed PINNs are validated through multiple numerical examples.
In this paper, two K-band compact dual-band ceramic-based half-mode substrate integrated waveguide (HMSIW) filters are proposed. In one of them, the planar microstrip modes and SIW cavity modes can be combined to construct a single-cavity hybrid four-mode resonator, and this hybrid resonator is used to design a compact second-/second-order filter, which exhibits significant advantages including ultra-small size, high unloaded quality factor and controllable modes. In the other, the half-wavelength microstrip resonator with interdigital capacitor is embedded within the HMSIW and simultaneously coupled to two cavity modes to form the cross-coupled topology, which realizes a high-selectivity third-order/second-order dual-band filter with very compact size. Finally, two dual-band filters are designed at 13.3 GHz/20.82 GHz (f1/f2) and 15.43 GHz/26.6 GHz (f1/f2) with 3-dB fractional bandwidth of 25.27%/27.81% and 25.2%/18%. In addition, the minimum insertion losses for the first and second passbands are 1.00 dB and 0.56 dB for the first filter, and 0.90 dB and 1.11 dB for the second filter, respectively. They occupy areas only 0.23λ12 and 0.27λ12, where λ1 is guided wavelength at f1.
