Junfa Mao

In 2.5-D packaging, through-silicon vias (TSVs) and micro-bump interconnects significantly influence system-level thermal behavior. However, their complex geometries lead to prohibitively high computational costs in direct finite element simulations. This work proposes a genetic algorithm–backpropagation (GA-BP) neural network approach to predict the equivalent thermal conductivity (ETC) of interconnect structures for efficient multiscale thermal modeling. High-fidelity finite element models were constructed at the representative volume element (RVE) scale, and a process-constrained Latin hypercube sampling strategy was employed to generate training datasets. The genetic algorithm (GA) was employed to optimize neural network parameters, enabling accurate nonlinear mapping between geometric parameters and ETC. The proposed model achieves high prediction accuracy. When applied to system-level thermal simulations, the predicted ETC yields temperature deviations of less than 1 K, while reducing the solver time of the simulation by a factor of 26.42 compared with direct finite element modeling. The proposed method provides an efficient and reliable surrogate modeling framework for rapid system-level thermal analysis of advanced 2.5-D packaging.

This work reports a wideband programmable multibit RF memristor monolithically implemented on a passive silicon interposer to enable nonvolatile, package-level impedance control for chiplet-based RF system-in-package (SiP). The device is based on a Cu/Nafion/Au conductive bridge random access memory (CBRAM) stack and is fabricated on a full 4-in high-resistivity silicon (HRSi) wafer using low-temperature and package-compatible back-end-of-line (BEOL) processes. A broadband equivalent-circuit model is established by first characterizing the electrical properties of the Nafion thin film using dedicated shunt metal–insulator–metal (MIM) structures, and then modeling the memristor as a programmable parallel network with lossy sections. Multilevel programming is achieved by compliance-current-controlled filament formation, providing low-resistance state (LRS) tuning in the range of 1– $80~\Omega $ and five representative RF states with extracted values of 1, 7, 13, 33, and $62~\Omega $ . It also demonstrates an 8000- $\Omega $ high-resistance state (HRS). On-wafer measurements demonstrate distinct transmission states from 1 to 50GHz with good agreement with circuit simulations, enabling broadband programmable attenuation referenced to the lowest-loss state. The demonstrated combination of 1–50GHz operation, nonvolatility with near-zero standby power, compact footprint (0.11mm², the core area is only $50~\mu $ m ${}^{2}$ ), and CMOS/packaging compatibility positions passive-interposer-integrated RF memristors as promising building blocks for per-channel gain trimming, interconnect equalization, and built-in-test in advanced RF SiP.

A broadband dual circularly polarized (DCP) filtering dielectric patch antenna (DPA) is proposed in this letter. Firstly, based on the coupling topology of a bandpass filter, the filtering response of the DPA is designed using its first two resonant modes (TM₁₀₁ and TM₁₂₁). Accordingly, the theoretical external quality factors are derived from the synthesized coupling matrix and are physically realized by constructing an air cavity and tuning the slot length. Subsequently, the four microstrip feedlines are replaced with two hybrid couplers to realize the DCP function. In order to study the isolation between the DCP ports, an anisotropic DP model is established, and an improvement scheme of isolation is provided. For demonstration, a prototype is designed, fabricated and measured. The proposed DPA shows a wide effective DCP bandwidth of 23.5% from 4.29 GHz to 5.43 GHz, which is the overlapped bandwidth of the −10 dB return loss, −15 dB DCP isolation, 3 dB axial ratio (AR), and 1 dB gain. Besides, the proposed DPA achieves a peak CP gain of 8.3 dBi with a low profile of 0.055λ₀.

An autoencoder (AE) is a classic machine learning (ML) method that has been widely applied to feature extraction in natural language processing and computer vision. In this work, a physics-informed AE (PIAE) is proposed for efficient and flexible radiation source reconstruction (RSR). The proposed PIAE has an encoder-decoder pair, in which the encoder converts measured near-field data into equivalent dipole moments and the decoder reconstructs the near field from the dipole moments. The encoder consists of a convolutional neural network, which reduces the dimensionality of the input field data by extracting key features, and a multilayer perceptron, which maps the reduced-dimensionality features to the dipole moments. The decoder analytically calculates the dipole radiation fields from the dipole moments. Finally, the proposed PIAE is trained by minimizing the loss between measured and reconstructed fields. Compared with conventional ML-based RSR methods, which usually need dual-plane phaseless data, the proposed PIAE only requires single-plane phaseless data, which significantly reduces measurement time. Moreover, the proposed PIAE directly outputs equivalent dipoles, enabling seamless integration into electromagnetic interference prediction workflows. Experimental results demonstrate that the proposed PIAE has both high accuracy and efficiency in RSR.

Multiple-input multiple-output (MIMO) frequency-modulated continuous wave (FMCW) radar systems have been widely adopted in advanced sensing applications due to affording superior resolution and accuracy. However, when both a wide modulation bandwidth and a large aperture are incorporated, which are recognized as essential features of the high-resolution radar systems, performance degradation is often introduced due to the range-angle coupling (RAC) effect. This phenomenon impairs resolution in both range and angular domains, with especially severe degradation occurring at off-boresight angles. Although various compensation techniques have been proposed, most of them involve substantial complexity and have limited practical applicability. To address these challenges, a low-complexity and efficient spatial resampling technique (SRT) is proposed, in which the discrete frequency components of the intermediate frequency (IF) signal are utilized to perform resampling in the virtual elements. For each frequency bin, the resampling process is conducted based on the corresponding wavelength, instead of the physical wavelength. By aligning the spatial sampling grid with the wavelength of each frequency across the entire modulation bandwidth, the inconsistency in phase differences across virtual array elements is effectively eliminated. Subsequent 2-D angle-range fast Fourier transforms (FFTs) are applied, ensuring each target to be concentrated within its correct cell and effectively mitigating resolution degradation. In contrast to the conventional techniques that typically introduce higher orders of complexity, the proposed technique computational burden is constituted almost entirely by highly efficient FFT operations, ensuring that an exceptionally low complexity is maintained, with an improvement of 100 times over existing methods. Experiments have been conducted to validate the proposed technique. The technique provides a practical solution for high-resolution, low-cost MIMO FMCW radar by addressing a fundamental wave-propagation issue inherent to wideband large-aperture radar systems.

Lead-free piezoelectric materials have garnered increasing attention as environmentally friendly alternatives to lead-based transducers. Among them, lithium niobate (LN) is a promising candidate. However, the transmit displacement sensitivity of existing LN piezoelectric micromachined ultrasonic transducers (PMUTs) remains significantly lower than that of PZT devices due to the limitation of lateral-field-excitation (LFE). In this work, an LN PMUT operating in the thickness-field- excitation (TFE) mode was designed and fabricated on a 41° Y-cut lithium niobate on insulator (LNOI) substrate. The bottom electrode was innovatively fabricated by deep reactive ion etching and magnetron sputtering, which simplifies the fabrication process and reduces cost compared with conventional methods. The device achieves a transmit sensitivity of 460.3nm/V at an operating frequency of 400kHz, surpassing previously reported LFE LN PMUTs. [2025-0247]

In this article, a flip-chip InP-based balanced ultrawideband (UWB) bandpass filter (BPF) is optimally designed and characterized under cryogenic conditions. The filter chip is designed using spiral inductors and metal–insulator–metal (MIM) capacitors. An electromagnetic (EM)-circuit co-optimization method is introduced to assist in the design process, and the gray wolf algorithm is adopted to iteratively refine variables during each optimization stage. The filter die is then flip-chip bonded to a microstrip-line (MSL) structure on an alumina ceramic substrate. The final filter module is experimentally characterized at 300, 70, and 5K. Measurements show minimal variations in differential-mode (DM) center frequency (15.45 GHz), bandwidth (101.9%), out-of-band suppression, and common-mode rejection ratio (CMRR) (20 dB in the range of 0–24.5 GHz) across different temperatures. Notably, the minimum DM in-band insertion loss (IL) decreases from 1.02 dB at 300 K to 0.72 dB at 70 K, with improvements of 0.3–0.8 dB across the passband. Analysis of the temperature distribution and deformation shows that the reduction in IL is primarily attributed to the increased conductivity of gold at low temperature, which significantly outweighs the effects of dielectric loss and structural deformation.

In this article, a low-loss power dividing transition from a microstrip line (ML) to two rectangular dielectric waveguides (RDWs) at W-band is proposed. The power dividing transition has a patch, a ridge waveguide magic-T (RWMT), three tapered ridge waveguides (TRWs), two dielectric-filled rectangular waveguides (RWGs), two tapered horns (THs), and a RWG. The patch and the RWMT are coupled to realize low insertion loss and high bandwidth. An equivalent circuit model of the transition is established to design the transition based on a synthesis method. The measurement results show that the proposed transition has an insertion loss of 0.49–1.5 dB in the frequency range of 85.1–110 GHz. Within the frequency range, the input return loss of an RDW with two transitions is more than 11.2 dB, the output return loss of the transition with two RDWs is greater than 10 dB, and the isolation between two outputs of the transition exceeds 10.6 dB. Due to its low loss and wide bandwidth, the proposed transition is suitable for RDW-based interconnects.