ABSTRACT. This paper presents a neural operator-based approach for circuit behavioral modeling using Deep Operator Network (DeepONet). We extend DeepONet by introducing a memory encoder network that incorporates sequential observers through Recurrent Neural Networks (RNNs) to capture long-term temporal dependencies of lossy transmission lines with impedance mismatches and delayed signals. Our Memory-Encoded DeepONet (M-DeepONet) demonstrates adaptive time-step prediction capabilities, showing significant performance improvements over classical DeepONet.

ABSTRACT. We present a signal and thermal integrity (SI/TI) co-optimization framework for multi-chiplet placement and routing using generative adversarial imitation learning (GAIL). Design of modern 2.5-D integrated circuits face the critical challenge simultaneously optimizing TI and SI. GAIL goes beyond simply mimicking expert strategies by learning policies that generate consistent optimal behaviors across diverse state distributions, making it particularly well-suited for chiplet placement problems requiring simultaneous consideration of conflicting TI and SI objectives. Our approach leverages GAIL to learn placement strategies that effectively balance these competing objectives, incorporating both eye diagram-based signal quality metrics with turn penalties and power-aware thermal modeling. Experimental results demonstrate that our framework successfully achieves balanced optimization between thermal and SI constraints, enabling adaptive placement decisions for heterogeneous chiplet systems without requiring objective-specific retraining.

ABSTRACT. In this paper, we propose a novel power distribution network (PDN) design agent that reduces power supply induced jitter (PSIJ) in high bandwidth memory using a Switch Transformer-based reinforcement learning (RL). The Switch Transformer architecture sparsely activates weights in feed-forward network of a conventional transformer model, allowing computational cost to remain low while scaling model size. As Switch Transformer may suffer from instabilities in training loss, proximal policy optimization method is adopted to maximize training efficiency while ensuring stability. The RL reward includes both PSIJ and amount of decoupling capacitor (decap), enabling the agent to satisfy target PSIJ while using minimal number of decap. For validation, performance of the proposed method is compared against random search, genetic algorithm, and Transformer-based RL, where the proposed method consistently outperformed all comparison methods.

ABSTRACT. In this paper, the reinforcement learning-based integrated design optimization methodology is proposed for package substrate interconnection in multiple power domain 3D-ICs. Although the use of multiple power domain in 3D-ICs enables power savings by selectively activating cores tailored to power requirement, it introduces design challenges such as complex domain-to-domain coupling and a vast combinatorial design space. To address these challenges and design objectives, reinforcement learning algorithm with hierarchical framework is introduced. The proposed method achieves SSN suppression and electromigration condition of ball grid array (BGA). Furthermore, the proposed method demonstrates generalization performance with random seed map and current profile. It outperforms the random search and flat reinforcement learning algorithm in terms both multi-objective optimality and execution time.

ABSTRACT. A tailored wirebonding process is proposed, conceived to minimize interconnect losses between active components and printed-circuit-board (PCB)-integrated structures. This is achieved by thinning the integrated circuit (IC) and embedding it into a milled cavity in the PCB carrier to level the IC and PCB surfaces, thereby minimizing bond wire length. In this way, a silicon-germanium BiCMOS D-band power amplifier is compactly integrated on the backside of a PCB-based 4×1 full D-band antenna array with only 1 dB of insertion loss. The effects of variations within expected tolerances in the different processing steps, being the thinning, milling and bondwiring are analyzed to ascertain the robustness of the proposed approach. The thinning and milling step is found to be the most crucial one, with the maximum considered deviations leading to a gain decrease by 2.5 dB. If this step is well controlled, we find that a decrease by 1.5 dB is expected, which agrees well with the measured decrease by under 2 dB.

ABSTRACT. This paper presents the design and analysis of a tunable terahertz bandstop filter based on a monolayer graphene-integrated structure, specifically engineered for operation at 2.3 THz. The filter incorporates a transmission-line loaded pair of stepped impedance resonators to achieve sharp resonance and strong signal suppression. Utilizing the unique tunable surface conductivity of graphene, the proposed filter supports both geometric and electrical tuning mechanisms. Full-wave simulations demonstrate a pronounced bandstop response with an absolute bandwidth of 0.41 THz and a return loss of –25 dB. Parametric analysis shows that varying the resonator physical dimension shifts the resonance frequency from 2.1 THz to 2.2 THz, while chemical potential variation between 0.08 eV and 0.11 eV enables dynamic frequency tuning from 2.19 THz to 2.40 THz. The filter exhibits stable bandwidth and high attenuation across all configurations. Its compact structure, planar integration, and reconfigurable operation make it highly suitable for advanced semiconductor packaging, signal integrity monitoring, and next-generation THz communication systems.

ABSTRACT. High-resolution 77 GHz millimeter-wave (mmWave) radar plays a vital role in reliable perception for autonomous driving. Its short wavelength causes strong clutter on rough surfaces such as curbs, while limited resolution leads to relatively sparse point clouds. The sparsity and non-uniformity of the 77 GHz mmWave radar point cloud present challenges for traditional clustering algorithms. In order to address these physics-induced artifacts, we propose RV-ADPC, a two-stage method. The first stage applies a lightweight U-Net to the RA map to filter curb clutter at the source. Then, an enhanced Density Peak Clustering (DPC) algorithm with a velocity-aware distance metric is applied to the point cloud. RV-ADPC is evaluated on a custom 77 GHz mmWave radar platform. Compared with DBSCAN and DPC, it improves clustering accuracy and robustness, especially in roadside clutter scenarios. In curb-included cases, RV-ADPC improves clustering accuracy by up to 15% compared to traditional DPC and DBSCAN. These results demonstrate its potential for robust, high-resolution perception in mmWave radar applications.

ABSTRACT. A compact slot bow-tie magnetic dipole antenna with horizontal polarization is proposed for Wi-Fi 7 applications. The antenna features a dual-slot array fed by a combined feed network, achieving wide impedance bandwidth, high gain, and broad beamwidth across 5.15–7.125 GHz. Simulations show a peak gain of 13.2 dBi, beamwidth over , and excellent polarization purity. Vertical metal walls are added along the antenna's narrow sides to suppress H-plane sidelobes, enhancing radiation pattern quality. The design offers strong integration potential for orthogonally polarized MIMO systems and is well-suited for complex indoor environments such as supermarkets and corridors, which provides a promising antenna solution for next-generation wireless communication.

ABSTRACT. A decoupling network (DN) is presented to suppress differential-mode (DM) noise radiated from high-speed connectors to nearby antennas. The proposed method utilizes two couplers and a phase delay line to generate a cancellation signal with a precisely controlled magnitude and phase. Theoretical analysis shows that for effective suppression, the phase error must be maintained within ±60° of the ideal 180° anti-phase condition. For experimental validation, a USB Type-A connector and a PIFA antenna were used as the aggressor and victim, respectively. Simulation results show the importance of choosing the effective electrical length of the delay line, and the implemented DN achieves a 18.6 dB DM noise suppression at 2.48 GHz, confirming the effectiveness of the design.

ABSTRACT. In this study, we proposed and measured an ultra-wideband bidirectional absorptive common mode filter(A-CMF) sing a defect ground structure. In order to achieve high-speed digital communication while maintaining compatibilitywith older standards, common mode filters are required to cover high-frequency and wide-band ranges. However, it is difficult to suppress common mode (CM) noise in a wide band while reducing the impact on differential mode (DM) signals. In particular, CMFs that can be adapted to frequencies over 15 GHz have not been studied until now. In addition, while reflective CMFs (R-CMFs) that return CM noise to the input side have been studied in the past, absorptive CMFs (A-CMFs) that absorb common mode noise within the filter have begun to be studied in recent years. And we proposed an ultra-wideband bidirectional A-CMF that attenuates CM by 10 dB over a wide frequency band from 2.5 GHz to 20 GHz and suppresses DM attenuation to less than 3 dB. We conducted prototyping and measured it. We confirmed that the performance was comparable to that of the design.

ABSTRACT. In this paper, a novel hybrid passive common-mode filter (CMF) is proposed to suppress wideband common-mode (CM) noise in high-speed differential signaling systems. The proposed filter integrates a defected ground structure (DGS) with an electromagnetic bandgap (EBG) cavity to achieve enhanced suppression performance over a broad frequency range. While conventional CMFs exhibit narrowband attenuation centered around a single resonant frequency, the proposed hybrid struc- ture achieves suppression exceeding -20 dB across a 3.2 GHz bandwidth. In the time domain, the filter reduces the peak- to-peak CM voltage from 134 mV to 35 mV, corresponding to nearly 75% improvement in CM noise reduction. These results, validated through full-wave 3D electromagnetic (EM) simulation, demonstrate that the proposed hybrid CMF effectively suppresses CM noise while maintaining signal integrity, making it suitable for high-speed and EMI-sensitive digital systems.

ABSTRACT. A design and validation methodology for an ultra-lightweight single-walled carbon nanotube (CNT)-aerogel absorber is evaluated for suppressing radiated noise from high-speed packages. The proposed material achieves a shielding effectiveness (SE) of over 10 dB above 900 MHz. This performance is comparable to a commercial product while being over 60 times lighter. A broadband material model was developed from high-frequency S-parameter measurements to predict the SE. While the prediction showed good agreement with 3-meter method measurements at higher frequencies, a significant difference was observed below 10 GHz. This is attributed to resonant cavity effects formed by the absorber, the circuit board, and its surrounding aluminum box with an aperture, which are not considered by standard material models. This work thus demonstrates the usability of the ultra-lightweight absorber for noise suppression in high-frequency circuits, while also highlighting the necessity of incorporating such structural effects into future predictive models.