The accuracy of RF transistor models is critical for reliable microwave circuit design, yet conventional modeling workflows rely heavily on expert experience for extraction sequencing, and iterative manual tuning, resulting in long cycles and limited reproducibility. This paper presents an AI agent-based method for automatic DC parameter extraction of RF transistors, supported by a dynamic retrieval-augmented generation (RAG) module that incrementally encodes domain knowledge from relevant documentation. The extraction task is decomposed and execution paths are generated by a large language model (LLM), the intelligent optimization algorithms are employed to iteratively update parameters and assess convergence, forming a closed-loop workflow of simulation, evaluation, and optimization. The method is validated on BSIM4 and HEMT model, achieving a reduction in extraction time from several days or weeks to a few hours, with average RMS error maintained below 10%. This method provides a new implementation path for modeling RF devices.

This article presents a synthesis method for on-chip fourth-order bandpass filters (BPFs) for 5G millimeter-wave (mmWave) applications. The method transforms the conventional $K$ -filter prototype into microstrip schematics comprising capacitively loaded microstrip (CLM) resonators and semi-lumped vias acting as impedance inverters. The approach enables rapid calculation of design parameters under varying bandwidth, center frequency, and harmonic suppression requirements. Two BPFs are designed and fabricated using the integrated passive device (IPD) process to validate the method. The first operates at 24 GHz with an insertion loss of 3.98 dB, a 3-dB fractional bandwidth (FBW) of 13.5% and 53-dB stopband attenuation. The second employs a cross-coupled structure at 26 GHz, achieving an insertion loss of 3.92 dB, a FBW of 12.5%, and transmission zeros (TZs) at 22.8 and 28.4 GHz, corresponding to a rectangular coefficient (20-dB BW/3-dB BW) of 1.53. The results validate the theoretical counterparts and demonstrate that the designed filters exhibit good passband selectivity and stopband suppression.

As semiconductor device models grow in complexity, conventional parameter extraction techniques struggle with heavy computational overhead and conflicting performance metrics. In this study, we introduce a highly efficient extraction approach utilizing a Surrogate-Assisted Multi-Objective Grey Wolf Optimizer (SAMOGWO). To minimize reliance on time-intensive simulator calls, the framework leverages a Kriging model to dynamically approximate the device's physical responses. This surrogate-guided search mechanism significantly accelerates the discovery of optimal parameter sets while balancing trade-offs between different electrical characteristics. Extensive benchmark testing proves the algorithm's robust search capabilities. Furthermore, when applied to GaN HEMT DC parameter extraction, the simulated curves show excellent agreement with measured data (RMSE < 7%).

Radio-frequency (RF) integrated circuit design requires reasoning across electromagnetics, device physics, impedance matching, stability analysis, and SPICE/netlist-level debugging. General-purpose large language models (LLMs) can support design analysis, but their responses remain sensitive to missing domain context and physically inconsistent numerical reasoning. This paper presents RF-LLM, a constraint-aware specialized LLM assistant for RF circuit reasoning and netlist analysis. The framework combines a hierarchical RF instruction corpus, progressive low-rank adaptation, and a confidence-gated retrieval/context augmentation mechanism that injects RF physical constraints when retrieval evidence is insufficient. RF-LLM is evaluated on a curated RF textual question-answering benchmark covering conceptual reasoning, numerical derivation, and netlist-aware analysis, together with representative RF circuit debugging cases. The results indicate that RF-domain adaptation and constraint-aware retrieval improve the reliability of LLM-assisted RF circuit analysis, particularly for calculation-heavy and netlist-level tasks.

With the increasing demand for higher bandwidth and frequency in high-speed digital systems, the interference of common-mode (CM) noise in differential signal transmission has become more severe. CM suppression filters (CMFs) were proposed to solve this issue, but their design process typically relies on empirical parameter tuning with extensive electromagnetic (EM) simulations, which not only increases design costs but also limits efficiency and flexibility. In this article, a transformer-based inverse design method for CMFs is proposed for the first time, and it can eliminate the need for empirical parameter adjustments by automatically predicting the targeted geometric parameters, thereby improving the design efficiency. In addition, to address the problem of imbalanced data distribution, the multilabel synthetic minority over-sampling technique (MLSMOTE), which can enhance the data representation in sparse sample regions, is implemented. Further validation on tunable CMFs confirms that the proposed inverse design method has broad applicability. The experimental results demonstrate that the proposed inverse design method can accurately predict the geometric parameters and improve the efficiency, thereby providing an innovative solution for the design and applications of CMFs.

A disconnect remains between high-fidelity physical-characteristic simulation and upper-level validation in RF system design. High-fidelity simulations can accurately characterize key physical effects, such as frequency response, noise, and nonlinearity, but their results are difficult to directly transform into executable models for upper-level validation. In contrast, upper-level validation often relies on idealized or empirical parameters rather than real hardware characteristics. To address this issue, this paper proposes a dataflow-driven behavioral modeling method for RF systems, with system input-output characteristics as the modeling core. A behavioral model is constructed using characteristic blocks representing frequency response, noise, coupling, nonlinearity, and phase shift. Model parameters are configured from high-fidelity simulation results and/or hardware measurement data, thereby establishing a parameter-transfer path from physical-characteristic results to the executable behavioral model. Driven by baseband-equivalent input data streams, the model generates output data streams containing key physical effects and provides a reusable RF-link model for upper-level validation. The proposed method is instantiated and validated on the receive (Rx) channel of an X-band eight-channel phased-array transmit/receive module. Comparisons with circuit-level benchmark results demonstrate that the proposed method can effectively inherit underlying physical characteristics and exhibits good accuracy and practical feasibility.

The absence of a high-quality and standardized data foundation severely restricts the pervasive application of Artificial Intelligence in Electronic Design Automation (AI for EDA). Existing RF design assets are highly heterogeneous and fragmented, which makes them difficult for AI agents to parse, retrieve, and reuse automatically. To address these challenges, this paper proposes a hierarchical governance and intelligent quality assurance (QA) framework for heterogeneous EDA assets across the entire RF toolchain. First, a multi-tier verification engine encompassing physical compliance and cross-file collaborative logic is constructed to guarantee the efficient and standardized quality verification of multi-level engineering assets. Second, a knowledge-graph-based semantic consistency verification module is introduced to elevate the verification process from local rule checking to global relational reasoning, enabling the detection of structural and logical inconsistencies across heterogeneous models. Furthermore, a cross-modal semantic consistency verification mechanism featuring schema-guided dual-branch collaboration is designed. Large multimodal models are further employed to verify semantic consistency between unstructured datasheets and structured engineering configurations. Experimental results on 665 heterogeneous RF assets show that only 33.38% of the original assets satisfy baseline compliance requirements. The proposed framework detects 1084 defects spanning physical integrity, semantic consistency, and cross-modal verification, demonstrating its effectiveness in improving the reliability and reusability of RF design assets for AI-assisted EDA workflows.

This paper proposes a physics-constrained S-parameter surrogate model based on a Bidirectional LongShort-Term Memory (BiLSTM) network for the rapid design of Wilkinson power dividers. A physics-constrained prior sampling strategy is proposed to cover the 2–20 GHz band to reduce the reliance oncomputationally expensive full-wave electromagnetic simulations. A composite physics-constrainedloss function, which incorporates frequency-aware weighting, low-magnitude response enhancementand electromagnetic physics-based constraints, is also proposed to improve the prediction accuracy ofthe surrogate model. To further evaluate the practical applicability of the proposed surrogate model,four Wilkinson power dividers with different center frequencies, which are not included in the originaldataset, are fabricated and measured according to the predictions of the surrogate model. The measuredresults exhibit good agreement with the predicted performance, confirming that the proposed surrogatemodel can be used to help the practical device fabrication. Consequently, the proposed frameworkprovides an efficient and reliable solution for the rapid design of equal-split Wilkinson power dividers.

A compact quasi-elliptic bandpass filter (BPF) based on self-coupled quarter-wavelength resonators (SCQWRs) is proposed. The SCQWR is realized by vertically folding a microstrip line to form a self-coupled structure, which significantly reduces the circuit footprint and enhances the mixed coupling between resonators, enabling a quasi-elliptic response with two transmission zeros (TZs). Source-load coupling is further introduced through layout optimization to generate an additional controllable TZ. An equivalent circuit model is developed to explain the operating mechanism of the filter. For experimental verification, a second-order quasi-elliptic BPF is designed and fabricated based on the GaAs-IPD process. The fabricated filter exhibits a center frequency of 28 GHz, an insertion loss of 2.16 dB, and a 3-dB fractional bandwidth (FBW) of 12.9%. The maximum stopband attenuation is 41 dB with a compact core area of 0.034 lambda(0)& times;0.025 lambda(0).

Although the learning-based surrogate model could achieve fast prediction, the model uncertainty in data and predictions remains a significant challenge. To alleviate the uncertainty problems for the surrogate model, a learning-based uncertainty analysis method for improving the stability of the inverse model of RF devices is proposed. The inherent uncertainty in the inverse model is analyzed by leveraging ensemble adversarial learning, enabling the prediction of confidence intervals for the inverse surrogate model output without changing the original neural network structure. Subsequently, the differential evolution (DE) algorithm is employed to optimize the geometric parameters of the inverse model output within the predicted interval. To validate feasibility and effectiveness of the method, RF hairpin filter is used for the design example via the neural networks. RF filters with multiple center frequencies and bandwidths are simulated, fabricated and measured. Both simulated and measured results demonstrate a notable enhancement, with a 37.22% predicted accuracy improvement in the degree of fitting between the optimized response and the label of the input network, compared to the circuit response (S-parameter) pre-optimization. The proposed method exhibits much better performance in terms of the predicted accuracy and computational time compared to direct optimization without inverse model. The learning-based method could also be applicable to the other RF devices, affirming the practical applicability and robustness of the approach.

In this article, we present a planar microstrip sensor based on unequal-width three-coupled lines (UWTCLs) for characterizing the relative permittivity of liquid materials. The sensor consists of three parallel coupled lines (PCLs), featuring a wider central line and narrower side lines. The interaction of these unequal-width coupled lines enhances the fringe field effect and electric field concentration, thereby improving sensitivity and resolution. The sensor's performance was evaluated by measuring its response in ethanolwater mixtures with varying volume fractions. The permittivity of ethanolwater mixtures ranges from 6 to 75 at 25 degrees C, and the corresponding resonance frequency varies from 3.7 to 5.65 GHz. For this range of permittivity, the proposed sensor achieves an average normalized sensitivity of 2.064%. Compared to previously reported sensors, the proposed sensor demonstrates superior sensitivity across a wide range of permittivity.

This article proposes a novel neural network (NN) architecture combining convolutional and transposed convolutional NNs (TCNNs) to accurately and efficiently model the S-parameter of interconnects for 3-D integration. The network incorporates physical consistency constraints, specifically causality and passivity, into its design to ensure the physical effectiveness of the output. The transposed convolutional network serves as a subnetwork to map the relationship between the geometrical parameters and the S-parameter for substructures. Then, the S-parameters of individual substructures are cascaded to deal with a complex structure composed of substructures. A coupling NN (CONN), with causality and passivity constraints, is developed to map the coarse cascaded S-parameters to the fine, accurate S-parameters. With the help of this high-dimensional space mapping, a small amount of electromagnetic (EM) simulation data of complex interconnect structures is sufficient to learn the relationship between cascaded and real S-parameters. To ensure the completeness of the training set distribution when training the CONN on small datasets, a sensitivity analysis-based training set screening method is proposed to enhance the training performance of the CONN. The proposed algorithm is demonstrated in two different assemble structure applications. The results highlight the effectiveness, flexibility, and versatility of the proposed architecture in modeling complex structures with small, costly simulation data while maintaining accuracy and physical consistency.

Modern integrated circuit (IC) design relies heavily on electronic design automation (EDA) tools. However, none currently offers efficient support for chip selection in system-level IC design. To address this challenge, this paper presents ChipMate, a system that rapidly and accurately generates chip recommendations in response to user requirements expressed in natural language. To realize this capability, a design-experience-driven multi-agent collaboration (DEMA) framework is proposed. Specifically, an agent-assisted dataset generation (AADG) method is introduced to improve the efficiency of generating domain-specific datasets enriched with structured design experience. Building upon this, a domain-specific retrieval-augmented generation (DSRAG) workflow is proposed, and a specialized chip database is constructed to achieve high accuracy in chip selection. Moreover, to continuously enhance the recommendation quality, a human-machine feedback-enhanced loop (HMFL) is incorporated into the DEMA framework, enabling iterative refinement and allowing recommendation accuracy and reliability to continuously evolve. Experimental results demonstrate that ChipMate significantly outperforms the state-of-the-art DeepSeek-R1 of comparable scale, achieving improvements of 3.03 & times;, 1.38 & times;, and 1.34 & times; in BLEU, METEOR, and ROUGE-L scores, respectively. Notably, in the reliability level assessment, it yields the highest proportion of "Highly Reliable" ratings, reaching 79.3%, confirming its overall effectiveness.

This paper proposes a Dual-Attention Physics-Informed Gated Recurrent Unit (DA-PIGRU) surrogate model for broadband branch-line coupler modeling over the 3--13 GHz frequency range. A Frequency-Driven Geometric Relevance (FDGR) method is proposed at the input stage, where a physics-informed mask enables adaptive feature reweighting within the operating frequency band. In addition, a physics-informed joint loss function, which combines asymmetric amplitude weighting and phase-gradient smoothing, is also presented to suppress nonphysical dispersion artifacts and improve resonance prediction accuracy in this paper. Ablation studies show that DA-PIGRU reduces the passband mean absolute error (MAE) by 47.0%, achieving 0.16 dB compared with purely data-driven baselines, while maintaining second-level inference time. Experimental validation using three fabricated prototypes demonstrates well agreement between predictions and measurements. The proposed surrogate model provides an accurate and computationally efficient solution for real-time optimization in RF electronic design automation (EDA).

The design of 5G/6G millimeter-wave filters confronts dual challenges: high-dimensional parameter optimization and computationally prohibitive full-wave simulations. To overcome these limitations, this paper presents an Engineering-Verified Adaptive Surrogate Optimization (EVASO) framework. The EVASO framework employs a high-fidelity artificial neural network (ANN) surrogate model for rapid design optimization, enhanced by an adaptive mechanism that intelligently invokes High Frequency Structure Simulator (HFSS) simulations to ensure accuracy. The core innovation lies in its integrated engineering verification loop, which utilizes Monte Carlo sampling to predict production yield and validates results through correlation with physical measurement data. In a 75-110 GHz filter design case study, EVASO achieved a remarkable reduction in optimization cycle time from 72 h to under 3 h while demonstrating a predicted yield of approximately 70% that showed strong alignment with measured performance. This work establishes a physically-verified paradigm for intelligent and highly reliable millimeter-wave device design.

In the context of process design kit (PDK) development, obtaining accurate broadband $S$ -parameters of interconnect structures is often challenged by limited measurement bandwidth and the high computational cost of wide-frequency electromagnetic (EM) simulations. To address these constraints, this article proposes a frequency-domain extrapolation method based on machine learning (ML). Unlike traditional approaches that rely on physical modeling or focus solely on local spectral patterns, the proposed multihead convolutional neural network-long short-term memory network transfer learning (MH-CNN-LSTM-TL) model establishes a nonlinear mapping between low- and high-frequency responses, with geometric parameters incorporated as auxiliary inputs to achieve accurate extrapolation of high-frequency $S$ -parameters. The model employs a multibranch architecture, enabling the parallel extrapolation of multiple $S$ -parameter frequency responses. To mitigate the high cost of EM simulation data acquisition, a transfer learning (TL) strategy from circuit simulation data to EM simulation data is introduced, which significantly improves generalization under small-sample conditions and reduces computational overhead. The effectiveness of the proposed approach is evaluated with three representative simulation cases: one on microstrip lines and two on complex through-silicon via (TSV) with redistribution layer (RDL) interconnect structures, in addition to validation with experimental data from differential microstrip lines. Results demonstrate superior prediction accuracy across a wide-frequency range, along with strong adaptability and generalization capability, highlighting its potential for practical engineering applications.

A physics-guided neural network (PGNN) for surface potential calculation in Gallium Nitride (GaN) highelectron-mobility transistors (HEMTs) is proposed in this paper. Since the discontinuities of piecewise analytical function, lack physical consistency in data-driven forms, or rely on empirical Drain-Induced Barrier Lowering (DIBL) parameters to indirectly capture drain-bias effects, we employ a PGNN that directly takes drain voltage (Vd) as input and replaces threshold voltage (Vth) with the physically derived Vth_dibl. Compared with purely datadriven neural networks and piecewise analytical solutions, the PGNN demonstrates superior physical consistency and numerical smoothness, closely matching reference numerical-solution accuracy.

An enhanced fin field-effect transistor (FinFET) model is proposed to characterize the quantum tunneling currents across ultrathin gate dielectrics under cryogenic conditions. Critically, the quantum tunneling effect persists even in the absence of an external bias voltage. To capture this effect, the model incorporates the work function difference between the materials into the conventional tunneling formula. Furthermore, a temperature-dependent correction term is integrated to extend the model’s applicability across the cryogenic-to-room temperature range. Experimental validation confirms that the proposed model accurately characterizes the cryogenic physical phenomena of FinFET devices, providing a theoretical foundation for further research on cryogenic integrated circuits. The validity of the model was ascertained by comparison with $S$ -parameter measurements up to 66.2 GHz.

This paper presents a machine-learning-assisted software framework for circuit and system-level electro-thermal co-design of three-dimensional integrated circuits. The framework targets power delivery networks (PDNs) with embedded TSV microchannel structures and unifies equivalent circuit modeling, surrogate-based performance prediction, and intelligent optimization into a design automation workflow. To enable fast power integrity evaluation, a physics-based matrix equivalent circuit formulation is developed to extract the frequency dependent RLGC parameters of densely coupled TSV arrays under coolant perturbation. The proposed model achieves approximately 90% runtime reduction over full-wave electromagnetic simulation while keeping the impedance error within 6.6%. For large scale PDN synthesis, a Transformer-Pointer reinforcement learning solver trained with proximal policy optimization is introduced for decoupling capacitor placement in high-dimensional combinatorial spaces, reducing the PDN impedance by 20%. For thermal design space exploration, a tri-model hybrid semi-supervised surrogate is integrated with NSGA-II to accelerate the multi-objective thermal integrity optimization process by predicting thermal performance indices. The proposed framework provides a scalable software methodology for integrated modeling, analysis, and optimization of advanced 3D power delivery and thermal management systems.

Deterministic optimization methods offer superior efficiency and reproducibility compared with stochastic approaches. This paper presents an improved deterministic single-objective (DSO) sizing algorithm that enhances computational efficiency while effectively handling design constraints. Building upon this framework, a deterministic multi-objective (DMO) sizing algorithm is further developed by integrating a modified weighted sum method to enable high-quality Pareto front exploration. Experimental results show that the proposed DSO sizing algorithm surpasses the best-performing baseline in optimization efficiency, yielding improvements of 21.48% and 76.97% for the two-stage and three-stage Op-Amps, respectively. Similarly, the proposed DMO sizing algorithm outperforms NSGA-II in the spacing metric, achieving improvements of 80.27% and 73.29% for the two-stage and three-stage Op-Amps, respectively, while also producing a higher-quality Pareto front. These results demonstrate that the proposed deterministic sizing algorithms achieve reliable convergence, efficient design space exploration, and strong competitiveness against existing optimization methods. By providing repeatable, physically interpretable, and computationally efficient optimization results, the proposed deterministic sizing algorithm enables designers to systematically explore performance trade-offs and make informed decisions in complex analog IC multi-objective optimization scenarios.