Digital predistortion (DPD) is a key technique for mitigating power amplifier (PA) nonlinearities in radio frequency (RF) systems. However, rapidly varying transmission configurations in emerging communication systems lead to configuration-dependent PA nonlinear behavior, challenging conventional fixed-state DPD models. This article proposes a multimodel residual structure neural network (MRS-NN). A physics-inspired low-complexity bandpass feedback recurrent neural network (BPFRNN) is developed as the core nonlinear modeling module, together with a dynamic feature extraction network. Furthermore, a systematic lightweighting strategy is developed by identifying structural redundancies in MRS-NN and applying corresponding optimization methods. Explicitly removable redundancies are eliminated through structural reconstruction, while knowledge distillation (KD) with a hybrid distillation target is employed to assist the training of lightweight structures with nonremovable redundancies. Experiments under diverse transmission configurations demonstrate that the proposed model achieves superior linearization performance compared with existing varying-configuration DPD approaches. The lightweight model reduces the parameter count by over 50% while maintaining comparable linearization capability.
This article presents a frequency-aware automated optimization framework for broadband filtering power amplifiers (FPAs) to overcome the limitations of conventional methods in capturing frequency-dependent performance. A frequency-domain objective function, constrained by the generalized Gaussian transition frequency response, is formulated to incorporate the distributed characteristics of key metrics, such as output power, efficiency, and S21, into a unified multiobjective model. This enables holistic shaping of the wideband frequency response. Utilizing this framework, the multiobjective black-winged kite algorithm (MOBKA), combined with a Pareto-based approach, is employed to globally co-optimize the broadband filtering matching network and circuit parameters, thereby seamlessly integrating power amplification and filtering functions, without relying on a specific circuit topology. Experimental results show that the fabricated broadband FPA achieves an average drain efficiency (DE) of 78.9% and an average saturated output power of 41.0 dBm across the 1.0–1.4 GHz band, with substantial out-of-band suppression, confirming the effectiveness of the proposed approach.
In this article, a wideband lumped-element bandpass filter (BPF) is presented based on the proposed generalized Chebyshev synthesis theory. First, the basic unit of the proposed BPF consists of three parallel $LC$ resonators, where two $LC$ resonators are connected in parallel on both sides of the two ports and one $LC$ resonator is cascaded between the two ports. Then, a $\Pi $ -type electric coupling resonator (ECR)/magnetic coupling resonator (MCR), capable of simultaneously generating one reflection zero (RZ) and one transmission zero (TZ), is introduced into the parallel resonators of the basic unit of the proposed BPF. The proposed BPFs can be extended to $N$ th-order configurations by cascading nonresonating nodes (NRNs) along the main path and then introducing a $\Pi $ -type resonator. The proposed synthesis theory proves that the proposed BPFs belong to the same theoretical generalized Chebyshev characteristic function. Furthermore, a summary and discussion of the design algorithm as well as relevant design charts are presented. Several design examples are selected to demonstrate that the proposed synthesis theory can flexibly control the design specifications of BPFs, such as fractional bandwidth (FBW), center frequency, return loss (RL), and the locations of TZs. Finally, two prototype circuits of BPFs with $\Pi $ -type ECR and MCR operating at X band were fabricated and measured based on the gallium arsenide (GaAs)-based integrated passive device (IPD) technology, respectively. The simulation results are in good agreement with the measured results.
This article presents a 20–28-GHz impedance-transformation-based balanced front end (FE) that fundamentally eliminates the conventional need for the series switch in the transmit (TX) mode. The TX path employs a balanced power amplifier (PA) architecture. By exploiting the impedance-transformation property of a 3-dB quadrature coupler and properly controlling the load conditions at the through (THRU) and coupled (CPL) ports, the PA output is transformed to a high equivalent impedance during the receive (RX) mode, thereby mitigating the loading effect of the TX path on the RX chain without the need for a TX-path series switch. As a result, the insertion loss in the TX path is significantly reduced, and the TX efficiency is improved. The chip is implemented in a 65-nm CMOS process with a compact core area of 0.275 mm2 (excluding pads). Measurement results demonstrate that the FE achieves a TX gain of 16 dB, a saturated output power of 15.9 dBm, and a peak power-added efficiency (PAE) of 26.4% over 20–28 GHz. In the RX mode, a gain of 11.4 dB and a minimum noise figure (NF) of 6.2 dB are obtained. The measured results validate the effectiveness of the proposed impedance-transformation-based balanced FE in achieving low TX-path loss, enhanced power efficiency, and sufficient receiver isolation within a compact CMOS implementation.
Yield estimation of microwave components under fabrication variations usually requires hundreds of full-wave electromagnetic (EM) simulations, which is too expensive for finite-element models. Existing surrogate-based acceleration methods can reduce the number of full-wave evaluations for yield estimation, but their model construction still requires substantial EM data and repeated simulation runs. To further reduce the repeated sample-evaluation cost, this article develops a common-basis model order reduction (CBMOR) framework for microwave yield estimation. A common reduced basis is constructed from a small number of representative samples under tolerance variations, each evaluated by full-wave EM simulation, so that the Monte Carlo population can be efficiently analyzed using common reduced-order systems. To eliminate the repeated full-system assembly and large matrix-basis multiplications that would otherwise be required for individual yield samples, a local matrix reconstruction and new projection formulation is developed. For each online sample, only the variable-related faces and nearby cells are deformed, only the corresponding variable-dependent matrix blocks are reconstructed, and their reduced contributions are calculated using the associated local rows of the common basis. This avoids repeated full-matrix assembly and repeated large matrix-basis multiplications in the Monte Carlo loop. Moreover, an adaptive common-basis sample-selection algorithm (CBSSA) is developed to adaptively select informative samples from the yield population. These selected samples are then used as representative samples to construct and enrich the common basis. Three numerical examples validate the proposed framework. The proposed method achieves acceptable yield-estimation accuracy with respect to the HFSS direct Monte Carlo while substantially reducing the computational cost.
